/
utils.py
321 lines (296 loc) · 11.2 KB
/
utils.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
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
"""
Disclaimer:
This code is based on code by Peter Ruch.
See his prunhild repository: https://github.com/gfrogat/prunhild
Snippets of code also borrowed from Arun Joseph pruning code.
https://github.com/00arun00/Pruning-Pytorch/blob/master/prune.py
"""
from __future__ import print_function
import torch
import torch.nn as nn
import torch.nn.functional as F
import torch.optim as optim
import numpy as np
from torchvision import datasets, transforms
from mtcnn_pytorch_master.src import detect_faces, show_bboxes
from PIL import Image
from LightCNN_master.load_imglist import ImageList
def load_training_set(batch_size = 20):
root_path = '/home/ubuntu/Project_Insight/Data/GOT/'
train_list = root_path+'train.txt'
train_loader = torch.utils.data.DataLoader(
ImageList(root=root_path, fileList=train_list,
transform=transforms.Compose([
transforms.CenterCrop(128),
transforms.ToTensor(),
])),
batch_size=batch_size, shuffle=False,
pin_memory=True)
return train_loader
def load_validation_set(batch_size = 20):
root_path = '/home/ubuntu/Project_Insight/Data/GOT/'
val_list = root_path+'val.txt'
val_loader = torch.utils.data.DataLoader(
ImageList(root=root_path, fileList=val_list,
transform=transforms.Compose([
transforms.CenterCrop(128),
transforms.ToTensor(),
])),
batch_size=batch_size, shuffle=False,
pin_memory=True)
return val_loader
def load_test_set(batch_size = 20):
root_path = '/home/ubuntu/Project_Insight/Data/GOT/'
test_list = root_path+'test.txt'
test_loader = torch.utils.data.DataLoader(
ImageList(root=root_path, fileList=test_list,
transform=transforms.Compose([
transforms.CenterCrop(128),
transforms.ToTensor(),
])),
batch_size=batch_size, shuffle=False,
#num_workers=args.workers,
pin_memory=True)
return test_loader
def setup_dataloaders(kwargs, dataset_to_use):
mnist_transform = transforms.Compose(
[transforms.ToTensor(), transforms.Normalize((0.1307,), (0.3081,))]
)
if dataset_to_use == 'GOT':
train_loader = load_training_set()
train_loader_eval = load_validation_set()
test_loader_eval = load_test_set()
else:
datafolder = "~/data/torch"
dataset = (datasets.FashionMNIST)#if dataset_to_use == "FashionMNIST" else datasets.MNIST)
train_loader = torch.utils.data.DataLoader(
dataset(datafolder, train=True, download=True, transform=mnist_transform),
batch_size=args.batch_size,
shuffle=True,
**kwargs
)
train_loader_eval = torch.utils.data.DataLoader(
dataset(datafolder, train=True, download=True, transform=mnist_transform),
batch_size=args.batch_size_eval,
shuffle=True,
**kwargs
)
test_loader_eval = torch.utils.data.DataLoader(
dataset(datafolder, train=False, transform=mnist_transform),
batch_size=args.batch_size_eval,
shuffle=True,
**kwargs
)
return train_loader, train_loader_eval, test_loader_eval
def train(
model,
device,
data_loaders,
optimizer,
epoch,
use_mask = False,
prune=False,
cut_ratio = .2
):
prune_interval = 400
print_interval = 200
eval_interval = 200
train_loader, train_loader_eval, test_loader_eval = data_loaders
logs = []
model.train()
for batch_idx, (data, target) in enumerate(train_loader):
data, target = data.to(device), target.to(device)
optimizer.zero_grad()
output = model(data)
if sum(sum(torch.isnan(output))).cpu().detach().numpy() > 0:
print(batch_idx)
loss = F.cross_entropy(output, target)
if (loss.cpu().detach().numpy()) > 1000: # For debugging purposes
print(loss)
print('=============')
break
loss.backward()
if use_mask or prune:
# Make sure the pruned parameters aren't updated
for index,params in enumerate(model.parameters()):
mask = model.prune_mask[index]
params.grad.data[mask]=0
optimizer.step()
# Pruning
if prune:
if (batch_idx % prune_interval == 0) and (batch_idx > 0):
print('-')
print(batch_idx)
model.prune_model(cut_ratio)
if batch_idx % print_interval == 0:
print(
"[Train] Epoch: {} [{}/{} ({:.0f}%)]\tLoss: {:.6f} Frac Zeros: {:.2f}".format(
epoch,
batch_idx * len(data),
len(train_loader.dataset),
100.0 * batch_idx / len(train_loader),
loss.item(),
model.get_prune_frac()
)
)
if batch_idx % eval_interval == 0:
ratio_zero = model.get_prune_frac()
acc_train = evaluate(model, device, train_loader_eval, training = True)
acc_test = evaluate(model, device, test_loader_eval)
logs.append((epoch, batch_idx, ratio_zero, acc_train, acc_test))
return logs
def evaluate(model, device, data_loader, training=False):
fold = "Train" if training else "Test"
model.eval()
test_loss = 0
correct = 0
with torch.no_grad():
for data, target in data_loader:
data, target = data.to(device), target.to(device)
output = model(data)
test_loss += F.cross_entropy(
output, target, reduction="sum"
).item() # sum up batch loss
pred = output.argmax(
dim=1, keepdim=True
) # get the index of the max log-probability
correct += pred.eq(target.view_as(pred)).sum().item()
test_loss /= len(data_loader.dataset)
accuracy = correct / len(data_loader.dataset)
print(
"[Eval] {} set: Average loss: {:.4f}, Accuracy: {}/{} ({:.2f})".format(
fold, test_loss, correct, len(data_loader.dataset), accuracy
)
)
return accuracy
def get_seed(i,j):
seeds = [[10,22000,13,154,65],[832,120,1294,138,4567], [43, 616, 94, 1732, 7253]]
return int(seeds[i][j])
def set_seed(seed):
torch.manual_seed(seed)
def prune_and_train(prune_frac = .9, seed_model = 1, max_epoch = 200, dir_out = '', save_all = False, cmdline_args=None):
if prune_frac >=1. :
print('prune_frac must be in [0,1[.')
return 0
model_to_use = 'LCNN'
batch_size_eval= 512
learning_rate= 5e-3
momentum= 0.8
cutoff_ratio= 0.15
cut_ratio = .2
if model_to_use == 'LCNN':
prune_interval = 10
prune_train = False
prune_epoch = True
else:
prune_interval = 0
prune_train = True
prune_epoch = False
use_cuda = True
device = torch.device("cuda" if use_cuda else "cpu")
kwargs = {"num_workers": 1, "pin_memory": True} if use_cuda else {}
data_loaders = setup_dataloaders(kwargs,dataset_to_use = "GOT")
model = TestNN().to(device) #LotteryLeNet().to(device)
optimizer = optim.SGD(
model.parameters(), lr=learning_rate, momentum=momentum
)
logs_prune = []
print("Pruning Start")
set_seed(seed_model)
model.apply(init_weights)
epoch = 1
while model.get_prune_frac() < prune_frac:
if prune_epoch:
if (epoch % prune_interval == 0):
print('-epoch-')
print(epoch)
model.prune_model(cut_ratio)
if save_all:
prune_percent = int(np.round(model.get_prune_frac()*100))
savefile_model = dir_out+'model'+model_to_use+'_prune_'+str(prune_percent)+'seed_'+str(seed_model)+'.pt'
torch.save(model.state_dict(), savefile_model)
savefile_mask = dir_out+'model'+model_to_use+'_prune_'+str(prune_percent)+'seed_'+str(seed_model)+'_mask.sav'
torch.save(model.prune_mask, savefile_mask)
logs_prune += train(
model, device, data_loaders, optimizer, epoch, prune=prune_train, use_mask = True, cut_ratio = cut_ratio
)
epoch += 1
if epoch >= max_epoch:
break
# Train the model after final pruning
for i in range(2*prune_interval):
logs_prune += train(
model, device, data_loaders, optimizer, epoch, prune=prune_train, use_mask = True, cut_ratio = cut_ratio
)
epoch += 1
print("\n\n\n")
prune_percent = int(np.round(model.get_prune_frac()*100))
savefile_model = dir_out+'model'+model_to_use+'_prune_'+str(prune_percent)+'seed_'+str(seed_model)+'.pt'
torch.save(model.state_dict(), savefile_model)
savefile_logs = dir_out+'model'+model_to_use+'_prune_'+str(prune_percent)+'seed_'+str(seed_model)+'_logs.sav'
savefile_mask = dir_out+'model'+model_to_use+'_prune_'+str(prune_percent)+'seed_'+str(seed_model)+'_mask.sav'
torch.save(logs_prune, savefile_logs)
torch.save(model.prune_mask, savefile_mask)
return logs_prune
def retrain(epochs=10, prune_percent = 0, seed_model=1, seed_baseline =0 ,cmdline_args=None):
dir_sav = 'saves/'
model_to_use = 'LCNN'
batch_size_eval= 512
learning_rate= 1e-2
momentum= 0.8
cutoff_ratio= 0.15
cut_ratio = .2
use_cuda = torch.cuda.is_available()
device = torch.device("cuda" if use_cuda else "cpu")
kwargs = {"num_workers": 1, "pin_memory": True} if use_cuda else {}
data_loaders = setup_dataloaders(kwargs,dataset_to_use = "GOT")
if model_to_use == 'LCNN':
prune_interval = 10
prune_train = False
prune_epoch = True
else:
prune_interval = 0
prune_train = True
prune_epoch = False
###########
# Load model
savefile_model = dir_sav+'model'+model_to_use+'_prune_'+str(prune_percent)+'seed_'+str(seed_model)+'.pt'
model = TestNN().to(device)
model.load_state_dict(torch.load(savefile_model))
savefile_mask = dir_sav+'model'+model_to_use+'_prune_'+str(prune_percent)+'seed_'+str(seed_model)+'_mask.sav'
model.prune_mask = torch.load(savefile_mask)
n_trials =2
nvar_log = 5
logs_retrain = torch.empty(n_trials, epochs, nvar_log, dtype=torch.double)
for i_test in range(n_trials):
logs_i = []
if model_to_use == 'LCNN':
model_retrain = LightCNN().to(device)
else:
model_retrain = PruneNN().to(device)
if i_test <1:
model_retrain.reinit(seed_model, seed_model, n_layer = i_test)
else:
set_seed(seed_baseline)
model_retrain.apply(init_weights)
optimizer_retrain = optim.SGD(
model_retrain.parameters(), lr=learning_rate, momentum=momentum
)
print("Retraining Start -- "+"str(i_test)")
# Restore mask
model_retrain.prune_mask = model.prune_mask
for epoch in range(1, epochs + 1):
logs_i += train(
model_retrain,
device,
data_loaders,
optimizer_retrain,
epoch,
use_mask = True,
prune=False
)
print("\n\n\n")
logs_retrain[i_test,:,:] = torch.DoubleTensor(logs_i)
savefile_retrain = dir_sav+'model'+model_to_use+'_prune_'+str(prune_percent)+'seed_'+str(seed_model)+'_retrain.sav'
torch.save(logs_retrain, savefile_retrain)
return logs_retrain