Full Code of garion9013/impl-pruning-TF for AI

Repository: garion9013/impl-pruning-TF
Branch: master
Commit: a185420c7bf8
Files: 16
Total size: 50.0 MB

Directory structure:
gitextract_z8dibw7e/

├── README.md
├── config.py
├── deploy_test.py
├── deploy_test_pruned.py
├── draw_histogram.py
├── model_ckpt_dense
├── model_ckpt_dense.meta
├── model_ckpt_dense_pruned
├── model_ckpt_sparse_retrained
├── papl.py
├── read_model.py
├── sparse_model_extreme/
│   ├── model_ckpt_dense_pruned
│   ├── model_ckpt_dense_retrained
│   └── model_ckpt_sparse_retrained
├── thspace.py
└── train.py

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

================================================
FILE: README.md
================================================
## TensorFlow implementation of "Iterative Pruning"

**CAUTION**: Out-of-date notices.

Currently, I've checked TF (>1.3) supports *sparse_matmul* and it seems that this is
more correct way to implement iterative pruning. This work is just naively done with quite old
versions (0.8.0) and thus, I do not recommend to consider these codes for your serious cases. And there will be no updates or maintenance either.

---

This work is based on "Learning both Weights and Connections for Efficient
Neural Network." [Song et al.](http://arxiv.org/pdf/1506.02626v3.pdf) @ NIPS '15.
Note that these works are just for quantifying its effectiveness on latency (within TensorFlow),
not a best optimal. Thus, some details are abbreviated for simplicity. (e.g. # of iterations, adjusted dropout ratio, etc.)

I applied Iterative Pruning on a small MNIST CNN model (13MB, originally), which can be
accessed from [TensorFlow Tutorials](https://www.tensorflow.org/versions/r0.8/tutorials/mnist/pros/index.html).
After pruning off some percentages of weights, I've simply retrained two epochs for
each case and got compressed models (minimum 2.6MB with 90% off) with minor loss of accuracy.
(99.17% -> 98.99% with 90% off and retraining) Again, this is not an optimal.

## Issues

Due to lack of supports on SparseTensor and its operations of TensorFlow (0.8.0),
this implementation has some limitations. This work uses [*embedding_lookup_sparse*](https://www.tensorflow.org/versions/r0.8/api_docs/python/nn.html#embedding_lookup_sparse) to compute sparse matrix-vector multiplication.
It is not solely for the purpose of sparse matrix vector multiplication, and thus its performance may be sub-optimal. (I'm not sure.)
Also, TensorFlow uses \<index, value\> pair for sparse matrix rather than
using typical CSR format which is more compact and performant.
In summary, because of the following reasons, I think this implementation has some limitations.

1. *embedding_lookup_sparse* doesn't support ```broadcasting```, which prohibits users to run test with normal test datasets.
2. Performance may be somewhat sub-optimal.
3. Because "Sparse Variable" is not supported, manual dense to sparse and sparse to dense transformation is required.
4. 4D Convolution Tensor may also be applicable, but bit tricky.
5. Current *embedding_lookup_sparse* forces additional matrix transpose, dimension squeeze and dimension reshape.

## File descriptions and usages

model_ckpt_dense: original model<br>
model_ckpt_dense_pruned: 90% pruned-only model<br>
model_ckpt_sparse_retrained: 90% pruned and retrained model<br>

#### Python package requirements
```bash
sudo apt-get install python-scipy python-numpy python-matplotlib
```

To regenerate these sparse model, edit ```config.py``` first as your threshold configuration,
and then run training with second (pruning and retraining) and third (generate sparse form of weight data) round options.

```bash
./train.py -2 -3
```

To inference single image (seven.png) and measure its latency,

```bash
./deploy_test.py -d -m model_ckpt_dense
./deploy_test_sparse.py -d -m model_ckpt_sparse_retrained
```

To test dense model,

```bash
./deploy_test.py -t -m model_ckpt_dense
./deploy_test.py -t -m model_ckpt_dense_pruned
./deploy_test.py -t -m model_ckpt_dense_retrained
```

To draw histogram that shows the weight distribution,

```bash
# After running train.py (it generates .dat files)
./draw_histogram.py
```

## Performance
Results are currently somewhat mediocre or degraded due to indirection and additional storage overhead originated from sparse matrix form.
Also, it may because model size is too small. (12.49MB)

#### Storage overhead
Baseline: 12.49 MB<br>
10 % pruned: 21.86 MB<br>
20 % pruned: 19.45 MB<br>
30 % pruned: 17.05 MB<br>
40 % pruned: 14.64 MB<br>
50 % pruned: 12.23 MB<br>
60 % pruned: 9.83 MB<br>
70 % pruned: 7.42 MB<br>
80 % pruned: 5.02 MB<br>
90 % pruned: 2.61 MB<br>

#### CPU performance (5 times averaged)
CPU: Intel Core i5-2500 @ 3.3 GHz,
LLC size: 6 MB

<img src=http://younghwanoh.github.io/images/cpu-desktop.png alt=http://younghwanoh.github.io/images/cpu-desktop.png>

Baseline: 0.01118040085 s<br>
10 % pruned: 1.919299984   s<br>
20 % pruned: 0.2325239658  s<br>
30 % pruned: 0.2111079693  s<br>
40 % pruned: 0.1982570648  s<br>
50 % pruned: 0.1691776752  s<br>
60 % pruned: 0.1305227757  s<br>
70 % pruned: 0.116039753   s<br>
80 % pruned: 0.103564167   s<br>
90 % pruned: 0.1058168888  s<br>

#### GPU performance (5 times averaged)
GPU: Nvidia Geforce GTX650 @ 1.058 GHz,
LLC size: 256 KB

<img src=http://younghwanoh.github.io/images/gpu-desktop.png alt=http://younghwanoh.github.io/images/gpu-desktop.png>

Baseline: 0.1475181845 s<br>
10 % pruned: 0.2954540253 s<br>
20 % pruned: 0.2665398121 s<br>
30 % pruned: 0.2585638046 s<br>
40 % pruned: 0.2090051651 s<br>
50 % pruned: 0.1995279789 s<br>
60 % pruned: 0.1815193653 s<br>
70 % pruned: 0.1436806202 s<br>
80 % pruned: 0.135668993  s<br>
90 % pruned: 0.1218701839 s<br>










================================================
FILE: config.py
================================================
#!/usr/bin/python
import thspace as ths

def _complex_concat(a, b):
    tmp = []
    for i in a:
        for j in b:
            tmp.append(i+j)
    return tmp

def _add_prefix(a):
    tmp = []
    for idx, val in enumerate(a):
        tmp.append("w_" + val)
        # tmp.append("b_" + val)
    return tmp

# Pruning threshold setting (90 % off)
th = ths.th90

# CNN settings for pruned training
target_layer = ["fc1", "fc2"]
retrain_iterations = 10

# Output data lists: do not change this
target_all_layer = _add_prefix(target_layer)

target_dat = _complex_concat(target_all_layer, [".dat"])
target_p_dat = _complex_concat(target_all_layer, ["_p.dat"])
target_tp_dat = _complex_concat(target_all_layer, ["_tp.dat"])

weight_all = target_dat + target_p_dat + target_tp_dat
syn_all = ["in_conv1.syn", "in_conv2.syn", "in_fc1.syn", "in_fc2.syn"]

# Data settings
show_zero = False

# Graph settings
alpha = 0.75
color = "green"
pdf_prefix = ""


================================================
FILE: deploy_test.py
================================================
#!/usr/bin/python

import sys
sys.dont_write_bytecode = True

import tensorflow as tf
import numpy as np
import argparse
import papl
import config

argparser = argparse.ArgumentParser()
argparser.add_argument("-t", "--test", action="store_true", help="Run test")
argparser.add_argument("-d", "--deploy", action="store_true", help="Run deploy with seven.png")
argparser.add_argument("-s", "--print_syn", action="store_true", help="Print synapses to .syn")
argparser.add_argument("-m", "--model", default="./model_ckpt_dense", help="Specify a target model file")
args = argparser.parse_args()

if (args.test or args.deploy or args.print_syn) == True:
    from tensorflow.examples.tutorials.mnist import input_data
    mnist = input_data.read_data_sets('/tmp/data/', one_hot=True)
else:
    argparser.print_help()
    sys.exit()

# sess = tf.InteractiveSession(config=tf.ConfigProto(gpu_options=tf.GPUOptions(allow_growth=True)))
sess = tf.InteractiveSession()
# sess = tf.Session()

def imgread(path):
    tmp = papl.imread(path)
    img = np.zeros((28,28,1))
    img[:,:,0]=tmp[:,:,0]
    img = np.reshape(img, img.size)
    return img

# Restore values of variables
saver = tf.train.import_meta_graph(args.model+'.meta')
saver.restore(sess, args.model)

# Calc results
if args.test == True:
    # Evaluate test sets
    import time
    accuracy = tf.get_collection("accuracy")[0]

    # To avoid OOM, run validation with 500/10000 test dataset
    b = time.time()
    result = 0
    for i in range(20):
        batch = mnist.test.next_batch(500)
        result += sess.run(accuracy, feed_dict={"x:0": batch[0],
                                                "y_:0": batch[1],
                                                "keep_prob:0": 1.0})
    result /= 20
    a = time.time()

    print("Test accuracy %g" % result)
    print "Time: %s s" % (a-b)
elif args.deploy == True:
    # Infer a single image & check its latency
    import time
    img = imgread('seven.png')
    y_conv = tf.get_collection("y_conv")[0]

    b = time.time()
    result = sess.run(tf.argmax(y_conv,1), feed_dict={"x:0":[img],
                                                      "y_:0":mnist.test.labels,
                                                      "keep_prob:0": 1.0})
    a = time.time()

    print "Output: %s" % result
    print "Time: %s s" % (a-b)
    papl.log("performance_ref.log", a-b)

elif args.print_syn == True:
    # Print synapses (Input data of each neuron)
    img = imgread('seven.png')
    target_syn = config.syn_all
    synapses = [ tf.get_collection(elem.split(".")[0])[0] for elem in target_syn ]
    for i,j in zip(synapses, config.syn_all):
        syn = sess.run(i, feed_dict={"x:0":[img],
                                     "y_:0":mnist.test.labels,
                                     "keep_prob:0": 1.0})
        papl.print_synapse_nps(syn, j)
    print "Done! Synapse data is printed to x.syn"


================================================
FILE: deploy_test_pruned.py
================================================
#!/usr/bin/python

import sys
sys.dont_write_bytecode = True

import tensorflow as tf
import numpy as np
import argparse
import config
import papl

argparser = argparse.ArgumentParser()
argparser.add_argument("-t", "--test", action="store_true", help="Run test")
argparser.add_argument("-d", "--deploy", action="store_true", help="Run deploy with seven.png")
argparser.add_argument("-m", "--model", default="./model_ckpt_sparse_retrained", help="Specify a target model file")
args = argparser.parse_args()

if (args.test) == True:
    print "Error: TensorFlow 0.8 doesn't support broadcasts on sparse operations, cannot run test set now"
    sys.exit()
elif (args.deploy) == True:
    from tensorflow.examples.tutorials.mnist import input_data
    mnist = input_data.read_data_sets('/tmp/data/', one_hot=True)
else:
    argparser.print_help()
    sys.exit()

# sess = tf.InteractiveSession(config=tf.ConfigProto(gpu_options=tf.GPUOptions(allow_growth=True)))
sess = tf.InteractiveSession(config=tf.ConfigProto(device_count={'GPU':0}))
# sess = tf.Session()

def imgread(path):
    tmp = papl.imread(path)
    img = np.zeros((28,28,1))
    img[:,:,0]=tmp[:,:,0]
    return img

# Declare weight variables
sparse_w={
    "w_conv1": tf.Variable(tf.truncated_normal([5, 5, 1, 32], stddev=0.1), name="w_conv1"),
    "b_conv1": tf.Variable(tf.constant(0.1, shape=[32]), name="b_conv1"),
    "w_conv2": tf.Variable(tf.truncated_normal([5, 5, 32, 64], stddev=0.1), name="w_conv2"),
    "b_conv2": tf.Variable(tf.constant(0.1, shape=[64]), name="b_conv2"),
    "w_fc1":      tf.Variable(tf.zeros([config.th["fc1_nnz"]],  dtype=tf.float32),name="w_fc1"),
    "w_fc1_idx":  tf.Variable(tf.zeros([config.th["fc1_nnz"],2],dtype=tf.int32),  name="w_fc1_idx"),
    "w_fc1_shape":tf.Variable(tf.zeros([2],     dtype=tf.int32),  name="w_fc1_shape"),
    "b_fc1":      tf.Variable(tf.zeros([1024], dtype=tf.float32), name="b_fc1"),
    "w_fc2":      tf.Variable(tf.zeros([config.th["fc2_nnz"]],  dtype=tf.float32),name="w_fc2"),
    "w_fc2_idx":  tf.Variable(tf.zeros([config.th["fc2_nnz"],2],dtype=tf.int32),  name="w_fc2_idx"),
    "w_fc2_shape":tf.Variable(tf.zeros([2],     dtype=tf.int32),  name="w_fc2_shape"),
    "b_fc2":      tf.Variable(tf.zeros([10], dtype=tf.float32), name="b_fc2"),
}

def sparse_cnn_model(weights):
    def conv2d(x, W):
        return tf.nn.conv2d(x, W, strides=[1, 1, 1, 1], padding='SAME')
    def max_pool_2x2(x):
        return tf.nn.max_pool(x, ksize=[1, 2, 2, 1],
                              strides=[1, 2, 2, 1], padding='SAME')

    h_conv1 = tf.nn.relu(conv2d(x_image, weights["w_conv1"]) + weights["b_conv1"])
    h_pool1 = max_pool_2x2(h_conv1)
    h_conv2 = tf.nn.relu(conv2d(h_pool1, weights["w_conv2"]) + weights["b_conv2"])
    h_pool2 = max_pool_2x2(h_conv2)
    h_pool2_flat = tf.squeeze(tf.reshape(h_pool2, [-1, 7*7*64]))
    h_fc1 = tf.nn.relu(tf.nn.embedding_lookup_sparse(h_pool2_flat, weights["w_fc1_ids"], weights["w_fc1"], combiner="sum") + weights["b_fc1"])
    h_fc1_drop = tf.squeeze(tf.nn.dropout(h_fc1, keep_prob))
    y_conv = tf.nn.relu(tf.nn.embedding_lookup_sparse(h_fc1_drop, weights["w_fc2_ids"], weights["w_fc2"], combiner="sum") + weights["b_fc2"])
    y_conv = tf.nn.softmax(tf.reshape(y_conv, [1, -1]))

    return y_conv

# Restore values of variables
saver = tf.train.Saver()
saver.restore(sess, args.model)

# Retrieve SparseTensor from serialized dense variables
sparse_w["w_fc1"] = tf.SparseTensor(sparse_w["w_fc1_idx"].eval(),
                                    sparse_w["w_fc1"].eval(),
                                    sparse_w["w_fc1_shape"].eval())
sparse_w["w_fc2"] = tf.SparseTensor(sparse_w["w_fc2_idx"].eval(),
                                    sparse_w["w_fc2"].eval(),
                                    sparse_w["w_fc2_shape"].eval())
sparse_w["w_fc1_ids"] = tf.SparseTensor(sparse_w["w_fc1_idx"].eval(),
                                    sparse_w["w_fc1_idx"].eval()[:,1],
                                    sparse_w["w_fc1_shape"].eval())
sparse_w["w_fc2_ids"] = tf.SparseTensor(sparse_w["w_fc2_idx"].eval(),
                                    sparse_w["w_fc2_idx"].eval()[:,1],
                                    sparse_w["w_fc2_shape"].eval())

# Construct a sparse model with retrieved variables
if args.test == True:
    x = tf.placeholder("float", shape=[None, 784])
    x_image = tf.reshape(x, [-1,28,28,1])
elif args.deploy == True:
    img = imgread("./seven.png")
    x = tf.placeholder("float", shape=[None, 28, 28, 1])
    x_image = x
y_ = tf.placeholder("float", shape=[None, 10])
keep_prob = tf.placeholder("float")

y_conv = sparse_cnn_model(sparse_w)

# Calc results
if args.test == True:
    # Evaluate test sets
    import time
    correct_prediction = tf.equal(tf.argmax(y_conv,1), tf.argmax(y_,1))
    accuracy = tf.reduce_mean(tf.cast(correct_prediction, "float"))

    # To avoid OOM, run validation with 500/10000 test dataset
    b = time.time()
    result = 0
    for i in range(20):
        batch = mnist.test.next_batch(500)
        result += accuracy.eval(feed_dict={x: batch[0],
                                          y_: batch[1],
                                          keep_prob: 1.0})
    result /= 20
    a = time.time()

    print("test accuracy %g" % result)
    print "time: %s s" % (a-b)
elif args.deploy == True:
    # Infer a single image & check its latency
    import time

    b = time.time()
    result = sess.run(tf.argmax(y_conv,1), feed_dict={x:[img], y_:mnist.test.labels, keep_prob: 1.0})
    a = time.time()

    print "output: %s" % result
    print "time: %s s" % (a-b)
    papl.log("performance_ref.log", a-b)


================================================
FILE: draw_histogram.py
================================================
#!/usr/bin/python

import sys
sys.dont_write_bytecode = True

import papl
import config

papl.draw_histogram(config.weight_all, step=0.01)
papl.draw_histogram(config.syn_all, step=1)


================================================
FILE: model_ckpt_dense
================================================
[File too large to display: 12.5 MB]

================================================
FILE: model_ckpt_dense_pruned
================================================
[File too large to display: 12.5 MB]

================================================
FILE: papl.py
================================================
import csv
import matplotlib
import matplotlib.pyplot as plt
from matplotlib.ticker import FuncFormatter
from matplotlib.backends.backend_pdf import PdfPages
import numpy as np
import sys
sys.dont_write_bytecode = True

import config

# =====================================================================================
# Private methods
# =====================================================================================

def _saveToPdf(output):
    pp = PdfPages(output)
    plt.savefig(pp, format='pdf')
    pp.close()
    plt.close()

# Manipulate y-axis of histogarm
def _to_percent(y, position):
    # tick locations calculated from fraction (global).
    s = str(y*100)
    # The percent symbol needs escaping in latex
    if matplotlib.rcParams['text.usetex'] == True:
        return s + r'$\%$'
    else:
        return s + '%'

# Calc min, max position of each histogram bin
def _minRuler(array):
    minimum = min(array)
    print " - min: ", minimum
    offset = minimum % step
    return minimum - offset

def _maxRuler(array):
    maximum = max(array)
    print " - max: ", maximum
    offset = maximum % step
    return maximum - offset + step

# =====================================================================================
# Start main methods (tools)
# =====================================================================================

# Input: x.dat from global variables (config) or arguments
# Output: histogram. x.pdf
# Histogram settings are configurable through config.py
def draw_histogram(*target, **kwargs):
    if len(target) == 1:
        target = target[0]
        assert type(target) == list
        file_list = target
    else:
        file_list = config.weight_all
    global step
    step = kwargs["step"]

    for target in file_list:
        print "Target: ", target
        try:
            with open(config.pdf_prefix+"%s" % target) as text:
                x = np.float32(text.read().rstrip("\n").split("\n"))

            # norm = np.ones_like(x) / float(len(x))
            norm = np.ones_like(x)
            binspace = np.arange(_minRuler(x), _maxRuler(x), step)
            n, bins, patches = plt.hist(x, bins=binspace, weights=norm,
                alpha=config.alpha, facecolor=config.color)

            # formatter = FuncFormatter(_to_percent)
            # plt.gca().yaxis.set_major_formatter(formatter)
            plt.grid(True)

            _saveToPdf(config.pdf_prefix+"%s.pdf" % target.split(".")[0])
        except IOError as e:
            print "Warning: I/O error({0}) - {1}".format(e.errno, e.strerror)
            pass
        except:
            print "Unexpected error:", sys.exc_info()[0]
            raise
    print "Graphs are drawned!"

# Input: model object list, Output: human-readable form of model as x.dat
def print_weight_vars(obj_dict, weight_obj_list, fname_list, show_zero=False):
    for elem, fname in zip(weight_obj_list, fname_list):
        weight_arr = obj_dict[elem].eval()
        ndim = weight_arr.size
        flat_weight_space = weight_arr.reshape(ndim)
        with open(fname, "w") as filelog:
            if show_zero == False:
                flat_weight_space = flat_weight_space[flat_weight_space != 0]
            writeLine = csv.writer(filelog, delimiter='\n')
            writeLine.writerow(flat_weight_space)

# Input: synapse, Output: human-readable form of model as x.syn
def print_synapse_nps(syn_arr, fname, show_zero=False):
    ndim = syn_arr.size
    flat_syn_space = syn_arr.reshape(ndim)
    with open(fname, "w") as filelog:
        if show_zero == False:
            flat_syn_space = flat_syn_space[flat_syn_space != 0]
        writeLine = csv.writer(filelog, delimiter='\n')
        writeLine.writerow(flat_syn_space)

# Input: sparse model object list, Output: human-readable form of model as x.dat
def print_sparse_weight_vars(obj_dict, weight_obj_list, fname_list):
    for elem, fname in zip(weight_obj_list, fname_list):
        weight_arr = obj_dict[elem].eval().values
        ndim = weight_arr.size
        flat_weight_space = weight_arr.reshape(ndim)
        with open(fname, "w") as filelog:
            writeLine = csv.writer(filelog, delimiter='\n')
            writeLine.writerow(flat_weight_space)

# Input: n-d dense array, Output: pruned array with threshold
def prune_dense(weight_arr, name="None", thresh=0.005, **kwargs):
    """Apply weight pruning with threshold """
    under_threshold = abs(weight_arr) < thresh
    weight_arr[under_threshold] = 0
    count = np.sum(under_threshold)
    print "Non-zero count (%s): %s" % (name, weight_arr.size - count)
    return weight_arr, -under_threshold, count

# Input: anonymous dimension array and its pruning threshold,
# Output: indices - index list of non-zero elements
#         values  - value list of non-zero elements
#         shape   - original shape of matrix
def prune_tf_sparse(weight_arr, name="None", thresh=0.005):
    assert isinstance(weight_arr, np.ndarray)

    under_threshold = abs(weight_arr) < thresh
    weight_arr[under_threshold] = 0
    values = weight_arr[weight_arr != 0]
    indices = np.transpose(np.nonzero(weight_arr))
    shape = list(weight_arr.shape)

    count = np.sum(under_threshold)
    print "Non-zero count (Sparse %s): %s" % (name, weight_arr.size - count)
    return [indices, values, shape]

# Input: file name and text, Output: log file
def log(fname, log):
   with open(fname, "a") as wobj:
        wobj.write(str(log)+"\n")

# Input: Path to target image, Output: ndarray resized to fixed (28,28)
def imread(path):
    import numpy as np
    import Image
    return np.array(Image.open(path).resize((28,28), resample=2))


================================================
FILE: read_model.py
================================================
#!/usr/bin/python

import sys
sys.dont_write_bytecode = True

import tensorflow as tf
import papl
import argparse

argparser = argparse.ArgumentParser()
argparser.add_argument("-m", "--model", required=True, help="Specify serialized input model")
argparser.add_argument("-r", "--ratio", help="Specify ratio")
args = argparser.parse_args()

def read_model_obj_with_sorted_ratio(fname, ratio):
    saver = tf.train.Saver()
    saver.restore(sess, fname)

    print str(ratio*100)+" %"
    target_obj_list = [weights[elem] for elem in papl.config.target_all_layer]
    for elem in target_obj_list:
        arr = elem.eval()
        arr = list(arr.reshape(arr.size))
        arr.sort(cmp=lambda x,y:cmp(abs(x), abs(y)))

        print "\""+elem.name[:-2]+"\": ", abs(arr[int(len(arr)*ratio)-1]), ","

def print_raw_matrix(fname):
    saver = tf.train.Saver()
    saver.restore(sess, fname)
    import numpy as np
    np.save("w_fc1.raw", weights["w_fc1"].eval())
    np.save("w_fc2.raw", weights["w_fc2"].eval())

def read_model_obj(fname):
    saver = tf.train.Saver()

    import os.path
    try:
        assert os.path.isfile(fname)
        saver.restore(sess, fname)
        switcher = {
            "model_ckpt_dense": papl.config.target_dat,
            "model_ckpt_dense_pruned": papl.config.target_p_dat,
            "model_ckpt_dense_retrained": papl.config.target_tp_dat
        }
        papl.print_weight_vars(weights, papl.config.target_all_layer, switcher.get(args.model))
    except AssertionError:
        print "Warning: No such files or directory\n"
        pass
    except:
        import sys
        print "Unexpected error:", sys.exc_info()[0]

weights = {
    "w_conv1": tf.Variable(tf.truncated_normal([5, 5, 1, 32], stddev=0.1), name="w_conv1"),
    "b_conv1": tf.Variable(tf.constant(0.1, shape=[32]), name="b_conv1"),
    "w_conv2": tf.Variable(tf.truncated_normal([5, 5, 32, 64], stddev=0.1), name="w_conv2"),
    "b_conv2": tf.Variable(tf.constant(0.1, shape=[64]), name="b_conv2"),
    "w_fc1": tf.Variable(tf.truncated_normal([7*7*64, 1024], stddev=0.1), name="w_fc1"),
    "b_fc1": tf.Variable(tf.constant(0.1, shape=[1024]), name="b_fc1"),
    "w_fc2": tf.Variable(tf.truncated_normal([1024, 10], stddev=0.1), name="w_fc2"),
    "b_fc2": tf.Variable(tf.constant(0.1, shape=[10]), name="b_fc2")
}

sess = tf.InteractiveSession()

if __name__ == "__main__":
    if bool(args.ratio) == False:
        read_model_obj(args.model)
    else:
        read_model_obj_with_sorted_ratio(args.model, float(args.ratio))
    # print_raw_matrix(args.model)


================================================
FILE: sparse_model_extreme/model_ckpt_dense_pruned
================================================
[File too large to display: 12.5 MB]

================================================
FILE: sparse_model_extreme/model_ckpt_dense_retrained
================================================
[File too large to display: 12.5 MB]

================================================
FILE: thspace.py
================================================
th10 = {
    "fc1_nnz":  2890138 ,
    "fc2_nnz":  9216 ,
    "w_conv1":  0.0143598 ,
    "b_conv1":  0.0639633 ,
    "w_conv2":  0.0122113 ,
    "b_conv2":  0.0764835 ,
    "w_fc1":  0.0121145 ,
    "b_fc1":  0.0907546 ,
    "w_fc2":  0.0132132 ,
    "b_fc2":  0.0911222
}

th20 = {
    "fc1_nnz":  2569011 ,
    "fc2_nnz":  8192 ,
    "w_conv1":  0.0303244 ,
    "b_conv1":  0.0703646 ,
    "w_conv2":  0.0243528 ,
    "b_conv2":  0.0805169 ,
    "w_fc1":  0.024393 ,
    "b_fc1":  0.0929042 ,
    "w_fc2":  0.0266463 ,
    "b_fc2":  0.0943206
}

th30 = { 
    "fc1_nnz":  2247885 ,
    "fc2_nnz":  7168 ,
    "w_conv1":  0.0473049 ,
    "b_conv1":  0.075084 ,
    "w_conv2":  0.0371282 ,
    "b_conv2":  0.0821582 ,
    "w_fc1":  0.0370279 ,
    "b_fc1":  0.0944582 ,
    "w_fc2":  0.0407012 ,
    "b_fc2":  0.0944
}

th40 = {
    "fc1_nnz":  1926757 ,
    "fc2_nnz":  6143 ,
    "w_conv1":  0.0619981 ,
    "b_conv1":  0.0783646 ,
    "w_conv2":  0.0506691 ,
    "b_conv2":  0.0849416 ,
    "w_fc1":  0.0503098 ,
    "b_fc1":  0.0957049 ,
    "w_fc2":  0.0552152 ,
    "b_fc2":  0.0960752
}

th50 = {
    "fc1_nnz":  1605631 ,
    "fc2_nnz":  5120 ,
    "w_conv1":  0.0762394 ,
    "b_conv1":  0.0791745 ,
    "w_conv2":  0.0650136 ,
    "b_conv2":  0.0858885 ,
    "w_fc1":  0.0645222 ,
    "b_fc1":  0.0967964 ,
    "w_fc2":  0.0705915 ,
    "b_fc2":  0.0978322
}

th60 = {
    "fc1_nnz":  1284506 ,
    "fc2_nnz":  4095 ,
    "w_conv1":  0.0936658 ,
    "b_conv1":  0.0817409 ,
    "w_conv2":  0.0805099 ,
    "b_conv2":  0.0873334 ,
    "w_fc1":  0.0801966 ,
    "b_fc1":  0.0979769 ,
    "w_fc2":  0.0870296 ,
    "b_fc2":  0.0996566
}

th70 = {
    "fc1_nnz":  963379 ,
    "fc2_nnz":  3071 ,
    "w_conv1":  0.110689 ,
    "b_conv1":  0.0830755 ,
    "w_conv2":  0.0988535 ,
    "b_conv2":  0.088152 ,
    "w_fc1":  0.0979934 ,
    "b_fc1":  0.0991785 ,
    "w_fc2":  0.105566 ,
    "b_fc2":  0.100518
}

th80 = {
    "fc1_nnz":  642247 ,
    "fc2_nnz":  2048 ,
    "w_conv1":  0.130691 ,
    "b_conv1":  0.0912763 ,
    "w_conv2":  0.120732 ,
    "b_conv2":  0.0898623 ,
    "w_fc1":  0.119522 ,
    "b_fc1":  0.100626 ,
    "w_fc2":  0.126458 ,
    "b_fc2":  0.112008
}

th90 = {
    "fc1_nnz":  320939 ,
    "fc2_nnz":  1014 ,
    "w_conv1":  0.162963 ,
    "b_conv1":  0.0956728 ,
    "w_conv2":  0.150202 ,
    "b_conv2":  0.0928398 ,
    "w_fc1":  0.148615 ,
    "b_fc1":  0.102556 ,
    "w_fc2":  0.15566 ,
    "b_fc2":  0.112008
}

th95 = {
    "fc1_nnz":  160566 ,
    "fc2_nnz":  513 ,
    "w_fc1":  0.169592 ,
    "b_fc1":  0.103936 ,
    "w_fc2":  0.17703 ,
    "b_fc2":  0.112008
}

th99 = {
    "fc1_nnz":  32111 ,
    "fc2_nnz":  103 ,
    "w_fc1":  0.1975 ,
    "b_fc1":  0.10679 ,
    "w_fc2":  0.21004 ,
    "b_fc2":  0.112008
}


================================================
FILE: train.py
================================================
#!/usr/bin/python

from __future__ import absolute_import
from __future__ import division
from __future__ import print_function

import sys
sys.dont_write_bytecode = True

import tensorflow as tf
import numpy as np
import argparse
import papl

import scipy.sparse as sp

argparser = argparse.ArgumentParser()
argparser.add_argument("-1", "--first_round", action="store_true",
    help="Run 1st-round: train with 20000 iterations")
argparser.add_argument("-2", "--second_round", action="store_true",
    help="Run 2nd-round: apply pruning and its additional training")
argparser.add_argument("-3", "--third_round", action="store_true",
    help="Run 3rd-round: transform model to a sparse format and save it")
argparser.add_argument("-m", "--checkpoint", default="./model_ckpt_dense",
    help="Target checkpoint model file for 2nd and 3rd round")
args = argparser.parse_args()

from tensorflow.examples.tutorials.mnist import input_data
mnist = input_data.read_data_sets('/tmp/data/', one_hot=True)
if (args.first_round or args.second_round or args.third_round) == False:
    argparser.print_help()
    sys.exit()

sess = tf.InteractiveSession()

def apply_prune(weights):
    total_fc_byte = 0
    total_fc_csr_byte = 0
    total_nnz_elem = 0
    total_origin_elem = 0

    dict_nzidx = {}

    for target in papl.config.target_layer:
        wl = "w_" + target
        print(wl + " threshold:\t" + str(papl.config.th[wl]))

        # Get target layer's weights
        weight_obj = weights[wl]
        weight_arr = weight_obj.eval()

        # Apply pruning
        weight_arr, w_nzidx, w_nnz = papl.prune_dense(weight_arr, name=wl,
                                            thresh=papl.config.th[wl])

        # Store pruned weights as tensorflow objects
        dict_nzidx[wl] = w_nzidx
        sess.run(weight_obj.assign(weight_arr))

    return dict_nzidx

def apply_prune_on_grads(grads_and_vars, dict_nzidx):
    # Mask gradients with pruned elements
    for key, nzidx in dict_nzidx.items():
        count = 0
        for grad, var in grads_and_vars:
            if var.name == key+":0":
                nzidx_obj = tf.cast(tf.constant(nzidx), tf.float32)
                grads_and_vars[count] = (tf.mul(nzidx_obj, grad), var)
            count += 1
    return grads_and_vars

def gen_sparse_dict(dense_w):
    sparse_w = dense_w
    for target in papl.config.target_all_layer:
        target_arr = np.transpose(dense_w[target].eval())
        sparse_arr = papl.prune_tf_sparse(target_arr, name=target)
        sparse_w[target+"_idx"]=tf.Variable(tf.constant(sparse_arr[0],dtype=tf.int32),
            name=target+"_idx")
        sparse_w[target]=tf.Variable(tf.constant(sparse_arr[1],dtype=tf.float32),
            name=target)
        sparse_w[target+"_shape"]=tf.Variable(tf.constant(sparse_arr[2],dtype=tf.int32),
            name=target+"_shape")
    return sparse_w

dense_w={
    "w_conv1": tf.Variable(tf.truncated_normal([5,5,1,32],stddev=0.1), name="w_conv1"),
    "b_conv1": tf.Variable(tf.constant(0.1,shape=[32]), name="b_conv1"),
    "w_conv2": tf.Variable(tf.truncated_normal([5,5,32,64],stddev=0.1), name="w_conv2"),
    "b_conv2": tf.Variable(tf.constant(0.1,shape=[64]), name="b_conv2"),
    "w_fc1": tf.Variable(tf.truncated_normal([7*7*64,1024],stddev=0.1), name="w_fc1"),
    "b_fc1": tf.Variable(tf.constant(0.1,shape=[1024]), name="b_fc1"),
    "w_fc2": tf.Variable(tf.truncated_normal([1024,10],stddev=0.1), name="w_fc2"),
    "b_fc2": tf.Variable(tf.constant(0.1,shape=[10]), name="b_fc2")
}

def dense_cnn_model(weights):
    def conv2d(x, W):
        return tf.nn.conv2d(x, W, strides=[1, 1, 1, 1], padding='SAME')

    def max_pool_2x2(x):
        return tf.nn.max_pool(x, ksize=[1, 2, 2, 1],
                              strides=[1, 2, 2, 1], padding='SAME')

    x_image = tf.reshape(x, [-1,28,28,1])
    h_conv1 = tf.nn.relu(conv2d(x_image, weights["w_conv1"]) + weights["b_conv1"])
    tf.add_to_collection("in_conv1", x_image)
    h_pool1 = max_pool_2x2(h_conv1)
    tf.add_to_collection("in_conv2", h_pool1)
    h_conv2 = tf.nn.relu(conv2d(h_pool1, weights["w_conv2"]) + weights["b_conv2"])
    h_pool2 = max_pool_2x2(h_conv2)
    h_pool2_flat = tf.reshape(h_pool2, [-1, 7*7*64])
    tf.add_to_collection("in_fc1", h_pool2_flat)
    h_fc1 = tf.nn.relu(tf.matmul(h_pool2_flat, weights["w_fc1"]) + weights["b_fc1"])
    h_fc1_drop = tf.nn.dropout(h_fc1, keep_prob)
    tf.add_to_collection("in_fc2", h_fc1_drop)
    y_conv=tf.nn.softmax(tf.matmul(h_fc1_drop, weights["w_fc2"]) + weights["b_fc2"])
    return y_conv

def test(y_infer, message="None."):
    correct_prediction = tf.equal(tf.argmax(y_infer,1), tf.argmax(y_,1))
    accuracy = tf.reduce_mean(tf.cast(correct_prediction, "float"))

    # To avoid OOM, run validation with 500/10000 test dataset
    result = 0
    for i in range(20):
        batch = mnist.test.next_batch(500)
        result += accuracy.eval(feed_dict={x: batch[0],
                                          y_: batch[1],
                                          keep_prob: 1.0})
    result /= 20

    print(message+" %g\n" % result)
    return result

def check_file_exists(key):
    import os
    fileList = os.listdir(".")
    count = 0
    for elem in fileList:
        if elem.find(key) >= 0:
            count += 1
    return key + ("-"+str(count) if count>0 else "")

# Construct a dense model
x = tf.placeholder("float", shape=[None, 784], name="x")
y_ = tf.placeholder("float", shape=[None, 10], name="y_")
keep_prob = tf.placeholder("float", name="keep_prob")

y_conv = dense_cnn_model(dense_w)
tf.add_to_collection("y_conv", y_conv)

saver = tf.train.Saver()

if args.first_round == True:
    # First round: Train baseline dense model
    cross_entropy = -tf.reduce_sum(y_*tf.log(tf.clip_by_value(y_conv,1e-10,1.0)))
    train_step = tf.train.AdamOptimizer(1e-4).minimize(cross_entropy)
    correct_prediction = tf.equal(tf.argmax(y_conv,1), tf.argmax(y_,1))
    accuracy = tf.reduce_mean(tf.cast(correct_prediction, "float"))
    tf.add_to_collection("accuracy", accuracy)

    sess.run(tf.initialize_all_variables())

    for i in range(20000):
        batch = mnist.train.next_batch(50)
        if i%100 == 0:
            train_accuracy = accuracy.eval(feed_dict={
                x:batch[0], y_: batch[1], keep_prob: 1.0})
            print("step %d, training accuracy %g"%(i, train_accuracy))
        train_step.run(feed_dict={x: batch[0], y_: batch[1], keep_prob: 0.5})

    # Test
    score = test(y_conv, message="First-round prune-only test accuracy")
    papl.log("baseline_accuracy.log", score)
    
    # Save model objects to readable format
    papl.print_weight_vars(dense_w, papl.config.target_all_layer,
                           papl.config.target_dat, show_zero=papl.config.show_zero)
    # Save model objects to serialized format
    saver.save(sess, "./model_ckpt_dense")

if args.second_round == True:
    # Second round: Retrain pruned model, start with default model: model_ckpt_dense
    saver.restore(sess, args.checkpoint)

    # Apply pruning on this context
    dict_nzidx = apply_prune(dense_w)

    # save model objects to readable format
    papl.print_weight_vars(dense_w, papl.config.target_all_layer,
                           papl.config.target_p_dat, show_zero=papl.config.show_zero)

    # Test prune-only networks
    score = test(y_conv, message="Second-round prune-only test accuracy")
    papl.log("prune_accuracy.log", score)

    # save model objects to serialized format
    saver.save(sess, "./model_ckpt_dense_pruned")

    # Retrain networks
    cross_entropy = -tf.reduce_sum(y_*tf.log(tf.clip_by_value(y_conv,1e-10,1.0)))
    trainer = tf.train.AdamOptimizer(1e-4)
    grads_and_vars = trainer.compute_gradients(cross_entropy)
    grads_and_vars = apply_prune_on_grads(grads_and_vars, dict_nzidx)
    train_step = trainer.apply_gradients(grads_and_vars)

    correct_prediction = tf.equal(tf.argmax(y_conv,1), tf.argmax(y_,1))
    accuracy = tf.reduce_mean(tf.cast(correct_prediction, "float"))

    # Initialize firstly touched variables (mostly from accuracy calc.)
    for var in tf.all_variables():
        if tf.is_variable_initialized(var).eval() == False:
            sess.run(tf.initialize_variables([var]))

    # Train x epochs additionally
    for i in range(papl.config.retrain_iterations):
        batch = mnist.train.next_batch(50)
        if i%100 == 0:
            train_accuracy = accuracy.eval(feed_dict={
                x:batch[0], y_: batch[1], keep_prob: 1.0})
            print("step %d, training accuracy %g"%(i, train_accuracy))
        train_step.run(feed_dict={x: batch[0], y_: batch[1], keep_prob: 0.5})

    # Save retrained variables to a desne form
    # key = check_file_exists("model_ckpt_dense_retrained")
    # saver.save(sess, key)
    saver.save(sess, "model_ckpt_dense_retrained")

    # Test the retrained model
    score = test(y_conv, message="Second-round final test accuracy")
    papl.log("final_accuracy.log", score)

if args.third_round == True:
    # Third round: Transform iteratively pruned model to a sparse format and save it
    if args.second_round == False:
        saver.restore(sess, "./model_ckpt_dense_pruned")

    # Transform final weights to a sparse form
    sparse_w = gen_sparse_dict(dense_w)

    # Initialize new variables in a sparse form
    for var in tf.all_variables():
        if tf.is_variable_initialized(var).eval() == False:
            sess.run(tf.initialize_variables([var]))

    # Save model objects to readable format
    papl.print_weight_vars(dense_w, papl.config.target_all_layer,
                           papl.config.target_tp_dat, show_zero=papl.config.show_zero)
    # Save model objects to serialized format
    final_saver = tf.train.Saver(sparse_w)
    final_saver.save(sess, "./model_ckpt_sparse_retrained")