-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathcompression.py
196 lines (163 loc) · 10.5 KB
/
compression.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
import os
import pickle
import numpy as np
os.environ['TF_CPP_MIN_LOG_LEVEL'] = '3'
import tensorflow as tf
import click
from datasets import SHIDC_B_Ki67
from sHAM import pruning, uCWS, uPWS
from sHAM import uUQ, uECSQ, pruning_uCWS, pruning_uPWS
from sHAM import pruning_uUQ, pruning_uECSQ
exec(open("../GPU.py").read())
import sys
sys.path.append('./PathoNet/')
import utils
# Set seed to make results reproducible
# SEED = 1
# from numpy.random import seed
# seed(SEED)
# tf.random.set_seed(SEED)
from tensorflow.python import _pywrap_util_port
print("MKL enabled:", _pywrap_util_port.IsMklEnabled())
# This script does not excercise old non-unified methods. Check https://github.com/giosumarin/ICPR2020_sHAM for those
def run(compression, net, data_path, output_path, learning_rate, epochs, lr_cumulative, minibatch, prfc, prcnn, clusterfc, clustercnn, tr, lambd, logger, ptnc, internal_ho):
# Load model
model = tf.keras.models.load_model(net)
# Compression definition
if compression == 'pr':
compression_model = pruning.pruning(model=model, perc_prun_for_dense=prfc, perc_prun_for_cnn=prcnn)
elif compression == 'uCWS':
compression_model = uCWS.uCWS(model=model, clusters_for_conv_layers=clustercnn, clusters_for_dense_layers=clusterfc)
elif compression == 'uPWS':
compression_model = uPWS.uPWS(model=model, clusters_for_conv_layers=clustercnn, clusters_for_dense_layers=clusterfc)
elif compression == 'uUQ':
compression_model = uUQ.uUQ(model=model, clusters_for_conv_layers=clustercnn, clusters_for_dense_layers=clusterfc)
elif compression == 'uECSQ':
compression_model = uECSQ.uECSQ(model=model, clusters_for_conv_layers=clustercnn, clusters_for_dense_layers=clusterfc, wanted_clusters_cnn=clustercnn, wanted_clusters_fc=clusterfc, tr=tr, lamb=lambd)
elif compression == 'pruCWS':
compression_model = pruning_uCWS.pruning_uCWS(model=model, perc_prun_for_dense=prfc, perc_prun_for_cnn=prcnn, clusters_for_conv_layers=clustercnn, clusters_for_dense_layers=clusterfc)
elif compression == 'pruPWS':
compression_model = pruning_uPWS.pruning_uPWS(model=model, perc_prun_for_dense=prfc, perc_prun_for_cnn=prcnn, clusters_for_conv_layers=clustercnn, clusters_for_dense_layers=clusterfc)
elif compression == 'pruUQ':
compression_model = pruning_uUQ.pruning_uUQ(model=model, perc_prun_for_dense=prfc, perc_prun_for_cnn=prcnn, clusters_for_conv_layers=clustercnn, clusters_for_dense_layers=clusterfc)
elif compression == 'pruECSQ':
compression_model = pruning_uECSQ.pruning_uECSQ(model=model, perc_prun_for_dense=prfc, perc_prun_for_cnn=prcnn, clusters_for_conv_layers=clustercnn, clusters_for_dense_layers=clusterfc, wanted_clusters_cnn=clustercnn, wanted_clusters_fc=clusterfc, tr=tr, lamb=lambd)
# Load dataset
print("\nLoading dataset (it takes several minutes)... \n")
if internal_ho == True:
# Evaluation on the validation set
dataset, x_train, y_train, x_val, y_val, _, _ = SHIDC_B_Ki67(minibatch, data_path)
x_train = tf.convert_to_tensor(x_train)
y_train = tf.convert_to_tensor(y_train)
x_test = tf.convert_to_tensor(x_val)
y_test = tf.convert_to_tensor(y_val)
else:
# Training on train+validation set and evaluation on test set
_, x_train, y_train, x_val, y_val, x_test, y_test = SHIDC_B_Ki67(minibatch, data_path) #34.3 GB
x_test = tf.convert_to_tensor(x_test)
y_test = tf.convert_to_tensor(y_test)
x_val_aug, y_val_aug = utils.dataAugmentation(x_val, y_val) #data augmentation validation set
x_train = tf.convert_to_tensor(np.concatenate((x_train, x_val_aug)))
y_train = tf.convert_to_tensor(np.concatenate((y_train, y_val_aug)))
dataset = tf.data.Dataset.from_tensor_slices((x_train, y_train))
dataset = dataset.shuffle(1000).batch(minibatch)
print("\nTraining set has shape: ", x_train.shape, "; with dtype: ", x_train.dtype, "\n")
print("\nTest set has shape: ", x_test.shape, "; with dtype: ", x_test.dtype, "\n")
# Compile the model
model.compile(optimizer = tf.keras.optimizers.Adam(lr=learning_rate), loss = 'mse', metrics=['MeanSquaredError']) # taken from PathoNet (except metric)
# Pre-compression prediction assessment
pre_compr_train = compression_model.model.evaluate(x_train, y_train, verbose = 0)[1] #batch size default 32
pre_compr_test = compression_model.model.evaluate(x_test, y_test, verbose = 0)[1] #batch size default 32
print("before compression, performance on train -->", pre_compr_train)
print("before compression, performance on test -->", pre_compr_test)
# Model compression
if compression == 'pr':
compression_model.apply_pruning()
elif compression == 'uCWS':
compression_model.apply_uCWS()
elif compression == 'uPWS':
compression_model.apply_uPWS()
elif compression == 'uUQ':
compression_model.apply_uUQ()
elif compression == 'uECSQ':
# lambdas = [1e-10, 5e-9, 1e-9, 1e-8, 1e-7, 5e-7, 1e-6, 5e-6, 7.5e-6, 1e-5, 2e-5][::-1]
# lambdas = [1e-12, 2.5e-12, 5e-12, 7.5e-12, 1e-11, 2.5e-11, 5e-11, 7.5e-11, 1e-10][::-1]
# compression_model.tune_lambda(lambdas)
compression_model.apply_uECSQ()
print("Lambda value is set to:", getattr(compression_model, 'lamb_cnn'))
elif compression == 'pruCWS':
compression_model.apply_pr_uCWS()
elif compression == 'pruPWS':
compression_model.apply_pr_uPWS()
elif compression == 'pruUQ':
compression_model.apply_pr_uUQ()
elif compression == 'pruECSQ':
# lambdas = [1e-10, 5e-9, 1e-9, 1e-8, 1e-7, 5e-7, 1e-6, 5e-6, 7.5e-6, 1e-5, 2e-5][::-1]
# lambdas = [1e-12, 2.5e-12, 5e-12, 7.5e-12, 1e-11, 2.5e-11, 5e-11, 7.5e-11, 1e-10][::-1]
# compression_model.tune_lambda(lambdas)
compression_model.apply_pr_uECSQ()
print("Lambda value is set to:", getattr(compression_model, 'lamb_cnn'))
# Post-compression prediction assessment
compression_model.set_loss(tf.keras.losses.MeanSquaredError())
compression_model.set_optimizer(tf.keras.optimizers.Adam(lr=learning_rate))
post_compr_train = compression_model.model.evaluate(x_train, y_train, verbose = 0)[1]
post_compr_test = compression_model.model.evaluate(x_test, y_test, verbose = 0)[1]
print("Applying initial compression setting before retrain, performance on train -->" , post_compr_train)
print("Applying initial compression setting before retrain, performance on test -->" , post_compr_test)
if compression == "pr":
compression_model.train_pr(epochs=epochs, dataset=dataset, X_train=x_train, y_train=y_train, X_test=x_test, y_test=y_test, step_per_epoch = 10000000, patience=0)
else:
compression_model.train_ws(epochs=epochs, lr=lr_cumulative, dataset=dataset, X_train=x_train, y_train=y_train, X_test=x_test, y_test=y_test, patience=ptnc, min_is_better=True, threshold=0.001, best_model=True)
# Model save
name_net = (net.split("/")[-1])[:-5] #Adapted to format ".hdf5"
TRAIN_RES = ([pre_compr_train] + [post_compr_train] + compression_model.acc_train)
TEST_RES = ([pre_compr_test] + [post_compr_test] + compression_model.acc_test)
if compression in ['uUQ', 'pruUQ', 'uECSQ', 'pruECSQ']:
temp_centers_fc = len(compression_model.centers_fc) if hasattr(compression_model, 'centers_fc') else 0
temp_centers_cnn = len(compression_model.centers_cnn) if hasattr(compression_model, 'centers_cnn') else 0
compression_param = "-".join([str(x) for x in [prfc, prcnn, temp_centers_fc, temp_centers_cnn, learning_rate, lr_cumulative, ptnc, epochs]]) + '-'
else:
compression_param = "-".join([str(x) for x in [prfc, prcnn, clusterfc, clustercnn, learning_rate, lr_cumulative, ptnc, epochs]]) + '-'
if logger:
with open("{}_{}.txt".format(name_net, compression), "a+") as tex:
tex.write("lr {} {} -->\n {}\n , {}\n\n".format(learning_rate, compression_param, TRAIN_RES, TEST_RES))
if internal_ho == False:
DIR="{}/{}".format(output_path, compression)
TO_SAVE = "{}{}".format(compression_param, round(TEST_RES[-1],5))
if not os.path.isdir(output_path):
os.mkdir(output_path)
if not os.path.isdir(DIR):
os.mkdir(DIR)
end_weights = compression_model.model.get_weights()
with open(DIR+"/"+TO_SAVE+".h5", "wb") as file:
pickle.dump(end_weights, file)
# Save model weights using keras
compression_model.model.save_weights(DIR+"/"+TO_SAVE+"_save_weights.h5")
compr_model_path = DIR+"/"+TO_SAVE+"_save_weights.h5"
# Save model weights+structure
compression_model.model.save(DIR+"/"+TO_SAVE+"_save_weights-struct.h5")
else:
compr_model_path = None
return TRAIN_RES, TEST_RES, compression_param, compr_model_path
@click.command()
@click.option('--compression', help='Type of compression')
@click.option('--net', help='original network datapath (path to .hdf5 file)')
@click.option('--data_path', help='path to preprocessed data (obtained using script preprocessing.py)')
@click.option('--output_path', help='path to folder to save reults')
@click.option('--learning_rate', default=0.001, help='learning rate')
@click.option('--epochs', default=30, help='epochs')
@click.option('--lr_cumulative', default=0.001, help='learning rate for cumulative gradient descent')
@click.option('--minibatch', default=8, help='size of minibatch')
@click.option('--prfc', default=0, help='percentage of pruned connection (dense layers)')
@click.option('--prcnn', default=0, help='percentage of pruned connection (convolutional layers)')
@click.option('--clusterfc', default=0, help='different values for all dense layers')
@click.option('--clustercnn', default=0, help='different values for all convolutional layers')
@click.option('--tr', default=0.001, help='treshold for ECSQ')
@click.option('--lambd', default=0., help='coefficient for entropy with ECSQ')
@click.option('--logger', default=False, help='set True for logging train into txt')
@click.option('--ptnc', default=0, help='patience (default 0)')
@click.option('--internal_ho', default=False, help='Test on validation set not on test set')
def main(compression, net, data_path, output_path, learning_rate, epochs, lr_cumulative, minibatch, prfc, prcnn, clusterfc, clustercnn, tr, lambd, logger, ptnc, internal_ho):
run(compression, net, data_path, output_path, learning_rate, epochs, lr_cumulative, minibatch, prfc, prcnn, clusterfc, clustercnn, tr, lambd, logger, ptnc, internal_ho)
if __name__ == '__main__':
main()