Full Code of YeeU/InverseRenderNet for AI

Repository: YeeU/InverseRenderNet
Branch: master
Commit: af96b366ffda
Files: 40
Total size: 71.8 KB

Directory structure:
gitextract_qj_wnxju/

├── Data/
│   ├── 037/
│   │   ├── 037_0000.pk
│   │   ├── 037_0001.pk
│   │   ├── 037_0002.pk
│   │   ├── 037_0003.pk
│   │   ├── 037_0004.pk
│   │   ├── 037_0005.pk
│   │   ├── 037_0006.pk
│   │   ├── 037_0007.pk
│   │   ├── 037_0008.pk
│   │   └── 037_0009.pk
│   └── 038/
│       ├── 038_0000.pk
│       ├── 038_0001.pk
│       ├── 038_0002.pk
│       ├── 038_0003.pk
│       ├── 038_0004.pk
│       ├── 038_0005.pk
│       ├── 038_0006.pk
│       ├── 038_0007.pk
│       ├── 038_0008.pk
│       └── 038_0009.pk
├── LICENSE
├── README.md
├── iiw_test_ids.npy
├── illu_pca/
│   ├── mean.npy
│   ├── pcaMean.npy
│   ├── pcaVariance.npy
│   └── pcaVector.npy
├── model/
│   ├── SfMNet.py
│   ├── dataloader.py
│   ├── lambSH_layer.py
│   ├── loss_layer.py
│   ├── pred_illuDecomp_layer.py
│   ├── reproj_layer.py
│   └── sup_illuDecomp_layer.py
├── pre_train_model/
│   └── .keep
├── test_demo.py
├── test_iiw.py
├── train.py
└── utils/
    ├── render_sphere_nm.py
    └── whdr.py

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

================================================
FILE: LICENSE
================================================
                                 Apache License
                           Version 2.0, January 2004
                        http://www.apache.org/licenses/

   TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION

   1. Definitions.

      "License" shall mean the terms and conditions for use, reproduction,
      and distribution as defined by Sections 1 through 9 of this document.

      "Licensor" shall mean the copyright owner or entity authorized by
      the copyright owner that is granting the License.

      "Legal Entity" shall mean the union of the acting entity and all
      other entities that control, are controlled by, or are under common
      control with that entity. For the purposes of this definition,
      "control" means (i) the power, direct or indirect, to cause the
      direction or management of such entity, whether by contract or
      otherwise, or (ii) ownership of fifty percent (50%) or more of the
      outstanding shares, or (iii) beneficial ownership of such entity.

      "You" (or "Your") shall mean an individual or Legal Entity
      exercising permissions granted by this License.

      "Source" form shall mean the preferred form for making modifications,
      including but not limited to software source code, documentation
      source, and configuration files.

      "Object" form shall mean any form resulting from mechanical
      transformation or translation of a Source form, including but
      not limited to compiled object code, generated documentation,
      and conversions to other media types.

      "Work" shall mean the work of authorship, whether in Source or
      Object form, made available under the License, as indicated by a
      copyright notice that is included in or attached to the work
      (an example is provided in the Appendix below).

      "Derivative Works" shall mean any work, whether in Source or Object
      form, that is based on (or derived from) the Work and for which the
      editorial revisions, annotations, elaborations, or other modifications
      represent, as a whole, an original work of authorship. For the purposes
      of this License, Derivative Works shall not include works that remain
      separable from, or merely link (or bind by name) to the interfaces of,
      the Work and Derivative Works thereof.

      "Contribution" shall mean any work of authorship, including
      the original version of the Work and any modifications or additions
      to that Work or Derivative Works thereof, that is intentionally
      submitted to Licensor for inclusion in the Work by the copyright owner
      or by an individual or Legal Entity authorized to submit on behalf of
      the copyright owner. For the purposes of this definition, "submitted"
      means any form of electronic, verbal, or written communication sent
      to the Licensor or its representatives, including but not limited to
      communication on electronic mailing lists, source code control systems,
      and issue tracking systems that are managed by, or on behalf of, the
      Licensor for the purpose of discussing and improving the Work, but
      excluding communication that is conspicuously marked or otherwise
      designated in writing by the copyright owner as "Not a Contribution."

      "Contributor" shall mean Licensor and any individual or Legal Entity
      on behalf of whom a Contribution has been received by Licensor and
      subsequently incorporated within the Work.

   2. Grant of Copyright License. Subject to the terms and conditions of
      this License, each Contributor hereby grants to You a perpetual,
      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
      copyright license to reproduce, prepare Derivative Works of,
      publicly display, publicly perform, sublicense, and distribute the
      Work and such Derivative Works in Source or Object form.

   3. Grant of Patent License. Subject to the terms and conditions of
      this License, each Contributor hereby grants to You a perpetual,
      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
      (except as stated in this section) patent license to make, have made,
      use, offer to sell, sell, import, and otherwise transfer the Work,
      where such license applies only to those patent claims licensable
      by such Contributor that are necessarily infringed by their
      Contribution(s) alone or by combination of their Contribution(s)
      with the Work to which such Contribution(s) was submitted. If You
      institute patent litigation against any entity (including a
      cross-claim or counterclaim in a lawsuit) alleging that the Work
      or a Contribution incorporated within the Work constitutes direct
      or contributory patent infringement, then any patent licenses
      granted to You under this License for that Work shall terminate
      as of the date such litigation is filed.

   4. Redistribution. You may reproduce and distribute copies of the
      Work or Derivative Works thereof in any medium, with or without
      modifications, and in Source or Object form, provided that You
      meet the following conditions:

      (a) You must give any other recipients of the Work or
          Derivative Works a copy of this License; and

      (b) You must cause any modified files to carry prominent notices
          stating that You changed the files; and

      (c) You must retain, in the Source form of any Derivative Works
          that You distribute, all copyright, patent, trademark, and
          attribution notices from the Source form of the Work,
          excluding those notices that do not pertain to any part of
          the Derivative Works; and

      (d) If the Work includes a "NOTICE" text file as part of its
          distribution, then any Derivative Works that You distribute must
          include a readable copy of the attribution notices contained
          within such NOTICE file, excluding those notices that do not
          pertain to any part of the Derivative Works, in at least one
          of the following places: within a NOTICE text file distributed
          as part of the Derivative Works; within the Source form or
          documentation, if provided along with the Derivative Works; or,
          within a display generated by the Derivative Works, if and
          wherever such third-party notices normally appear. The contents
          of the NOTICE file are for informational purposes only and
          do not modify the License. You may add Your own attribution
          notices within Derivative Works that You distribute, alongside
          or as an addendum to the NOTICE text from the Work, provided
          that such additional attribution notices cannot be construed
          as modifying the License.

      You may add Your own copyright statement to Your modifications and
      may provide additional or different license terms and conditions
      for use, reproduction, or distribution of Your modifications, or
      for any such Derivative Works as a whole, provided Your use,
      reproduction, and distribution of the Work otherwise complies with
      the conditions stated in this License.

   5. Submission of Contributions. Unless You explicitly state otherwise,
      any Contribution intentionally submitted for inclusion in the Work
      by You to the Licensor shall be under the terms and conditions of
      this License, without any additional terms or conditions.
      Notwithstanding the above, nothing herein shall supersede or modify
      the terms of any separate license agreement you may have executed
      with Licensor regarding such Contributions.

   6. Trademarks. This License does not grant permission to use the trade
      names, trademarks, service marks, or product names of the Licensor,
      except as required for reasonable and customary use in describing the
      origin of the Work and reproducing the content of the NOTICE file.

   7. Disclaimer of Warranty. Unless required by applicable law or
      agreed to in writing, Licensor provides the Work (and each
      Contributor provides its Contributions) on an "AS IS" BASIS,
      WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or
      implied, including, without limitation, any warranties or conditions
      of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A
      PARTICULAR PURPOSE. You are solely responsible for determining the
      appropriateness of using or redistributing the Work and assume any
      risks associated with Your exercise of permissions under this License.

   8. Limitation of Liability. In no event and under no legal theory,
      whether in tort (including negligence), contract, or otherwise,
      unless required by applicable law (such as deliberate and grossly
      negligent acts) or agreed to in writing, shall any Contributor be
      liable to You for damages, including any direct, indirect, special,
      incidental, or consequential damages of any character arising as a
      result of this License or out of the use or inability to use the
      Work (including but not limited to damages for loss of goodwill,
      work stoppage, computer failure or malfunction, or any and all
      other commercial damages or losses), even if such Contributor
      has been advised of the possibility of such damages.

   9. Accepting Warranty or Additional Liability. While redistributing
      the Work or Derivative Works thereof, You may choose to offer,
      and charge a fee for, acceptance of support, warranty, indemnity,
      or other liability obligations and/or rights consistent with this
      License. However, in accepting such obligations, You may act only
      on Your own behalf and on Your sole responsibility, not on behalf
      of any other Contributor, and only if You agree to indemnify,
      defend, and hold each Contributor harmless for any liability
      incurred by, or claims asserted against, such Contributor by reason
      of your accepting any such warranty or additional liability.

   END OF TERMS AND CONDITIONS

   APPENDIX: How to apply the Apache License to your work.

      To apply the Apache License to your work, attach the following
      boilerplate notice, with the fields enclosed by brackets "[]"
      replaced with your own identifying information. (Don't include
      the brackets!)  The text should be enclosed in the appropriate
      comment syntax for the file format. We also recommend that a
      file or class name and description of purpose be included on the
      same "printed page" as the copyright notice for easier
      identification within third-party archives.

   Copyright [yyyy] [name of copyright owner]

   Licensed under the Apache License, Version 2.0 (the "License");
   you may not use this file except in compliance with the License.
   You may obtain a copy of the License at

       http://www.apache.org/licenses/LICENSE-2.0

   Unless required by applicable law or agreed to in writing, software
   distributed under the License is distributed on an "AS IS" BASIS,
   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
   See the License for the specific language governing permissions and
   limitations under the License.


================================================
FILE: README.md
================================================
# InverseRenderNet: Learning single image inverse rendering

***!! Check out our new work InverseRenderNet++ [paper](https://arxiv.org/abs/2102.06591) and [code](https://github.com/YeeU/InverseRenderNet_v2), which improves the inverse rendering results and shadow handling.***

This is the implementation of the paper "InverseRenderNet: Learning single image inverse rendering". The model is implemented in tensorflow.

If you use our code, please cite the following paper:

    @inproceedings{yu19inverserendernet,
        title={InverseRenderNet: Learning single image inverse rendering},
        author={Yu, Ye and Smith, William AP},
        booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
        year={2019}
    }

## Evaluation

#### Dependencies
To run our evaluation code, please create your environment based on following dependencies:

    tensorflow 1.12.0
    python 3.6
    skimage
    cv2
    numpy

#### Pretrained model
* Download our pretrained model from: [Link](https://drive.google.com/uc?export=download&id=1VKeByvprmWWXSig-7-fxfXs3KA-HG_-P)
* Unzip the downloaded file 
* Make sure the model files are placed in a folder named "irn_model"


#### Test on demo image
You can perform inverse rendering on random RGB image by our pretrained model. To run the demo code, you need to specify the path to pretrained model, path to RGB image and corresponding mask which masked out sky in the image. The mask can be generated by PSPNet, which you can find on https://github.com/hszhao/PSPNet. Finally inverse rendering results will be saved to the output folder named by your argument.

```bash
python3 test_demo.py --model /PATH/TO/irn_model --image demo.jpg --mask demo_mask.jpg --output test_results
```


#### Test on IIW
* IIW dataset should be downloaded firstly from http://opensurfaces.cs.cornell.edu/publications/intrinsic/#download 

* Run testing code where you need to specify the path to model and IIW data:
```bash
python3 test_iiw.py --model /PATH/TO/irn_model --iiw /PATH/TO/iiw-dataset
```

## Training

#### Train from scratch
The training for InverseRenderNet contains two stages: pre-train and self-train.
* To begin with pre-train stage, you need to use training command specifying option `-m` to `pre-train`. 
* After finishing pre-train stage, you can run self-train by specifying option `-m` to `self-train`. 

In addition, you can control the size of batch in training, and the path to training data should be specified.

An example for training command:
```bash
python3 train.py -n 2 -p Data -m pre-train
```

#### Data for training
To directly use our code for training, you need to pre-process the training data to match the data format as shown in examples in `Data` folder. 

In particular, we pre-process the data before training, such that five images with great overlaps are bundled up into one mini-batch, and images are resized and cropped to a shape of 200 * 200 pixels. Along with input images associated depth maps, camera parameters, sky masks and normal maps are stored in the same mini-batch. For efficiency, every mini-batch containing all training elements for 5 involved images are saved as a pickle file. While training the data feeding thread directly load each mini-batch from corresponding pickle file.







================================================
FILE: model/SfMNet.py
================================================
import importlib
import tensorflow as tf
import numpy as np
import tensorflow.contrib.layers as layers

def SfMNet(inputs, height, width, name='', n_layers=12, n_pools=2, is_training=True, depth_base=64):
	conv_layers = np.int32(n_layers/2) -1
	deconv_layers = np.int32(n_layers/2)
	# number of layers before perform pooling
	nlayers_befPool = np.int32(np.ceil((conv_layers-1)/n_pools)-1)

	max_depth = 512

	if depth_base*2**n_pools < max_depth:
		tail = conv_layers - nlayers_befPool*n_pools


		tail_deconv = deconv_layers - nlayers_befPool*n_pools
	else:
		maxNum_pool = np.log2(max_depth / depth_base)
		tail = np.int32(conv_layers - nlayers_befPool * maxNum_pool)
		tail_deconv = np.int32(deconv_layers - nlayers_befPool * maxNum_pool)

	f_in_conv = [3] + [np.int32(depth_base*2**(np.ceil(i/nlayers_befPool)-1)) for i in range(1, conv_layers-tail+1)] + [np.int32(depth_base*2**maxNum_pool) for i in range(conv_layers-tail+1, conv_layers+1)]
	f_out_conv = [64] + [np.int32(depth_base*2**(np.floor(i/nlayers_befPool))) for i in range(1, conv_layers-tail+1)] + [np.int32(depth_base*2**maxNum_pool) for i in range(conv_layers-tail+1, conv_layers+1)]

	f_in_deconv = f_out_conv[:0:-1] + [64]
	f_out_amDeconv = f_in_conv[:0:-1] + [3]
	f_out_MaskDeconv = f_in_conv[:0:-1] + [2]
	f_out_nmDeconv = f_in_conv[:0:-1] + [2]



	batch_norm_params = {'decay':0.9, 'center':True, 'scale':True, 'epsilon':1e-4, 'param_initializers':{'beta_initializer':tf.zeros_initializer(),'gamma_initializer':tf.ones_initializer(),'moving_variance_initializer':tf.ones_initializer(),'moving_average_initializer':tf.zeros_initializer()}, 'param_regularizers':{'beta_regularizer':None,'gamma_regularizer':layers.l2_regularizer(scale=1e-5)},'is_training':is_training,'trainable':is_training} 

	### contractive conv_layer block
	conv_out = inputs
	conv_out_list = []
	for i,f_in,f_out in zip(range(1,conv_layers+2),f_in_conv,f_out_conv):
		scope = name+'conv'+str(i)

		if np.mod(i-1,nlayers_befPool)==0 and i<=n_pools*nlayers_befPool+1 and i != 1:
			conv_out_list.append(conv_out)
			conv_out = layers.conv2d(conv_out,num_outputs=f_out,kernel_size=[3,3],stride=[1,1],padding='SAME',normalizer_fn=layers.batch_norm, normalizer_params=batch_norm_params,weights_initializer=tf.random_normal_initializer(mean=0,stddev=np.sqrt(2/9/f_in)),weights_regularizer=layers.l2_regularizer(scale=1e-5),biases_initializer=None,scope=scope, trainable=is_training)
			conv_out = tf.nn.max_pool(conv_out, ksize=[1,2,2,1], strides=[1,2,2,1], padding='SAME')

		else:

			conv_out = layers.conv2d(conv_out,num_outputs=f_out,kernel_size=[3,3],stride=[1,1],padding='SAME',normalizer_fn=layers.batch_norm, normalizer_params=batch_norm_params,weights_initializer=tf.random_normal_initializer(mean=0,stddev=np.sqrt(2/9/f_in)),weights_regularizer=layers.l2_regularizer(scale=1e-5),biases_initializer=None,scope=scope, trainable=is_training)


	### expanding deconv_layer block succeeding conv_layer block
	am_deconv_out = conv_out
	for i,f_in,f_out in zip(range(1,deconv_layers+1),f_in_deconv,f_out_amDeconv):
		scope = name+'am/am_deconv'+str(i)

		# expand resolution every after nlayers_befPool deconv_layer
		if np.mod(i,nlayers_befPool)==0 and i<=n_pools*nlayers_befPool:
			with tf.variable_scope(scope):
				W = tf.get_variable(regularizer=layers.l2_regularizer(scale=1e-5),initializer=get_bilinear_filter([3,3,f_out,f_in],2),shape=[3,3,f_out,f_in],name='filter', trainable=is_training)	
				# import ipdb; ipdb.set_trace()
				# attach previous convolutional output to upsampling/deconvolutional output
				tmp = conv_out_list[-np.int32(i/nlayers_befPool)]
				output_shape = tf.shape(tmp)
				am_deconv_out = tf.nn.conv2d_transpose(am_deconv_out,filter=W,output_shape=output_shape,strides=[1,2,2,1],padding='SAME')
				am_deconv_out = layers.batch_norm(scope=scope,activation_fn=tf.nn.relu,inputs=am_deconv_out,decay=0.9, center=True, scale=True, param_initializers={'beta_initializer':tf.zeros_initializer(),'gamma_initializer':tf.ones_initializer(),'moving_variance_initializer':tf.ones_initializer(),'moving_average_initializer':tf.zeros_initializer()}, param_regularizers={'beta_regularizer':None,'gamma_regularizer':layers.l2_regularizer(scale=1e-5)},is_training=is_training,trainable=is_training)


			tmp = layers.conv2d(tmp,num_outputs=f_out,kernel_size=[3,3],stride=[1,1],padding='SAME',normalizer_fn=layers.batch_norm, normalizer_params=batch_norm_params,weights_initializer=tf.random_normal_initializer(mean=0,stddev=np.sqrt(2/9/f_in)),weights_regularizer=layers.l2_regularizer(scale=1e-5),biases_initializer=None,scope=scope,trainable=is_training)
			am_deconv_out = tmp + am_deconv_out


		elif i==deconv_layers:
			am_deconv_out = layers.conv2d(am_deconv_out,num_outputs=f_out,kernel_size=[3,3],stride=[1,1],padding='SAME',normalizer_fn=None,activation_fn=None,weights_initializer=tf.random_normal_initializer(mean=0,stddev=np.sqrt(2/9/f_in)),weights_regularizer=layers.l2_regularizer(scale=1e-5),scope=scope,trainable=is_training)


		else:
			am_deconv_out = layers.conv2d(am_deconv_out,num_outputs=f_out,kernel_size=[3,3],stride=[1,1],padding='SAME',normalizer_fn=layers.batch_norm, normalizer_params=batch_norm_params,weights_initializer=tf.random_normal_initializer(mean=0,stddev=np.sqrt(2/9/f_in)),weights_regularizer=layers.l2_regularizer(scale=1e-5),biases_initializer=None,scope=scope,trainable=is_training)



	### deconvolution net for nm estimates
	nm_deconv_out = conv_out
	for i,f_in,f_out in zip(range(1,deconv_layers+1),f_in_deconv,f_out_nmDeconv):
		scope = name+'nm/nm'+str(i)

		# expand resolution every after nlayers_befPool deconv_layer
		if np.mod(i,nlayers_befPool)==0 and i<=n_pools*nlayers_befPool:
			with tf.variable_scope(scope):
				W = tf.get_variable(regularizer=layers.l2_regularizer(scale=1e-5),initializer=get_bilinear_filter([3,3,f_out,f_in],2),shape=[3,3,f_out,f_in],name='filter',trainable=is_training)	

				# attach previous convolutional output to upsampling/deconvolutional output
				tmp = conv_out_list[-np.int32(i/nlayers_befPool)]
				output_shape = tf.shape(tmp)
				nm_deconv_out = tf.nn.conv2d_transpose(nm_deconv_out,filter=W,output_shape=output_shape,strides=[1,2,2,1],padding='SAME')
				nm_deconv_out = layers.batch_norm(scope=scope,activation_fn=tf.nn.relu,inputs=nm_deconv_out,decay=0.9, center=True, scale=True, epsilon=1e-4, param_initializers={'beta_initializer':tf.zeros_initializer(),'gamma_initializer':tf.ones_initializer(),'moving_variance_initializer':tf.ones_initializer(),'moving_average_initializer':tf.zeros_initializer()}, param_regularizers={'beta_regularizer':None,'gamma_regularizer':layers.l2_regularizer(scale=1e-5)},is_training=is_training,trainable=is_training)


			tmp = layers.conv2d(tmp,num_outputs=f_out,kernel_size=[3,3],stride=[1,1],padding='SAME',normalizer_fn=layers.batch_norm, normalizer_params=batch_norm_params,weights_initializer=tf.random_normal_initializer(mean=0,stddev=np.sqrt(2/9/f_in)),weights_regularizer=layers.l2_regularizer(scale=1e-5),biases_initializer=None,scope=scope,trainable=is_training)
			nm_deconv_out = tmp + nm_deconv_out


		elif i==deconv_layers:
			nm_deconv_out = layers.conv2d(nm_deconv_out,num_outputs=f_out,kernel_size=[3,3],stride=[1,1],padding='SAME',normalizer_fn=None,activation_fn=None,weights_initializer=tf.random_normal_initializer(mean=0,stddev=np.sqrt(2/9/f_in)),weights_regularizer=layers.l2_regularizer(scale=1e-5),biases_initializer=None,scope=scope,trainable=is_training)


		else:
			nm_deconv_out = layers.conv2d(nm_deconv_out,num_outputs=f_out,kernel_size=[3,3],stride=[1,1],padding='SAME',normalizer_fn=layers.batch_norm, normalizer_params=batch_norm_params,weights_initializer=tf.random_normal_initializer(mean=0,stddev=np.sqrt(2/9/f_in)),weights_regularizer=layers.l2_regularizer(scale=1e-5),biases_initializer=None,scope=scope,trainable=is_training)



	return am_deconv_out, nm_deconv_out

 

def get_bilinear_filter(filter_shape, upscale_factor):
    ##filter_shape is [width, height, num_in_channels, num_out_channels]
    kernel_size = filter_shape[1]
    ### Centre location of the filter for which value is calculated
    if kernel_size % 2 == 1:
        centre_location = upscale_factor - 1
    else:
        centre_location = upscale_factor - 0.5

    x,y = np.meshgrid(np.arange(kernel_size),np.arange(kernel_size))
    bilinear = (1 - abs((x - centre_location)/ upscale_factor)) * (1 - abs((y - centre_location)/ upscale_factor))
    weights = np.tile(bilinear[:,:,None,None],(1,1,filter_shape[2],filter_shape[3]))

    return tf.constant_initializer(weights)












================================================
FILE: model/dataloader.py
================================================
import pickle as pk
import os
import numpy as np
import tensorflow as tf
import skimage.transform as imgTform
import glob
from scipy import io


def megaDepth_dataPipeline(num_subbatch_input, dir):
	# import ipdb; ipdb.set_trace()
	# locate all scenes 
	data_scenes1 = np.array(sorted(glob.glob(os.path.join(dir, '*'))))

	# scan scenes
	# sort scenes by number of training images in each
	scenes_size1 = np.array([len(os.listdir(i)) for i in data_scenes1])
	scenes_sorted1 = np.argsort(scenes_size1)

	# define scenes for training and testing
	train_scenes = data_scenes1[scenes_sorted1]


	# load data from each scene
	# locate each data minibatch in each sorted sc
	train_scenes_items = [sorted(glob.glob(os.path.join(sc, '*.pk'))) for sc in train_scenes]
	train_scenes_items = np.concatenate(train_scenes_items, axis=0)

	train_items = train_scenes_items

	### contruct training data pipeline
	# remove residual data over number of data in one epoch
	res_train_items = len(train_items) - (len(train_items) % num_subbatch_input)
	train_items = train_items[:res_train_items]
	train_data = md_construct_inputPipeline(train_items, flag_shuffle=True, batch_size=num_subbatch_input)

	# define re-initialisable iterator
	iterator = tf.data.Iterator.from_structure(train_data.output_types, train_data.output_shapes)
	next_element = iterator.get_next()

	# define initialisation for each iterator
	trainData_init_op = iterator.make_initializer(train_data)

	return next_element, trainData_init_op, len(train_items)


def _read_pk_function(filename):
	with open(filename, 'rb') as f:
		batch_data = pk.load(f)
	input = np.float32(batch_data['input'])
	dm = batch_data['dm']
	nm = np.float32(batch_data['nm'])
	cam = np.float32(batch_data['cam'])
	scaleX= batch_data['scaleX']
	scaleY = batch_data['scaleY']
	mask = np.float32(batch_data['mask'])

	return input, dm, nm, cam, scaleX, scaleY, mask

def md_read_func(filename):

	input, dm, nm, cam, scaleX, scaleY, mask = tf.py_func(_read_pk_function, [filename], [tf.float32, tf.float32, tf.float32, tf.float32, tf.float32, tf.float32, tf.float32])	

	input = tf.data.Dataset.from_tensor_slices(input[None])
	dm = tf.data.Dataset.from_tensor_slices(dm[None])
	nm = tf.data.Dataset.from_tensor_slices(nm[None])
	cam = tf.data.Dataset.from_tensor_slices(cam[None])
	scaleX = tf.data.Dataset.from_tensor_slices(scaleX[None])
	scaleY = tf.data.Dataset.from_tensor_slices(scaleY[None])
	mask = tf.data.Dataset.from_tensor_slices(mask[None])

	return tf.data.Dataset.zip((input, dm, nm, cam, scaleX, scaleY, mask))


def md_preprocess_func(input, dm, nm, cam, scaleX, scaleY, mask):

	input = input/255.

	nm = nm/127

	return input, dm, nm, cam, scaleX, scaleY, mask


def md_construct_inputPipeline(items, batch_size, flag_shuffle=True):
	data = tf.data.Dataset.from_tensor_slices(items)
	if flag_shuffle:
		data = data.apply(tf.contrib.data.shuffle_and_repeat(buffer_size=100000))
	else:
		data = data.repeat()
	data = data.apply(tf.contrib.data.parallel_interleave(md_read_func, cycle_length=batch_size, block_length=1, sloppy=False ))
	data = data.map(md_preprocess_func, num_parallel_calls=8 )
	data = data.batch(batch_size).prefetch(4)

	return data




================================================
FILE: model/lambSH_layer.py
================================================
import tensorflow as tf


# am is the albedo map, which has shape (batch, height, width, 3[rgb]) 
# nm is the sparse normal map, which has shape (batch, height, width, 3[x,y,z])
# L_SHcoeff contains the SH coefficients for environment illumination, using 2nd order SH. L_SHcoeff has shape (batch, 9, 3[rgb])
def lambSH_layer(am, nm, L_SHcoeffs, gamma):

	""" 
	i = albedo * irradiance
	the multiplication is elementwise
	albedo is given
	irraidance = n.T * M * n, where n is (x,y,z,1)
	M is contructed from some precomputed constants and L_SHcoeff, where M contains information about illuminations, clamped cosine and SH basis
	"""

	# M is only related with lighting
	c1 = tf.constant(0.429043,dtype=tf.float32)
	c2 = tf.constant(0.511664,dtype=tf.float32)
	c3 = tf.constant(0.743125,dtype=tf.float32)
	c4 = tf.constant(0.886227,dtype=tf.float32)
	c5 = tf.constant(0.247708,dtype=tf.float32)

	# each row have shape (batch, 4, 3)
	M_row1 = tf.stack([c1*L_SHcoeffs[:,8,:], c1*L_SHcoeffs[:,4,:], c1*L_SHcoeffs[:,7,:], c2*L_SHcoeffs[:,3,:]],axis=1)
	M_row2 = tf.stack([c1*L_SHcoeffs[:,4,:], -c1*L_SHcoeffs[:,8,:], c1*L_SHcoeffs[:,5,:], c2*L_SHcoeffs[:,1,:]],axis=1)
	M_row3 = tf.stack([c1*L_SHcoeffs[:,7,:], c1*L_SHcoeffs[:,5,:], c3*L_SHcoeffs[:,6,:], c2*L_SHcoeffs[:,2,:]],axis=1)
	M_row4 = tf.stack([c2*L_SHcoeffs[:,3,:], c2*L_SHcoeffs[:,1,:], c2*L_SHcoeffs[:,2,:], c4*L_SHcoeffs[:,0,:]-c5*L_SHcoeffs[:,6,:]],axis=1)

	# M is a 5d tensot with shape (batch,4,4,3[rgb]), the axis 1 and 2 are transposely equivalent
	M = tf.stack([M_row1,M_row2,M_row3,M_row4], axis=1)

	# find batch-spatial three dimensional mask of defined normals over nm
	# mask = tf.logical_not(tf.is_nan(nm[:,:,:,0]))
	mask = tf.not_equal(tf.reduce_sum(nm,axis=-1),0)


	# extend Cartesian to homogeneous coords and extend its last for rgb individual multiplication dimension, nm_homo have shape (total_npix, 4)
	total_npix = tf.shape(nm)[:3]
	ones = tf.ones(total_npix)
	nm_homo = tf.concat([nm,tf.expand_dims(ones,axis=-1)], axis=-1)

	# contruct batch-wise flatten M corresponding with nm_homo, such that multiplication between them is batch-wise
	M = tf.expand_dims(tf.expand_dims(M,axis=1),axis=1)


	# expand M for broadcasting, such that M has shape (npix,4,4,3)
	# expand nm_homo, such that nm_homo has shape (npix,4,1,1)
	nm_homo = tf.expand_dims(tf.expand_dims(nm_homo,axis=-1),axis=-1)
	# tmp have shape (npix, 4, 3[rgb])
	tmp = tf.reduce_sum(nm_homo*M,axis=-3)
	# E has shape (npix, 3[rbg])
	E = tf.reduce_sum(tmp*nm_homo[:,:,:,:,0,:],axis=-2)


	# compute intensity by product between irradiance and albedo
	i = E*am

	# gamma correction
	i = tf.clip_by_value(i, 0., 1.) + tf.constant(1e-4)
	i = tf.pow(i,1./gamma)

	return i, mask















================================================
FILE: model/loss_layer.py
================================================
# formulate loss function based on supplied ground truth and outputs from network

import importlib
import tensorflow as tf
import numpy as np
import os
from model import SfMNet, lambSH_layer, pred_illuDecomp_layer, sup_illuDecomp_layer, reproj_layer

def loss_formulate(albedos, nm_pred, am_sup, nm_gt, inputs, dms, cams, scale_xs, scale_ys, masks, pair_label, preTrain_flag, am_smt_w_var, reproj_w_var, reg_loss_flag=True):

	# define gamma nonlinear mapping factor
	gamma = tf.constant(2.2)

	albedos = tf.nn.sigmoid(albedos) * masks + tf.constant(1e-4)

	### pre-process nm_pred such that in range (-1,1)
	nm_pred_norm = tf.sqrt(tf.reduce_sum(nm_pred**2, axis=-1, keepdims=True)+tf.constant(1.))
	nm_pred_xy = nm_pred / nm_pred_norm
	nm_pred_z = tf.constant(1.) / nm_pred_norm
	nm_pred_xyz = tf.concat([nm_pred_xy, nm_pred_z], axis=-1) * masks

	# selete normal map used in rendering - gt or pred
	normals = nm_gt if preTrain_flag else nm_pred_xyz


	# reconstruct SH lightings from predicted statistical SH lighting model
	lighting_model = '../hdr_illu_pca'
	lighting_vectors = tf.constant(np.load(os.path.join(lighting_model,'pcaVector.npy')),dtype=tf.float32)
	lighting_means = tf.constant(np.load(os.path.join(lighting_model,'mean.npy')),dtype=tf.float32)
	lightings_var = tf.constant(np.load(os.path.join(lighting_model,'pcaVariance.npy')),dtype=tf.float32)
	
	if preTrain_flag:
		lightings = sup_illuDecomp_layer.illuDecomp(inputs,albedos,nm_gt,gamma)
	else:
		lightings =pred_illuDecomp_layer.illuDecomp(inputs,albedos,nm_pred_xyz,gamma,masks)

	lightings_pca = tf.matmul((lightings - lighting_means), pinv(lighting_vectors))

	# recompute lightings from lightins_pca which could add weak constraint on lighting reconstruction 
	lightings = tf.matmul(lightings_pca,lighting_vectors) + lighting_means 

	# reshape 27-D lightings to 9*3 lightings
	lightings = tf.reshape(lightings,[tf.shape(lightings)[0],9,3])


	### lighting prior loss
	var = tf.reduce_mean(lightings_pca**2,axis=0)

	illu_prior_loss = tf.losses.absolute_difference(var, lightings_var)

	illu_prior_loss = tf.log(illu_prior_loss + 1.)


	### stereo supervision based on albedos reprojection consistancy
	reproj_tb = tf.to_float(tf.equal(pair_label,tf.transpose(pair_label)))
	reproj_tb = tf.cast(tf.matrix_set_diag(reproj_tb, tf.zeros([tf.shape(inputs)[0]])),tf.bool)
	reproj_list = tf.where(reproj_tb)
	img1_inds = tf.expand_dims(reproj_list[:,0],axis=-1)
	img2_inds = tf.expand_dims(reproj_list[:,1],axis=-1)
	albedo1 = tf.gather_nd(albedos,img1_inds)
	dms1 = tf.gather_nd(dms,img1_inds)
	cams1 = tf.gather_nd(cams,img1_inds)
	albedo2 = tf.gather_nd(albedos,img2_inds)
	cams2 = tf.gather_nd(cams,img2_inds)
	scale_xs1 = tf.gather_nd(scale_xs, img1_inds)
	scale_xs2 = tf.gather_nd(scale_xs, img2_inds)
	scale_ys1 = tf.gather_nd(scale_ys, img1_inds)
	scale_ys2 = tf.gather_nd(scale_ys, img2_inds)

	input1 = tf.gather_nd(inputs, img1_inds)

	# mask_indices contains indices for image index inside batch and spatial locations, and ignores the rgb channel index
	reproj_albedo1, reproj_mask = reproj_layer.map_reproj(dms1,albedo2,cams1,cams2,scale_xs1,scale_xs2,scale_ys1,scale_ys2)

	reproj_albedo1 = reproj_albedo1+tf.constant(1e-4) # numerical stable constant



	### scale intensities for each image
	num_imgs = tf.shape(reproj_mask)[0]
	im_ = tf.constant(0)
	output = tf.TensorArray(dtype=tf.float32,size=num_imgs)	

	def body(im_, output):
		reproj_mask_ = reproj_mask[im_]
		albedo1_ = tf.boolean_mask(albedo1[im_],reproj_mask_)
		reproj_albedo1_ = tf.boolean_mask(reproj_albedo1[im_],reproj_mask_)


		k = tf.reduce_sum(albedo1_*reproj_albedo1_,keepdims=True)/(tf.reduce_sum(reproj_albedo1_**2,keepdims=True)+tf.constant(1e-4))

		output = output.write(im_,k)
		im_ += tf.constant(1)

		return im_, output

	def condition(im_, output):
		return tf.less(im_,num_imgs)

	_,output = tf.while_loop(condition, body, loop_vars=[im_, output])


	ks = tf.expand_dims(output.stack(), axis=-1)



	albedo1_pixels = tf.boolean_mask(albedo1, reproj_mask)
	reproj_albedo1_pixels = tf.boolean_mask(reproj_albedo1*ks, reproj_mask)
	reproj_err = tf.losses.mean_squared_error(cvtLab(albedo1_pixels), cvtLab(reproj_albedo1_pixels))


	### formulate loss based on paired batches ###
	# self-supervision based on intensity reconstruction
	shadings, renderings_mask = lambSH_layer.lambSH_layer(tf.ones_like(albedos), normals, lightings, 1.)

	# compare rendering intensity by Lab
	inputs_pixels = cvtLab(tf.boolean_mask(inputs,renderings_mask))
	renderings = cvtLab(tf.boolean_mask(tf.pow(albedos*shadings,1./gamma),renderings_mask))
	render_err = tf.losses.mean_squared_error(inputs_pixels,renderings)


	### compute rendering loss from cross-projected alebdo map
	cross_shadings = tf.gather_nd(shadings, img1_inds)
	inputs_pixels = cvtLab(tf.boolean_mask(input1,reproj_mask))
	cross_renderings = cvtLab(tf.boolean_mask(tf.pow(tf.nn.relu(cross_shadings*reproj_albedo1*ks), 1./gamma),reproj_mask))
	cross_render_err = tf.losses.mean_squared_error(inputs_pixels,cross_renderings)


	### measure smoothness of albedo map
	Gx = tf.constant(1/2)*tf.expand_dims(tf.expand_dims(tf.constant([[-1,1]], dtype=tf.float32), axis=-1), axis=-1)
	Gy = tf.constant(1/2)*tf.expand_dims(tf.expand_dims(tf.constant([[-1],[1]], dtype=tf.float32), axis=-1), axis=-1)
	Gx_3 = tf.tile(Gx, multiples=(1,1,3,1))	
	Gy_3 = tf.tile(Gy, multiples=(1,1,3,1))	
	albedo_lab = tf.reshape(cvtLab(tf.reshape(albedos,[-1,3])),[-1,200,200,3])

	aGx = tf.nn.conv2d(albedos, Gx_3, padding='SAME', strides=(1,1,1,1))
	aGy = tf.nn.conv2d(albedos, Gy_3, padding='SAME', strides=(1,1,1,1))
	aGxy = tf.concat([aGx,aGy], axis=-1)


	# compute pixel-wise smoothness weights by angle distance between neighbour pixels' chromaticities
	inputs_pad = tf.pad(inputs, paddings=tf.constant([[0,0], [0,1], [0,1], [0,0]]))
	chroma_pad = tf.nn.l2_normalize(inputs_pad, axis=-1)

	chroma = chroma_pad[:,:-1,:-1,:]
	chroma_X = chroma_pad[:,:-1,1:,:]
	chroma_Y = chroma_pad[:,1:,:-1,:]
	chroma_Gx = tf.reduce_sum(chroma*chroma_X, axis=-1, keepdims=True)**tf.constant(2.) - tf.constant(1.)
	chroma_Gy = tf.reduce_sum(chroma*chroma_Y, axis=-1, keepdims=True)**tf.constant(2.) - tf.constant(1.)
	chroma_Gx = tf.exp(chroma_Gx / tf.constant(0.0001))
	chroma_Gy = tf.exp(chroma_Gy / tf.constant(0.0001))
	chroma_Gxy = tf.concat([chroma_Gx, chroma_Gy], axis=-1)

	int_pad = tf.reduce_sum(inputs_pad**tf.constant(2.), axis=-1, keepdims=True)
	int = int_pad[:,:-1,:-1,:]
	int_X = int_pad[:,:-1,1:,:]
	int_Y = int_pad[:,1:,:-1,:]

	int_Gx = tf.where(condition=int < int_X, x=int, y=int_X)
	int_Gy = tf.where(condition=int < int_Y, x=int, y=int_Y)
	int_Gx = tf.constant(1.) + tf.exp(- int_Gx / tf.constant(.8))
	int_Gy = tf.constant(1.) + tf.exp(- int_Gy / tf.constant(.8))
	int_Gxy = tf.concat([int_Gx, int_Gy], axis=-1)

	Gxy_weights = int_Gxy * chroma_Gxy
	albedo_smt_error = tf.reduce_mean(tf.abs(aGxy)*Gxy_weights)


	### albedo map pseudo-supervision loss
	if preTrain_flag:
		am_loss = tf.constant(0.)
	else:
		amSup_mask = tf.not_equal(tf.reduce_sum(nm_gt,axis=-1),0)
		am_sup_pixel = cvtLab(tf.boolean_mask(am_sup, amSup_mask))
		albedos_pixel = cvtLab(tf.boolean_mask(albedos, amSup_mask))
		am_loss = tf.losses.mean_squared_error(am_sup_pixel, albedos_pixel)



	### regualarisation loss
	reg_loss = sum(tf.get_collection(tf.GraphKeys.REGULARIZATION_LOSSES))


	### compute nm_pred error
	nmSup_mask = tf.not_equal(tf.reduce_sum(nm_gt,axis=-1),0)
	nm_gt_pixel = tf.boolean_mask(nm_gt, nmSup_mask)
	nm_pred_pixel = tf.boolean_mask(nm_pred_xyz, nmSup_mask)
	nm_prod = tf.reduce_sum(nm_pred_pixel * nm_gt_pixel, axis=-1, keepdims=True)	
	nm_cosValue = tf.constant(0.9999)
	nm_prod = tf.clip_by_value(nm_prod, -nm_cosValue, nm_cosValue)
	nm_angle = tf.acos(nm_prod) + tf.constant(1e-4)
	nm_loss = tf.reduce_mean(nm_angle**2)



	### compute gradient loss
	nm_pred_Gx = conv2d_nosum(nm_pred_xyz, Gx)
	nm_pred_Gy = conv2d_nosum(nm_pred_xyz, Gy)
	nm_pred_Gxy = tf.concat([nm_pred_Gx, nm_pred_Gy], axis=-1)
	normals_Gx = conv2d_nosum(nm_gt, Gx)
	normals_Gy = conv2d_nosum(nm_gt, Gy)
	normals_Gxy = tf.concat([normals_Gx, normals_Gy], axis=-1)

	nm_pred_smt_error = tf.losses.mean_squared_error(nm_pred_Gxy, normals_Gxy)


	### total loss
	render_err *= tf.constant(.1)
	reproj_err *= tf.constant(.05) * reproj_w_var
	cross_render_err *= tf.constant(.1)
	am_loss *= tf.constant(.1)
	illu_prior_loss *= tf.constant(.01)
	albedo_smt_error *= tf.constant(50.) * am_smt_w_var
	nm_pred_smt_error *= tf.constant(1.)
	nm_loss *= tf.constant(1.)



	if reg_loss_flag == True:
		loss = render_err + reproj_err + cross_render_err + reg_loss + illu_prior_loss + albedo_smt_error + nm_pred_smt_error + nm_loss + am_loss
	else:
		loss = render_err + reproj_err + cross_render_err + illu_prior_loss + albedo_smt_error + nm_pred_smt_error + nm_loss + am_loss

	return lightings, albedos, nm_pred_xyz, loss, render_err, reproj_err, cross_render_err, reg_loss, illu_prior_loss, albedo_smt_error, nm_pred_smt_error, nm_loss, am_loss



# input RGB is 2d tensor with shape (n_pix, 3)
def cvtLab(RGB):

	# threshold definition
	T = tf.constant(0.008856)

	# matrix for converting RGB to LUV color space
	cvt_XYZ = tf.constant([[0.412453,0.35758,0.180423],[0.212671,0.71516,0.072169],[0.019334,0.119193,0.950227]])

	# convert RGB to XYZ
	XYZ = tf.matmul(RGB,tf.transpose(cvt_XYZ))

	# normalise for D65 white point
	XYZ /= tf.constant([[0.950456, 1., 1.088754]])*100

	mask = tf.to_float(tf.greater(XYZ,T))

	fXYZ = XYZ**(1/3)*mask + (1.-mask)*(tf.constant(7.787)*XYZ + tf.constant(0.137931))

	M_cvtLab = tf.constant([[0., 116., 0.], [500., -500., 0.], [0., 200., -200.]])

	Lab = tf.matmul(fXYZ, tf.transpose(M_cvtLab)) + tf.constant([[-16., 0., 0.]])
	mask = tf.to_float(tf.equal(Lab, tf.constant(0.)))

	Lab += mask * tf.constant(1e-4)

	return Lab





# compute pseudo inverse for input matrix
def pinv(A, reltol=1e-6):
	# compute SVD of input A
	s, u, v = tf.svd(A)

	# invert s and clear entries lower than reltol*s_max
	atol = tf.reduce_max(s) * reltol
	s = tf.where(s>atol, s, atol*tf.ones_like(s))
	s_inv = tf.diag(1./s)

	# compute v * s_inv * u_t as psuedo inverse
	return tf.matmul(v, tf.matmul(s_inv, tf.transpose(u)))



# compute regular 2d convolution on 3d data
def conv2d_nosum(input, kernel):
	input_x = input[:,:,:,0:1]
	input_y = input[:,:,:,1:2]
	input_z = input[:,:,:,2:3]

	output_x = tf.nn.conv2d(input_x, kernel, strides=(1,1,1,1), padding='SAME')
	output_y = tf.nn.conv2d(input_y, kernel, strides=(1,1,1,1), padding='SAME')
	output_z = tf.nn.conv2d(input_z, kernel, strides=(1,1,1,1), padding='SAME')

	return tf.concat([output_x,output_y,output_z], axis=-1)



# compute regular 2d convolution on 3d data
def conv2d_nosum_2ch(input, kernel):
	input_x = input[:,:,:,0:1]
	input_y = input[:,:,:,1:2]

	output_x = tf.nn.conv2d(input_x, kernel, strides=(1,1,1,1), padding='SAME')
	output_y = tf.nn.conv2d(input_y, kernel, strides=(1,1,1,1), padding='SAME')

	return tf.concat([output_x,output_y], axis=-1)










================================================
FILE: model/pred_illuDecomp_layer.py
================================================
import tensorflow as tf


# am is the albedo map, which has shape (batch, height, width, 3[rgb]) 
# nm is the sparse normal map, which has shape (batch, height, width, 3[x,y,z])
# L_SHcoeff contains the SH coefficients for environment illumination, using 2nd order SH. L_SHcoeff has shape (batch, 9, 3[rgb])
def illuDecomp(input, am, nm, gamma, masks):

	""" 
	i = albedo * irradiance
	the multiplication is elementwise
	albedo is given
	irraidance = n.T * M * n, where n is (x,y,z,1)
	M is contructed from some precomputed constants and L_SHcoeff, where M contains information about illuminations, clamped cosine and SH basis
	"""

	# compute shading by dividing input by albedo
	shadings = tf.pow(input,gamma)/am
	# perform clamping on resulted shading to guarantee its numerical range
	shadings = (tf.clip_by_value(shadings, 0., 1.) + tf.constant(1e-4)) * masks


	# compute shading by linear equation regarding nm and L_SHcoeffs
	# E(n) = c1*L22*(x**2-y**2) + (c3*z**2 - c5)*L20 + c4*L00 + 2*c1*L2-2*x*y + 2*c1*L21*x*z + 2*c1*L2-1*y*z + 2*c2*L11*x + 2*c2*L1-1*y + 2*c2*L10*z
	# E(n) = c4*L00 + 2*c2*y*L1-1 + 2*c2*z*L10 + 2*c2*x*L11 + 2*c1*x*y*L2-2 + 2*c1*y*z*L2-1 + (c3*z**2 - c5)*L20 + 2*c1*x*z*L21 + c1*(x**2-y**2)*L22 
	c1 = tf.constant(0.429043,dtype=tf.float32)
	c2 = tf.constant(0.511664,dtype=tf.float32)
	c3 = tf.constant(0.743125,dtype=tf.float32)
	c4 = tf.constant(0.886227,dtype=tf.float32)
	c5 = tf.constant(0.247708,dtype=tf.float32)


	# find defined pixels
	num_iter = tf.shape(nm)[0]
	output = tf.TensorArray(dtype=tf.float32, size=num_iter)
	i = tf.constant(0)

	def condition(i, output):
		return i<num_iter

	def body(i, output):
		shadings_ = shadings[i]
		nm_ = nm[i]
		shadings_pixel = tf.reshape(shadings_, (-1,3))
		nm_ = tf.reshape(nm_, (-1,3))

		# E(n) = A*L_SHcoeffs
		total_npix = tf.shape(nm_)[0:1]
		ones = tf.ones(total_npix)
		A = tf.stack([c4*ones, 2*c2*nm_[:,1], 2*c2*nm_[:,2], 2*c2*nm_[:,0], 2*c1*nm_[:,0]*nm_[:,1], 2*c1*nm_[:,1]*nm_[:,2], c3*nm_[:,2]**2-c5, 2*c1*nm_[:,2]*nm_[:,0], c1*(nm_[:,0]**2-nm_[:,1]**2)], axis=-1)
		output = output.write(i, tf.matmul(pinv(A), shadings_pixel))
		i += tf.constant(1)

		return i, output

	_, output = tf.while_loop(condition, body, loop_vars=[i,output])
	L_SHcoeffs = output.stack()


	return tf.reshape(L_SHcoeffs, [-1,27])



def pinv(A, reltol=1e-6):
	# compute SVD of input A
	s, u, v = tf.svd(A)

	# invert s and clear entries lower than reltol*s_max
	atol = tf.reduce_max(s) * reltol
	s = tf.boolean_mask(s, s>atol)
	s_inv = tf.diag(1./s)

	# compute v * s_inv * u_t as psuedo inverse
	return tf.matmul(v, tf.matmul(s_inv, tf.transpose(u)))






================================================
FILE: model/reproj_layer.py
================================================
# apply error mask in albedo reprojection


# no rotation involved


#### directly output flatten reprojected pixels and the reconstruction mask

# the differentiable layer performing reprojection

import tensorflow as tf
import numpy as np

# pc is n-by-3 matrix containing point could three locations
# cam is the new camera parameters, whose f and p_a have shape (batch) and c has shape (batch, 2)
# dm1 is the depth map associated with cam1 that is camera for output image, which has shape (batch, height, width)
# img2 is the input image that acts as source image for reprojection, which has shape (batch, height, width, 3)
def map_reproj(dm1,map2,cam1,cam2,scale_x1,scale_x2,scale_y1,scale_y2):
	batch_size = tf.shape(dm1)[0]

	# read camera parameters
	c1 = cam1[:,2:4]
	f1 = cam1[:,0]
	p_a1 = cam1[:,1] # ratio is width divided by height
	R1 = tf.reshape(cam1[:,4:13],[-1,3,3])
	t1 = cam1[:,13:]

	c2 = cam2[:,2:4]
	f2 = cam2[:,0]
	p_a2 = cam2[:,1]
	R2 = tf.reshape(cam2[:,4:13],[-1,3,3])
	t2 = cam2[:,13:]

	# project pixel points back to camera coords
	# u is the height and v is the width
	# u and v are scalars
	u1 = tf.shape(dm1)[1]
	v1 = tf.shape(dm1)[2]

	# convert u1 and v1 to float, convenient for computation
	u1 = tf.to_float(u1)
	v1 = tf.to_float(v1)

	### regular grid in output image
	# x increase towards right, y increase toward down
	vm,um = tf.meshgrid(tf.range(1.,v1+1.), tf.range(1.,u1+1.))


	# apply scaling factors on f
	# f1 = f1/(scale_x1+scale_y1)*2
	# f1 = tf.stack([f1, f1*p_a1],axis=-1)
	f1 = tf.stack([f1/scale_x1, f1/scale_y1*p_a1],axis=-1)

	# expand f1 (batch,2,1,1), to be consistant with dm 
	f1 = tf.expand_dims(tf.expand_dims(f1,axis=-1),axis=-1)
	# expand c1 dimension (batch,2,1,1)
	c1 = tf.expand_dims(tf.expand_dims(c1,axis=-1),axis=-1)
	# expand vm and um to have shape (1,height,width)
	vm = tf.expand_dims(vm,axis=0)	
	um = tf.expand_dims(um,axis=0)	

	# compute 3D point x and y coordinates
	# Xm and Ym have shape (batch, height, width)
	Xm = (vm-c1[:,0])/f1[:,0]*dm1
	Ym = (um-c1[:,1])/f1[:,1]*dm1

	# the point cloud is (batch, 3, npix) matrix, each row is XYZ cam coords for one point
	pc = tf.stack([tf.contrib.layers.flatten(Xm), tf.contrib.layers.flatten(Ym), tf.contrib.layers.flatten(dm1)], axis=1)

	### transfer pc from coords of cam1 to cam2
	# construct homogeneous point cloud with shape batch-4-by-num_pix
	num_pix = tf.shape(pc)[-1]
	homo_pc_c1 = tf.concat([pc, tf.ones((batch_size,1,num_pix), dtype=tf.float32)], axis=1)

	# both transformation matrix have shape batch-by-4-by-4, valid for multiplication with defined homogeneous point cloud
	last_row = tf.tile(tf.constant([[[0,0,0,1]]],dtype=tf.float32), multiples=[batch_size,1,1])
	W_C_R_t1 = tf.concat([R1,tf.expand_dims(t1,axis=2)],axis=2)
	W_C_trans1 = tf.concat([W_C_R_t1, last_row], axis=1)
	W_C_R_t2 = tf.concat([R2,tf.expand_dims(t2,axis=2)],axis=2)
	W_C_trans2 = tf.concat([W_C_R_t2, last_row], axis=1)

	# batch dot product, output has shape (batch, 4, npix)	
	homo_pc_c2 = tf.matmul(W_C_trans2, tf.matmul(tf.matrix_inverse(W_C_trans1), homo_pc_c1))

	### project point cloud to cam2 pixel coordinates
	# u in vertical and v in horizontal
	u2 = tf.shape(map2)[1]
	v2 = tf.shape(map2)[2]

	# convert u2 and v2 to float
	u2 = tf.to_float(u2)
	v2 = tf.to_float(v2)
	
	# f2 = f2/(scale_x2+scale_y2)*2
	# f2 = tf.stack([f2, f2*p_a2],axis=-1)
	f2 = tf.stack([f2/scale_x2, f2/scale_y2*p_a2],axis=-1)

	# construct intrics matrics, which has shape (batch, 3, 4)
	zeros = tf.zeros_like(f2[:,0],dtype=tf.float32)	
	ones = tf.ones_like(f2[:,0],tf.float32)
	k2 = tf.stack([tf.stack([f2[:,0],zeros,c2[:,0],zeros],axis=1), tf.stack([zeros,f2[:,1],c2[:,1],zeros],axis=1), tf.stack([zeros,zeros,ones,zeros],axis=1)],axis=1)

	## manual batch dot product
	k2 = tf.expand_dims(k2,axis=-1)
	homo_pc_c2 = tf.expand_dims(homo_pc_c2,axis=1)
	# homo_uv2 has shape (batch, 3, npix)
	homo_uv2 = tf.reduce_sum(k2*homo_pc_c2,axis=2)

	# the reprojected locations of regular grid in output image
	# both have shape (batch, npix)
	v_reproj = homo_uv2[:,0,:]/homo_uv2[:,2,:]
	u_reproj = homo_uv2[:,1,:]/homo_uv2[:,2,:]

	# u and v are flatten vector containing reprojected pixel locations
	# the u and v on same index compose one pixel 
	u_valid = tf.logical_and(tf.logical_and(tf.logical_not(tf.is_nan(u_reproj)), u_reproj>0), u_reproj<u2-1)
	v_valid = tf.logical_and(tf.logical_and(tf.logical_not(tf.is_nan(v_reproj)), v_reproj>0), v_reproj<v2-1)
	# pixels has shape (batch, npix), indicating available reprojected pixels
	pixels = tf.logical_and(u_valid,v_valid)

	# pixels is bool indicator over original regular grid
	# v_reproj and u_reproj is x and y coordinates in source image
	# pixels, v_reproj and u_reproj are corresponded with each other by their indices

	### interpolation function based on source image img2
	# it has shape (total_npix, 3), the second dimension contains [img_inds, x, y]; we need to use img_inds to distinguish each pixel's request image
	# img_inds is 2d matrix with shape (batch, npix), containing img_ind for each (x,y) location
	img_inds = tf.tile(tf.expand_dims(tf.to_float(tf.range(batch_size)), axis=1), multiples=[1,num_pix])
	request_points1 = tf.stack([tf.boolean_mask(img_inds,pixels), tf.boolean_mask(v_reproj,pixels), tf.boolean_mask(u_reproj,pixels)], axis=1)



	# the output is stacked flatten pixel values for channels
	re_proj_pixs = interpImg(request_points1, map2)	

	# reconstruct original shaped re-projection map
	ndims = tf.shape(map2)[3]
	shape = [batch_size, tf.to_int32(u1), tf.to_int32(v1),3]

	pixels = tf.reshape(pixels,shape=tf.stack([batch_size, tf.to_int32(u1), tf.to_int32(v1)],axis=0))
	indices = tf.to_int32(tf.where(tf.equal(pixels,True)))

	re_proj_pixs = tf.scatter_nd(updates=re_proj_pixs, indices=indices, shape=shape)

	# re_proj_pix is flatten reprojection results with shape (total_npix, 3)
	# indices contains first three indices in original image shape for each pixel in re_proj_pixs
	return re_proj_pixs, pixels



def interpImg(unknown,data):
	# interpolate unknown data on pixel locations defined in unknown from known data with location defined in on regular grid

	# find neighbour pixels on regular grid
	# x is horizontal, y is vertical
	img_inds = tf.to_int32(unknown[:,0])
	x = unknown[:,1]
	y = unknown[:,2]
	# rgb_inds = tf.to_int32(unknown[:,3])

	low_x = tf.to_int32(tf.floor(x))
	high_x = tf.to_int32(tf.ceil(x))
	low_y = tf.to_int32(tf.floor(y))
	high_y = tf.to_int32(tf.ceil(y))

	# measure the weights for neighbourhood average based on distance
	dist_low_x = tf.expand_dims(x - tf.to_float(low_x), axis=-1)
	dist_high_x = tf.expand_dims(tf.to_float(high_x) - x, axis=-1)
	dist_low_y = tf.expand_dims(y - tf.to_float(low_y), axis=-1)
	dist_high_y = tf.expand_dims(tf.to_float(high_y) - y, axis=-1)

	# compute horizontal avarage
	avg_low_y = dist_low_x*tf.gather_nd(data, indices=tf.stack([img_inds,low_y,low_x],axis=1)) + dist_high_x*tf.gather_nd(data, indices=tf.stack([img_inds,low_y,high_x],axis=1))
	avg_high_y = dist_low_x*tf.gather_nd(data, indices=tf.stack([img_inds,high_y,low_x],axis=1)) + dist_high_x*tf.gather_nd(data, indices=tf.stack([img_inds,high_y,high_x],axis=1))

	# compute vertical average
	avg = dist_low_y*avg_low_y + dist_high_y*avg_high_y

	return avg








================================================
FILE: model/sup_illuDecomp_layer.py
================================================
import tensorflow as tf


# am is the albedo map, which has shape (batch, height, width, 3[rgb]) 
# nm is the sparse normal map, which has shape (batch, height, width, 3[x,y,z])
# L_SHcoeff contains the SH coefficients for environment illumination, using 2nd order SH. L_SHcoeff has shape (batch, 9, 3[rgb])
def illuDecomp(input, am, nm, gamma):

	""" 
	i = albedo * irradiance
	the multiplication is elementwise
	albedo is given
	irraidance = n.T * M * n, where n is (x,y,z,1)
	M is contructed from some precomputed constants and L_SHcoeff, where M contains information about illuminations, clamped cosine and SH basis
	"""

	# compute shading by dividing input by albedo
	shadings = tf.pow(input,gamma)/(am)
	# perform clamping on resulted shading to guarantee its numerical range
	shadings = tf.clip_by_value(shadings, 0., 1.) + tf.constant(1e-4)


	# compute shading by linear equation regarding nm and L_SHcoeffs
	c1 = tf.constant(0.429043,dtype=tf.float32)
	c2 = tf.constant(0.511664,dtype=tf.float32)
	c3 = tf.constant(0.743125,dtype=tf.float32)
	c4 = tf.constant(0.886227,dtype=tf.float32)
	c5 = tf.constant(0.247708,dtype=tf.float32)


	# find defined pixels
	mask = tf.not_equal(tf.reduce_sum(nm,axis=-1),0)
	num_iter = tf.shape(mask)[0]
	output = tf.TensorArray(dtype=tf.float32, size=num_iter)
	i = tf.constant(0)

	def condition(i, output):
		return i<num_iter

	def body(i, output):
		mask_ = mask[i]
		shadings_ = shadings[i]
		nm_ = nm[i]
		shadings_pixel = tf.boolean_mask(shadings_, mask_)
		nm_ = tf.boolean_mask(nm_, mask_)

		# E(n) = A*L_SHcoeffs
		total_npix = tf.shape(nm_)[0:1]
		ones = tf.ones(total_npix)
		A = tf.stack([c4*ones, 2*c2*nm_[:,1], 2*c2*nm_[:,2], 2*c2*nm_[:,0], 2*c1*nm_[:,0]*nm_[:,1], 2*c1*nm_[:,1]*nm_[:,2], c3*nm_[:,2]**2-c5, 2*c1*nm_[:,2]*nm_[:,0], c1*(nm_[:,0]**2-nm_[:,1]**2)], axis=-1)
		output = output.write(i, tf.matmul(pinv(A), shadings_pixel))
		i += tf.constant(1)

		return i, output

	_, output = tf.while_loop(condition, body, loop_vars=[i,output])
	L_SHcoeffs = output.stack()

	return tf.reshape(L_SHcoeffs, [-1,27])



def pinv(A, reltol=1e-6):
	# compute SVD of input A
	s, u, v = tf.svd(A)

	# invert s and clear entries lower than reltol*s_max
	atol = tf.reduce_max(s) * reltol
	s = tf.boolean_mask(s, s>atol)
	s_inv = tf.diag(1./s)

	# compute v * s_inv * u_t as psuedo inverse
	return tf.matmul(v, tf.matmul(s_inv, tf.transpose(u)))





================================================
FILE: pre_train_model/.keep
================================================


================================================
FILE: test_demo.py
================================================
import os
import numpy as np
import tensorflow as tf
import cv2
from skimage import io
import argparse
from model import SfMNet, lambSH_layer, pred_illuDecomp_layer
from utils import render_sphere_nm


parser = argparse.ArgumentParser(description='InverseRenderNet')
parser.add_argument('--image', help='Path to test image')
parser.add_argument('--mask', help='Path to image mask')
parser.add_argument('--model', help='Path to trained model')
parser.add_argument('--output', help='Folder saving outputs')


args = parser.parse_args()

img_path = args.image
mask_path = args.mask

img = io.imread(img_path)
mask = io.imread(mask_path)


dst_dir = args.output
os.makedirs(dst_dir)

input_height = 200
input_width = 200
ori_height, ori_width = img.shape[:2]

if ori_height / ori_width >1:
    scale = ori_width / 200
    input_height = np.int32(scale * 200)
else:
    scale = ori_height / 200
    input_width = np.int32(scale * 200)


# compute pseudo inverse for input matrix
def pinv(A, reltol=1e-6):
	# compute SVD of input A
	s, u, v = tf.svd(A)

	# invert s and clear entries lower than reltol*s_max
	atol = tf.reduce_max(s) * reltol
	s = tf.boolean_mask(s, s>atol)
	s_inv = tf.diag(1./s)

	# compute v * s_inv * u_t as psuedo inverse
	return tf.matmul(v, tf.matmul(s_inv, tf.transpose(u)))


import ipdb; ipdb.set_trace()
inputs_var = tf.placeholder(tf.float32, (None, input_height, input_width, 3))
masks_var = tf.placeholder(tf.float32, (None, input_height, input_width, 1))
am_deconvOut, nm_deconvOut = SfMNet.SfMNet(inputs=inputs_var,is_training=False, height=input_height, width=input_width, n_layers=30, n_pools=4, depth_base=32)


# separate albedo, error mask and shadow mask from deconvolutional output
albedos = am_deconvOut
nm_pred = nm_deconvOut

gamma = tf.constant(2.2)

# post-process on raw albedo and nm_pred
albedos = tf.nn.sigmoid(albedos) * masks_var + tf.constant(1e-4)

nm_pred_norm = tf.sqrt(tf.reduce_sum(nm_pred**2, axis=-1, keepdims=True)+tf.constant(1.))
nm_pred_xy = nm_pred / nm_pred_norm
nm_pred_z = tf.constant(1.) / nm_pred_norm
nm_pred_xyz = tf.concat([nm_pred_xy, nm_pred_z], axis=-1) * masks_var


# compute illumination
lighting_model = 'illu_pca'
lighting_vectors = tf.constant(np.load(os.path.join(lighting_model,'pcaVector.npy')),dtype=tf.float32)
lighting_means = tf.constant(np.load(os.path.join(lighting_model,'mean.npy')),dtype=tf.float32)	
lightings = pred_illuDecomp_layer.illuDecomp(inputs_var, albedos, nm_pred_xyz, gamma, masks_var)


lightings_pca = tf.matmul((lightings - lighting_means), pinv(lighting_vectors))
lightings = tf.matmul(lightings_pca,lighting_vectors) + lighting_means 
# reshape 27-D lightings to 9*3 lightings
lightings = tf.reshape(lightings,[tf.shape(lightings)[0],9,3])

# visualisations
shading, _ = lambSH_layer.lambSH_layer(tf.ones_like(albedos), nm_pred_xyz, lightings, 1.)
nm_sphere = tf.constant(render_sphere_nm.render_sphere_nm(100,1),dtype=tf.float32)
nm_sphere = tf.tile(nm_sphere, (tf.shape(inputs_var)[0],1,1,1))
lighting_recon, _ = lambSH_layer.lambSH_layer(tf.ones_like(nm_sphere), nm_sphere, lightings, 1.)


irn_vars = tf.get_collection(tf.GraphKeys.GLOBAL_VARIABLES, scope='conv') + tf.get_collection(tf.GraphKeys.GLOBAL_VARIABLES, scope='am') + tf.get_collection(tf.GraphKeys.GLOBAL_VARIABLES, scope='nm')
model_path = tf.train.get_checkpoint_state(args.model).model_checkpoint_path

total_loss = 0 
sess = tf.InteractiveSession()
saver = tf.train.Saver(irn_vars)
saver.restore(sess, model_path)


# evaluation
ori_img = img
ori_height, ori_width = ori_img.shape[:2]
img = cv2.resize(img, (input_width, input_height))
img = np.float32(img)/255.
img = img[None, :, :, :]
mask = cv2.resize(mask, (input_width, input_height), cv2.INTER_NEAREST)
mask = np.float32(mask==255)[None,:,:,None]

[albedos_val, nm_pred_val, lighting_recon_val, shading_val] = sess.run([albedos, nm_pred_xyz, lighting_recon, shading], feed_dict={inputs_var:img, masks_var:mask})


# post-process results
nm_pred_val = (nm_pred_val+1.)/2.

albedos_val = cv2.resize(albedos_val[0], (ori_width, ori_height))
shading_val = cv2.resize(shading_val[0], (ori_width, ori_height))
lighting_recon_val = lighting_recon_val[0]
nm_pred_val = cv2.resize(nm_pred_val[0], (ori_width, ori_height))


albedos_val = (albedos_val-albedos_val.min()) / (albedos_val.max()-albedos_val.min())

albedos_val = np.uint8(albedos_val*255.)
shading_val = np.uint8(shading_val*255.)
lighting_recon_val = np.uint8(lighting_recon_val*255.)
nm_pred_val = np.uint8(nm_pred_val*255.)

input_path = os.path.join(dst_dir, 'img.png')
io.imsave(input_path, ori_img)
albedo_path = os.path.join(dst_dir, 'albedo.png')
io.imsave(albedo_path, albedos_val)
shading_path = os.path.join(dst_dir, 'shading.png')
io.imsave(shading_path, shading_val)
nm_pred_path = os.path.join(dst_dir, 'nm_pred.png')
io.imsave(nm_pred_path, nm_pred_val)
lighting_path = os.path.join(dst_dir, 'lighting.png')
io.imsave(lighting_path, lighting_recon_val)




================================================
FILE: test_iiw.py
================================================
import json
import os
import numpy as np
import tensorflow as tf
import importlib
import cv2
from skimage import io
import argparse
from model import SfMNet, lambSH_layer, pred_illuDecomp_layer
from glob import glob
from utils.whdr import compute_whdr


parser = argparse.ArgumentParser(description='InverseRenderNet')
parser.add_argument('--iiw', help='Root directory for iiw-dataset')
parser.add_argument('--model', help='Path to trained model')


args = parser.parse_args()

iiw = args.iiw
test_ids = np.load('iiw_test_ids.npy')



input_height = 200
input_width = 200



# compute pseudo inverse for input matrix
def pinv(A, reltol=1e-6):
	# compute SVD of input A
	s, u, v = tf.svd(A)

	# invert s and clear entries lower than reltol*s_max
	atol = tf.reduce_max(s) * reltol
	s = tf.boolean_mask(s, s>atol)
	s_inv = tf.diag(1./s)

	# compute v * s_inv * u_t as psuedo inverse
	return tf.matmul(v, tf.matmul(s_inv, tf.transpose(u)))



inputs_var = tf.placeholder(tf.float32, (None, input_height, input_width, 3))
masks_var = tf.placeholder(tf.float32, (None, input_height, input_width, 1))
train_flag = tf.placeholder(tf.bool, ())
am_deconvOut, _ = SfMNet.SfMNet(inputs=inputs_var,is_training=train_flag, height=input_height, width=input_width, n_layers=30, n_pools=4, depth_base=32)


# separate albedo, error mask and shadow mask from deconvolutional output
albedos = am_deconvOut

# post-process on raw albedo and nm_pred
albedos = tf.nn.sigmoid(albedos) * masks_var + tf.constant(1e-4)

irn_vars = tf.get_collection(tf.GraphKeys.GLOBAL_VARIABLES, scope='conv') + tf.get_collection(tf.GraphKeys.GLOBAL_VARIABLES, scope='am') + tf.get_collection(tf.GraphKeys.GLOBAL_VARIABLES, scope='nm')
model_path = tf.train.get_checkpoint_state(args.model).model_checkpoint_path

total_loss = 0 
sess = tf.InteractiveSession()
saver = tf.train.Saver(irn_vars)
saver.restore(sess, model_path)


for counter, test_id in enumerate(test_ids):
    img_file = str(test_id)+'.png'
    judgement_file = str(test_id)+'.json'

    img_path = os.path.join(iiw, 'data', img_file)
    judgement_path = os.path.join(iiw, 'data', judgement_file)

    img = io.imread(img_path)
    judgement = json.load(open(judgement_path))

    ori_width, ori_height = img.shape[:2]

    img = cv2.resize(img, (input_width, input_height))
    img = np.float32(img)/255.
    img = img[None, :, :, :]
    mask = np.ones((1, input_height, input_width, 1), np.bool)


    [albedos_val] = sess.run([albedos], feed_dict={train_flag:False, inputs_var:img, masks_var:mask})

    albedos_val = cv2.resize(albedos_val[0], (ori_width, ori_height))

    albedos_val = (albedos_val-albedos_val.min()) / (albedos_val.max()-albedos_val.min())
    albedos_val = albedos_val/2+.5


    loss = compute_whdr(albedos_val, judgement)
    total_loss += loss
    print('whdr:{:f}\twhdr_avg:{:f}'.format(loss, total_loss/(counter+1)))


print("IIW TEST WHDR %f"%(total_loss/len(test_ids)))




================================================
FILE: train.py
================================================
# also predict shadow mask and error mask

# no rotation


#### compute albedo reproj loss only on reprojection available area; compute reconstruction and its loss only based on defined area


import tensorflow as tf
import importlib
import os
import pickle as pk
import sys
import numpy as np
import time
import argparse
from PIL import Image
import glob
from model import SfMNet, lambSH_layer, pred_illuDecomp_layer, loss_layer, dataloader


parser = argparse.ArgumentParser(description='InverseRenderNet')
parser.add_argument('--n_batch', '-n', help='number of minibatch', type=int)
parser.add_argument('--data_path', '-p', help='Path to training data')
parser.add_argument('--train_mode', '-m', help='specify the phase for training (pre-train/self-train)', choices={'pre-train', 'self-train'})


args = parser.parse_args()

def main():

	inputs_shape = (5,200,200,3)

	next_element, trainData_init_op, num_train_batches = dataloader.megaDepth_dataPipeline(args.n_batch, args.data_path)

	inputs_var = tf.reshape(next_element[0], (-1, inputs_shape[1], inputs_shape[2], inputs_shape[3]))
	dms_var = tf.reshape(next_element[1], (-1, inputs_shape[1], inputs_shape[2]))
	nms_var = tf.reshape(next_element[2], (-1, inputs_shape[1], inputs_shape[2], 3))
	cams_var = tf.reshape(next_element[3], (-1, 16))
	scaleXs_var = tf.reshape(next_element[4], (-1,))
	scaleYs_var = tf.reshape(next_element[5], (-1,))
	masks_var = tf.reshape(next_element[6], (-1, inputs_shape[1], inputs_shape[2]))

	# var helping cross projection
	pair_label_var = tf.constant(np.repeat(np.arange(args.n_batch),inputs_shape[0])[:,None], dtype=tf.float32)
	# weights for smooth loss and am_consistency loss
	am_smt_w_var = tf.placeholder(tf.float32, ())
	reproj_w_var = tf.placeholder(tf.float32, ())

	# mask out sky in inputs and nms
	masks_var_4d = tf.expand_dims(masks_var, axis=-1)
	inputs_var *= masks_var_4d
	nms_var *= masks_var_4d

	# inverserendernet
	if args.train_mode == 'pre-train':
		am_deconvOut, nm_deconvOut = SfMNet.SfMNet(inputs=inputs_var,is_training=True, height=inputs_shape[1], width=inputs_shape[2], name='pre_train_IRN/', n_layers=30, n_pools=4, depth_base=32)

		am_sup = tf.zeros_like(am_deconvOut)
		preTrain_flag = True


	elif args.train_mode == 'self-train':
		am_deconvOut, nm_deconvOut = SfMNet.SfMNet(inputs=inputs_var,is_training=True, height=inputs_shape[1], width=inputs_shape[2], name='IRN/', n_layers=30, n_pools=4, depth_base=32)

		am_sup, _ = SfMNet.SfMNet(inputs=inputs_var,is_training=False, height=inputs_shape[1], width=inputs_shape[2], name='pre_train_IRN/', n_layers=30, n_pools=4, depth_base=32)
		am_sup = tf.nn.sigmoid(am_sup) * masks_var_4d + tf.constant(1e-4)

		preTrain_flag = False

	# separate albedo, error mask and shadow mask from deconvolutional output
	albedoMaps = am_deconvOut[:,:,:,:3]

	# formulate loss
	light_SHCs, albedoMaps, nm_preds, loss, render_err, reproj_err, cross_render_err, reg_loss, illu_prior_loss, albedo_smt_error, nm_smt_loss, nm_loss, am_loss = loss_layer.loss_formulate(albedoMaps, nm_deconvOut, am_sup, nms_var, inputs_var, dms_var, cams_var, scaleXs_var, scaleYs_var, masks_var_4d, pair_label_var, True, am_smt_w_var, reproj_w_var, reg_loss_flag=True)


	# defined traning loop
	epochs = 30
	num_batches = num_train_batches
	num_subbatch = args.n_batch
	num_iters = np.int32(np.ceil(num_batches/num_subbatch))


	# training op
	global_step = tf.Variable(1,name='global_step',trainable=False)

	train_step = tf.contrib.layers.optimize_loss(loss, optimizer=tf.train.AdamOptimizer(learning_rate=.05, epsilon=1e-1), learning_rate=None, global_step=global_step)

	# define saver for saving and restoring
	irn_vars = tf.get_collection(tf.GraphKeys.GLOBAL_VARIABLES, scope='IRN') if args.train_mode == 'self-train' else tf.get_collection(tf.GraphKeys.GLOBAL_VARIABLES, scope='pre_train_IRN')
	saver = tf.train.Saver(irn_vars)

	# define session
	config = tf.ConfigProto(allow_soft_placement=True)
	config.gpu_options.allow_growth = True
	sess = tf.InteractiveSession(config=config)

	# train from scratch or keep training trained model
	tf.local_variables_initializer().run()
	tf.global_variables_initializer().run()

	assignOps = []
	if args.train_mode == 'self-train':
		# load am_sup net
		preTrain_irn_vars = tf.get_collection(tf.GraphKeys.GLOBAL_VARIABLES, scope='pre_train_IRN')
		saver_loadOldVar = tf.train.Saver(preTrain_irn_vars)
		saver_loadOldVar.restore(sess, 'pre_train_model/model.ckpt')

		# import ipdb; ipdb.set_trace()
		# duplicate pre_train model
		with tf.variable_scope(tf.get_variable_scope(), reuse=tf.AUTO_REUSE):
			vars = tf.contrib.framework.list_variables('pre_train_model')
			for var_name, _ in vars:
				var = tf.contrib.framework.load_variable('pre_train_model', var_name)
				new_var_name = var_name.replace('pre_train_IRN', 'IRN')

				new_var = tf.get_variable(name=new_var_name)
				assignOps += [new_var.assign(var)]

		sess.run(assignOps)




	# start training
	trainData_init_op.run()
	dst_dir = 'irn_model' if args.train_mode == 'self-train' else 'pre_train_model'
	for i in range(1,epochs+1):

		loss_avg = 0
		f = open('cost.txt','a')

		# graduately update weights if pre-training
		reproj_weight = .2 + np.clip(.8 * (i-16)/14, 0., .8) if args.train_mode == 'pre-train' else 1.
		am_smt_weight = .2 + np.clip(.8 * (i-1)/14, 0., .8) if args.train_mode == 'pre-train' else 1.

		for j in range(1,num_iters+1):
			start_time = time.time()

			# train
			[loss_val, reg_loss_val, render_err_val, reproj_err_val, cross_render_err_val, illu_prior_val, albedo_smt_error_val, nm_smt_loss_val, nm_loss_val, am_loss_val] = sess.run([train_step, reg_loss, render_err, reproj_err, cross_render_err, illu_prior_loss, albedo_smt_error, nm_smt_loss, nm_loss, am_loss], feed_dict={am_smt_w_var:am_smt_weight, reproj_w_var:reproj_weight})
			loss_avg += loss_val

			# log
			if j % 1 == 0:
				print('iter %d/%d loop %d/%d took %.3fs' % (i,epochs,j,num_iters,time.time()-start_time))
				print('\tloss_avg = %f, loss = %f' % (loss_avg / j,loss_val))
				print('\t\treg_loss = %f, render_err = %f, reproj_err = %f, cross_render_err = %f, illu_prior = %f, albedo_smt_error = %f, nm_smt_loss = %f, nm_loss = %f, am_loss = %f' % (reg_loss_val, render_err_val, reproj_err_val, cross_render_err_val, illu_prior_val, albedo_smt_error_val, nm_smt_loss_val, nm_loss_val, am_loss_val))

				f.write('iter %d/%d loop %d/%d took %.3fs\n\tloss_avg = %f, loss = %f\n\t\treg_loss = %f, render_err = %f, reproj_err = %f, cross_render_err = %f, illu_prior = %f, albedo_smt_error = %f, nm_smt_loss = %f, nm_loss = %f, am_loss = %f\n' % (i,epochs,j,num_iters,time.time()-start_time,loss_avg/j, loss_val, reg_loss_val, render_err_val, reproj_err_val, cross_render_err_val, illu_prior_val, albedo_smt_error_val, nm_smt_loss_val, nm_loss_val, am_loss_val))

		f.close()

		# save model every 10 iterations
		saver.save(sess,os.path.join(dst_dir, 'model.ckpt'))


if __name__ == '__main__':
	main()







================================================
FILE: utils/render_sphere_nm.py
================================================
import numpy as np

def render_sphere_nm(radius, num):
	# nm is a batch of normal maps
	nm = []

	for i in range(num):
		### hemisphere
		height = 2*radius
		width = 2*radius
		centre = radius
		x_grid, y_grid = np.meshgrid(np.arange(1.,2*radius+1), np.arange(1.,2*radius+1))
		# grids are (-radius, radius)
		x_grid -= centre
		# y_grid -= centre
		y_grid = centre - y_grid
		# scale range of h and w grid in (-1,1)
		x_grid /= radius
		y_grid /= radius
		dist = 1 - (x_grid**2+y_grid**2)
		mask = dist > 0
		z_grid = np.ones_like(mask) * np.nan
		z_grid[mask] = np.sqrt(dist[mask])

		# remove xs and ys by masking out nans in zs
		x_grid[~(mask)] = np.nan
		y_grid[~(mask)] = np.nan

		# concatenate normal map
		nm.append(np.stack([x_grid,y_grid,z_grid],axis=2))



		### sphere 
		# span the regular grid for computing azimuth and zenith angular map
		# height = 2*radius
		# width = 2*radius
		# centre = radius
		# h_grid, v_grid = np.meshgrid(np.arange(1.,2*radius+1), np.arange(1.,2*radius+1))
		# # grids are (-radius, radius)
		# h_grid -= centre
		# # v_grid -= centre
		# v_grid = centre - v_grid
		# # scale range of h and v grid in (-1,1)
		# h_grid /= radius
		# v_grid /= radius

		# # z_grid is linearly spread along theta/zenith in range (0,pi)
		# dist_grid = np.sqrt(h_grid**2+v_grid**2)
		# dist_grid[dist_grid>1] = np.nan
		# theta_grid = dist_grid * np.pi
		# z_grid = np.cos(theta_grid)

		# rho_grid = np.arctan2(v_grid,h_grid)
		# x_grid = np.sin(theta_grid)*np.cos(rho_grid)
		# y_grid = np.sin(theta_grid)*np.sin(rho_grid)

		# # concatenate normal map
		# nm.append(np.stack([x_grid,y_grid,z_grid],axis=2))


	# construct batch
	nm = np.stack(nm,axis=0)



	return nm



================================================
FILE: utils/whdr.py
================================================
#!/usr/bin/env python2.7
#
# This is an implementation of the WHDR metric proposed in this paper:
#
#     Sean Bell, Kavita Bala, Noah Snavely. "Intrinsic Images in the Wild". ACM
#     Transactions on Graphics (SIGGRAPH 2014). http://intrinsic.cs.cornell.edu.
#
# Please cite the above paper if you find this code useful.  This code is
# released under the MIT license (http://opensource.org/licenses/MIT).
#


import sys
import json
import argparse
import numpy as np
from PIL import Image


def compute_whdr(reflectance, judgements, delta=0.10):
    """ Return the WHDR score for a reflectance image, evaluated against human
    judgements.  The return value is in the range 0.0 to 1.0, or None if there
    are no judgements for the image.  See section 3.5 of our paper for more
    details.

    :param reflectance: a numpy array containing the linear RGB
    reflectance image.

    :param judgements: a JSON object loaded from the Intrinsic Images in
    the Wild dataset.

    :param delta: the threshold where humans switch from saying "about the
    same" to "one point is darker."
    """

    points = judgements['intrinsic_points']
    comparisons = judgements['intrinsic_comparisons']
    id_to_points = {p['id']: p for p in points}
    rows, cols = reflectance.shape[0:2]

    error_sum = 0.0
    weight_sum = 0.0

    for c in comparisons:
        # "darker" is "J_i" in our paper
        darker = c['darker']
        if darker not in ('1', '2', 'E'):
            continue

        # "darker_score" is "w_i" in our paper
        weight = c['darker_score']
        if weight <= 0 or weight is None:
            continue

        point1 = id_to_points[c['point1']]
        point2 = id_to_points[c['point2']]
        if not point1['opaque'] or not point2['opaque']:
            continue

        # convert to grayscale and threshold
        l1 = max(1e-10, np.mean(reflectance[
            int(point1['y'] * rows), int(point1['x'] * cols), ...]))
        l2 = max(1e-10, np.mean(reflectance[
            int(point2['y'] * rows), int(point2['x'] * cols), ...]))

        # convert algorithm value to the same units as human judgements
        if l2 / l1 > 1.0 + delta:
            alg_darker = '1'
        elif l1 / l2 > 1.0 + delta:
            alg_darker = '2'
        else:
            alg_darker = 'E'

        if darker != alg_darker:
            error_sum += weight
        weight_sum += weight

    if weight_sum:
        return error_sum / weight_sum
    else:
        return None


def load_image(filename, is_srgb=True):
    """ Load an image that is either linear or sRGB-encoded. """

    if not filename:
        raise ValueError("Empty filename")
    image = np.asarray(Image.open(filename)).astype(np.float) / 255.0
    if is_srgb:
        return srgb_to_rgb(image)
    else:
        return image


def srgb_to_rgb(srgb):
    """ Convert an sRGB image to a linear RGB image """

    ret = np.zeros_like(srgb)
    idx0 = srgb <= 0.04045
    idx1 = srgb > 0.04045
    ret[idx0] = srgb[idx0] / 12.92
    ret[idx1] = np.power((srgb[idx1] + 0.055) / 1.055, 2.4)
    return ret


if __name__ == "__main__":
    parser = argparse.ArgumentParser(
        description=(
            'Evaluate an intrinsic image decomposition using the WHDR metric presented in:\n'
            '    Sean Bell, Kavita Bala, Noah Snavely. "Intrinsic Images in the Wild".\n'
            '    ACM Transactions on Graphics (SIGGRAPH 2014).\n'
            '    http://intrinsic.cs.cornell.edu.\n'
            '\n'
            'The output is in the range 0.0 to 1.0.'
        )
    )

    parser.add_argument(
        'reflectance', metavar='<reflectance.png>',
        help='reflectance image to be evaluated')

    parser.add_argument(
        'judgements', metavar='<judgements.json>',
        help='human judgements JSON file')

    parser.add_argument(
        '-l', '--linear', action='store_true', required=False,
        help='assume the reflectance image is linear, otherwise assume sRGB')

    parser.add_argument(
        '-d', '--delta', metavar='<float>', type=float, required=False, default=0.10,
        help='delta threshold (default 0.10)')

    if len(sys.argv) < 2:
        parser.print_help()
        sys.exit(1)

    args = parser.parse_args()
    reflectance = load_image(filename=args.reflectance, is_srgb=(not args.linear))
    judgements = json.load(open(args.judgements))

    whdr = compute_whdr(reflectance, judgements, args.delta)
    print(whdr)