Full Code of doubledaibo/drnet_cvpr2017 for AI

Repository: doubledaibo/drnet_cvpr2017
Branch: master
Commit: a83aa5768b91
Files: 22
Total size: 123.2 KB

Directory structure:
gitextract_pa_vjz81/

├── .gitignore
├── LICENSE
├── README.md
├── lib/
│   ├── customize_layers/
│   │   ├── __init__.py
│   │   └── concat_layer.py
│   ├── rel_data_layer/
│   │   ├── __init__.py
│   │   └── layer.py
│   └── utils/
│       ├── __init__.py
│       └── eval_utils.py
├── prototxts/
│   ├── drnet_8units_linear_shareweight.prototxt
│   ├── drnet_8units_relu_shareweight.prototxt
│   ├── drnet_8units_softmax.prototxt
│   ├── test_drnet_8units_linear_shareweight.prototxt
│   ├── test_drnet_8units_relu_shareweight.prototxt
│   └── test_drnet_8units_softmax.prototxt
├── snapshots/
│   └── README.md
└── tools/
    ├── _init_paths.py
    ├── eval_triplet_recall.py
    ├── eval_union_recall.py
    ├── prepare_data.py
    ├── test_predicate_recognition.py
    └── test_triplet_detection.py

================================================
FILE CONTENTS
================================================

================================================
FILE: .gitignore
================================================
snapshots/*.caffemodel
svg


================================================
FILE: LICENSE
================================================
Copyright the Chinese University of Hong Kong. All rights reserved.

Contact persons:
Bo Dai (db014 [at] ie [dot] cuhk [dot] edu [dot] hk)

This software is being made available for research use only.
Any commercial use or redistribution of this software requires a license from
the Chinese University of Hong Kong.

You may use this work subject to the following conditions:

1. This work is provided "as is" by the copyright holder, with
absolutely no warranties of correctness, fitness, intellectual property
ownership, or anything else whatsoever.  You use the work
entirely at your own risk.  The copyright holder will not be liable for
any legal damages whatsoever connected with the use of this work.

2. The copyright holder retain all copyright to the work. All copies of
the work and all works derived from it must contain (1) this copyright
notice, and (2) additional notices describing the content, dates and
copyright holder of modifications or additions made to the work, if
any, including distribution and use conditions and intellectual property
claims.  Derived works must be clearly distinguished from the original
work, both by name and by the prominent inclusion of explicit
descriptions of overlaps and differences.

3. The names and trademarks of the copyright holder may not be used in
advertising or publicity related to this work without specific prior
written permission. 

4. In return for the free use of this work, you are requested, but not
legally required, to do the following:

* If you become aware of factors that may significantly affect other
  users of the work, for example major bugs or
  deficiencies or possible intellectual property issues, you are
  requested to report them to the copyright holder, if possible
  including redistributable fixes or workarounds.

* If you use the work in scientific research or as part of a larger
  software system, you are requested to cite the use in any related
  publications or technical documentation. The work is based upon:
 
 
  "Detecting Visual Relationship with Deep Relational Networks"
  Bo Dai, Yuqi Zhang, Dahua Lin, Computer Vision and Pattern Recognition,
  (CVPR 2017), 2017 (oral).

this copyright notice must be retained with all copies of the software,
including any modified or derived versions.


================================================
FILE: README.md
================================================
# Code of [Detecting Visual Relationships with Deep Relational Networks](https://arxiv.org/abs/1704.03114)

The code is written in python, and all networks are implemented using [Caffe](https://github.com/BVLC/caffe).

## Datasets 

* [VRD](http://cs.stanford.edu/people/ranjaykrishna/vrd/dataset.zip)
* sVG: subset of [Visual Genome](https://visualgenome.org/)
	- [Link](https://drive.google.com/file/d/0B5RJWjAhdT04SXRfVHBKZ0dOTzQ/view?usp=sharing&resourcekey=0-bW_W0QVJOfaNs5NyGjDjbQ)
	- Images can be downloaded from the website of Visual Genome
	- Remarks: eventually I found no time to further clean it. This subset has a manually cleaned list for relationship predicates. The list for objects may need further cleaning, although Faster-RCNN can get a recall@20 around 50%.
	- Using our method, you can get the corresponding results reported in the paper on this dataset.

## Networks

This repo contains three kinds of networks. And all of them get the raw response for predicate based on both appearance cues and spatial cues,
followed by a refinement according to responses of the subject, the object and the predicate.
The networks are designed for the task of predicate recognition, 
where ground-truth labels of the subject and the object are provided as inputs.
Therefore, in these networks, responses of the subject and the object are replaced with indicator vectors,
and only response of the predicate will be refined.

In these networks, the subnet for appearance cues is VGG16, and the subnet for spatial cues consists of three conv layers.
And outputs of both subnets are combined via a customized concatenate layer,
followed by two fc layers to generate raw response for the predicate.

The customized concatenate layer is used for combining the output of a fc layer and channels of the output of a conv layer,
which can be replaced with caffe's Concat layer
if the last conv layer in spatial subnet (conv3_p) is equivalently replaced with a fc layer.

The details of these networks are

* drnet_8units_softmax: it has 8 inference units with softmax function as the activation function.

* drnet_8units_linear_shareweight: it has 8 inference units with no activation function, and the weights are shared across units.

* drnet_8units_relu_shareweight: it has 8 inference units with relu function as the activation function, and the weights are shared across units.

### Training

The training procedure is component-by-component. 
Specifically, a network usually contains three components, 
namely the subnet for appearance (A), the subnet for spatial cues (S), and the drnet for statistical dependencies (D).
In training, we train the network as follow:
* train A in isolation
* train S in isolation
* train A + S in isolation, with weights initialized from previous steps
* train A + S + D jointly, with weights initialized from previous steps

Each step we use the same loss, and we use dropout to avoid overfit.

### Recalls on Predicate Recognition

| Networks | Recall@50 | Recall@100 |
| --- | :---: | :---: |
| drnet_8units_softmax | 75.22 | 77.55 |
| drnet_8units_linear_shareweight | 78.57 | 79.94 |
| drnet_8units_relu_shareweight | 80.86 | 81.83 |

## Codes

* lib/: python layers, as well as auxiliary files for evaluation
* prototxts/: training and testing prototxts
* tools/: python codes for preparing data and evaluation
* snapshots/: pretrain models

## Finetune or Evaluate

1. Download the dataset [VRD](https://github.com/Prof-Lu-Cewu/Visual-Relationship-Detection)
2. Preprocess the dataset using tools/preprare_data.py
3. Download one pretrain model in snapshots/
4. Finetune or Evaluate using corresponding prototxts in prototxts/

## Pair Filter

### Structure

![Structure](https://raw.github.com/doubledaibo/drnet/master/imgs/pair_filter.jpg)

### Training

To train this network, we randomly sample pairs of bounding boxes (with labels) from 
each training image, treating those with 0.5 IoU (or above) with any ground-truth pairs (with same labels)
as positive samples, and the rest as negative samples.

## Citation

If you use this code, please cite the following paper(s):

	@article{dai2017detecting,
		title={Detecting Visual Relationships with Deep Relational Networks},
		author={Dai, Bo and Zhang, Yuqi and Lin, Dahua},
  		booktitle={Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition},
  		year={2017}
	}

## License

This code is used for research only. See LICENSE for details.


================================================
FILE: lib/customize_layers/__init__.py
================================================


================================================
FILE: lib/customize_layers/concat_layer.py
================================================
import caffe
import numpy as np
import json

class Layer(caffe.Layer):
	def setup(self, bottom, top):
		self._sum = 0
		for i in xrange(len(bottom)):
			self._sum += bottom[i].data.shape[1]

	def reshape(self, bottom, top):
		top[0].reshape(bottom[0].data.shape[0], self._sum)

	def forward(self, bottom, top):
		offset = 0
		for i in xrange(len(bottom)):
			if len(bottom[i].data.shape) == 2:
				top[0].data[:, offset : offset + bottom[i].data.shape[1]] = bottom[i].data
			else:
				top[0].data[:, offset : offset + bottom[i].data.shape[1]] = bottom[i].data[:, :, 0, 0]
			offset += bottom[i].data.shape[1]
	
	def backward(self, top, propagate_down, bottom):
		offset = 0
		for i in xrange(len(bottom)):
			if len(bottom[i].data.shape) == 2:
				bottom[i].diff[...] = top[0].diff[:, offset : offset + bottom[i].data.shape[1]]
			else:
				bottom[i].diff[:, :, 0, 0] = top[0].diff[:, offset : offset + bottom[i].data.shape[1]]
			offset += bottom[i].data.shape[1]

		


================================================
FILE: lib/rel_data_layer/__init__.py
================================================


================================================
FILE: lib/rel_data_layer/layer.py
================================================
import caffe
import numpy as np
import json
import math
import os.path as osp
import os
import cv2
import h5py

class RelDataLayer(caffe.Layer):
	def _shuffle_inds(self):
		self._perm = np.random.permutation(np.arange(self._num_instance))
		self._cur = 0

	def _get_next_batch_ids(self):
		if self._cur + self._batch_size > self._num_instance:	
			self._shuffle_inds()
		ids = self._perm[self._cur : self._cur + self._batch_size]
		self._cur += self._batch_size	
		return ids

	def _getAppr(self, im, bb):
		subim = im[bb[1] : bb[3], bb[0] : bb[2], :]
		subim = cv2.resize(subim, None, None, 224.0 / subim.shape[1], 224.0 / subim.shape[0], interpolation=cv2.INTER_LINEAR)
		pixel_means = np.array([[[103.939, 116.779, 123.68]]])
		subim -= pixel_means
		subim = subim.transpose((2, 0, 1))
		return subim

	def _getDualMask(self, ih, iw, bb):
		rh = 32.0 / ih
		rw = 32.0 / iw
		x1 = max(0, int(math.floor(bb[0] * rw)))
		x2 = min(32, int(math.ceil(bb[2] * rw)))
		y1 = max(0, int(math.floor(bb[1] * rh)))
		y2 = min(32, int(math.ceil(bb[3] * rh)))
		mask = np.zeros((32, 32))
		mask[y1 : y2, x1 : x2] = 1
		assert(mask.sum() == (y2 - y1) * (x2 - x1))
		return mask		
	
	def _get_next_batch(self):
		ids = self._get_next_batch_ids()
		qas = []
		qbs = []
		ims = []
		poses = []
		labels = []
		for id in ids:
			sample = self._samples[id]
			im = cv2.imread(sample["imPath"]).astype(np.float32, copy=False)
			ih = im.shape[0]	
			iw = im.shape[1]
			qa = np.zeros(self._nclass)
			qa[sample["aLabel"] - 1] = 1
			qas.append(qa)
			qb = np.zeros(self._nclass)
			qb[sample["bLabel"] - 1] = 1
			qbs.append(qb)
			ims.append(self._getAppr(im, sample["rBBox"]))
			poses.append([self._getDualMask(ih, iw, sample["aBBox"]), \
					self._getDualMask(ih, iw, sample["bBBox"])])
			labels.append(sample["rLabel"])
		return {"qa": np.array(qas), "qb": np.array(qbs), "im": np.array(ims), "posdata": np.array(poses), "labels": np.array(labels)}
	
	def setup(self, bottom, top):
		layer_params = json.loads(self.param_str)
		self._samples = json.load(open(layer_params["dataset"]))
		self._num_instance = len(self._samples)
		self._batch_size = layer_params["batch_size"]
		self._nclass = layer_params["nclass"]
		
		self._name_to_top_map = {"qa": 0, "qb": 1, "im": 2, "posdata": 3, "labels": 4}
		self._shuffle_inds()
		top[0].reshape(self._batch_size, self._nclass)
		top[1].reshape(self._batch_size, self._nclass)
		top[2].reshape(self._batch_size, 3, 224, 224)
		top[3].reshape(self._batch_size, 2, 32, 32)
		top[4].reshape(self._batch_size)
		
	def forward(self, bottom, top):
		batch = self._get_next_batch()
		for blob_name, blob in batch.iteritems():
			idx = self._name_to_top_map[blob_name]	
			top[idx].reshape(*(blob.shape))
			top[idx].data[...] = blob.astype(np.float32, copy=False)
	
	def backward(self, top, propagate_down, bottom):
		pass

	def reshape(self, bottom, top):
		pass



================================================
FILE: lib/utils/__init__.py
================================================


================================================
FILE: lib/utils/eval_utils.py
================================================
def computeArea(bb):
	return max(0, bb[2] - bb[0] + 1) * max(0, bb[3] - bb[1] + 1)

def computeIoU(bb1, bb2):
	ibb = [max(bb1[0], bb2[0]), \
		max(bb1[1], bb2[1]), \
		min(bb1[2], bb2[2]), \
		min(bb1[3], bb2[3])]
	iArea = computeArea(ibb)
	uArea = computeArea(bb1) + computeArea(bb2) - iArea
	return (iArea + 0.0) / uArea



================================================
FILE: prototxts/drnet_8units_linear_shareweight.prototxt
================================================
layer {
        name: "data"
        type: "Python"
        top: "qa"
        top: "qb"
        top: "im"
        top: "posdata"
        top: "labels"
        include {
                phase: TRAIN     
        }         
        python_param {                 
                module: 'rel_data_layer.layer'              
                layer: 'RelDataLayer'                 
                param_str: '{"dataset": "reltrain.json", "batch_size": 25, "nclass": 100}'   
        }         
}                    
layer {           
        name: "data"                   
        type: "Python"                 
        top: "qa" 
        top: "qb" 
        top: "im" 
        top: "posdata"                 
        top: "labels"                  
        include { 
                phase: TEST            
        }         
        python_param {                 
                module: 'rel_data_layer.layer'              
                layer: 'RelDataLayer'
                param_str: '{"dataset": "reltest.json", "batch_size": 25, "nclass": 100}'
        }         
}
# Appearance Subnet
layer {
	name: "conv1_1"
	type: "Convolution"
	bottom: "im"
	top: "conv1_1"
	param {
		lr_mult: 0
		decay_mult: 0
	} 
	param {
		lr_mult: 0
		decay_mult: 0
	}
	convolution_param {
		num_output: 64
		pad: 1
		kernel_size: 3
	}
}
layer {
	name: "relu1_1"
	type: "ReLU"
	bottom: "conv1_1"
	top: "conv1_1"
}
layer {
	name: "conv1_2"
	type: "Convolution"
	bottom: "conv1_1"
	top: "conv1_2"
	param {
		lr_mult: 0
		decay_mult: 0
	}
	param {
		lr_mult: 0
		decay_mult: 0
	}
	convolution_param {
		num_output: 64
		pad: 1
		kernel_size: 3
	}
}
layer {
	name: "relu1_2"
	type: "ReLU"
	bottom: "conv1_2"
	top: "conv1_2"
}
layer {
	name: "pool1"
	type: "Pooling"
	bottom: "conv1_2"
	top: "pool1"
	pooling_param {
		pool: MAX
		kernel_size: 2
		stride: 2
	}
}
layer {
	name: "conv2_1"
	type: "Convolution"
	bottom: "pool1"
	top: "conv2_1"
	param {
		lr_mult: 0
		decay_mult: 0
	}
	param {
		lr_mult: 0
		decay_mult: 0
	}
	convolution_param {
		num_output: 128
		pad: 1
		kernel_size: 3
	}
}
layer {
	name: "relu2_1"
	type: "ReLU"
	bottom: "conv2_1"
	top: "conv2_1"
}
layer {
	name: "conv2_2"
	type: "Convolution"
	bottom: "conv2_1"
	top: "conv2_2"
	param {
		lr_mult: 0
		decay_mult: 0
	}
	param {
		lr_mult: 0
		decay_mult: 0
	}
	convolution_param {
		num_output: 128
		pad: 1
		kernel_size: 3
	}
}
layer {
	name: "relu2_2"
	type: "ReLU"
	bottom: "conv2_2"
	top: "conv2_2"
}
layer {
	name: "pool2"	
	type: "Pooling"
	bottom: "conv2_2"
	top: "pool2"
	pooling_param {
		pool: MAX
		kernel_size: 2
		stride: 2
	}
}
layer {
	name: "conv3_1"
	type: "Convolution"
	bottom: "pool2"
	top: "conv3_1"
	param {
		lr_mult: 0
		decay_mult: 0
	}
	param {
		lr_mult: 0
		decay_mult: 0
	}
	convolution_param {
		num_output: 256
		pad: 1
		kernel_size: 3
	}
}
layer {
	name: "relu3_1"
	type: "ReLU"
	bottom: "conv3_1"
	top: "conv3_1"
}
layer {
	name: "conv3_2"
	type: "Convolution"
	bottom: "conv3_1"
	top: "conv3_2"
	param {
		lr_mult: 0
		decay_mult: 0
	}
	param {
		lr_mult: 2
		decay_mult: 0
	}
	convolution_param {
		num_output: 256
		pad: 1
		kernel_size: 3
	}
}
layer {
	name: "relu3_2"
	type: "ReLU"
	bottom: "conv3_2"
	top: "conv3_2"
}
layer {
	name: "conv3_3"
	type: "Convolution"
	bottom: "conv3_2"
	top: "conv3_3"
	param {
		lr_mult: 0
		decay_mult: 0
	}
	param {
		lr_mult: 0
		decay_mult: 0
	}
	convolution_param {
		num_output: 256
		pad: 1
		kernel_size: 3
	}
}
layer {
	name: "relu3_3"
	type: "ReLU"
	bottom: "conv3_3"
	top: "conv3_3"
}
layer {
	name: "pool3"
	type: "Pooling"
	bottom: "conv3_3"
	top: "pool3"
	pooling_param {
		pool: MAX
		kernel_size: 2
		stride: 2
	}
}
layer {
	name: "conv4_1"
	type: "Convolution"
	bottom: "pool3"
	top: "conv4_1"
	param {
		lr_mult: 0
		decay_mult: 0
	}
	param {
		lr_mult: 0
		decay_mult: 0
	}
	convolution_param {
		num_output: 512
		pad: 1
		kernel_size: 3
	}
}
layer {
	name: "relu4_1"
	type: "ReLU"
	bottom: "conv4_1"
	top: "conv4_1"
}
layer {
	name: "conv4_2"
	type: "Convolution"
	bottom: "conv4_1"
	top: "conv4_2"
	param {
		lr_mult: 0
		decay_mult: 0
	}
	param {
		lr_mult: 0
		decay_mult: 0
	}
	convolution_param {
		num_output: 512
		pad: 1
		kernel_size: 3
	}
}
layer {
	name: "relu4_2"
	type: "ReLU"
	bottom: "conv4_2"
	top: "conv4_2"
}
layer {
	name: "conv4_3"
	type: "Convolution"
	bottom: "conv4_2"
	top: "conv4_3"
	param {
		lr_mult: 0
		decay_mult: 0
	}
	param {
		lr_mult: 0
		decay_mult: 0
	}
	convolution_param {
		num_output: 512
		pad: 1
		kernel_size: 3
	}
}
layer {
	name: "relu4_3"
	type: "ReLU"
	bottom: "conv4_3"
	top: "conv4_3"
}
layer {
	name: "pool4"
	type: "Pooling"
	bottom: "conv4_3"
	top: "pool4"
	pooling_param {
		pool: MAX
		kernel_size: 2
		stride: 2
	}
}
layer {
	name: "conv5_1"
	type: "Convolution"
	bottom: "pool4"
	top: "conv5_1"
	param {
		lr_mult: 0
		decay_mult: 0
	}
	param {
		lr_mult: 0
		decay_mult: 0
	}
	convolution_param {
		num_output: 512
		pad: 1
		kernel_size: 3
	}
}
layer {
	name: "relu5_1"
	type: "ReLU"
	bottom: "conv5_1"
	top: "conv5_1"
}
layer {
	name: "conv5_2"
	type: "Convolution"
	bottom: "conv5_1"
	top: "conv5_2"
	param {
		lr_mult: 0
		decay_mult: 0
	}
	param {
		lr_mult: 0
		decay_mult: 0
	}
	convolution_param {
		num_output: 512
		pad: 1
		kernel_size: 3
	}
}
layer {
	name: "relu5_2"
	type: "ReLU"
	bottom: "conv5_2"
	top: "conv5_2"
}
layer {
	name: "conv5_3"
	type: "Convolution"
	bottom: "conv5_2"
	top: "conv5_3"
	param {
		lr_mult: 0
		decay_mult: 0
	}
	param {
		lr_mult: 0
		decay_mult: 0
	}
	convolution_param {
		num_output: 512
		pad: 1
		kernel_size: 3
	}
}
layer {
	name: "relu5_3"
	type: "ReLU"
	bottom: "conv5_3"
	top: "conv5_3"
}
layer {
	name: "pool5"
	type: "Pooling"
	bottom: "conv5_3"
	top: "pool5"
	pooling_param {
		pool: MAX
		kernel_size: 2
		stride: 2
	}
}
layer {
	name: "fc6"
	type: "InnerProduct"
	bottom: "pool5"
	top: "fc6"
	param {
		lr_mult: 0
		decay_mult: 0
	}
	param {
		lr_mult: 0
		decay_mult: 0
	}
	inner_product_param {
		num_output: 4096
	}
}
layer {
	name: "relu6"
	type: "ReLU"
	bottom: "fc6"
	top: "fc6"
}
layer {
	name: "fc7"
	type: "InnerProduct"
	bottom: "fc6"
	top: "fc7"
	param {
		lr_mult: 0
		decay_mult: 0
	}
	param {
		lr_mult: 0
		decay_mult: 0
	}
	inner_product_param {
		num_output: 4096
	}
}
layer {
	name: "relu7"
	type: "ReLU"
	bottom: "fc7"
	top: "fc7"
}
layer {
	name: "fc8"
	type: "InnerProduct"
	bottom: "fc7"
	top: "fc8"
	param {
		lr_mult: 0
		decay_mult: 0
	}
	param {
		lr_mult: 0
		decay_mult: 0
	}
	inner_product_param {
		num_output: 256
		weight_filler {
			type: "xavier"
		}
		bias_filler {
			type: "constant"
			value: 0
		}
	}
}
layer {
	name: "relu8"
	type: "ReLU"
	bottom: "fc8"
	top: "fc8"
}
# Spatial Cfg Subnet
layer {
	name: "conv1_p"
	type: "Convolution"
	bottom: "posdata"
	top: "conv1_p"
	param {
		lr_mult: 0
		decay_mult: 0
	}
	param {
		lr_mult: 0
		decay_mult: 0
	}
	convolution_param {
		num_output: 96
		kernel_size: 5
		pad: 2
		stride: 2
	}
}
layer {
	name: "relu1_p"
	type: "ReLU"
	bottom: "conv1_p"
	top: "conv1_p"
}
layer {
	name: "conv2_p"
	type: "Convolution"
	bottom: "conv1_p"
	top: "conv2_p"
	param {
		lr_mult: 0
		decay_mult: 0
	}
	param {
		lr_mult: 0
		decay_mult: 0
	}
	convolution_param {
		num_output: 128
		kernel_size: 5
		pad: 2
		stride: 2
	}
}
layer {
	name: "conv3_p"
	type: "Convolution"
	bottom: "conv2_p"
	top: "conv3_p"
	param {
		lr_mult: 0
		decay_mult: 0
	}
	param {
		lr_mult: 0
		decay_mult: 0
	}
	convolution_param {
		num_output: 64
		kernel_size: 8
	}
}
layer {
	name: "relu3_p"
	type: "ReLU"
	bottom: "conv3_p"
	top: "conv3_p"
}
# Combine features from subnets
layer {
	name: "concat1_c"
	type: "Python"
	bottom: "fc8"
	bottom: "conv3_p"
	top: "concat1_c"
	python_param {
		module: "customize_layers.concat_layer"
		layer: "Layer"
	}
}
layer {
	name: "fc2_c"
	type: "InnerProduct"
	bottom: "concat1_c"
	top: "fc2_c"
	param {
		lr_mult: 0
		decay_mult: 0
	}
	param {
		lr_mult: 0
		decay_mult: 0
	}
	inner_product_param {
		num_output: 128
	}
}
layer {
	name: "relu2_c"
	type: "ReLU"
	bottom: "fc2_c"
	top: "fc2_c"
}
layer {
	name: "PhiR_0"
	type: "InnerProduct"
	bottom: "fc2_c"
	top: "qr0"
	param {
		lr_mult: 0
		decay_mult: 0
	}
	param {
		lr_mult: 0
		decay_mult: 0
	}
	inner_product_param {
		num_output: 70
		weight_filler {
			type: "xavier"
		}
		bias_filler {
			type: "constant"
			value: 0
		}
	}
}
#DR-Net
layer {
	name: "PhiA_1"
	type: "InnerProduct"
	bottom: "qa"
	top: "qar1"
	param {
		name: "qar_w"
		lr_mult: 1
		decay_mult: 1
	}
	param {
		name: "qar_b"
		lr_mult: 2
		decay_mult: 0
	}
	inner_product_param {
		num_output: 70
		weight_filler {
			type: "xavier"
		}
		bias_filler {
			type: "constant"
			value: 0
		}
	}
}
layer {
	name: "PhiB_1"
	type: "InnerProduct"
	bottom: "qb"
	top: "qbr1"
	param {
		name: "qbr_w"
		lr_mult: 1
		decay_mult: 1
	}
	param {
		name: "qbr_b"
		lr_mult: 2
		decay_mult: 0
	}
	inner_product_param {
		num_output: 70
		weight_filler {
			type: "xavier"
		}
		bias_filler {
			type: "constant"
			value: 0
		}
	}
}
layer {
	name: "PhiR_1"
	type: "InnerProduct"
	bottom: "qr0"
	top: "q1r"
	param {
		name: "qr_w"
		lr_mult: 1
		decay_mult: 1
	}
	param {	
		name: "qr_b"
		lr_mult: 2
		decay_mult: 0
	}
	inner_product_param {
		num_output: 70
		weight_filler {
			type: "xavier"
		}
		bias_filler {
			type: "constant"
			value: 0
		}
	}
}
layer {
	name: "QSum_1"
	type: "Eltwise"
	bottom: "qar1"
	bottom: "qbr1"
	bottom: "q1r"
	top: "qr1"
	eltwise_param { operation: SUM }
}
layer {
	name: "PhiA_2"
	type: "InnerProduct"
	bottom: "qa"
	top: "qar2"
	param {
		name: "qar_w"
		lr_mult: 1
		decay_mult: 1
	}
	param {
		name: "qar_b"
		lr_mult: 2
		decay_mult: 0
	}
	inner_product_param {
		num_output: 70
	}
}
layer {
	name: "PhiB_2"
	type: "InnerProduct"
	bottom: "qb"
	top: "qbr2"
	param {
		name: "qbr_w"
		lr_mult: 1
		decay_mult: 1
	}
	param {
		name: "qbr_b"
		lr_mult: 2
		decay_mult: 0
	}
	inner_product_param {
		num_output: 70
	}
}
layer {
	name: "PhiR_2"
	type: "InnerProduct"
	bottom: "qr1"
	top: "q2r"
	param {
		name: "qr_w"
		lr_mult: 1
		decay_mult: 1
	}
	param {
		name: "qr_b"
		lr_mult: 2
		decay_mult: 0
	}
	inner_product_param {
		num_output: 70
	}
}
layer {
	name: "QSum_2"
	type: "Eltwise"
	bottom: "qar2"
	bottom: "qbr2"
	bottom: "q2r"
	top: "qr2"
	eltwise_param { operation: SUM }
}
layer {
	name: "PhiA_3"
	type: "InnerProduct"
	bottom: "qa"
	top: "qar3"
	param {
		name: "qar_w"
		lr_mult: 1
		decay_mult: 1
	}
	param {
		name: "qar_b"
		lr_mult: 2
		decay_mult: 0
	}
	inner_product_param {
		num_output: 70
	}
}
layer {
	name: "PhiB_3"
	type: "InnerProduct"
	bottom: "qb"
	top: "qbr3"
	param {
		name: "qbr_w"
		lr_mult: 1
		decay_mult: 1
	}
	param {
		name: "qbr_b"
		lr_mult: 2
		decay_mult: 0
	}
	inner_product_param {
		num_output: 70
	}
}
layer {
	name: "PhiR_3"
	type: "InnerProduct"
	bottom: "qr2"
	top: "q3r"
	param {
		name: "qr_w"
		lr_mult: 1
		decay_mult: 1
	}
	param {
		name: "qr_b"
		lr_mult: 2
		decay_mult: 0
	}
	inner_product_param {
		num_output: 70
	}
}
layer {
	name: "QSum_3"
	type: "Eltwise"
	bottom: "qar3"
	bottom: "qbr3"
	bottom: "q3r"
	top: "qr3"
	eltwise_param { operation: SUM }
}
layer {
	name: "PhiA_4"
	type: "InnerProduct"
	bottom: "qa"
	top: "qar4"
	param {
		name: "qar_w"
		lr_mult: 1
		decay_mult: 1
	}
	param {
		name: "qar_b"
		lr_mult: 2
		decay_mult: 0
	}
	inner_product_param {
		num_output: 70
	}
}
layer {
	name: "PhiB_4"
	type: "InnerProduct"
	bottom: "qb"
	top: "qbr4"
	param {
		name: "qbr_w"
		lr_mult: 1
		decay_mult: 1
	}
	param {
		name: "qbr_b"
		lr_mult: 2
		decay_mult: 0
	}
	inner_product_param {
		num_output: 70
	}
}
layer {
	name: "PhiR_4"
	type: "InnerProduct"
	bottom: "qr3"
	top: "q4r"
	param {
		name: "qr_w"
		lr_mult: 1
		decay_mult: 1
	}
	param {
		name: "qr_b"
		lr_mult: 2
		decay_mult: 0
	}
	inner_product_param {
		num_output: 70
	}
}
layer {
	name: "QSum_4"
	type: "Eltwise"
	bottom: "qar4"
	bottom: "qbr4"
	bottom: "q4r"
	top: "qr4"
	eltwise_param { operation: SUM }
}
layer {
	name: "PhiA_5"
	type: "InnerProduct"
	bottom: "qa"
	top: "qar5"
	param {
		name: "qar_w"
		lr_mult: 1
		decay_mult: 1
	}
	param {
		name: "qar_b"
		lr_mult: 2
		decay_mult: 0
	}
	inner_product_param {
		num_output: 70
	}
}
layer {
	name: "PhiB_5"
	type: "InnerProduct"
	bottom: "qb"
	top: "qbr5"
	param {
		name: "qbr_w"
		lr_mult: 1
		decay_mult: 1
	}
	param {
		name: "qbr_b"
		lr_mult: 2
		decay_mult: 0
	}
	inner_product_param {
		num_output: 70
	}
}
layer {
	name: "PhiR_5"
	type: "InnerProduct"
	bottom: "qr4"
	top: "q5r"
	param {
		name: "qr_w"
		lr_mult: 1
		decay_mult: 1
	}
	param {
		name: "qr_b"
		lr_mult: 2
		decay_mult: 0
	}
	inner_product_param {
		num_output: 70
	}
}
layer {
	name: "QSum_5"
	type: "Eltwise"
	bottom: "qar5"
	bottom: "qbr5"
	bottom: "q5r"
	top: "qr5"
	eltwise_param { operation: SUM }
}
layer {
	name: "PhiA_6"
	type: "InnerProduct"
	bottom: "qa"
	top: "qar6"
	param {
		name: "qar_w"
		lr_mult: 1
		decay_mult: 1
	}
	param {
		name: "qar_b"
		lr_mult: 2
		decay_mult: 0
	}
	inner_product_param {
		num_output: 70
	}
}
layer {
	name: "PhiB_6"
	type: "InnerProduct"
	bottom: "qb"
	top: "qbr6"
	param {
		name: "qbr_w"
		lr_mult: 1
		decay_mult: 1
	}
	param {
		name: "qbr_b"
		lr_mult: 2
		decay_mult: 0
	}
	inner_product_param {
		num_output: 70
	}
}
layer {
	name: "PhiR_6"
	type: "InnerProduct"
	bottom: "qr5"
	top: "q6r"
	param {
		name: "qr_w"
		lr_mult: 1
		decay_mult: 1
	}
	param {
		name: "qr_b"
		lr_mult: 2
		decay_mult: 0
	}
	inner_product_param {
		num_output: 70
	}
}
layer {
	name: "QSum_6"
	type: "Eltwise"
	bottom: "qar6"
	bottom: "qbr6"
	bottom: "q6r"
	top: "qr6"
	eltwise_param { operation: SUM }
}
layer {
	name: "PhiA_7"
	type: "InnerProduct"
	bottom: "qa"
	top: "qar7"
	param {
		name: "qar_w"
		lr_mult: 1
		decay_mult: 1
	}
	param {
		name: "qar_b"
		lr_mult: 2
		decay_mult: 0
	}
	inner_product_param {
		num_output: 70
	}
}
layer {
	name: "PhiB_7"
	type: "InnerProduct"
	bottom: "qb"
	top: "qbr7"
	param {
		name: "qbr_w"
		lr_mult: 1
		decay_mult: 1
	}
	param {
		name: "qbr_b"
		lr_mult: 2
		decay_mult: 0
	}
	inner_product_param {
		num_output: 70
	}
}
layer {
	name: "PhiR_7"
	type: "InnerProduct"
	bottom: "qr6"
	top: "q7r"
	param {
		name: "qr_w"
		lr_mult: 1
		decay_mult: 1
	}
	param {
		name: "qr_b"
		lr_mult: 2
		decay_mult: 0
	}
	inner_product_param {
		num_output: 70
	}
}
layer {
	name: "QSum_7"
	type: "Eltwise"
	bottom: "qar7"
	bottom: "qbr7"
	bottom: "q7r"
	top: "qr7"
	eltwise_param { operation: SUM }
}
layer {
	name: "PhiA_8"
	type: "InnerProduct"
	bottom: "qa"
	top: "qar8"
	param {
		name: "qar_w"
		lr_mult: 1
		decay_mult: 1
	}
	param {
		name: "qar_b"
		lr_mult: 2
		decay_mult: 0
	}
	inner_product_param {
		num_output: 70
	}
}
layer {
	name: "PhiB_8"
	type: "InnerProduct"
	bottom: "qb"
	top: "qbr8"
	param {
		name: "qbr_w"
		lr_mult: 1
		decay_mult: 1
	}
	param {
		name: "qbr_b"
		lr_mult: 2
		decay_mult: 0
	}
	inner_product_param {
		num_output: 70
	}
}
layer {
	name: "PhiR_8"
	type: "InnerProduct"
	bottom: "qr7"
	top: "q8r"
	param {
		name: "qr_w"
		lr_mult: 1
		decay_mult: 1
	}
	param {
		name: "qr_b"
		lr_mult: 2
		decay_mult: 0
	}
	inner_product_param {
		num_output: 70
	}
}
layer {
	name: "QSum_8"
	type: "Eltwise"
	bottom: "qar8"
	bottom: "qbr8"
	bottom: "q8r"
	top: "qr8"
	eltwise_param { operation: SUM }
}

layer {
	name: "loss3"
	type: "SoftmaxWithLoss"
	top: "loss3"
	bottom: "qr8"
	bottom: "labels"
	include {
		phase: TRAIN
	}
	loss_weight: 1
}
layer {
	name: "softmax"
	type: "Softmax"
	bottom: "qr8"
	top: "pred"
	include {
		phase: TEST
	}
}
layer {
	name: "accuracy"
	type: "Accuracy"
	bottom: "pred"
	bottom: "labels"
	top: "top-1"
	include {
		phase: TEST
	}
	accuracy_param {
		top_k: 1
	}
}
layer {
	name: "accuracy_top5"
	type: "Accuracy"
	bottom: "pred"
	bottom: "labels"
	top: "top-5"
	include {
		phase: TEST
	}
	accuracy_param {
		top_k: 5
	}
}


================================================
FILE: prototxts/drnet_8units_relu_shareweight.prototxt
================================================
layer {
        name: "data"
        type: "Python"
        top: "qa"
        top: "qb"
        top: "im"
        top: "posdata"
        top: "labels"
        include {
                phase: TRAIN     
        }         
        python_param {                 
                module: 'rel_data_layer.layer'              
                layer: 'RelDataLayer'                 
                param_str: '{"dataset": "reltrain.json", "batch_size": 25, "nclass": 100}'   
        }         
}                    
layer {           
        name: "data"                   
        type: "Python"                 
        top: "qa" 
        top: "qb" 
        top: "im" 
        top: "posdata"                 
        top: "labels"                  
        include { 
                phase: TEST            
        }         
        python_param {                 
                module: 'rel_data_layer.layer'              
                layer: 'RelDataLayer'
                param_str: '{"dataset": "reltest.json", "batch_size": 25, "nclass": 100}'
        }         
}
# Appearance Subnet
layer {
	name: "conv1_1"
	type: "Convolution"
	bottom: "im"
	top: "conv1_1"
	param {
		lr_mult: 0
		decay_mult: 0
	} 
	param {
		lr_mult: 0
		decay_mult: 0
	}
	convolution_param {
		num_output: 64
		pad: 1
		kernel_size: 3
	}
}
layer {
	name: "relu1_1"
	type: "ReLU"
	bottom: "conv1_1"
	top: "conv1_1"
}
layer {
	name: "conv1_2"
	type: "Convolution"
	bottom: "conv1_1"
	top: "conv1_2"
	param {
		lr_mult: 0
		decay_mult: 0
	}
	param {
		lr_mult: 0
		decay_mult: 0
	}
	convolution_param {
		num_output: 64
		pad: 1
		kernel_size: 3
	}
}
layer {
	name: "relu1_2"
	type: "ReLU"
	bottom: "conv1_2"
	top: "conv1_2"
}
layer {
	name: "pool1"
	type: "Pooling"
	bottom: "conv1_2"
	top: "pool1"
	pooling_param {
		pool: MAX
		kernel_size: 2
		stride: 2
	}
}
layer {
	name: "conv2_1"
	type: "Convolution"
	bottom: "pool1"
	top: "conv2_1"
	param {
		lr_mult: 0
		decay_mult: 0
	}
	param {
		lr_mult: 0
		decay_mult: 0
	}
	convolution_param {
		num_output: 128
		pad: 1
		kernel_size: 3
	}
}
layer {
	name: "relu2_1"
	type: "ReLU"
	bottom: "conv2_1"
	top: "conv2_1"
}
layer {
	name: "conv2_2"
	type: "Convolution"
	bottom: "conv2_1"
	top: "conv2_2"
	param {
		lr_mult: 0
		decay_mult: 0
	}
	param {
		lr_mult: 0
		decay_mult: 0
	}
	convolution_param {
		num_output: 128
		pad: 1
		kernel_size: 3
	}
}
layer {
	name: "relu2_2"
	type: "ReLU"
	bottom: "conv2_2"
	top: "conv2_2"
}
layer {
	name: "pool2"	
	type: "Pooling"
	bottom: "conv2_2"
	top: "pool2"
	pooling_param {
		pool: MAX
		kernel_size: 2
		stride: 2
	}
}
layer {
	name: "conv3_1"
	type: "Convolution"
	bottom: "pool2"
	top: "conv3_1"
	param {
		lr_mult: 0
		decay_mult: 0
	}
	param {
		lr_mult: 0
		decay_mult: 0
	}
	convolution_param {
		num_output: 256
		pad: 1
		kernel_size: 3
	}
}
layer {
	name: "relu3_1"
	type: "ReLU"
	bottom: "conv3_1"
	top: "conv3_1"
}
layer {
	name: "conv3_2"
	type: "Convolution"
	bottom: "conv3_1"
	top: "conv3_2"
	param {
		lr_mult: 0
		decay_mult: 0
	}
	param {
		lr_mult: 2
		decay_mult: 0
	}
	convolution_param {
		num_output: 256
		pad: 1
		kernel_size: 3
	}
}
layer {
	name: "relu3_2"
	type: "ReLU"
	bottom: "conv3_2"
	top: "conv3_2"
}
layer {
	name: "conv3_3"
	type: "Convolution"
	bottom: "conv3_2"
	top: "conv3_3"
	param {
		lr_mult: 0
		decay_mult: 0
	}
	param {
		lr_mult: 0
		decay_mult: 0
	}
	convolution_param {
		num_output: 256
		pad: 1
		kernel_size: 3
	}
}
layer {
	name: "relu3_3"
	type: "ReLU"
	bottom: "conv3_3"
	top: "conv3_3"
}
layer {
	name: "pool3"
	type: "Pooling"
	bottom: "conv3_3"
	top: "pool3"
	pooling_param {
		pool: MAX
		kernel_size: 2
		stride: 2
	}
}
layer {
	name: "conv4_1"
	type: "Convolution"
	bottom: "pool3"
	top: "conv4_1"
	param {
		lr_mult: 0
		decay_mult: 0
	}
	param {
		lr_mult: 0
		decay_mult: 0
	}
	convolution_param {
		num_output: 512
		pad: 1
		kernel_size: 3
	}
}
layer {
	name: "relu4_1"
	type: "ReLU"
	bottom: "conv4_1"
	top: "conv4_1"
}
layer {
	name: "conv4_2"
	type: "Convolution"
	bottom: "conv4_1"
	top: "conv4_2"
	param {
		lr_mult: 0
		decay_mult: 0
	}
	param {
		lr_mult: 0
		decay_mult: 0
	}
	convolution_param {
		num_output: 512
		pad: 1
		kernel_size: 3
	}
}
layer {
	name: "relu4_2"
	type: "ReLU"
	bottom: "conv4_2"
	top: "conv4_2"
}
layer {
	name: "conv4_3"
	type: "Convolution"
	bottom: "conv4_2"
	top: "conv4_3"
	param {
		lr_mult: 0
		decay_mult: 0
	}
	param {
		lr_mult: 0
		decay_mult: 0
	}
	convolution_param {
		num_output: 512
		pad: 1
		kernel_size: 3
	}
}
layer {
	name: "relu4_3"
	type: "ReLU"
	bottom: "conv4_3"
	top: "conv4_3"
}
layer {
	name: "pool4"
	type: "Pooling"
	bottom: "conv4_3"
	top: "pool4"
	pooling_param {
		pool: MAX
		kernel_size: 2
		stride: 2
	}
}
layer {
	name: "conv5_1"
	type: "Convolution"
	bottom: "pool4"
	top: "conv5_1"
	param {
		lr_mult: 0
		decay_mult: 0
	}
	param {
		lr_mult: 0
		decay_mult: 0
	}
	convolution_param {
		num_output: 512
		pad: 1
		kernel_size: 3
	}
}
layer {
	name: "relu5_1"
	type: "ReLU"
	bottom: "conv5_1"
	top: "conv5_1"
}
layer {
	name: "conv5_2"
	type: "Convolution"
	bottom: "conv5_1"
	top: "conv5_2"
	param {
		lr_mult: 0
		decay_mult: 0
	}
	param {
		lr_mult: 0
		decay_mult: 0
	}
	convolution_param {
		num_output: 512
		pad: 1
		kernel_size: 3
	}
}
layer {
	name: "relu5_2"
	type: "ReLU"
	bottom: "conv5_2"
	top: "conv5_2"
}
layer {
	name: "conv5_3"
	type: "Convolution"
	bottom: "conv5_2"
	top: "conv5_3"
	param {
		lr_mult: 0
		decay_mult: 0
	}
	param {
		lr_mult: 0
		decay_mult: 0
	}
	convolution_param {
		num_output: 512
		pad: 1
		kernel_size: 3
	}
}
layer {
	name: "relu5_3"
	type: "ReLU"
	bottom: "conv5_3"
	top: "conv5_3"
}
layer {
	name: "pool5"
	type: "Pooling"
	bottom: "conv5_3"
	top: "pool5"
	pooling_param {
		pool: MAX
		kernel_size: 2
		stride: 2
	}
}
layer {
	name: "fc6"
	type: "InnerProduct"
	bottom: "pool5"
	top: "fc6"
	param {
		lr_mult: 0
		decay_mult: 0
	}
	param {
		lr_mult: 0
		decay_mult: 0
	}
	inner_product_param {
		num_output: 4096
	}
}
layer {
	name: "relu6"
	type: "ReLU"
	bottom: "fc6"
	top: "fc6"
}
layer {
	name: "fc7"
	type: "InnerProduct"
	bottom: "fc6"
	top: "fc7"
	param {
		lr_mult: 0
		decay_mult: 0
	}
	param {
		lr_mult: 0
		decay_mult: 0
	}
	inner_product_param {
		num_output: 4096
	}
}
layer {
	name: "relu7"
	type: "ReLU"
	bottom: "fc7"
	top: "fc7"
}
layer {
	name: "fc8"
	type: "InnerProduct"
	bottom: "fc7"
	top: "fc8"
	param {
		lr_mult: 0
		decay_mult: 0
	}
	param {
		lr_mult: 0
		decay_mult: 0
	}
	inner_product_param {
		num_output: 256
		weight_filler {
			type: "xavier"
		}
		bias_filler {
			type: "constant"
			value: 0
		}
	}
}
layer {
	name: "relu8"
	type: "ReLU"
	bottom: "fc8"
	top: "fc8"
}

# Spatial Cfg Subnet
layer {
	name: "conv1_p"
	type: "Convolution"
	bottom: "posdata"
	top: "conv1_p"
	param {
		lr_mult: 0
		decay_mult: 0
	}
	param {
		lr_mult: 0
		decay_mult: 0
	}
	convolution_param {
		num_output: 96
		kernel_size: 5
		pad: 2
		stride: 2
	}
}
layer {
	name: "relu1_p"
	type: "ReLU"
	bottom: "conv1_p"
	top: "conv1_p"
}
layer {
	name: "conv2_p"
	type: "Convolution"
	bottom: "conv1_p"
	top: "conv2_p"
	param {
		lr_mult: 0
		decay_mult: 0
	}
	param {
		lr_mult: 0
		decay_mult: 0
	}
	convolution_param {
		num_output: 128
		kernel_size: 5
		pad: 2
		stride: 2
	}
}
layer {
	name: "conv3_p"
	type: "Convolution"
	bottom: "conv2_p"
	top: "conv3_p"
	param {
		lr_mult: 0
		decay_mult: 0
	}
	param {
		lr_mult: 0
		decay_mult: 0
	}
	convolution_param {
		num_output: 64
		kernel_size: 8
	}
}
layer {
	name: "relu3_p"
	type: "ReLU"
	bottom: "conv3_p"
	top: "conv3_p"
}
# Combine features from subnets
layer {
	name: "concat1_c"
	type: "Python"
	bottom: "fc8"
	bottom: "conv3_p"
	top: "concat1_c"
	python_param {
		module: "customize_layers.concat_layer"
		layer: "Layer"
	}
}
layer {
	name: "fc2_c"
	type: "InnerProduct"
	bottom: "concat1_c"
	top: "fc2_c"
	param {
		lr_mult: 0
		decay_mult: 0
	}
	param {
		lr_mult: 0
		decay_mult: 0
	}
	inner_product_param {
		num_output: 128
	}
}
layer {
	name: "relu2_c"
	type: "ReLU"
	bottom: "fc2_c"
	top: "fc2_c"
}
layer {
	name: "PhiR_0"
	type: "InnerProduct"
	bottom: "fc2_c"
	top: "q0r"
	param {
		lr_mult: 0
		decay_mult: 0
	}
	param {
		lr_mult: 0
		decay_mult: 0
	}
	inner_product_param {
		num_output: 70
		weight_filler {
			type: "xavier"
		}
		bias_filler {
			type: "constant"
			value: 0
		}
	}
}
layer {
	name: "relu_0"
	type: "ReLU"
	bottom: "q0r"
	top: "qr0"
}
#DR-Net
layer {
	name: "PhiA_1"
	type: "InnerProduct"
	bottom: "qa"
	top: "qar1"
	param {
		name: "qar_w"
		lr_mult: 1
		decay_mult: 1
	}
	param {
		name: "qar_b"
		lr_mult: 2
		decay_mult: 0
	}
	inner_product_param {
		num_output: 70
		weight_filler {
			type: "xavier"
		}
		bias_filler {
			type: "constant"
			value: 0
		}
	}
}
layer {
	name: "PhiB_1"
	type: "InnerProduct"
	bottom: "qb"
	top: "qbr1"
	param {
		name: "qbr_w"
		lr_mult: 1
		decay_mult: 1
	}
	param {
		name: "qbr_b"
		lr_mult: 2
		decay_mult: 0
	}
	inner_product_param {
		num_output: 70
		weight_filler {
			type: "xavier"
		}
		bias_filler {
			type: "constant"
			value: 0
		}
	}
}
layer {
	name: "PhiR_1"
	type: "InnerProduct"
	bottom: "qr0"
	top: "q1r"
	param {
		name: "qr_w"
		lr_mult: 1
		decay_mult: 1
	}
	param {	
		name: "qr_b"
		lr_mult: 2
		decay_mult: 0
	}
	inner_product_param {
		num_output: 70
		weight_filler {
			type: "xavier"
		}
		bias_filler {
			type: "constant"
			value: 0
		}
	}
}
layer {
	name: "QSum_1"
	type: "Eltwise"
	bottom: "qar1"
	bottom: "qbr1"
	bottom: "q1r"
	top: "qr1un"
	eltwise_param { operation: SUM }
}
layer {
	name: "relu_1"
	type: "ReLU"
	bottom: "qr1un"
	top: "qr1"
}
layer {
	name: "PhiA_2"
	type: "InnerProduct"
	bottom: "qa"
	top: "qar2"
	param {
		name: "qar_w"
		lr_mult: 1
		decay_mult: 1
	}
	param {
		name: "qar_b"
		lr_mult: 2
		decay_mult: 0
	}
	inner_product_param {
		num_output: 70
	}
}
layer {
	name: "PhiB_2"
	type: "InnerProduct"
	bottom: "qb"
	top: "qbr2"
	param {
		name: "qbr_w"
		lr_mult: 1
		decay_mult: 1
	}
	param {
		name: "qbr_b"
		lr_mult: 2
		decay_mult: 0
	}
	inner_product_param {
		num_output: 70
	}
}
layer {
	name: "PhiR_2"
	type: "InnerProduct"
	bottom: "qr1"
	top: "q2r"
	param {
		name: "qr_w"
		lr_mult: 1
		decay_mult: 1
	}
	param {
		name: "qr_b"
		lr_mult: 2
		decay_mult: 0
	}
	inner_product_param {
		num_output: 70
	}
}
layer {
	name: "QSum_2"
	type: "Eltwise"
	bottom: "qar2"
	bottom: "qbr2"
	bottom: "q2r"
	top: "qr2un"
	eltwise_param { operation: SUM }
}
layer {
	name: "relu_2"
	type: "ReLU"
	bottom: "qr2un"
	top: "qr2"
}
layer {
	name: "PhiA_3"
	type: "InnerProduct"
	bottom: "qa"
	top: "qar3"
	param {
		name: "qar_w"
		lr_mult: 1
		decay_mult: 1
	}
	param {
		name: "qar_b"
		lr_mult: 2
		decay_mult: 0
	}
	inner_product_param {
		num_output: 70
	}
}
layer {
	name: "PhiB_3"
	type: "InnerProduct"
	bottom: "qb"
	top: "qbr3"
	param {
		name: "qbr_w"
		lr_mult: 1
		decay_mult: 1
	}
	param {
		name: "qbr_b"
		lr_mult: 2
		decay_mult: 0
	}
	inner_product_param {
		num_output: 70
	}
}
layer {
	name: "PhiR_3"
	type: "InnerProduct"
	bottom: "qr2"
	top: "q3r"
	param {
		name: "qr_w"
		lr_mult: 1
		decay_mult: 1
	}
	param {
		name: "qr_b"
		lr_mult: 2
		decay_mult: 0
	}
	inner_product_param {
		num_output: 70
	}
}
layer {
	name: "QSum_3"
	type: "Eltwise"
	bottom: "qar3"
	bottom: "qbr3"
	bottom: "q3r"
	top: "qr3un"
	eltwise_param { operation: SUM }
}
layer {
	name: "relu_3"
	type: "ReLU"
	bottom: "qr3un"
	top: "qr3"
}
layer {
	name: "PhiA_4"
	type: "InnerProduct"
	bottom: "qa"
	top: "qar4"
	param {
		name: "qar_w"
		lr_mult: 1
		decay_mult: 1
	}
	param {
		name: "qar_b"
		lr_mult: 2
		decay_mult: 0
	}
	inner_product_param {
		num_output: 70
	}
}
layer {
	name: "PhiB_4"
	type: "InnerProduct"
	bottom: "qb"
	top: "qbr4"
	param {
		name: "qbr_w"
		lr_mult: 1
		decay_mult: 1
	}
	param {
		name: "qbr_b"
		lr_mult: 2
		decay_mult: 0
	}
	inner_product_param {
		num_output: 70
	}
}
layer {
	name: "PhiR_4"
	type: "InnerProduct"
	bottom: "qr3"
	top: "q4r"
	param {
		name: "qr_w"
		lr_mult: 1
		decay_mult: 1
	}
	param {
		name: "qr_b"
		lr_mult: 2
		decay_mult: 0
	}
	inner_product_param {
		num_output: 70
	}
}
layer {
	name: "QSum_4"
	type: "Eltwise"
	bottom: "qar4"
	bottom: "qbr4"
	bottom: "q4r"
	top: "qr4un"
	eltwise_param { operation: SUM }
}
layer {
	name: "relu_4"
	type: "ReLU"
	bottom: "qr4un"
	top: "qr4"
}
layer {
	name: "PhiA_5"
	type: "InnerProduct"
	bottom: "qa"
	top: "qar5"
	param {
		name: "qar_w"
		lr_mult: 1
		decay_mult: 1
	}
	param {
		name: "qar_b"
		lr_mult: 2
		decay_mult: 0
	}
	inner_product_param {
		num_output: 70
	}
}
layer {
	name: "PhiB_5"
	type: "InnerProduct"
	bottom: "qb"
	top: "qbr5"
	param {
		name: "qbr_w"
		lr_mult: 1
		decay_mult: 1
	}
	param {
		name: "qbr_b"
		lr_mult: 2
		decay_mult: 0
	}
	inner_product_param {
		num_output: 70
	}
}
layer {
	name: "PhiR_5"
	type: "InnerProduct"
	bottom: "qr4"
	top: "q5r"
	param {
		name: "qr_w"
		lr_mult: 1
		decay_mult: 1
	}
	param {
		name: "qr_b"
		lr_mult: 2
		decay_mult: 0
	}
	inner_product_param {
		num_output: 70
	}
}
layer {
	name: "QSum_5"
	type: "Eltwise"
	bottom: "qar5"
	bottom: "qbr5"
	bottom: "q5r"
	top: "qr5un"
	eltwise_param { operation: SUM }
}
layer {
	name: "relu_5"
	type: "ReLU"
	bottom: "qr5un"
	top: "qr5"
}

layer {
	name: "PhiA_6"
	type: "InnerProduct"
	bottom: "qa"
	top: "qar6"
	param {
		name: "qar_w"
		lr_mult: 1
		decay_mult: 1
	}
	param {
		name: "qar_b"
		lr_mult: 2
		decay_mult: 0
	}
	inner_product_param {
		num_output: 70
	}
}
layer {
	name: "PhiB_6"
	type: "InnerProduct"
	bottom: "qb"
	top: "qbr6"
	param {
		name: "qbr_w"
		lr_mult: 1
		decay_mult: 1
	}
	param {
		name: "qbr_b"
		lr_mult: 2
		decay_mult: 0
	}
	inner_product_param {
		num_output: 70
	}
}
layer {
	name: "PhiR_6"
	type: "InnerProduct"
	bottom: "qr5"
	top: "q6r"
	param {
		name: "qr_w"
		lr_mult: 1
		decay_mult: 1
	}
	param {
		name: "qr_b"
		lr_mult: 2
		decay_mult: 0
	}
	inner_product_param {
		num_output: 70
	}
}
layer {
	name: "QSum_6"
	type: "Eltwise"
	bottom: "qar6"
	bottom: "qbr6"
	bottom: "q6r"
	top: "qr6un"
	eltwise_param { operation: SUM }
}
layer {
	name: "relu_6"
	type: "ReLU"
	bottom: "qr6un"
	top: "qr6"
}
layer {
	name: "PhiA_7"
	type: "InnerProduct"
	bottom: "qa"
	top: "qar7"
	param {
		name: "qar_w"
		lr_mult: 1
		decay_mult: 1
	}
	param {
		name: "qar_b"
		lr_mult: 2
		decay_mult: 0
	}
	inner_product_param {
		num_output: 70
	}
}
layer {
	name: "PhiB_7"
	type: "InnerProduct"
	bottom: "qb"
	top: "qbr7"
	param {
		name: "qbr_w"
		lr_mult: 1
		decay_mult: 1
	}
	param {
		name: "qbr_b"
		lr_mult: 2
		decay_mult: 0
	}
	inner_product_param {
		num_output: 70
	}
}
layer {
	name: "PhiR_7"
	type: "InnerProduct"
	bottom: "qr6"
	top: "q7r"
	param {
		name: "qr_w"
		lr_mult: 1
		decay_mult: 1
	}
	param {
		name: "qr_b"
		lr_mult: 2
		decay_mult: 0
	}
	inner_product_param {
		num_output: 70
	}
}
layer {
	name: "QSum_7"
	type: "Eltwise"
	bottom: "qar7"
	bottom: "qbr7"
	bottom: "q7r"
	top: "qr7un"
	eltwise_param { operation: SUM }
}
layer {
	name: "relu_7"
	type: "ReLU"
	bottom: "qr7un"
	top: "qr7"
}
layer {
	name: "PhiA_8"
	type: "InnerProduct"
	bottom: "qa"
	top: "qar8"
	param {
		name: "qar_w"
		lr_mult: 1
		decay_mult: 1
	}
	param {
		name: "qar_b"
		lr_mult: 2
		decay_mult: 0
	}
	inner_product_param {
		num_output: 70
	}
}
layer {
	name: "PhiB_8"
	type: "InnerProduct"
	bottom: "qb"
	top: "qbr8"
	param {
		name: "qbr_w"
		lr_mult: 1
		decay_mult: 1
	}
	param {
		name: "qbr_b"
		lr_mult: 2
		decay_mult: 0
	}
	inner_product_param {
		num_output: 70
	}
}
layer {
	name: "PhiR_8"
	type: "InnerProduct"
	bottom: "qr7"
	top: "q8r"
	param {
		name: "qr_w"
		lr_mult: 1
		decay_mult: 1
	}
	param {
		name: "qr_b"
		lr_mult: 2
		decay_mult: 0
	}
	inner_product_param {
		num_output: 70
	}
}
layer {
	name: "QSum_8"
	type: "Eltwise"
	bottom: "qar8"
	bottom: "qbr8"
	bottom: "q8r"
	top: "qr8"
	eltwise_param { operation: SUM }
}
layer {
	name: "loss3"
	type: "SoftmaxWithLoss"
	top: "loss3"
	bottom: "qr8"
	bottom: "labels"
	include {
		phase: TRAIN
	}
	loss_weight: 1
}
layer {
	name: "softmax"
	type: "Softmax"
	bottom: "qr8"
	top: "pred"
	include {
		phase: TEST
	}
}
layer {
	name: "accuracy"
	type: "Accuracy"
	bottom: "pred"
	bottom: "labels"
	top: "top-1"
	include {
		phase: TEST
	}
	accuracy_param {
		top_k: 1
	}
}
layer {
	name: "accuracy_top5"
	type: "Accuracy"
	bottom: "pred"
	bottom: "labels"
	top: "top-5"
	include {
		phase: TEST
	}
	accuracy_param {
		top_k: 5
	}
}


================================================
FILE: prototxts/drnet_8units_softmax.prototxt
================================================
layer {
        name: "data"
        type: "Python"
        top: "qa"
        top: "qb"
        top: "im"
        top: "posdata"
        top: "labels"
        include {
                phase: TRAIN     
        }         
        python_param {                 
                module: 'rel_data_layer.layer'              
                layer: 'RelDataLayer'                 
                param_str: '{"dataset": "reltrain.json", "batch_size": 25, "nclass": 100}'   
        }         
}                    
layer {           
        name: "data"                   
        type: "Python"                 
        top: "qa" 
        top: "qb" 
        top: "im" 
        top: "posdata"                 
        top: "labels"                  
        include { 
                phase: TEST            
        }         
        python_param {                 
                module: 'rel_data_layer.layer'              
                layer: 'RelDataLayer'
                param_str: '{"dataset": "reltest.json", "batch_size": 25, "nclass": 100}'
        }         
}
# Appearance Subnet 
layer {
	name: "conv1_1"
	type: "Convolution"
	bottom: "im"
	top: "conv1_1"
	param {
		lr_mult: 0
		decay_mult: 0
	} 
	param {
		lr_mult: 0
		decay_mult: 0
	}
	convolution_param {
		num_output: 64
		pad: 1
		kernel_size: 3
	}
}
layer {
	name: "relu1_1"
	type: "ReLU"
	bottom: "conv1_1"
	top: "conv1_1"
}
layer {
	name: "conv1_2"
	type: "Convolution"
	bottom: "conv1_1"
	top: "conv1_2"
	param {
		lr_mult: 0
		decay_mult: 0
	}
	param {
		lr_mult: 0
		decay_mult: 0
	}
	convolution_param {
		num_output: 64
		pad: 1
		kernel_size: 3
	}
}
layer {
	name: "relu1_2"
	type: "ReLU"
	bottom: "conv1_2"
	top: "conv1_2"
}
layer {
	name: "pool1"
	type: "Pooling"
	bottom: "conv1_2"
	top: "pool1"
	pooling_param {
		pool: MAX
		kernel_size: 2
		stride: 2
	}
}
layer {
	name: "conv2_1"
	type: "Convolution"
	bottom: "pool1"
	top: "conv2_1"
	param {
		lr_mult: 0
		decay_mult: 0
	}
	param {
		lr_mult: 0
		decay_mult: 0
	}
	convolution_param {
		num_output: 128
		pad: 1
		kernel_size: 3
	}
}
layer {
	name: "relu2_1"
	type: "ReLU"
	bottom: "conv2_1"
	top: "conv2_1"
}
layer {
	name: "conv2_2"
	type: "Convolution"
	bottom: "conv2_1"
	top: "conv2_2"
	param {
		lr_mult: 0
		decay_mult: 0
	}
	param {
		lr_mult: 0
		decay_mult: 0
	}
	convolution_param {
		num_output: 128
		pad: 1
		kernel_size: 3
	}
}
layer {
	name: "relu2_2"
	type: "ReLU"
	bottom: "conv2_2"
	top: "conv2_2"
}
layer {
	name: "pool2"	
	type: "Pooling"
	bottom: "conv2_2"
	top: "pool2"
	pooling_param {
		pool: MAX
		kernel_size: 2
		stride: 2
	}
}
layer {
	name: "conv3_1"
	type: "Convolution"
	bottom: "pool2"
	top: "conv3_1"
	param {
		lr_mult: 0
		decay_mult: 0
	}
	param {
		lr_mult: 0
		decay_mult: 0
	}
	convolution_param {
		num_output: 256
		pad: 1
		kernel_size: 3
	}
}
layer {
	name: "relu3_1"
	type: "ReLU"
	bottom: "conv3_1"
	top: "conv3_1"
}
layer {
	name: "conv3_2"
	type: "Convolution"
	bottom: "conv3_1"
	top: "conv3_2"
	param {
		lr_mult: 0
		decay_mult: 0
	}
	param {
		lr_mult: 2
		decay_mult: 0
	}
	convolution_param {
		num_output: 256
		pad: 1
		kernel_size: 3
	}
}
layer {
	name: "relu3_2"
	type: "ReLU"
	bottom: "conv3_2"
	top: "conv3_2"
}
layer {
	name: "conv3_3"
	type: "Convolution"
	bottom: "conv3_2"
	top: "conv3_3"
	param {
		lr_mult: 0
		decay_mult: 0
	}
	param {
		lr_mult: 0
		decay_mult: 0
	}
	convolution_param {
		num_output: 256
		pad: 1
		kernel_size: 3
	}
}
layer {
	name: "relu3_3"
	type: "ReLU"
	bottom: "conv3_3"
	top: "conv3_3"
}
layer {
	name: "pool3"
	type: "Pooling"
	bottom: "conv3_3"
	top: "pool3"
	pooling_param {
		pool: MAX
		kernel_size: 2
		stride: 2
	}
}
layer {
	name: "conv4_1"
	type: "Convolution"
	bottom: "pool3"
	top: "conv4_1"
	param {
		lr_mult: 0
		decay_mult: 0
	}
	param {
		lr_mult: 0
		decay_mult: 0
	}
	convolution_param {
		num_output: 512
		pad: 1
		kernel_size: 3
	}
}
layer {
	name: "relu4_1"
	type: "ReLU"
	bottom: "conv4_1"
	top: "conv4_1"
}
layer {
	name: "conv4_2"
	type: "Convolution"
	bottom: "conv4_1"
	top: "conv4_2"
	param {
		lr_mult: 0
		decay_mult: 0
	}
	param {
		lr_mult: 0
		decay_mult: 0
	}
	convolution_param {
		num_output: 512
		pad: 1
		kernel_size: 3
	}
}
layer {
	name: "relu4_2"
	type: "ReLU"
	bottom: "conv4_2"
	top: "conv4_2"
}
layer {
	name: "conv4_3"
	type: "Convolution"
	bottom: "conv4_2"
	top: "conv4_3"
	param {
		lr_mult: 0
		decay_mult: 0
	}
	param {
		lr_mult: 0
		decay_mult: 0
	}
	convolution_param {
		num_output: 512
		pad: 1
		kernel_size: 3
	}
}
layer {
	name: "relu4_3"
	type: "ReLU"
	bottom: "conv4_3"
	top: "conv4_3"
}
layer {
	name: "pool4"
	type: "Pooling"
	bottom: "conv4_3"
	top: "pool4"
	pooling_param {
		pool: MAX
		kernel_size: 2
		stride: 2
	}
}
layer {
	name: "conv5_1"
	type: "Convolution"
	bottom: "pool4"
	top: "conv5_1"
	param {
		lr_mult: 0
		decay_mult: 0
	}
	param {
		lr_mult: 0
		decay_mult: 0
	}
	convolution_param {
		num_output: 512
		pad: 1
		kernel_size: 3
	}
}
layer {
	name: "relu5_1"
	type: "ReLU"
	bottom: "conv5_1"
	top: "conv5_1"
}
layer {
	name: "conv5_2"
	type: "Convolution"
	bottom: "conv5_1"
	top: "conv5_2"
	param {
		lr_mult: 0
		decay_mult: 0
	}
	param {
		lr_mult: 0
		decay_mult: 0
	}
	convolution_param {
		num_output: 512
		pad: 1
		kernel_size: 3
	}
}
layer {
	name: "relu5_2"
	type: "ReLU"
	bottom: "conv5_2"
	top: "conv5_2"
}
layer {
	name: "conv5_3"
	type: "Convolution"
	bottom: "conv5_2"
	top: "conv5_3"
	param {
		lr_mult: 0
		decay_mult: 0
	}
	param {
		lr_mult: 0
		decay_mult: 0
	}
	convolution_param {
		num_output: 512
		pad: 1
		kernel_size: 3
	}
}
layer {
	name: "relu5_3"
	type: "ReLU"
	bottom: "conv5_3"
	top: "conv5_3"
}
layer {
	name: "pool5"
	type: "Pooling"
	bottom: "conv5_3"
	top: "pool5"
	pooling_param {
		pool: MAX
		kernel_size: 2
		stride: 2
	}
}
layer {
	name: "fc6"
	type: "InnerProduct"
	bottom: "pool5"
	top: "fc6"
	param {
		lr_mult: 0
		decay_mult: 0
	}
	param {
		lr_mult: 0
		decay_mult: 0
	}
	inner_product_param {
		num_output: 4096
	}
}
layer {
	name: "relu6"
	type: "ReLU"
	bottom: "fc6"
	top: "fc6"
}
layer {
	name: "fc7"
	type: "InnerProduct"
	bottom: "fc6"
	top: "fc7"
	param {
		lr_mult: 0
		decay_mult: 0
	}
	param {
		lr_mult: 0
		decay_mult: 0
	}
	inner_product_param {
		num_output: 4096
	}
}
layer {
	name: "relu7"
	type: "ReLU"
	bottom: "fc7"
	top: "fc7"
}
layer {
	name: "fc8"
	type: "InnerProduct"
	bottom: "fc7"
	top: "fc8"
	param {
		lr_mult: 0
		decay_mult: 0
	}
	param {
		lr_mult: 0
		decay_mult: 0
	}
	inner_product_param {
		num_output: 256
		weight_filler {
			type: "xavier"
		}
		bias_filler {
			type: "constant"
			value: 0
		}
	}
}
layer {
	name: "relu8"
	type: "ReLU"
	bottom: "fc8"
	top: "fc8"
}

# Spatial Cfg Subnet
layer {
	name: "conv1_p"
	type: "Convolution"
	bottom: "posdata"
	top: "conv1_p"
	param {
		lr_mult: 0
		decay_mult: 0
	}
	param {
		lr_mult: 0
		decay_mult: 0
	}
	convolution_param {
		num_output: 96
		kernel_size: 5
		pad: 2
		stride: 2
	}
}
layer {
	name: "relu1_p"
	type: "ReLU"
	bottom: "conv1_p"
	top: "conv1_p"
}
layer {
	name: "conv2_p"
	type: "Convolution"
	bottom: "conv1_p"
	top: "conv2_p"
	param {
		lr_mult: 0
		decay_mult: 0
	}
	param {
		lr_mult: 0
		decay_mult: 0
	}
	convolution_param {
		num_output: 128
		kernel_size: 5
		pad: 2
		stride: 2
	}
}
layer {
	name: "conv3_p"
	type: "Convolution"
	bottom: "conv2_p"
	top: "conv3_p"
	param {
		lr_mult: 0
		decay_mult: 0
	}
	param {
		lr_mult: 0
		decay_mult: 0
	}
	convolution_param {
		num_output: 64
		kernel_size: 8
	}
}
layer {
	name: "relu3_p"
	type: "ReLU"
	bottom: "conv3_p"
	top: "conv3_p"
}

# Combine features from subnets
layer {
	name: "concat1_c"
	type: "Python"
	bottom: "fc8"
	bottom: "conv3_p"
	top: "concat1_c"
	python_param {
		module: "customize_layers.concat_layer"
		layer: "Layer"
	}
}
layer {
	name: "fc2_c"
	type: "InnerProduct"
	bottom: "concat1_c"
	top: "fc2_c"
	param {
		lr_mult: 0
		decay_mult: 0
	}
	param {
		lr_mult: 0
		decay_mult: 0
	}
	inner_product_param {
		num_output: 128
	}
}
layer {
	name: "relu2_c"
	type: "ReLU"
	bottom: "fc2_c"
	top: "fc2_c"
}
layer {
	name: "PhiR_0"
	type: "InnerProduct"
	bottom: "fc2_c"
	top: "q0r"
	param {
		lr_mult: 0
		decay_mult: 0
	}
	param {
		lr_mult: 0
		decay_mult: 0
	}
	inner_product_param {
		num_output: 70
		weight_filler {
			type: "xavier"
		}
		bias_filler {
			type: "constant"
			value: 0
		}
	}
}
layer {
	name: "Softmax_0"
	type: "Softmax"
	bottom: "q0r"
	top: "qr0"
}

#DR-Net
layer {
	name: "PhiA_1"
	type: "InnerProduct"
	bottom: "qa"
	top: "qar1"
	param {
		lr_mult: 1
		decay_mult: 1
	}
	param {
		lr_mult: 2
		decay_mult: 0
	}
	inner_product_param {
		num_output: 70
		weight_filler {
			type: "xavier"
		}
		bias_filler {
			type: "constant"
			value: 0
		}
	}
}
layer {
	name: "PhiB_1"
	type: "InnerProduct"
	bottom: "qb"
	top: "qbr1"
	param {
		lr_mult: 1
		decay_mult: 1
	}
	param {
		lr_mult: 2
		decay_mult: 0
	}
	inner_product_param {
		num_output: 70
		weight_filler {
			type: "xavier"
		}
		bias_filler {
			type: "constant"
			value: 0
		}
	}
}
layer {
	name: "PhiR_1"
	type: "InnerProduct"
	bottom: "qr0"
	top: "q1r"
	param {
		lr_mult: 1
		decay_mult: 1
	}
	param {	
		lr_mult: 2
		decay_mult: 0
	}
	inner_product_param {
		num_output: 70
		weight_filler {
			type: "xavier"
		}
		bias_filler {
			type: "constant"
			value: 0
		}
	}
}
layer {
	name: "QSum_1"
	type: "Eltwise"
	bottom: "qar1"
	bottom: "qbr1"
	bottom: "q1r"
	top: "qr1un"
	eltwise_param { operation: SUM }
}
layer {
	name: "Softmax_1"
	type: "Softmax"
	bottom: "qr1un"
	top: "qr1"
}
layer {
	name: "PhiA_2"
	type: "InnerProduct"
	bottom: "qa"
	top: "qar2"
	param {
		lr_mult: 1
		decay_mult: 1
	}
	param {
		lr_mult: 2
		decay_mult: 0
	}
	inner_product_param {
		num_output: 70
	}
}
layer {
	name: "PhiB_2"
	type: "InnerProduct"
	bottom: "qb"
	top: "qbr2"
	param {
		lr_mult: 1
		decay_mult: 1
	}
	param {
		lr_mult: 2
		decay_mult: 0
	}
	inner_product_param {
		num_output: 70
	}
}
layer {
	name: "PhiR_2"
	type: "InnerProduct"
	bottom: "qr1"
	top: "q2r"
	param {
		lr_mult: 1
		decay_mult: 1
	}
	param {
		lr_mult: 2
		decay_mult: 0
	}
	inner_product_param {
		num_output: 70
	}
}
layer {
	name: "QSum_2"
	type: "Eltwise"
	bottom: "qar2"
	bottom: "qbr2"
	bottom: "q2r"
	top: "qr2un"
	eltwise_param { operation: SUM }
}
layer {
	name: "Softmax_2"
	type: "Softmax"
	bottom: "qr2un"
	top: "qr2"
}
layer {
	name: "PhiA_3"
	type: "InnerProduct"
	bottom: "qa"
	top: "qar3"
	param {
		lr_mult: 1
		decay_mult: 1
	}
	param {
		lr_mult: 2
		decay_mult: 0
	}
	inner_product_param {
		num_output: 70
	}
}
layer {
	name: "PhiB_3"
	type: "InnerProduct"
	bottom: "qb"
	top: "qbr3"
	param {
		lr_mult: 1
		decay_mult: 1
	}
	param {
		lr_mult: 2
		decay_mult: 0
	}
	inner_product_param {
		num_output: 70
	}
}
layer {
	name: "PhiR_3"
	type: "InnerProduct"
	bottom: "qr2"
	top: "q3r"
	param {
		lr_mult: 1
		decay_mult: 1
	}
	param {
		lr_mult: 2
		decay_mult: 0
	}
	inner_product_param {
		num_output: 70
	}
}
layer {
	name: "QSum_3"
	type: "Eltwise"
	bottom: "qar3"
	bottom: "qbr3"
	bottom: "q3r"
	top: "qr3un"
	eltwise_param { operation: SUM }
}
layer {
	name: "Softmax_3"
	type: "Softmax"
	bottom: "qr3un"
	top: "qr3"
}
layer {
	name: "PhiA_4"
	type: "InnerProduct"
	bottom: "qa"
	top: "qar4"
	param {
		lr_mult: 1
		decay_mult: 1
	}
	param {
		lr_mult: 2
		decay_mult: 0
	}
	inner_product_param {
		num_output: 70
	}
}
layer {
	name: "PhiB_4"
	type: "InnerProduct"
	bottom: "qb"
	top: "qbr4"
	param {
		lr_mult: 1
		decay_mult: 1
	}
	param {
		lr_mult: 2
		decay_mult: 0
	}
	inner_product_param {
		num_output: 70
	}
}
layer {
	name: "PhiR_4"
	type: "InnerProduct"
	bottom: "qr3"
	top: "q4r"
	param {
		lr_mult: 1
		decay_mult: 1
	}
	param {
		lr_mult: 2
		decay_mult: 0
	}
	inner_product_param {
		num_output: 70
	}
}
layer {
	name: "QSum_4"
	type: "Eltwise"
	bottom: "qar4"
	bottom: "qbr4"
	bottom: "q4r"
	top: "qr4un"
	eltwise_param { operation: SUM }
}
layer {
	name: "Softmax_4"
	type: "Softmax"
	bottom: "qr4un"
	top: "qr4"
}
layer {
	name: "PhiA_5"
	type: "InnerProduct"
	bottom: "qa"
	top: "qar5"
	param {
		lr_mult: 1
		decay_mult: 1
	}
	param {
		lr_mult: 2
		decay_mult: 0
	}
	inner_product_param {
		num_output: 70
	}
}
layer {
	name: "PhiB_5"
	type: "InnerProduct"
	bottom: "qb"
	top: "qbr5"
	param {
		lr_mult: 1
		decay_mult: 1
	}
	param {
		lr_mult: 2
		decay_mult: 0
	}
	inner_product_param {
		num_output: 70
	}
}
layer {
	name: "PhiR_5"
	type: "InnerProduct"
	bottom: "qr4"
	top: "q5r"
	param {
		lr_mult: 1
		decay_mult: 1
	}
	param {
		lr_mult: 2
		decay_mult: 0
	}
	inner_product_param {
		num_output: 70
	}
}
layer {
	name: "QSum_5"
	type: "Eltwise"
	bottom: "qar5"
	bottom: "qbr5"
	bottom: "q5r"
	top: "qr5un"
	eltwise_param { operation: SUM }
}
layer {
	name: "Softmax_5"
	type: "Softmax"
	bottom: "qr5un"
	top: "qr5"
}

layer {
	name: "PhiA_6"
	type: "InnerProduct"
	bottom: "qa"
	top: "qar6"
	param {
		lr_mult: 1
		decay_mult: 1
	}
	param {
		lr_mult: 2
		decay_mult: 0
	}
	inner_product_param {
		num_output: 70
	}
}
layer {
	name: "PhiB_6"
	type: "InnerProduct"
	bottom: "qb"
	top: "qbr6"
	param {
		lr_mult: 1
		decay_mult: 1
	}
	param {
		lr_mult: 2
		decay_mult: 0
	}
	inner_product_param {
		num_output: 70
	}
}
layer {
	name: "PhiR_6"
	type: "InnerProduct"
	bottom: "qr5"
	top: "q6r"
	param {
		lr_mult: 1
		decay_mult: 1
	}
	param {
		lr_mult: 2
		decay_mult: 0
	}
	inner_product_param {
		num_output: 70
	}
}
layer {
	name: "QSum_6"
	type: "Eltwise"
	bottom: "qar6"
	bottom: "qbr6"
	bottom: "q6r"
	top: "qr6un"
	eltwise_param { operation: SUM }
}
layer {
	name: "Softmax_6"
	type: "Softmax"
	bottom: "qr6un"
	top: "qr6"
}
layer {
	name: "PhiA_7"
	type: "InnerProduct"
	bottom: "qa"
	top: "qar7"
	param {
		lr_mult: 1
		decay_mult: 1
	}
	param {
		lr_mult: 2
		decay_mult: 0
	}
	inner_product_param {
		num_output: 70
	}
}
layer {
	name: "PhiB_7"
	type: "InnerProduct"
	bottom: "qb"
	top: "qbr7"
	param {
		lr_mult: 1
		decay_mult: 1
	}
	param {
		lr_mult: 2
		decay_mult: 0
	}
	inner_product_param {
		num_output: 70
	}
}
layer {
	name: "PhiR_7"
	type: "InnerProduct"
	bottom: "qr6"
	top: "q7r"
	param {
		lr_mult: 1
		decay_mult: 1
	}
	param {
		lr_mult: 2
		decay_mult: 0
	}
	inner_product_param {
		num_output: 70
	}
}
layer {
	name: "QSum_7"
	type: "Eltwise"
	bottom: "qar7"
	bottom: "qbr7"
	bottom: "q7r"
	top: "qr7un"
	eltwise_param { operation: SUM }
}
layer {
	name: "Softmax_7"
	type: "Softmax"
	bottom: "qr7un"
	top: "qr7"
}
layer {
	name: "PhiA_8"
	type: "InnerProduct"
	bottom: "qa"
	top: "qar8"
	param {
		lr_mult: 1
		decay_mult: 1
	}
	param {
		lr_mult: 2
		decay_mult: 0
	}
	inner_product_param {
		num_output: 70
	}
}
layer {
	name: "PhiB_8"
	type: "InnerProduct"
	bottom: "qb"
	top: "qbr8"
	param {
		lr_mult: 1
		decay_mult: 1
	}
	param {
		lr_mult: 2
		decay_mult: 0
	}
	inner_product_param {
		num_output: 70
	}
}
layer {
	name: "PhiR_8"
	type: "InnerProduct"
	bottom: "qr7"
	top: "q8r"
	param {
		lr_mult: 1
		decay_mult: 1
	}
	param {
		lr_mult: 2
		decay_mult: 0
	}
	inner_product_param {
		num_output: 70
	}
}
layer {
	name: "QSum_8"
	type: "Eltwise"
	bottom: "qar8"
	bottom: "qbr8"
	bottom: "q8r"
	top: "qr8"
	eltwise_param { operation: SUM }
}
layer {
	name: "loss3"
	type: "SoftmaxWithLoss"
	top: "loss3"
	bottom: "qr8"
	bottom: "labels"
	include {
		phase: TRAIN
	}
	loss_weight: 1
}
layer {
	name: "softmax"
	type: "Softmax"
	bottom: "qr8"
	top: "pred"
	include {
		phase: TEST
	}
}
layer {
	name: "accuracy"
	type: "Accuracy"
	bottom: "pred"
	bottom: "labels"
	top: "top-1"
	include {
		phase: TEST
	}
	accuracy_param {
		top_k: 1
	}
}
layer {
	name: "accuracy_top5"
	type: "Accuracy"
	bottom: "pred"
	bottom: "labels"
	top: "top-5"
	include {
		phase: TEST
	}
	accuracy_param {
		top_k: 5
	}
}


================================================
FILE: prototxts/test_drnet_8units_linear_shareweight.prototxt
================================================
input: "qa"
input_shape {	
	dim: 1
	dim: 100
}

input: "qb"
input_shape {
	dim: 1
	dim: 100
}

input: "im"
input_shape {
	dim: 1
	dim: 3
	dim: 224
	dim: 224
}

input: "posdata"
input_shape {
	dim: 1
	dim: 2
	dim: 32
	dim: 32
}

# Appearance Subnet
layer {
	name: "conv1_1"
	type: "Convolution"
	bottom: "im"
	top: "conv1_1"
	param {
		lr_mult: 0
		decay_mult: 0
	} 
	param {
		lr_mult: 0
		decay_mult: 0
	}
	convolution_param {
		num_output: 64
		pad: 1
		kernel_size: 3
	}
}
layer {
	name: "relu1_1"
	type: "ReLU"
	bottom: "conv1_1"
	top: "conv1_1"
}
layer {
	name: "conv1_2"
	type: "Convolution"
	bottom: "conv1_1"
	top: "conv1_2"
	param {
		lr_mult: 0
		decay_mult: 0
	}
	param {
		lr_mult: 0
		decay_mult: 0
	}
	convolution_param {
		num_output: 64
		pad: 1
		kernel_size: 3
	}
}
layer {
	name: "relu1_2"
	type: "ReLU"
	bottom: "conv1_2"
	top: "conv1_2"
}
layer {
	name: "pool1"
	type: "Pooling"
	bottom: "conv1_2"
	top: "pool1"
	pooling_param {
		pool: MAX
		kernel_size: 2
		stride: 2
	}
}
layer {
	name: "conv2_1"
	type: "Convolution"
	bottom: "pool1"
	top: "conv2_1"
	param {
		lr_mult: 0
		decay_mult: 0
	}
	param {
		lr_mult: 0
		decay_mult: 0
	}
	convolution_param {
		num_output: 128
		pad: 1
		kernel_size: 3
	}
}
layer {
	name: "relu2_1"
	type: "ReLU"
	bottom: "conv2_1"
	top: "conv2_1"
}
layer {
	name: "conv2_2"
	type: "Convolution"
	bottom: "conv2_1"
	top: "conv2_2"
	param {
		lr_mult: 0
		decay_mult: 0
	}
	param {
		lr_mult: 0
		decay_mult: 0
	}
	convolution_param {
		num_output: 128
		pad: 1
		kernel_size: 3
	}
}
layer {
	name: "relu2_2"
	type: "ReLU"
	bottom: "conv2_2"
	top: "conv2_2"
}
layer {
	name: "pool2"	
	type: "Pooling"
	bottom: "conv2_2"
	top: "pool2"
	pooling_param {
		pool: MAX
		kernel_size: 2
		stride: 2
	}
}
layer {
	name: "conv3_1"
	type: "Convolution"
	bottom: "pool2"
	top: "conv3_1"
	param {
		lr_mult: 0
		decay_mult: 0
	}
	param {
		lr_mult: 0
		decay_mult: 0
	}
	convolution_param {
		num_output: 256
		pad: 1
		kernel_size: 3
	}
}
layer {
	name: "relu3_1"
	type: "ReLU"
	bottom: "conv3_1"
	top: "conv3_1"
}
layer {
	name: "conv3_2"
	type: "Convolution"
	bottom: "conv3_1"
	top: "conv3_2"
	param {
		lr_mult: 0
		decay_mult: 0
	}
	param {
		lr_mult: 2
		decay_mult: 0
	}
	convolution_param {
		num_output: 256
		pad: 1
		kernel_size: 3
	}
}
layer {
	name: "relu3_2"
	type: "ReLU"
	bottom: "conv3_2"
	top: "conv3_2"
}
layer {
	name: "conv3_3"
	type: "Convolution"
	bottom: "conv3_2"
	top: "conv3_3"
	param {
		lr_mult: 0
		decay_mult: 0
	}
	param {
		lr_mult: 0
		decay_mult: 0
	}
	convolution_param {
		num_output: 256
		pad: 1
		kernel_size: 3
	}
}
layer {
	name: "relu3_3"
	type: "ReLU"
	bottom: "conv3_3"
	top: "conv3_3"
}
layer {
	name: "pool3"
	type: "Pooling"
	bottom: "conv3_3"
	top: "pool3"
	pooling_param {
		pool: MAX
		kernel_size: 2
		stride: 2
	}
}
layer {
	name: "conv4_1"
	type: "Convolution"
	bottom: "pool3"
	top: "conv4_1"
	param {
		lr_mult: 0
		decay_mult: 0
	}
	param {
		lr_mult: 0
		decay_mult: 0
	}
	convolution_param {
		num_output: 512
		pad: 1
		kernel_size: 3
	}
}
layer {
	name: "relu4_1"
	type: "ReLU"
	bottom: "conv4_1"
	top: "conv4_1"
}
layer {
	name: "conv4_2"
	type: "Convolution"
	bottom: "conv4_1"
	top: "conv4_2"
	param {
		lr_mult: 0
		decay_mult: 0
	}
	param {
		lr_mult: 0
		decay_mult: 0
	}
	convolution_param {
		num_output: 512
		pad: 1
		kernel_size: 3
	}
}
layer {
	name: "relu4_2"
	type: "ReLU"
	bottom: "conv4_2"
	top: "conv4_2"
}
layer {
	name: "conv4_3"
	type: "Convolution"
	bottom: "conv4_2"
	top: "conv4_3"
	param {
		lr_mult: 0
		decay_mult: 0
	}
	param {
		lr_mult: 0
		decay_mult: 0
	}
	convolution_param {
		num_output: 512
		pad: 1
		kernel_size: 3
	}
}
layer {
	name: "relu4_3"
	type: "ReLU"
	bottom: "conv4_3"
	top: "conv4_3"
}
layer {
	name: "pool4"
	type: "Pooling"
	bottom: "conv4_3"
	top: "pool4"
	pooling_param {
		pool: MAX
		kernel_size: 2
		stride: 2
	}
}
layer {
	name: "conv5_1"
	type: "Convolution"
	bottom: "pool4"
	top: "conv5_1"
	param {
		lr_mult: 0
		decay_mult: 0
	}
	param {
		lr_mult: 0
		decay_mult: 0
	}
	convolution_param {
		num_output: 512
		pad: 1
		kernel_size: 3
	}
}
layer {
	name: "relu5_1"
	type: "ReLU"
	bottom: "conv5_1"
	top: "conv5_1"
}
layer {
	name: "conv5_2"
	type: "Convolution"
	bottom: "conv5_1"
	top: "conv5_2"
	param {
		lr_mult: 0
		decay_mult: 0
	}
	param {
		lr_mult: 0
		decay_mult: 0
	}
	convolution_param {
		num_output: 512
		pad: 1
		kernel_size: 3
	}
}
layer {
	name: "relu5_2"
	type: "ReLU"
	bottom: "conv5_2"
	top: "conv5_2"
}
layer {
	name: "conv5_3"
	type: "Convolution"
	bottom: "conv5_2"
	top: "conv5_3"
	param {
		lr_mult: 0
		decay_mult: 0
	}
	param {
		lr_mult: 0
		decay_mult: 0
	}
	convolution_param {
		num_output: 512
		pad: 1
		kernel_size: 3
	}
}
layer {
	name: "relu5_3"
	type: "ReLU"
	bottom: "conv5_3"
	top: "conv5_3"
}
layer {
	name: "pool5"
	type: "Pooling"
	bottom: "conv5_3"
	top: "pool5"
	pooling_param {
		pool: MAX
		kernel_size: 2
		stride: 2
	}
}
layer {
	name: "fc6"
	type: "InnerProduct"
	bottom: "pool5"
	top: "fc6"
	param {
		lr_mult: 0
		decay_mult: 0
	}
	param {
		lr_mult: 0
		decay_mult: 0
	}
	inner_product_param {
		num_output: 4096
	}
}
layer {
	name: "relu6"
	type: "ReLU"
	bottom: "fc6"
	top: "fc6"
}
layer {
	name: "fc7"
	type: "InnerProduct"
	bottom: "fc6"
	top: "fc7"
	param {
		lr_mult: 0
		decay_mult: 0
	}
	param {
		lr_mult: 0
		decay_mult: 0
	}
	inner_product_param {
		num_output: 4096
	}
}
layer {
	name: "relu7"
	type: "ReLU"
	bottom: "fc7"
	top: "fc7"
}
layer {
	name: "fc8"
	type: "InnerProduct"
	bottom: "fc7"
	top: "fc8"
	param {
		lr_mult: 0
		decay_mult: 0
	}
	param {
		lr_mult: 0
		decay_mult: 0
	}
	inner_product_param {
		num_output: 256
		weight_filler {
			type: "xavier"
		}
		bias_filler {
			type: "constant"
			value: 0
		}
	}
}
layer {
	name: "relu8"
	type: "ReLU"
	bottom: "fc8"
	top: "fc8"
}
# Spatial Cfg Subnet
layer {
	name: "conv1_p"
	type: "Convolution"
	bottom: "posdata"
	top: "conv1_p"
	param {
		lr_mult: 0
		decay_mult: 0
	}
	param {
		lr_mult: 0
		decay_mult: 0
	}
	convolution_param {
		num_output: 96
		kernel_size: 5
		pad: 2
		stride: 2
	}
}
layer {
	name: "relu1_p"
	type: "ReLU"
	bottom: "conv1_p"
	top: "conv1_p"
}
layer {
	name: "conv2_p"
	type: "Convolution"
	bottom: "conv1_p"
	top: "conv2_p"
	param {
		lr_mult: 0
		decay_mult: 0
	}
	param {
		lr_mult: 0
		decay_mult: 0
	}
	convolution_param {
		num_output: 128
		kernel_size: 5
		pad: 2
		stride: 2
	}
}
layer {
	name: "conv3_p"
	type: "Convolution"
	bottom: "conv2_p"
	top: "conv3_p"
	param {
		lr_mult: 0
		decay_mult: 0
	}
	param {
		lr_mult: 0
		decay_mult: 0
	}
	convolution_param {
		num_output: 64
		kernel_size: 8
	}
}
layer {
	name: "relu3_p"
	type: "ReLU"
	bottom: "conv3_p"
	top: "conv3_p"
}
# Combine features from subnets
layer {
	name: "concat1_c"
	type: "Python"
	bottom: "fc8"
	bottom: "conv3_p"
	top: "concat1_c"
	python_param {
		module: "customize_layers.concat_layer"
		layer: "Layer"
	}
}
layer {
	name: "fc2_c"
	type: "InnerProduct"
	bottom: "concat1_c"
	top: "fc2_c"
	param {
		lr_mult: 0
		decay_mult: 0
	}
	param {
		lr_mult: 0
		decay_mult: 0
	}
	inner_product_param {
		num_output: 128
	}
}
layer {
	name: "relu2_c"
	type: "ReLU"
	bottom: "fc2_c"
	top: "fc2_c"
}
layer {
	name: "PhiR_0"
	type: "InnerProduct"
	bottom: "fc2_c"
	top: "qr0"
	param {
		lr_mult: 0
		decay_mult: 0
	}
	param {
		lr_mult: 0
		decay_mult: 0
	}
	inner_product_param {
		num_output: 70
		weight_filler {
			type: "xavier"
		}
		bias_filler {
			type: "constant"
			value: 0
		}
	}
}
#DR-Net
layer {
	name: "PhiA_1"
	type: "InnerProduct"
	bottom: "qa"
	top: "qar1"
	param {
		name: "qar_w"
		lr_mult: 1
		decay_mult: 1
	}
	param {
		name: "qar_b"
		lr_mult: 2
		decay_mult: 0
	}
	inner_product_param {
		num_output: 70
		weight_filler {
			type: "xavier"
		}
		bias_filler {
			type: "constant"
			value: 0
		}
	}
}
layer {
	name: "PhiB_1"
	type: "InnerProduct"
	bottom: "qb"
	top: "qbr1"
	param {
		name: "qbr_w"
		lr_mult: 1
		decay_mult: 1
	}
	param {
		name: "qbr_b"
		lr_mult: 2
		decay_mult: 0
	}
	inner_product_param {
		num_output: 70
		weight_filler {
			type: "xavier"
		}
		bias_filler {
			type: "constant"
			value: 0
		}
	}
}
layer {
	name: "PhiR_1"
	type: "InnerProduct"
	bottom: "qr0"
	top: "q1r"
	param {
		name: "qr_w"
		lr_mult: 1
		decay_mult: 1
	}
	param {	
		name: "qr_b"
		lr_mult: 2
		decay_mult: 0
	}
	inner_product_param {
		num_output: 70
		weight_filler {
			type: "xavier"
		}
		bias_filler {
			type: "constant"
			value: 0
		}
	}
}
layer {
	name: "QSum_1"
	type: "Eltwise"
	bottom: "qar1"
	bottom: "qbr1"
	bottom: "q1r"
	top: "qr1"
	eltwise_param { operation: SUM }
}
layer {
	name: "PhiA_2"
	type: "InnerProduct"
	bottom: "qa"
	top: "qar2"
	param {
		name: "qar_w"
		lr_mult: 1
		decay_mult: 1
	}
	param {
		name: "qar_b"
		lr_mult: 2
		decay_mult: 0
	}
	inner_product_param {
		num_output: 70
	}
}
layer {
	name: "PhiB_2"
	type: "InnerProduct"
	bottom: "qb"
	top: "qbr2"
	param {
		name: "qbr_w"
		lr_mult: 1
		decay_mult: 1
	}
	param {
		name: "qbr_b"
		lr_mult: 2
		decay_mult: 0
	}
	inner_product_param {
		num_output: 70
	}
}
layer {
	name: "PhiR_2"
	type: "InnerProduct"
	bottom: "qr1"
	top: "q2r"
	param {
		name: "qr_w"
		lr_mult: 1
		decay_mult: 1
	}
	param {
		name: "qr_b"
		lr_mult: 2
		decay_mult: 0
	}
	inner_product_param {
		num_output: 70
	}
}
layer {
	name: "QSum_2"
	type: "Eltwise"
	bottom: "qar2"
	bottom: "qbr2"
	bottom: "q2r"
	top: "qr2"
	eltwise_param { operation: SUM }
}
layer {
	name: "PhiA_3"
	type: "InnerProduct"
	bottom: "qa"
	top: "qar3"
	param {
		name: "qar_w"
		lr_mult: 1
		decay_mult: 1
	}
	param {
		name: "qar_b"
		lr_mult: 2
		decay_mult: 0
	}
	inner_product_param {
		num_output: 70
	}
}
layer {
	name: "PhiB_3"
	type: "InnerProduct"
	bottom: "qb"
	top: "qbr3"
	param {
		name: "qbr_w"
		lr_mult: 1
		decay_mult: 1
	}
	param {
		name: "qbr_b"
		lr_mult: 2
		decay_mult: 0
	}
	inner_product_param {
		num_output: 70
	}
}
layer {
	name: "PhiR_3"
	type: "InnerProduct"
	bottom: "qr2"
	top: "q3r"
	param {
		name: "qr_w"
		lr_mult: 1
		decay_mult: 1
	}
	param {
		name: "qr_b"
		lr_mult: 2
		decay_mult: 0
	}
	inner_product_param {
		num_output: 70
	}
}
layer {
	name: "QSum_3"
	type: "Eltwise"
	bottom: "qar3"
	bottom: "qbr3"
	bottom: "q3r"
	top: "qr3"
	eltwise_param { operation: SUM }
}
layer {
	name: "PhiA_4"
	type: "InnerProduct"
	bottom: "qa"
	top: "qar4"
	param {
		name: "qar_w"
		lr_mult: 1
		decay_mult: 1
	}
	param {
		name: "qar_b"
		lr_mult: 2
		decay_mult: 0
	}
	inner_product_param {
		num_output: 70
	}
}
layer {
	name: "PhiB_4"
	type: "InnerProduct"
	bottom: "qb"
	top: "qbr4"
	param {
		name: "qbr_w"
		lr_mult: 1
		decay_mult: 1
	}
	param {
		name: "qbr_b"
		lr_mult: 2
		decay_mult: 0
	}
	inner_product_param {
		num_output: 70
	}
}
layer {
	name: "PhiR_4"
	type: "InnerProduct"
	bottom: "qr3"
	top: "q4r"
	param {
		name: "qr_w"
		lr_mult: 1
		decay_mult: 1
	}
	param {
		name: "qr_b"
		lr_mult: 2
		decay_mult: 0
	}
	inner_product_param {
		num_output: 70
	}
}
layer {
	name: "QSum_4"
	type: "Eltwise"
	bottom: "qar4"
	bottom: "qbr4"
	bottom: "q4r"
	top: "qr4"
	eltwise_param { operation: SUM }
}
layer {
	name: "PhiA_5"
	type: "InnerProduct"
	bottom: "qa"
	top: "qar5"
	param {
		name: "qar_w"
		lr_mult: 1
		decay_mult: 1
	}
	param {
		name: "qar_b"
		lr_mult: 2
		decay_mult: 0
	}
	inner_product_param {
		num_output: 70
	}
}
layer {
	name: "PhiB_5"
	type: "InnerProduct"
	bottom: "qb"
	top: "qbr5"
	param {
		name: "qbr_w"
		lr_mult: 1
		decay_mult: 1
	}
	param {
		name: "qbr_b"
		lr_mult: 2
		decay_mult: 0
	}
	inner_product_param {
		num_output: 70
	}
}
layer {
	name: "PhiR_5"
	type: "InnerProduct"
	bottom: "qr4"
	top: "q5r"
	param {
		name: "qr_w"
		lr_mult: 1
		decay_mult: 1
	}
	param {
		name: "qr_b"
		lr_mult: 2
		decay_mult: 0
	}
	inner_product_param {
		num_output: 70
	}
}
layer {
	name: "QSum_5"
	type: "Eltwise"
	bottom: "qar5"
	bottom: "qbr5"
	bottom: "q5r"
	top: "qr5"
	eltwise_param { operation: SUM }
}
layer {
	name: "PhiA_6"
	type: "InnerProduct"
	bottom: "qa"
	top: "qar6"
	param {
		name: "qar_w"
		lr_mult: 1
		decay_mult: 1
	}
	param {
		name: "qar_b"
		lr_mult: 2
		decay_mult: 0
	}
	inner_product_param {
		num_output: 70
	}
}
layer {
	name: "PhiB_6"
	type: "InnerProduct"
	bottom: "qb"
	top: "qbr6"
	param {
		name: "qbr_w"
		lr_mult: 1
		decay_mult: 1
	}
	param {
		name: "qbr_b"
		lr_mult: 2
		decay_mult: 0
	}
	inner_product_param {
		num_output: 70
	}
}
layer {
	name: "PhiR_6"
	type: "InnerProduct"
	bottom: "qr5"
	top: "q6r"
	param {
		name: "qr_w"
		lr_mult: 1
		decay_mult: 1
	}
	param {
		name: "qr_b"
		lr_mult: 2
		decay_mult: 0
	}
	inner_product_param {
		num_output: 70
	}
}
layer {
	name: "QSum_6"
	type: "Eltwise"
	bottom: "qar6"
	bottom: "qbr6"
	bottom: "q6r"
	top: "qr6"
	eltwise_param { operation: SUM }
}
layer {
	name: "PhiA_7"
	type: "InnerProduct"
	bottom: "qa"
	top: "qar7"
	param {
		name: "qar_w"
		lr_mult: 1
		decay_mult: 1
	}
	param {
		name: "qar_b"
		lr_mult: 2
		decay_mult: 0
	}
	inner_product_param {
		num_output: 70
	}
}
layer {
	name: "PhiB_7"
	type: "InnerProduct"
	bottom: "qb"
	top: "qbr7"
	param {
		name: "qbr_w"
		lr_mult: 1
		decay_mult: 1
	}
	param {
		name: "qbr_b"
		lr_mult: 2
		decay_mult: 0
	}
	inner_product_param {
		num_output: 70
	}
}
layer {
	name: "PhiR_7"
	type: "InnerProduct"
	bottom: "qr6"
	top: "q7r"
	param {
		name: "qr_w"
		lr_mult: 1
		decay_mult: 1
	}
	param {
		name: "qr_b"
		lr_mult: 2
		decay_mult: 0
	}
	inner_product_param {
		num_output: 70
	}
}
layer {
	name: "QSum_7"
	type: "Eltwise"
	bottom: "qar7"
	bottom: "qbr7"
	bottom: "q7r"
	top: "qr7"
	eltwise_param { operation: SUM }
}
layer {
	name: "PhiA_8"
	type: "InnerProduct"
	bottom: "qa"
	top: "qar8"
	param {
		name: "qar_w"
		lr_mult: 1
		decay_mult: 1
	}
	param {
		name: "qar_b"
		lr_mult: 2
		decay_mult: 0
	}
	inner_product_param {
		num_output: 70
	}
}
layer {
	name: "PhiB_8"
	type: "InnerProduct"
	bottom: "qb"
	top: "qbr8"
	param {
		name: "qbr_w"
		lr_mult: 1
		decay_mult: 1
	}
	param {
		name: "qbr_b"
		lr_mult: 2
		decay_mult: 0
	}
	inner_product_param {
		num_output: 70
	}
}
layer {
	name: "PhiR_8"
	type: "InnerProduct"
	bottom: "qr7"
	top: "q8r"
	param {
		name: "qr_w"
		lr_mult: 1
		decay_mult: 1
	}
	param {
		name: "qr_b"
		lr_mult: 2
		decay_mult: 0
	}
	inner_product_param {
		num_output: 70
	}
}
layer {
	name: "QSum_8"
	type: "Eltwise"
	bottom: "qar8"
	bottom: "qbr8"
	bottom: "q8r"
	top: "qr8"
	eltwise_param { operation: SUM }
}

layer {
	name: "softmax"
	type: "Softmax"
	bottom: "qr8"
	top: "pred"
}


================================================
FILE: prototxts/test_drnet_8units_relu_shareweight.prototxt
================================================
input: "qa"
input_shape {	
	dim: 1
	dim: 100
}

input: "qb"
input_shape {
	dim: 1
	dim: 100
}

input: "im"
input_shape {
	dim: 1
	dim: 3
	dim: 224
	dim: 224
}

input: "posdata"
input_shape {
	dim: 1
	dim: 2
	dim: 32
	dim: 32
}

# Appearance Subnet
layer {
	name: "conv1_1"
	type: "Convolution"
	bottom: "im"
	top: "conv1_1"
	param {
		lr_mult: 0
		decay_mult: 0
	} 
	param {
		lr_mult: 0
		decay_mult: 0
	}
	convolution_param {
		num_output: 64
		pad: 1
		kernel_size: 3
	}
}
layer {
	name: "relu1_1"
	type: "ReLU"
	bottom: "conv1_1"
	top: "conv1_1"
}
layer {
	name: "conv1_2"
	type: "Convolution"
	bottom: "conv1_1"
	top: "conv1_2"
	param {
		lr_mult: 0
		decay_mult: 0
	}
	param {
		lr_mult: 0
		decay_mult: 0
	}
	convolution_param {
		num_output: 64
		pad: 1
		kernel_size: 3
	}
}
layer {
	name: "relu1_2"
	type: "ReLU"
	bottom: "conv1_2"
	top: "conv1_2"
}
layer {
	name: "pool1"
	type: "Pooling"
	bottom: "conv1_2"
	top: "pool1"
	pooling_param {
		pool: MAX
		kernel_size: 2
		stride: 2
	}
}
layer {
	name: "conv2_1"
	type: "Convolution"
	bottom: "pool1"
	top: "conv2_1"
	param {
		lr_mult: 0
		decay_mult: 0
	}
	param {
		lr_mult: 0
		decay_mult: 0
	}
	convolution_param {
		num_output: 128
		pad: 1
		kernel_size: 3
	}
}
layer {
	name: "relu2_1"
	type: "ReLU"
	bottom: "conv2_1"
	top: "conv2_1"
}
layer {
	name: "conv2_2"
	type: "Convolution"
	bottom: "conv2_1"
	top: "conv2_2"
	param {
		lr_mult: 0
		decay_mult: 0
	}
	param {
		lr_mult: 0
		decay_mult: 0
	}
	convolution_param {
		num_output: 128
		pad: 1
		kernel_size: 3
	}
}
layer {
	name: "relu2_2"
	type: "ReLU"
	bottom: "conv2_2"
	top: "conv2_2"
}
layer {
	name: "pool2"	
	type: "Pooling"
	bottom: "conv2_2"
	top: "pool2"
	pooling_param {
		pool: MAX
		kernel_size: 2
		stride: 2
	}
}
layer {
	name: "conv3_1"
	type: "Convolution"
	bottom: "pool2"
	top: "conv3_1"
	param {
		lr_mult: 0
		decay_mult: 0
	}
	param {
		lr_mult: 0
		decay_mult: 0
	}
	convolution_param {
		num_output: 256
		pad: 1
		kernel_size: 3
	}
}
layer {
	name: "relu3_1"
	type: "ReLU"
	bottom: "conv3_1"
	top: "conv3_1"
}
layer {
	name: "conv3_2"
	type: "Convolution"
	bottom: "conv3_1"
	top: "conv3_2"
	param {
		lr_mult: 0
		decay_mult: 0
	}
	param {
		lr_mult: 2
		decay_mult: 0
	}
	convolution_param {
		num_output: 256
		pad: 1
		kernel_size: 3
	}
}
layer {
	name: "relu3_2"
	type: "ReLU"
	bottom: "conv3_2"
	top: "conv3_2"
}
layer {
	name: "conv3_3"
	type: "Convolution"
	bottom: "conv3_2"
	top: "conv3_3"
	param {
		lr_mult: 0
		decay_mult: 0
	}
	param {
		lr_mult: 0
		decay_mult: 0
	}
	convolution_param {
		num_output: 256
		pad: 1
		kernel_size: 3
	}
}
layer {
	name: "relu3_3"
	type: "ReLU"
	bottom: "conv3_3"
	top: "conv3_3"
}
layer {
	name: "pool3"
	type: "Pooling"
	bottom: "conv3_3"
	top: "pool3"
	pooling_param {
		pool: MAX
		kernel_size: 2
		stride: 2
	}
}
layer {
	name: "conv4_1"
	type: "Convolution"
	bottom: "pool3"
	top: "conv4_1"
	param {
		lr_mult: 0
		decay_mult: 0
	}
	param {
		lr_mult: 0
		decay_mult: 0
	}
	convolution_param {
		num_output: 512
		pad: 1
		kernel_size: 3
	}
}
layer {
	name: "relu4_1"
	type: "ReLU"
	bottom: "conv4_1"
	top: "conv4_1"
}
layer {
	name: "conv4_2"
	type: "Convolution"
	bottom: "conv4_1"
	top: "conv4_2"
	param {
		lr_mult: 0
		decay_mult: 0
	}
	param {
		lr_mult: 0
		decay_mult: 0
	}
	convolution_param {
		num_output: 512
		pad: 1
		kernel_size: 3
	}
}
layer {
	name: "relu4_2"
	type: "ReLU"
	bottom: "conv4_2"
	top: "conv4_2"
}
layer {
	name: "conv4_3"
	type: "Convolution"
	bottom: "conv4_2"
	top: "conv4_3"
	param {
		lr_mult: 0
		decay_mult: 0
	}
	param {
		lr_mult: 0
		decay_mult: 0
	}
	convolution_param {
		num_output: 512
		pad: 1
		kernel_size: 3
	}
}
layer {
	name: "relu4_3"
	type: "ReLU"
	bottom: "conv4_3"
	top: "conv4_3"
}
layer {
	name: "pool4"
	type: "Pooling"
	bottom: "conv4_3"
	top: "pool4"
	pooling_param {
		pool: MAX
		kernel_size: 2
		stride: 2
	}
}
layer {
	name: "conv5_1"
	type: "Convolution"
	bottom: "pool4"
	top: "conv5_1"
	param {
		lr_mult: 0
		decay_mult: 0
	}
	param {
		lr_mult: 0
		decay_mult: 0
	}
	convolution_param {
		num_output: 512
		pad: 1
		kernel_size: 3
	}
}
layer {
	name: "relu5_1"
	type: "ReLU"
	bottom: "conv5_1"
	top: "conv5_1"
}
layer {
	name: "conv5_2"
	type: "Convolution"
	bottom: "conv5_1"
	top: "conv5_2"
	param {
		lr_mult: 0
		decay_mult: 0
	}
	param {
		lr_mult: 0
		decay_mult: 0
	}
	convolution_param {
		num_output: 512
		pad: 1
		kernel_size: 3
	}
}
layer {
	name: "relu5_2"
	type: "ReLU"
	bottom: "conv5_2"
	top: "conv5_2"
}
layer {
	name: "conv5_3"
	type: "Convolution"
	bottom: "conv5_2"
	top: "conv5_3"
	param {
		lr_mult: 0
		decay_mult: 0
	}
	param {
		lr_mult: 0
		decay_mult: 0
	}
	convolution_param {
		num_output: 512
		pad: 1
		kernel_size: 3
	}
}
layer {
	name: "relu5_3"
	type: "ReLU"
	bottom: "conv5_3"
	top: "conv5_3"
}
layer {
	name: "pool5"
	type: "Pooling"
	bottom: "conv5_3"
	top: "pool5"
	pooling_param {
		pool: MAX
		kernel_size: 2
		stride: 2
	}
}
layer {
	name: "fc6"
	type: "InnerProduct"
	bottom: "pool5"
	top: "fc6"
	param {
		lr_mult: 0
		decay_mult: 0
	}
	param {
		lr_mult: 0
		decay_mult: 0
	}
	inner_product_param {
		num_output: 4096
	}
}
layer {
	name: "relu6"
	type: "ReLU"
	bottom: "fc6"
	top: "fc6"
}
layer {
	name: "fc7"
	type: "InnerProduct"
	bottom: "fc6"
	top: "fc7"
	param {
		lr_mult: 0
		decay_mult: 0
	}
	param {
		lr_mult: 0
		decay_mult: 0
	}
	inner_product_param {
		num_output: 4096
	}
}
layer {
	name: "relu7"
	type: "ReLU"
	bottom: "fc7"
	top: "fc7"
}
layer {
	name: "fc8"
	type: "InnerProduct"
	bottom: "fc7"
	top: "fc8"
	param {
		lr_mult: 0
		decay_mult: 0
	}
	param {
		lr_mult: 0
		decay_mult: 0
	}
	inner_product_param {
		num_output: 256
		weight_filler {
			type: "xavier"
		}
		bias_filler {
			type: "constant"
			value: 0
		}
	}
}
layer {
	name: "relu8"
	type: "ReLU"
	bottom: "fc8"
	top: "fc8"
}

# Spatial Cfg Subnet
layer {
	name: "conv1_p"
	type: "Convolution"
	bottom: "posdata"
	top: "conv1_p"
	param {
		lr_mult: 0
		decay_mult: 0
	}
	param {
		lr_mult: 0
		decay_mult: 0
	}
	convolution_param {
		num_output: 96
		kernel_size: 5
		pad: 2
		stride: 2
	}
}
layer {
	name: "relu1_p"
	type: "ReLU"
	bottom: "conv1_p"
	top: "conv1_p"
}
layer {
	name: "conv2_p"
	type: "Convolution"
	bottom: "conv1_p"
	top: "conv2_p"
	param {
		lr_mult: 0
		decay_mult: 0
	}
	param {
		lr_mult: 0
		decay_mult: 0
	}
	convolution_param {
		num_output: 128
		kernel_size: 5
		pad: 2
		stride: 2
	}
}
layer {
	name: "conv3_p"
	type: "Convolution"
	bottom: "conv2_p"
	top: "conv3_p"
	param {
		lr_mult: 0
		decay_mult: 0
	}
	param {
		lr_mult: 0
		decay_mult: 0
	}
	convolution_param {
		num_output: 64
		kernel_size: 8
	}
}
layer {
	name: "relu3_p"
	type: "ReLU"
	bottom: "conv3_p"
	top: "conv3_p"
}
# Combine features from subnets
layer {
	name: "concat1_c"
	type: "Python"
	bottom: "fc8"
	bottom: "conv3_p"
	top: "concat1_c"
	python_param {
		module: "customize_layers.concat_layer"
		layer: "Layer"
	}
}
layer {
	name: "fc2_c"
	type: "InnerProduct"
	bottom: "concat1_c"
	top: "fc2_c"
	param {
		lr_mult: 0
		decay_mult: 0
	}
	param {
		lr_mult: 0
		decay_mult: 0
	}
	inner_product_param {
		num_output: 128
	}
}
layer {
	name: "relu2_c"
	type: "ReLU"
	bottom: "fc2_c"
	top: "fc2_c"
}
layer {
	name: "PhiR_0"
	type: "InnerProduct"
	bottom: "fc2_c"
	top: "q0r"
	param {
		lr_mult: 0
		decay_mult: 0
	}
	param {
		lr_mult: 0
		decay_mult: 0
	}
	inner_product_param {
		num_output: 70
		weight_filler {
			type: "xavier"
		}
		bias_filler {
			type: "constant"
			value: 0
		}
	}
}
layer {
	name: "relu_0"
	type: "ReLU"
	bottom: "q0r"
	top: "qr0"
}
#DR-Net
layer {
	name: "PhiA_1"
	type: "InnerProduct"
	bottom: "qa"
	top: "qar1"
	param {
		name: "qar_w"
		lr_mult: 1
		decay_mult: 1
	}
	param {
		name: "qar_b"
		lr_mult: 2
		decay_mult: 0
	}
	inner_product_param {
		num_output: 70
		weight_filler {
			type: "xavier"
		}
		bias_filler {
			type: "constant"
			value: 0
		}
	}
}
layer {
	name: "PhiB_1"
	type: "InnerProduct"
	bottom: "qb"
	top: "qbr1"
	param {
		name: "qbr_w"
		lr_mult: 1
		decay_mult: 1
	}
	param {
		name: "qbr_b"
		lr_mult: 2
		decay_mult: 0
	}
	inner_product_param {
		num_output: 70
		weight_filler {
			type: "xavier"
		}
		bias_filler {
			type: "constant"
			value: 0
		}
	}
}
layer {
	name: "PhiR_1"
	type: "InnerProduct"
	bottom: "qr0"
	top: "q1r"
	param {
		name: "qr_w"
		lr_mult: 1
		decay_mult: 1
	}
	param {	
		name: "qr_b"
		lr_mult: 2
		decay_mult: 0
	}
	inner_product_param {
		num_output: 70
		weight_filler {
			type: "xavier"
		}
		bias_filler {
			type: "constant"
			value: 0
		}
	}
}
layer {
	name: "QSum_1"
	type: "Eltwise"
	bottom: "qar1"
	bottom: "qbr1"
	bottom: "q1r"
	top: "qr1un"
	eltwise_param { operation: SUM }
}
layer {
	name: "relu_1"
	type: "ReLU"
	bottom: "qr1un"
	top: "qr1"
}
layer {
	name: "PhiA_2"
	type: "InnerProduct"
	bottom: "qa"
	top: "qar2"
	param {
		name: "qar_w"
		lr_mult: 1
		decay_mult: 1
	}
	param {
		name: "qar_b"
		lr_mult: 2
		decay_mult: 0
	}
	inner_product_param {
		num_output: 70
	}
}
layer {
	name: "PhiB_2"
	type: "InnerProduct"
	bottom: "qb"
	top: "qbr2"
	param {
		name: "qbr_w"
		lr_mult: 1
		decay_mult: 1
	}
	param {
		name: "qbr_b"
		lr_mult: 2
		decay_mult: 0
	}
	inner_product_param {
		num_output: 70
	}
}
layer {
	name: "PhiR_2"
	type: "InnerProduct"
	bottom: "qr1"
	top: "q2r"
	param {
		name: "qr_w"
		lr_mult: 1
		decay_mult: 1
	}
	param {
		name: "qr_b"
		lr_mult: 2
		decay_mult: 0
	}
	inner_product_param {
		num_output: 70
	}
}
layer {
	name: "QSum_2"
	type: "Eltwise"
	bottom: "qar2"
	bottom: "qbr2"
	bottom: "q2r"
	top: "qr2un"
	eltwise_param { operation: SUM }
}
layer {
	name: "relu_2"
	type: "ReLU"
	bottom: "qr2un"
	top: "qr2"
}
layer {
	name: "PhiA_3"
	type: "InnerProduct"
	bottom: "qa"
	top: "qar3"
	param {
		name: "qar_w"
		lr_mult: 1
		decay_mult: 1
	}
	param {
		name: "qar_b"
		lr_mult: 2
		decay_mult: 0
	}
	inner_product_param {
		num_output: 70
	}
}
layer {
	name: "PhiB_3"
	type: "InnerProduct"
	bottom: "qb"
	top: "qbr3"
	param {
		name: "qbr_w"
		lr_mult: 1
		decay_mult: 1
	}
	param {
		name: "qbr_b"
		lr_mult: 2
		decay_mult: 0
	}
	inner_product_param {
		num_output: 70
	}
}
layer {
	name: "PhiR_3"
	type: "InnerProduct"
	bottom: "qr2"
	top: "q3r"
	param {
		name: "qr_w"
		lr_mult: 1
		decay_mult: 1
	}
	param {
		name: "qr_b"
		lr_mult: 2
		decay_mult: 0
	}
	inner_product_param {
		num_output: 70
	}
}
layer {
	name: "QSum_3"
	type: "Eltwise"
	bottom: "qar3"
	bottom: "qbr3"
	bottom: "q3r"
	top: "qr3un"
	eltwise_param { operation: SUM }
}
layer {
	name: "relu_3"
	type: "ReLU"
	bottom: "qr3un"
	top: "qr3"
}
layer {
	name: "PhiA_4"
	type: "InnerProduct"
	bottom: "qa"
	top: "qar4"
	param {
		name: "qar_w"
		lr_mult: 1
		decay_mult: 1
	}
	param {
		name: "qar_b"
		lr_mult: 2
		decay_mult: 0
	}
	inner_product_param {
		num_output: 70
	}
}
layer {
	name: "PhiB_4"
	type: "InnerProduct"
	bottom: "qb"
	top: "qbr4"
	param {
		name: "qbr_w"
		lr_mult: 1
		decay_mult: 1
	}
	param {
		name: "qbr_b"
		lr_mult: 2
		decay_mult: 0
	}
	inner_product_param {
		num_output: 70
	}
}
layer {
	name: "PhiR_4"
	type: "InnerProduct"
	bottom: "qr3"
	top: "q4r"
	param {
		name: "qr_w"
		lr_mult: 1
		decay_mult: 1
	}
	param {
		name: "qr_b"
		lr_mult: 2
		decay_mult: 0
	}
	inner_product_param {
		num_output: 70
	}
}
layer {
	name: "QSum_4"
	type: "Eltwise"
	bottom: "qar4"
	bottom: "qbr4"
	bottom: "q4r"
	top: "qr4un"
	eltwise_param { operation: SUM }
}
layer {
	name: "relu_4"
	type: "ReLU"
	bottom: "qr4un"
	top: "qr4"
}
layer {
	name: "PhiA_5"
	type: "InnerProduct"
	bottom: "qa"
	top: "qar5"
	param {
		name: "qar_w"
		lr_mult: 1
		decay_mult: 1
	}
	param {
		name: "qar_b"
		lr_mult: 2
		decay_mult: 0
	}
	inner_product_param {
		num_output: 70
	}
}
layer {
	name: "PhiB_5"
	type: "InnerProduct"
	bottom: "qb"
	top: "qbr5"
	param {
		name: "qbr_w"
		lr_mult: 1
		decay_mult: 1
	}
	param {
		name: "qbr_b"
		lr_mult: 2
		decay_mult: 0
	}
	inner_product_param {
		num_output: 70
	}
}
layer {
	name: "PhiR_5"
	type: "InnerProduct"
	bottom: "qr4"
	top: "q5r"
	param {
		name: "qr_w"
		lr_mult: 1
		decay_mult: 1
	}
	param {
		name: "qr_b"
		lr_mult: 2
		decay_mult: 0
	}
	inner_product_param {
		num_output: 70
	}
}
layer {
	name: "QSum_5"
	type: "Eltwise"
	bottom: "qar5"
	bottom: "qbr5"
	bottom: "q5r"
	top: "qr5un"
	eltwise_param { operation: SUM }
}
layer {
	name: "relu_5"
	type: "ReLU"
	bottom: "qr5un"
	top: "qr5"
}

layer {
	name: "PhiA_6"
	type: "InnerProduct"
	bottom: "qa"
	top: "qar6"
	param {
		name: "qar_w"
		lr_mult: 1
		decay_mult: 1
	}
	param {
		name: "qar_b"
		lr_mult: 2
		decay_mult: 0
	}
	inner_product_param {
		num_output: 70
	}
}
layer {
	name: "PhiB_6"
	type: "InnerProduct"
	bottom: "qb"
	top: "qbr6"
	param {
		name: "qbr_w"
		lr_mult: 1
		decay_mult: 1
	}
	param {
		name: "qbr_b"
		lr_mult: 2
		decay_mult: 0
	}
	inner_product_param {
		num_output: 70
	}
}
layer {
	name: "PhiR_6"
	type: "InnerProduct"
	bottom: "qr5"
	top: "q6r"
	param {
		name: "qr_w"
		lr_mult: 1
		decay_mult: 1
	}
	param {
		name: "qr_b"
		lr_mult: 2
		decay_mult: 0
	}
	inner_product_param {
		num_output: 70
	}
}
layer {
	name: "QSum_6"
	type: "Eltwise"
	bottom: "qar6"
	bottom: "qbr6"
	bottom: "q6r"
	top: "qr6un"
	eltwise_param { operation: SUM }
}
layer {
	name: "relu_6"
	type: "ReLU"
	bottom: "qr6un"
	top: "qr6"
}
layer {
	name: "PhiA_7"
	type: "InnerProduct"
	bottom: "qa"
	top: "qar7"
	param {
		name: "qar_w"
		lr_mult: 1
		decay_mult: 1
	}
	param {
		name: "qar_b"
		lr_mult: 2
		decay_mult: 0
	}
	inner_product_param {
		num_output: 70
	}
}
layer {
	name: "PhiB_7"
	type: "InnerProduct"
	bottom: "qb"
	top: "qbr7"
	param {
		name: "qbr_w"
		lr_mult: 1
		decay_mult: 1
	}
	param {
		name: "qbr_b"
		lr_mult: 2
		decay_mult: 0
	}
	inner_product_param {
		num_output: 70
	}
}
layer {
	name: "PhiR_7"
	type: "InnerProduct"
	bottom: "qr6"
	top: "q7r"
	param {
		name: "qr_w"
		lr_mult: 1
		decay_mult: 1
	}
	param {
		name: "qr_b"
		lr_mult: 2
		decay_mult: 0
	}
	inner_product_param {
		num_output: 70
	}
}
layer {
	name: "QSum_7"
	type: "Eltwise"
	bottom: "qar7"
	bottom: "qbr7"
	bottom: "q7r"
	top: "qr7un"
	eltwise_param { operation: SUM }
}
layer {
	name: "relu_7"
	type: "ReLU"
	bottom: "qr7un"
	top: "qr7"
}
layer {
	name: "PhiA_8"
	type: "InnerProduct"
	bottom: "qa"
	top: "qar8"
	param {
		name: "qar_w"
		lr_mult: 1
		decay_mult: 1
	}
	param {
		name: "qar_b"
		lr_mult: 2
		decay_mult: 0
	}
	inner_product_param {
		num_output: 70
	}
}
layer {
	name: "PhiB_8"
	type: "InnerProduct"
	bottom: "qb"
	top: "qbr8"
	param {
		name: "qbr_w"
		lr_mult: 1
		decay_mult: 1
	}
	param {
		name: "qbr_b"
		lr_mult: 2
		decay_mult: 0
	}
	inner_product_param {
		num_output: 70
	}
}
layer {
	name: "PhiR_8"
	type: "InnerProduct"
	bottom: "qr7"
	top: "q8r"
	param {
		name: "qr_w"
		lr_mult: 1
		decay_mult: 1
	}
	param {
		name: "qr_b"
		lr_mult: 2
		decay_mult: 0
	}
	inner_product_param {
		num_output: 70
	}
}
layer {
	name: "QSum_8"
	type: "Eltwise"
	bottom: "qar8"
	bottom: "qbr8"
	bottom: "q8r"
	top: "qr8"
	eltwise_param { operation: SUM }
}
layer {
	name: "softmax"
	type: "Softmax"
	bottom: "qr8"
	top: "pred"
}


================================================
FILE: prototxts/test_drnet_8units_softmax.prototxt
================================================
input: "qa"
input_shape {	
	dim: 1
	dim: 100
}

input: "qb"
input_shape {
	dim: 1
	dim: 100
}

input: "im"
input_shape {
	dim: 1
	dim: 3
	dim: 224
	dim: 224
}

input: "posdata"
input_shape {
	dim: 1
	dim: 2
	dim: 32
	dim: 32
}

# Appearance Subnet 
layer {
	name: "conv1_1"
	type: "Convolution"
	bottom: "im"
	top: "conv1_1"
	param {
		lr_mult: 0
		decay_mult: 0
	} 
	param {
		lr_mult: 0
		decay_mult: 0
	}
	convolution_param {
		num_output: 64
		pad: 1
		kernel_size: 3
	}
}
layer {
	name: "relu1_1"
	type: "ReLU"
	bottom: "conv1_1"
	top: "conv1_1"
}
layer {
	name: "conv1_2"
	type: "Convolution"
	bottom: "conv1_1"
	top: "conv1_2"
	param {
		lr_mult: 0
		decay_mult: 0
	}
	param {
		lr_mult: 0
		decay_mult: 0
	}
	convolution_param {
		num_output: 64
		pad: 1
		kernel_size: 3
	}
}
layer {
	name: "relu1_2"
	type: "ReLU"
	bottom: "conv1_2"
	top: "conv1_2"
}
layer {
	name: "pool1"
	type: "Pooling"
	bottom: "conv1_2"
	top: "pool1"
	pooling_param {
		pool: MAX
		kernel_size: 2
		stride: 2
	}
}
layer {
	name: "conv2_1"
	type: "Convolution"
	bottom: "pool1"
	top: "conv2_1"
	param {
		lr_mult: 0
		decay_mult: 0
	}
	param {
		lr_mult: 0
		decay_mult: 0
	}
	convolution_param {
		num_output: 128
		pad: 1
		kernel_size: 3
	}
}
layer {
	name: "relu2_1"
	type: "ReLU"
	bottom: "conv2_1"
	top: "conv2_1"
}
layer {
	name: "conv2_2"
	type: "Convolution"
	bottom: "conv2_1"
	top: "conv2_2"
	param {
		lr_mult: 0
		decay_mult: 0
	}
	param {
		lr_mult: 0
		decay_mult: 0
	}
	convolution_param {
		num_output: 128
		pad: 1
		kernel_size: 3
	}
}
layer {
	name: "relu2_2"
	type: "ReLU"
	bottom: "conv2_2"
	top: "conv2_2"
}
layer {
	name: "pool2"	
	type: "Pooling"
	bottom: "conv2_2"
	top: "pool2"
	pooling_param {
		pool: MAX
		kernel_size: 2
		stride: 2
	}
}
layer {
	name: "conv3_1"
	type: "Convolution"
	bottom: "pool2"
	top: "conv3_1"
	param {
		lr_mult: 0
		decay_mult: 0
	}
	param {
		lr_mult: 0
		decay_mult: 0
	}
	convolution_param {
		num_output: 256
		pad: 1
		kernel_size: 3
	}
}
layer {
	name: "relu3_1"
	type: "ReLU"
	bottom: "conv3_1"
	top: "conv3_1"
}
layer {
	name: "conv3_2"
	type: "Convolution"
	bottom: "conv3_1"
	top: "conv3_2"
	param {
		lr_mult: 0
		decay_mult: 0
	}
	param {
		lr_mult: 2
		decay_mult: 0
	}
	convolution_param {
		num_output: 256
		pad: 1
		kernel_size: 3
	}
}
layer {
	name: "relu3_2"
	type: "ReLU"
	bottom: "conv3_2"
	top: "conv3_2"
}
layer {
	name: "conv3_3"
	type: "Convolution"
	bottom: "conv3_2"
	top: "conv3_3"
	param {
		lr_mult: 0
		decay_mult: 0
	}
	param {
		lr_mult: 0
		decay_mult: 0
	}
	convolution_param {
		num_output: 256
		pad: 1
		kernel_size: 3
	}
}
layer {
	name: "relu3_3"
	type: "ReLU"
	bottom: "conv3_3"
	top: "conv3_3"
}
layer {
	name: "pool3"
	type: "Pooling"
	bottom: "conv3_3"
	top: "pool3"
	pooling_param {
		pool: MAX
		kernel_size: 2
		stride: 2
	}
}
layer {
	name: "conv4_1"
	type: "Convolution"
	bottom: "pool3"
	top: "conv4_1"
	param {
		lr_mult: 0
		decay_mult: 0
	}
	param {
		lr_mult: 0
		decay_mult: 0
	}
	convolution_param {
		num_output: 512
		pad: 1
		kernel_size: 3
	}
}
layer {
	name: "relu4_1"
	type: "ReLU"
	bottom: "conv4_1"
	top: "conv4_1"
}
layer {
	name: "conv4_2"
	type: "Convolution"
	bottom: "conv4_1"
	top: "conv4_2"
	param {
		lr_mult: 0
		decay_mult: 0
	}
	param {
		lr_mult: 0
		decay_mult: 0
	}
	convolution_param {
		num_output: 512
		pad: 1
		kernel_size: 3
	}
}
layer {
	name: "relu4_2"
	type: "ReLU"
	bottom: "conv4_2"
	top: "conv4_2"
}
layer {
	name: "conv4_3"
	type: "Convolution"
	bottom: "conv4_2"
	top: "conv4_3"
	param {
		lr_mult: 0
		decay_mult: 0
	}
	param {
		lr_mult: 0
		decay_mult: 0
	}
	convolution_param {
		num_output: 512
		pad: 1
		kernel_size: 3
	}
}
layer {
	name: "relu4_3"
	type: "ReLU"
	bottom: "conv4_3"
	top: "conv4_3"
}
layer {
	name: "pool4"
	type: "Pooling"
	bottom: "conv4_3"
	top: "pool4"
	pooling_param {
		pool: MAX
		kernel_size: 2
		stride: 2
	}
}
layer {
	name: "conv5_1"
	type: "Convolution"
	bottom: "pool4"
	top: "conv5_1"
	param {
		lr_mult: 0
		decay_mult: 0
	}
	param {
		lr_mult: 0
		decay_mult: 0
	}
	convolution_param {
		num_output: 512
		pad: 1
		kernel_size: 3
	}
}
layer {
	name: "relu5_1"
	type: "ReLU"
	bottom: "conv5_1"
	top: "conv5_1"
}
layer {
	name: "conv5_2"
	type: "Convolution"
	bottom: "conv5_1"
	top: "conv5_2"
	param {
		lr_mult: 0
		decay_mult: 0
	}
	param {
		lr_mult: 0
		decay_mult: 0
	}
	convolution_param {
		num_output: 512
		pad: 1
		kernel_size: 3
	}
}
layer {
	name: "relu5_2"
	type: "ReLU"
	bottom: "conv5_2"
	top: "conv5_2"
}
layer {
	name: "conv5_3"
	type: "Convolution"
	bottom: "conv5_2"
	top: "conv5_3"
	param {
		lr_mult: 0
		decay_mult: 0
	}
	param {
		lr_mult: 0
		decay_mult: 0
	}
	convolution_param {
		num_output: 512
		pad: 1
		kernel_size: 3
	}
}
layer {
	name: "relu5_3"
	type: "ReLU"
	bottom: "conv5_3"
	top: "conv5_3"
}
layer {
	name: "pool5"
	type: "Pooling"
	bottom: "conv5_3"
	top: "pool5"
	pooling_param {
		pool: MAX
		kernel_size: 2
		stride: 2
	}
}
layer {
	name: "fc6"
	type: "InnerProduct"
	bottom: "pool5"
	top: "fc6"
	param {
		lr_mult: 0
		decay_mult: 0
	}
	param {
		lr_mult: 0
		decay_mult: 0
	}
	inner_product_param {
		num_output: 4096
	}
}
layer {
	name: "relu6"
	type: "ReLU"
	bottom: "fc6"
	top: "fc6"
}
layer {
	name: "fc7"
	type: "InnerProduct"
	bottom: "fc6"
	top: "fc7"
	param {
		lr_mult: 0
		decay_mult: 0
	}
	param {
		lr_mult: 0
		decay_mult: 0
	}
	inner_product_param {
		num_output: 4096
	}
}
layer {
	name: "relu7"
	type: "ReLU"
	bottom: "fc7"
	top: "fc7"
}
layer {
	name: "fc8"
	type: "InnerProduct"
	bottom: "fc7"
	top: "fc8"
	param {
		lr_mult: 0
		decay_mult: 0
	}
	param {
		lr_mult: 0
		decay_mult: 0
	}
	inner_product_param {
		num_output: 256
		weight_filler {
			type: "xavier"
		}
		bias_filler {
			type: "constant"
			value: 0
		}
	}
}
layer {
	name: "relu8"
	type: "ReLU"
	bottom: "fc8"
	top: "fc8"
}

# Spatial Cfg Subnet
layer {
	name: "conv1_p"
	type: "Convolution"
	bottom: "posdata"
	top: "conv1_p"
	param {
		lr_mult: 0
		decay_mult: 0
	}
	param {
		lr_mult: 0
		decay_mult: 0
	}
	convolution_param {
		num_output: 96
		kernel_size: 5
		pad: 2
		stride: 2
	}
}
layer {
	name: "relu1_p"
	type: "ReLU"
	bottom: "conv1_p"
	top: "conv1_p"
}
layer {
	name: "conv2_p"
	type: "Convolution"
	bottom: "conv1_p"
	top: "conv2_p"
	param {
		lr_mult: 0
		decay_mult: 0
	}
	param {
		lr_mult: 0
		decay_mult: 0
	}
	convolution_param {
		num_output: 128
		kernel_size: 5
		pad: 2
		stride: 2
	}
}
layer {
	name: "conv3_p"
	type: "Convolution"
	bottom: "conv2_p"
	top: "conv3_p"
	param {
		lr_mult: 0
		decay_mult: 0
	}
	param {
		lr_mult: 0
		decay_mult: 0
	}
	convolution_param {
		num_output: 64
		kernel_size: 8
	}
}
layer {
	name: "relu3_p"
	type: "ReLU"
	bottom: "conv3_p"
	top: "conv3_p"
}

# Combine features from subnets
layer {
	name: "concat1_c"
	type: "Python"
	bottom: "fc8"
	bottom: "conv3_p"
	top: "concat1_c"
	python_param {
		module: "customize_layers.concat_layer"
		layer: "Layer"
	}
}
layer {
	name: "fc2_c"
	type: "InnerProduct"
	bottom: "concat1_c"
	top: "fc2_c"
	param {
		lr_mult: 0
		decay_mult: 0
	}
	param {
		lr_mult: 0
		decay_mult: 0
	}
	inner_product_param {
		num_output: 128
	}
}
layer {
	name: "relu2_c"
	type: "ReLU"
	bottom: "fc2_c"
	top: "fc2_c"
}
layer {
	name: "PhiR_0"
	type: "InnerProduct"
	bottom: "fc2_c"
	top: "q0r"
	param {
		lr_mult: 0
		decay_mult: 0
	}
	param {
		lr_mult: 0
		decay_mult: 0
	}
	inner_product_param {
		num_output: 70
		weight_filler {
			type: "xavier"
		}
		bias_filler {
			type: "constant"
			value: 0
		}
	}
}
layer {
	name: "Softmax_0"
	type: "Softmax"
	bottom: "q0r"
	top: "qr0"
}

#DR-Net
layer {
	name: "PhiA_1"
	type: "InnerProduct"
	bottom: "qa"
	top: "qar1"
	param {
		lr_mult: 1
		decay_mult: 1
	}
	param {
		lr_mult: 2
		decay_mult: 0
	}
	inner_product_param {
		num_output: 70
		weight_filler {
			type: "xavier"
		}
		bias_filler {
			type: "constant"
			value: 0
		}
	}
}
layer {
	name: "PhiB_1"
	type: "InnerProduct"
	bottom: "qb"
	top: "qbr1"
	param {
		lr_mult: 1
		decay_mult: 1
	}
	param {
		lr_mult: 2
		decay_mult: 0
	}
	inner_product_param {
		num_output: 70
		weight_filler {
			type: "xavier"
		}
		bias_filler {
			type: "constant"
			value: 0
		}
	}
}
layer {
	name: "PhiR_1"
	type: "InnerProduct"
	bottom: "qr0"
	top: "q1r"
	param {
		lr_mult: 1
		decay_mult: 1
	}
	param {	
		lr_mult: 2
		decay_mult: 0
	}
	inner_product_param {
		num_output: 70
		weight_filler {
			type: "xavier"
		}
		bias_filler {
			type: "constant"
			value: 0
		}
	}
}
layer {
	name: "QSum_1"
	type: "Eltwise"
	bottom: "qar1"
	bottom: "qbr1"
	bottom: "q1r"
	top: "qr1un"
	eltwise_param { operation: SUM }
}
layer {
	name: "Softmax_1"
	type: "Softmax"
	bottom: "qr1un"
	top: "qr1"
}
layer {
	name: "PhiA_2"
	type: "InnerProduct"
	bottom: "qa"
	top: "qar2"
	param {
		lr_mult: 1
		decay_mult: 1
	}
	param {
		lr_mult: 2
		decay_mult: 0
	}
	inner_product_param {
		num_output: 70
	}
}
layer {
	name: "PhiB_2"
	type: "InnerProduct"
	bottom: "qb"
	top: "qbr2"
	param {
		lr_mult: 1
		decay_mult: 1
	}
	param {
		lr_mult: 2
		decay_mult: 0
	}
	inner_product_param {
		num_output: 70
	}
}
layer {
	name: "PhiR_2"
	type: "InnerProduct"
	bottom: "qr1"
	top: "q2r"
	param {
		lr_mult: 1
		decay_mult: 1
	}
	param {
		lr_mult: 2
		decay_mult: 0
	}
	inner_product_param {
		num_output: 70
	}
}
layer {
	name: "QSum_2"
	type: "Eltwise"
	bottom: "qar2"
	bottom: "qbr2"
	bottom: "q2r"
	top: "qr2un"
	eltwise_param { operation: SUM }
}
layer {
	name: "Softmax_2"
	type: "Softmax"
	bottom: "qr2un"
	top: "qr2"
}
layer {
	name: "PhiA_3"
	type: "InnerProduct"
	bottom: "qa"
	top: "qar3"
	param {
		lr_mult: 1
		decay_mult: 1
	}
	param {
		lr_mult: 2
		decay_mult: 0
	}
	inner_product_param {
		num_output: 70
	}
}
layer {
	name: "PhiB_3"
	type: "InnerProduct"
	bottom: "qb"
	top: "qbr3"
	param {
		lr_mult: 1
		decay_mult: 1
	}
	param {
		lr_mult: 2
		decay_mult: 0
	}
	inner_product_param {
		num_output: 70
	}
}
layer {
	name: "PhiR_3"
	type: "InnerProduct"
	bottom: "qr2"
	top: "q3r"
	param {
		lr_mult: 1
		decay_mult: 1
	}
	param {
		lr_mult: 2
		decay_mult: 0
	}
	inner_product_param {
		num_output: 70
	}
}
layer {
	name: "QSum_3"
	type: "Eltwise"
	bottom: "qar3"
	bottom: "qbr3"
	bottom: "q3r"
	top: "qr3un"
	eltwise_param { operation: SUM }
}
layer {
	name: "Softmax_3"
	type: "Softmax"
	bottom: "qr3un"
	top: "qr3"
}
layer {
	name: "PhiA_4"
	type: "InnerProduct"
	bottom: "qa"
	top: "qar4"
	param {
		lr_mult: 1
		decay_mult: 1
	}
	param {
		lr_mult: 2
		decay_mult: 0
	}
	inner_product_param {
		num_output: 70
	}
}
layer {
	name: "PhiB_4"
	type: "InnerProduct"
	bottom: "qb"
	top: "qbr4"
	param {
		lr_mult: 1
		decay_mult: 1
	}
	param {
		lr_mult: 2
		decay_mult: 0
	}
	inner_product_param {
		num_output: 70
	}
}
layer {
	name: "PhiR_4"
	type: "InnerProduct"
	bottom: "qr3"
	top: "q4r"
	param {
		lr_mult: 1
		decay_mult: 1
	}
	param {
		lr_mult: 2
		decay_mult: 0
	}
	inner_product_param {
		num_output: 70
	}
}
layer {
	name: "QSum_4"
	type: "Eltwise"
	bottom: "qar4"
	bottom: "qbr4"
	bottom: "q4r"
	top: "qr4un"
	eltwise_param { operation: SUM }
}
layer {
	name: "Softmax_4"
	type: "Softmax"
	bottom: "qr4un"
	top: "qr4"
}
layer {
	name: "PhiA_5"
	type: "InnerProduct"
	bottom: "qa"
	top: "qar5"
	param {
		lr_mult: 1
		decay_mult: 1
	}
	param {
		lr_mult: 2
		decay_mult: 0
	}
	inner_product_param {
		num_output: 70
	}
}
layer {
	name: "PhiB_5"
	type: "InnerProduct"
	bottom: "qb"
	top: "qbr5"
	param {
		lr_mult: 1
		decay_mult: 1
	}
	param {
		lr_mult: 2
		decay_mult: 0
	}
	inner_product_param {
		num_output: 70
	}
}
layer {
	name: "PhiR_5"
	type: "InnerProduct"
	bottom: "qr4"
	top: "q5r"
	param {
		lr_mult: 1
		decay_mult: 1
	}
	param {
		lr_mult: 2
		decay_mult: 0
	}
	inner_product_param {
		num_output: 70
	}
}
layer {
	name: "QSum_5"
	type: "Eltwise"
	bottom: "qar5"
	bottom: "qbr5"
	bottom: "q5r"
	top: "qr5un"
	eltwise_param { operation: SUM }
}
layer {
	name: "Softmax_5"
	type: "Softmax"
	bottom: "qr5un"
	top: "qr5"
}

layer {
	name: "PhiA_6"
	type: "InnerProduct"
	bottom: "qa"
	top: "qar6"
	param {
		lr_mult: 1
		decay_mult: 1
	}
	param {
		lr_mult: 2
		decay_mult: 0
	}
	inner_product_param {
		num_output: 70
	}
}
layer {
	name: "PhiB_6"
	type: "InnerProduct"
	bottom: "qb"
	top: "qbr6"
	param {
		lr_mult: 1
		decay_mult: 1
	}
	param {
		lr_mult: 2
		decay_mult: 0
	}
	inner_product_param {
		num_output: 70
	}
}
layer {
	name: "PhiR_6"
	type: "InnerProduct"
	bottom: "qr5"
	top: "q6r"
	param {
		lr_mult: 1
		decay_mult: 1
	}
	param {
		lr_mult: 2
		decay_mult: 0
	}
	inner_product_param {
		num_output: 70
	}
}
layer {
	name: "QSum_6"
	type: "Eltwise"
	bottom: "qar6"
	bottom: "qbr6"
	bottom: "q6r"
	top: "qr6un"
	eltwise_param { operation: SUM }
}
layer {
	name: "Softmax_6"
	type: "Softmax"
	bottom: "qr6un"
	top: "qr6"
}
layer {
	name: "PhiA_7"
	type: "InnerProduct"
	bottom: "qa"
	top: "qar7"
	param {
		lr_mult: 1
		decay_mult: 1
	}
	param {
		lr_mult: 2
		decay_mult: 0
	}
	inner_product_param {
		num_output: 70
	}
}
layer {
	name: "PhiB_7"
	type: "InnerProduct"
	bottom: "qb"
	top: "qbr7"
	param {
		lr_mult: 1
		decay_mult: 1
	}
	param {
		lr_mult: 2
		decay_mult: 0
	}
	inner_product_param {
		num_output: 70
	}
}
layer {
	name: "PhiR_7"
	type: "InnerProduct"
	bottom: "qr6"
	top: "q7r"
	param {
		lr_mult: 1
		decay_mult: 1
	}
	param {
		lr_mult: 2
		decay_mult: 0
	}
	inner_product_param {
		num_output: 70
	}
}
layer {
	name: "QSum_7"
	type: "Eltwise"
	bottom: "qar7"
	bottom: "qbr7"
	bottom: "q7r"
	top: "qr7un"
	eltwise_param { operation: SUM }
}
layer {
	name: "Softmax_7"
	type: "Softmax"
	bottom: "qr7un"
	top: "qr7"
}
layer {
	name: "PhiA_8"
	type: "InnerProduct"
	bottom: "qa"
	top: "qar8"
	param {
		lr_mult: 1
		decay_mult: 1
	}
	param {
		lr_mult: 2
		decay_mult: 0
	}
	inner_product_param {
		num_output: 70
	}
}
layer {
	name: "PhiB_8"
	type: "InnerProduct"
	bottom: "qb"
	top: "qbr8"
	param {
		lr_mult: 1
		decay_mult: 1
	}
	param {
		lr_mult: 2
		decay_mult: 0
	}
	inner_product_param {
		num_output: 70
	}
}
layer {
	name: "PhiR_8"
	type: "InnerProduct"
	bottom: "qr7"
	top: "q8r"
	param {
		lr_mult: 1
		decay_mult: 1
	}
	param {
		lr_mult: 2
		decay_mult: 0
	}
	inner_product_param {
		num_output: 70
	}
}
layer {
	name: "QSum_8"
	type: "Eltwise"
	bottom: "qar8"
	bottom: "qbr8"
	bottom: "q8r"
	top: "qr8"
	eltwise_param { operation: SUM }
}
layer {
	name: "softmax"
	type: "Softmax"
	bottom: "qr8"
	top: "pred"
}


================================================
FILE: snapshots/README.md
================================================
## Download Links

* [drnet_8units_softmax](https://drive.google.com/file/d/0B5RJWjAhdT04Y2pJelpxSEhYWTQ/view?usp=sharing)

* [drnet_8units_linear_shareweight](https://drive.google.com/file/d/0B5RJWjAhdT04X24xWFY1aHhLY0U/view?usp=sharing)

* [drnet_8units_relu_shareweight](https://drive.google.com/file/d/0B5RJWjAhdT04MTVvanRHcWtlODg/view?usp=sharing)


================================================
FILE: tools/_init_paths.py
================================================
import os.path as osp
import sys

def add_path(path):
    if path not in sys.path:
        sys.path.insert(0, path)

this_dir = osp.dirname(__file__)

# Add caffe to PYTHONPATH
caffe_path = osp.join(this_dir, '..', '..', '..', 'caffes', 'caffe-newest', 'python')
add_path(caffe_path)

# Add lib to PYTHONPATH
lib_path = osp.join(this_dir, '..', 'lib')
add_path(lib_path)


================================================
FILE: tools/eval_triplet_recall.py
================================================
#!/usr/bin/env python

import _init_paths
import argparse

import time, os, sys
import json
import cv2
import cPickle as cp
import numpy as np
import math
from utils.eval_utils import computeIoU

def parse_args():
	"""
	Parse input arguments
	"""
	parser = argparse.ArgumentParser()
	# gt file format: [ gt_label, gt_box ]
	# 	gt_label: list [ gt_label(image_i) for image_i in images ]
	# 		gt_label(image_i): numpy.array of size: num_instance x 3
	# 			instance: [ label_s, label_r, label_o ]
	# 	gt_box: list [ gt_box(image_i) for image_i in images ]
	#		gt_box(image_i): numpy.array of size: num_instance x 2 x 4
	#			instance: [ [x1_s, y1_s, x2_s, y2_s], 
	#				    [x1_o, y1_o, x2_o, y2_o]]
	parser.add_argument('--gt_file', dest='gt_file',
						help='file containing gts',
						default=None, type=str)
	parser.add_argument('--num_dets', dest='num_dets',
						help='max number of detections per image',
						default=50, type=int)
	# det file format: [ det_label, det_box ]
	# 	det_label: list [ det_label(image_i) for image_i in images ]
	# 		det_label(image_i): numpy.array of size: num_instance x 6
	# 			instance: [ prob_s, prob_r, prob_o, label_s, label_r, label_o ]
	# 	det_box: list [ det_box(image_i) for image_i in images ]
	#		det_box(image_i): numpy.array of size: num_instance x 2 x 4
	#			instance: [ [x1_s, y1_s, x2_s, y2_s], 
	#				    [x1_o, y1_o, x2_o, y2_o]]
	parser.add_argument('--det_file', dest='det_file', 
						help='file containing triplet detections',
						default=None, type=str)
	
	parser.add_argument('--min_overlap', dest='ov_thresh',
						help='minimum overlap for a correct detection',
						default=0.5, type=float)


	if len(sys.argv) == 1:
		parser.print_help()
		sys.exit(1)

	args = parser.parse_args()
	return args

def computeOverlap(detBBs, gtBBs):
	aIoU = computeIoU(detBBs[0, :], gtBBs[0, :])
	bIoU = computeIoU(detBBs[1, :], gtBBs[1, :])
	return min(aIoU, bIoU)		

def eval_recall(args):
	f = open(args.det_file, "r")
	dets, det_bboxes = cp.load(f)
	f.close()
	f = open(args.gt_file, "r")
	all_gts, all_gt_bboxes = cp.load(f)
	f.close()
	num_img = len(dets)
	tp = []
	fp = []
	score = []
	total_num_gts = 0
	for i in xrange(num_img):
		gts = all_gts[i]
		gt_bboxes = all_gt_bboxes[i]
		num_gts = gts.shape[0]
		total_num_gts += num_gts
		gt_detected = np.zeros(num_gts)
		if isinstance(dets[i], np.ndarray) and dets[i].shape[0] > 0:
			det_score = np.log(dets[i][:, 0]) + np.log(dets[i][:, 1]) + np.log(dets[i][:, 2])
			inds = np.argsort(det_score)[::-1]
			if args.num_dets > 0 and args.num_dets < len(inds):
				inds = inds[:args.num_dets]
			top_dets = dets[i][inds, 3:]
			top_scores = det_score[inds]
			top_det_bboxes = det_bboxes[i][inds, :]
			num_dets = len(inds)
			for j in xrange(num_dets):
				ov_max = 0
				arg_max = -1
				for k in xrange(num_gts):
					if gt_detected[k] == 0 and top_dets[j, 0] == gts[k, 0] and top_dets[j, 1] == gts[k, 1] and top_dets[j, 2] == gts[k, 2]:
						ov = computeOverlap(top_det_bboxes[j, :, :], gt_bboxes[k, :, :])
						if ov >= args.ov_thresh and ov > ov_max:
							ov_max = ov
							arg_max = k
				if arg_max != -1:
					gt_detected[arg_max] = 1
					tp.append(1)
					fp.append(0)
				else:
					tp.append(0)
					fp.append(1)
				score.append(top_scores[j])
	score = np.array(score)
	tp = np.array(tp)
	fp = np.array(fp)
	inds = np.argsort(score)
	inds = inds[::-1]
	tp = tp[inds]
	fp = fp[inds]
	tp = np.cumsum(tp)
	fp = np.cumsum(fp)
	recall = (tp + 0.0) / total_num_gts
	top_recall = recall[-1] 
	print top_recall						

if __name__ == '__main__':
	args = parse_args()

	print('Called with args:')
	print(args)

	eval_recall(args)
	


================================================
FILE: tools/eval_union_recall.py
================================================
#!/usr/bin/env python

import _init_paths
import argparse

import time, os, sys
import json
import cv2
import cPickle as cp
import numpy as np
import math
from utils.eval_utils import computeIoU

def parse_args():
	"""
	Parse input arguments
	"""
	parser = argparse.ArgumentParser()
	# gt file format: [ gt_label, gt_box ]
	# 	gt_label: list [ gt_label(image_i) for image_i in images ]
	# 		gt_label(image_i): numpy.array of size: num_instance x 3
	# 			instance: [ label_s, label_r, label_o ]
	# 	gt_box: list [ gt_box(image_i) for image_i in images ]
	#		gt_box(image_i): numpy.array of size: num_instance x 2 x 4
	#			instance: [ [x1_s, y1_s, x2_s, y2_s], 
	#				    [x1_o, y1_o, x2_o, y2_o]]
	parser.add_argument('--gt_file', dest='gt_file',
						help='file containing gts',
						default=None, type=str)
	parser.add_argument('--num_dets', dest='num_dets',
						help='max number of detections per image',
						default=50, type=int)
	# det file format: [ det_label, det_box ]
	# 	det_label: list [ det_label(image_i) for image_i in images ]
	# 		det_label(image_i): numpy.array of size: num_instance x 6
	# 			instance: [ prob_s, prob_r, prob_o, label_s, label_r, label_o ]
	# 	det_box: list [ det_box(image_i) for image_i in images ]
	#		det_box(image_i): numpy.array of size: num_instance x 2 x 4
	#			instance: [ [x1_s, y1_s, x2_s, y2_s], 
	#				    [x1_o, y1_o, x2_o, y2_o]]
	parser.add_argument('--det_file', dest='det_file', 
						help='file containing triplet detections',
						default=None, type=str)
	
	parser.add_argument('--min_overlap', dest='ov_thresh',
						help='minimum overlap for a correct detection',
						default=0.5, type=float)


	if len(sys.argv) == 1:
		parser.print_help()
		sys.exit(1)

	args = parser.parse_args()
	return args

def getUnionBB(aBB, bBB):
	return [min(aBB[0], bBB[0]), \
		min(aBB[1], bBB[1]), \
		max(aBB[2], bBB[2]), \
		max(aBB[3], bBB[3])]	

def computeOverlap(detBBs, gtBBs):
	aIoU = computeIoU(detBBs[0, :], gtBBs[0, :])
	bIoU = computeIoU(detBBs[1, :], gtBBs[1, :])
	return min(aIoU, bIoU)		

def eval_recall(args):
	f = open(args.det_file, "r")
	dets, det_bboxes = cp.load(f)
	f.close()
	f = open(args.gt_file, "r")
	all_gts, all_gt_bboxes = cp.load(f)
	f.close()
	num_img = len(dets)
	tp = []
	fp = []
	score = []
	total_num_gts = 0
	for i in xrange(num_img):
		gts = all_gts[i]
		gt_bboxes = all_gt_bboxes[i]
		gt_ubbs = [] 	
		num_gts = gts.shape[0]
		for j in xrange(num_gts):
			gt_ubbs.append(getUnionBB(gt_bboxes[j, 0, :], gt_bboxes[j, 1, :]))
		total_num_gts += num_gts
		gt_detected = np.zeros(num_gts)
		if isinstance(dets[i], np.ndarray) and dets[i].shape[0] > 0:
			det_score = np.log(dets[i][:, 0]) + np.log(dets[i][:, 1]) + np.log(dets[i][:, 2])
			inds = np.argsort(det_score)[::-1]
			if args.num_dets > 0 and args.num_dets < len(inds):
				inds = inds[:args.num_dets]
			top_dets = dets[i][inds, 3:]
			top_scores = det_score[inds]
			top_det_bboxes = det_bboxes[i][inds, :]
			top_det_ubbs = []
			num_dets = len(inds)
			for j in xrange(num_dets):
				top_det_ubbs.append(getUnionBB(top_det_bboxes[j, 0, :], top_det_bboxes[j, 1, :]))
			for j in xrange(num_dets):
				ov_max = 0
				arg_max = -1
				for k in xrange(num_gts):
					if gt_detected[k] == 0 and top_dets[j, 0] == gts[k, 0] and top_dets[j, 1] == gts[k, 1] and top_dets[j, 2] == gts[k, 2]:
						ov = computeIoU(top_det_ubbs[j], gt_ubbs[k])
						if ov >= args.ov_thresh and ov > ov_max:
							ov_max = ov
							arg_max = k
				if arg_max != -1:
					gt_detected[arg_max] = 1
					tp.append(1)
					fp.append(0)
				else:
					tp.append(0)
					fp.append(1)
				score.append(top_scores[j])
	score = np.array(score)
	tp = np.array(tp)
	fp = np.array(fp)
	inds = np.argsort(score)
	inds = inds[::-1]
	tp = tp[inds]
	fp = fp[inds]
	tp = np.cumsum(tp)
	fp = np.cumsum(fp)
	recall = (tp + 0.0) / total_num_gts
	top_recall = recall[-1] 
	print top_recall						

if __name__ == '__main__':
	args = parse_args()

	print('Called with args:')
	print(args)

	eval_recall(args)
	


================================================
FILE: tools/prepare_data.py
================================================
import os.path as osp
import mat4py as mp
import cv2
import numpy as np
import math
import json

class DataLoader:
	def __init__(self, datasetRoot, split):
		self._root = datasetRoot
		self._split = split
		self._loadLabels()
		self._loadAnnotation(split)

	def _loadLabels(self):
		mat = mp.loadmat(osp.join(self._root, "predicate.mat"))
		self._relList = mat["predicate"]
		self._numRelClass = len(self._relList)
		self._relMapping = {}
		for i in xrange(len(self._relList)):
			self._relMapping[self._relList[i]] = i	
		mat = mp.loadmat(osp.join(self._root, "objectListN.mat"))
		self._objList = mat["objectListN"]
		self._numObjClass = len(self._objList) + 1
		self._objMapping = {}
		self._objMapping["__BG"] = 0
		for i in xrange(len(self._objList)):
			self._objMapping[self._objList[i]] = i + 1	

	def _loadAnnotation(self, split):
		mat = mp.loadmat(osp.join(self._root, "annotation_" + split + ".mat"))
		self._annotations = mat["annotation_" + split]

	def _getNumImgs(self):
		return len(self._annotations)

	def _getImPath(self, idx):
		return osp.join(self._root, "images", self._split, self._annotations[idx]["filename"])
	
	def _getNumRel(self):
		numRels = 0
		n = self._getNumImgs()
		for i in xrange(n):
			rels = self._getRels(i)
			numRels += len(rels)	
		return numRels

	def _getRels(self, idx):
		if "relationship" in self._annotations[idx]:
			rels = self._annotations[idx]["relationship"] 
			if isinstance(rels, dict):
				rels = [rels]	
			return rels
		else:
			return []

	def _outputDB(self, type, data):
		json.dump(data, open(type + self._split + ".json", "w"))

	def _bboxTransform(self, bbox, ih, iw): #[x1, y1, x2, y2]
		return [max(bbox[2], 0), max(bbox[0], 0), min(bbox[3] + 1, iw), min(bbox[1] + 1, ih)]
	
	def _getRelLabel(self, predicate):
		if not (predicate in self._relMapping):
			return -1
		return self._relMapping[predicate]

	def _getObjLabel(self, predicate):
		if not (predicate in self._objMapping):
			return -1
		return self._objMapping[predicate]

	def _getUnionBBox(self, aBB, bBB, ih, iw, margin = 10):
		return [max(0, min(aBB[0], bBB[0]) - margin), \
			max(0, min(aBB[1], bBB[1]) - margin), \
			min(iw, max(aBB[2], bBB[2]) + margin), \
			min(ih, max(aBB[3], bBB[3]) + margin)]

	def _getRelSamplesSingle(self):
		n = self._getNumImgs()
		self._sampleIdx = 0
		samples = []
		for i in xrange(n):
			rels = self._getRels(i)
			if len(rels) == 0:
				continue
			path = self._getImPath(i)
			im = cv2.imread(path)
			ih = im.shape[0]	
			iw = im.shape[1]
			for rel in rels:
				phrase = rel["phrase"]
				rLabel = self._getRelLabel(phrase[1])
				aLabel = self._getObjLabel(phrase[0])
				bLabel = self._getObjLabel(phrase[2])
				aBBox = self._bboxTransform(rel["subBox"], ih, iw)
				bBBox = self._bboxTransform(rel["objBox"], ih, iw)
				rBBox = self._getUnionBBox(aBBox, bBBox, ih, iw)	
				samples.append({"imPath": path, "rLabel": rLabel, "aLabel": aLabel, "bLabel": bLabel, "rBBox": rBBox, "aBBox": aBBox, "bBBox": bBBox})
				self._sampleIdx += 1	
				if self._sampleIdx % 100 == 0:
					print self._sampleIdx
		self._outputDB("rel", samples)	
		
if __name__ == "__main__":
	loader = DataLoader("/datasets/vrd", "test")
	loader._getRelSamplesSingle()


================================================
FILE: tools/test_predicate_recognition.py
================================================
#!/usr/bin/env python

import _init_paths
import caffe
import argparse
import time, os, sys
import json
import cv2
import cPickle as cp
import numpy as np
import math

def parse_args():
	"""
	Parse input arguments
	"""
	parser = argparse.ArgumentParser()
	# image_paths file format: [ path of image_i for image_i in images ]
	# the order of images in image_paths should be the same with gt_file
	parser.add_argument('--image_paths', dest='image_paths', help='file containing image paths',
						default='', type=str)

	parser.add_argument('--gpu', dest='gpu_id', help='GPU id to use',
						default=0, type=int)
	parser.add_argument('--def', dest='prototxt',
						help='prototxt file defining the network',
						default=None, type=str)
	parser.add_argument('--net', dest='caffemodel',
						help='model to test',
						default=None, type=str)
	# gt file format: [ gt_label, gt_box ]
	# 	gt_label: list [ gt_label(image_i) for image_i in images ]
	# 		gt_label(image_i): numpy.array of size: num_instance x 3
	# 			instance: [ label_s, label_r, label_o ]
	# 	gt_box: list [ gt_box(image_i) for image_i in images ]
	#		gt_box(image_i): numpy.array of size: num_instance x 2 x 4
	#			instance: [ [x1_s, y1_s, x2_s, y2_s], 
	#				    [x1_o, y1_o, x2_o, y2_o]]
	parser.add_argument('--gt_file', dest='gt_file', 
						help='file containing ground truth pairs',
						default=None, type=str)

	parser.add_argument('--ncls', dest='ncls', help='number of object classes', default=101, type=int)

	parser.add_argument('--input_type', dest='type',
						help='type of input sets',
						default=0, type=int)

	parser.add_argument('--out', dest='out', help='name of output file', default='', type=str)

	if len(sys.argv) == 1:
		parser.print_help()
		sys.exit(1)

	args = parser.parse_args()
	return args


def getUnionBBox(aBB, bBB, ih, iw):
	margin = 10
	return [max(0, min(aBB[0], bBB[0]) - margin), \
		max(0, min(aBB[1], bBB[1]) - margin), \
		min(iw, max(aBB[2], bBB[2]) + margin), \
		min(ih, max(aBB[3], bBB[3]) + margin)]

def getAppr(im, bb):
	subim = im[bb[1] : bb[3], bb[0] : bb[2], :]
	subim = cv2.resize(subim, None, None, 224.0 / subim.shape[1], 224.0 / subim.shape[0], interpolation=cv2.INTER_LINEAR)
	pixel_means = np.array([[[103.939, 116.779, 123.68]]])
	subim -= pixel_means
	subim = subim.transpose((2, 0, 1))
	return subim	

def getDualMask(ih, iw, bb):
	rh = 32.0 / ih
	rw = 32.0 / iw	
	x1 = max(0, int(math.floor(bb[0] * rw)))
	x2 = min(32, int(math.ceil(bb[2] * rw)))
	y1 = max(0, int(math.floor(bb[1] * rh)))
	y2 = min(32, int(math.ceil(bb[3] * rh)))
	mask = np.zeros((32, 32))
	mask[y1 : y2, x1 : x2] = 1
	assert(mask.sum() == (y2 - y1) * (x2 - x1))
	return mask

def forward_batch(net, ims, poses, qas, qbs, args):
	forward_args = {}
	if args.type != 1:
		net.blobs["im"].reshape(*(ims.shape))
		forward_args["im"] = ims.astype(np.float32, copy=False)
	if args.type != 0:
		net.blobs["posdata"].reshape(*(poses.shape))
		forward_args["posdata"] = poses.astype(np.float32, copy=False)
	if args.type == 3:
		net.blobs["qa"].reshape(*(qas.shape))
		forward_args["qa"] = qas.astype(np.float32, copy=False)
		net.blobs["qb"].reshape(*(qbs.shape))
		forward_args["qb"] = qbs.astype(np.float32, copy=False)

	net_out = net.forward(**forward_args)
	itr_pred = net_out["pred"].copy()
	return itr_pred

def test_net(net, image_paths, args):
	f = open(args.gt_file, "r")
	all_gts, all_gt_bboxes = cp.load(f)
	f.close()
	num_img = len(image_paths)
	num_class = args.ncls
	thresh = 0.05
	batch_size = 20
	pred = []
	pred_bboxes = []
	for i in xrange(num_img):
		print str(i) + " / " + str(num_img)
		im = cv2.imread(image_paths[i]).astype(np.float32, copy=False)
		ih = im.shape[0]
		iw = im.shape[1]
		gts = all_gts[i]
		gt_bboxes = all_gt_bboxes[i]
		num_gts = gts.shape[0]
		pred.append([])
		pred_bboxes.append([])
		ims = []
		poses = []
		qas = []
		qbs = []
		for j in xrange(num_gts):	
			sub = gt_bboxes[j, 0, :]
			obj = gt_bboxes[j, 1, :]
			rBB = getUnionBBox(sub, obj, ih, iw)
			rAppr = getAppr(im, rBB)	
			rMask = np.array([getDualMask(ih, iw, sub), getDualMask(ih, iw, obj)])
			ims.append(rAppr)
			poses.append(rMask)
			qa = np.zeros(num_class - 1)
			qa[gts[j, 0] - 1] = 1
			qb = np.zeros(num_class - 1)
			qb[gts[j, 2] - 1] = 1
			qas.append(qa)
			qbs.append(qb)
		if len(ims) == 0:
			continue
		ims = np.array(ims)
		poses = np.array(poses)
		qas = np.array(qas)
		qbs = np.array(qbs)
		_cursor = 0
		itr_pred = None
		num_ins = ims.shape[0]
		while _cursor < num_ins:
			_end_batch = min(_cursor + batch_size, num_ins)
			itr_pred_batch = forward_batch(net, ims[_cursor : _end_batch], poses[_cursor : _end_batch], qas[_cursor : _end_batch], qbs[_cursor : _end_batch], args)
			if itr_pred is None:
				itr_pred = itr_pred_batch
			else:
				itr_pred = np.vstack((itr_pred, itr_pred_batch))
			_cursor = _end_batch
		
		for j in xrange(num_gts):	
			sub = gt_bboxes[j, 0, :]
			obj = gt_bboxes[j, 1, :]
			for k in xrange(itr_pred.shape[1]):
				if itr_pred[j, k] < thresh: 
					continue
				pred[i].append([itr_pred[j, k], 1, 1, gts[j, 0], k, gts[j, 2]])
				pred_bboxes[i].append([sub, obj])
		pred[i] = np.array(pred[i])
		pred_bboxes[i] = np.array(pred_bboxes[i])	
	print "writing file.."
	f = open(args.out, "wb")
	cp.dump([pred, pred_bboxes], f, cp.HIGHEST_PROTOCOL)			
	f.close()

if __name__ == '__main__':
	args = parse_args()

	print('Called with args:')
	print(args)

	caffe.set_mode_gpu()
	caffe.set_device(args.gpu_id)
	net = caffe.Net(args.prototxt, args.caffemodel, caffe.TEST)

	test_image_paths = json.load(open(args.image_paths))

	test_net(net, test_image_paths, args)
	


================================================
FILE: tools/test_triplet_detection.py
================================================
#!/usr/bin/env python

import _init_paths
import caffe
import argparse
import time, os, sys
import json
import cv2
import cPickle as cp
import numpy as np
import math

def parse_args():
	"""
	Parse input arguments
	"""
	parser = argparse.ArgumentParser()
	# image_paths file format: [ path of image_i for image_i in images ]
	# the order of images in image_paths should be the same with obj_dets_file
	parser.add_argument('--image_paths', dest='image_paths', help='file containing test dataset',
						default='', type=str)
	parser.add_argument('--gpu', dest='gpu_id', help='GPU id to use',
						default=0, type=int)
	parser.add_argument('--def', dest='prototxt',
						help='prototxt file defining the network',
						default=None, type=str)
	parser.add_argument('--net', dest='caffemodel',
						help='model to test',
						default=None, type=str)
	parser.add_argument('--num_dets', dest='max_det',
						help='max number of detections per image',
						default=100, type=int)
	# obj_dets_file format: [ obj_dets of image_i for image_i in images ]
	# 	obj_dets: numpy.array of size: num_instance x 5
	# 		instance: [x1, y1, x2, y2, prob, label]
	parser.add_argument('--obj_dets_file', dest='obj_dets_file', 
						help='file containing object detections',
						default=None, type=str)
	# type 0: im only
	# type 1: pos only
	# type 2: im + pos
	# type 3: im + pos + qa + qb
	parser.add_argument('--input_type', dest='type',
						help='type of input sets',
						default=0, type=int)

	parser.add_argument('--ncls', dest='num_class', help='number of object classes', default=101, type=int)

	parser.add_argument('--out', dest='out', help='name of output file', default='', type=str)

	if len(sys.argv) == 1:
		parser.print_help()
		sys.exit(1)

	args = parser.parse_args()
	return args


def getPred(pred, max_num_det):
	if pred.shape[0] == 0:
		return pred
	inds = np.argsort(pred[:, 4])
	inds = inds[::-1]
	if len(inds) > max_num_det:
		inds = inds[:max_num_det]			
	return pred[inds, :]


def getUnionBBox(aBB, bBB, ih, iw):
	margin = 10
	return [max(0, min(aBB[0], bBB[0]) - margin), \
		max(0, min(aBB[1], bBB[1]) - margin), \
		min(iw, max(aBB[2], bBB[2]) + margin), \
		min(ih, max(aBB[3], bBB[3]) + margin)]

def getAppr(im, bb):
	subim = im[bb[1] : bb[3], bb[0] : bb[2], :]
	subim = cv2.resize(subim, None, None, 224.0 / subim.shape[1], 224.0 / subim.shape[0], interpolation=cv2.INTER_LINEAR)
	pixel_means = np.array([[[103.939, 116.779, 123.68]]])
	subim -= pixel_means
	subim = subim.transpose((2, 0, 1))
	return subim	

def getDualMask(ih, iw, bb):
	rh = 32.0 / ih
	rw = 32.0 / iw	
	x1 = max(0, int(math.floor(bb[0] * rw)))
	x2 = min(32, int(math.ceil(bb[2] * rw)))
	y1 = max(0, int(math.floor(bb[1] * rh)))
	y2 = min(32, int(math.ceil(bb[3] * rh)))
	mask = np.zeros((32, 32))
	mask[y1 : y2, x1 : x2] = 1
	assert(mask.sum() == (y2 - y1) * (x2 - x1))
	return mask

def forward_batch(net, ims, poses, qas, qbs, args):
	forward_args = {}
	if args.type != 1:
		net.blobs["im"].reshape(*(ims.shape))
		forward_args["im"] = ims.astype(np.float32, copy=False)
	if args.type != 0:
		net.blobs["posdata"].reshape(*(poses.shape))
		forward_args["posdata"] = poses.astype(np.float32, copy=False)
	if args.type == 3:		
		net.blobs["qa"].reshape(*(qas.shape))
		forward_args["qa"] = qas.astype(np.float32, copy=False)
		net.blobs["qb"].reshape(*(qbs.shape))
		forward_args["qb"] = qbs.astype(np.float32, copy=False)	
	net_out = net.forward(**forward_args)
	itr_pred = net_out["pred"].copy()
	return itr_pred

def test_net(net, image_paths, args):
	f = open(args.obj_dets_file, "r")
	all_dets = cp.load(f)
	f.close()
	num_img = len(image_paths)
	num_class = args.num_class
	thresh = 0.05
	max_num_det = args.max_det
	batch_size = 30
	pred = []
	pred_bboxes = []
	for i in xrange(num_img):
		im = cv2.imread(image_paths[i]).astype(np.float32, copy=False)
		ih = im.shape[0]
		iw = im.shape[1]
			
		dets = getPred(all_dets[i], max_num_det)
		num_dets = dets.shape[0]
		pred.append([])
		pred_bboxes.append([])
		for subIdx in xrange(num_dets):
			ims = []
			poses = []
			qas = []
			qbs = []
			for objIdx in xrange(num_dets):
				if subIdx != objIdx: 
					sub = dets[subIdx, 0: 4]
					obj = dets[objIdx, 0: 4]
					rBB = getUnionBBox(sub, obj, ih, iw)
					rAppr = getAppr(im, rBB)	
					rMask = np.array([getDualMask(ih, iw, sub), getDualMask(ih, iw, obj)])
					ims.append(rAppr)
					poses.append(rMask)
					qa = np.zeros(num_class - 1)
					qa[dets[subIdx, 5] - 1] = 1
					qb = np.zeros(num_class - 1)
					qb[dets[objIdx, 5] - 1] = 1
					qas.append(qa)
					qbs.append(qb)
			if len(ims) == 0:
				break
			ims = np.array(ims)
			poses = np.array(poses)
			qas = np.array(qas)
			qbs = np.array(qbs)
			_cursor = 0
			itr_pred = None
			num_ins = ims.shape[0]
			while _cursor < num_ins:
				_end_batch = min(_cursor + batch_size, num_ins)
				itr_pred_batch = forward_batch(net, ims[_cursor : _end_batch] if ims.shape[0] > 0 else None, poses[_cursor : _end_batch] if poses.shape[0] > 0 else None, qas[_cursor : _end_batch] if qas.shape[0] > 0 else None, qbs[_cursor : _end_batch] if qbs.shape[0] > 0 else None, args)
				if itr_pred is None:
					itr_pred = itr_pred_batch
				else:
					itr_pred = np.vstack((itr_pred, itr_pred_batch))
				_cursor = _end_batch

			
			cur = 0
			for objIdx in xrange(num_dets):	
				if subIdx != objIdx:
					sub = dets[subIdx, 0: 4]
					obj = dets[objIdx, 0: 4]
					for j in xrange(itr_pred.shape[1]):
						if itr_pred[cur, j] < thresh: 
							continue
						pred[i].append([itr_pred[cur, j], dets[subIdx, 4], dets[objIdx, 4], dets[subIdx, 5], j, dets[objIdx, 5]])
						pred_bboxes[i].append([sub, obj])
					cur += 1
			assert(cur == itr_pred.shape[0])
		pred[i] = np.array(pred[i])
		pred_bboxes[i] = np.array(pred_bboxes[i])	
	print "writing file.."
	f = open(args.out, "wb")
	cp.dump([pred, pred_bboxes], f, cp.HIGHEST_PROTOCOL)			
	f.close()

if __name__ == '__main__':
	args = parse_args()

	print('Called with args:')
	print(args)

	caffe.set_mode_gpu()
	caffe.set_device(args.gpu_id)
	net = caffe.Net(args.prototxt, args.caffemodel, caffe.TEST)

	test_image_paths = json.load(open(args.image_paths))

	test_net(net, test_image_paths, args)