-
Notifications
You must be signed in to change notification settings - Fork 0
/
train.py
341 lines (311 loc) · 12.8 KB
/
train.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
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
import pickle
import argparse
import os
import copy
import time
from datetime import timedelta
from pathlib import Path
import torch
import torch.utils
from torch import nn,optim
import numpy as np
from dataset.load_dataset import LoadDatasetFromFolder,CreateTrainValDatasets
from models.resnet50 import resnet50
import sys
def load_init_weights(file_name=None,file_type="pth",model=None,optimizer=None,ignore_weights=['fc.bias','fc.weight'],resume=False,test=False):
"""
Used to load a weights into a model
Attributes
-----------
file_name:
The name of the weight file
file_type:
The type of the weight file
model:
The model to which the weights will be loaded
optimizer:
The optimizer to which weights will be loaded
ignore_weights:
The weights that won't be loaded from the weight file
resume:
Flag used to resume training
test:
Used during test time model init
"""
if model is None:
raise ValueError("Model cannot be empty")
assert os.path.exists(file_name), 'file: {} not found.'.format(file_name)
if file_type == 'pth':
saved_model = torch.load(file_name)
if resume:
model.load_state_dict(saved_model['weights'])
if optimizer is not None:
optimizer.load_state_dict(saved_model['optimizer_state_dict'])
return model
weights = saved_model
elif file_type == 'pkl':
with open(file_name, 'rb') as f:
weights = pickle.load(f, encoding='latin1')
else:
raise ValueError('weight file type must be pkl or pth, given',file_type)
ignore = ignore_weights
if test:
ignore = []
own_state = model.state_dict()
parameters = model.named_parameters()
copied_params = []
for name, param in weights.items():
if (name in own_state) and (name not in ignore):
#print(name)
try:
if torch.is_tensor(param):
own_state[name].copy_(param)
else:
own_state[name].copy_(torch.from_numpy(param))
copied_params.append(name)
except Exception as e:
#print(name)
print(e)
raise RuntimeError('While copying the parameter named {}, whose dimensions in the model are {} and whose dimensions in the checkpoint are {}.'.format(name, own_state[name].size(), param.shape))
for param in parameters:
if param[0] in copied_params:
param[1].requires_grad = False
return model
def create_batched_loader(dataset=None,batch_size=16,shuffle=True):
"""
Create a dataloader
Attributes
----------
dataset:
The dataset for which the dataloader must be created
batch_size:
The batch size of the dataloader
shuffle:
Whether to shuffle the dataset
"""
if dataset is None:
raise ValueError("Dataset cannot be empty")
loader = torch.utils.data.DataLoader(dataset,batch_size=batch_size,shuffle=shuffle)
return loader
def val_evaluation(dataloader, model,loss_fn, device='cpu',test=False):
"""
Perform evaluation on the validation dataset
Attributes
----------
dataloader:
The validation dataloader
model:
The model to run eval on
device:
The device to use for inference
"""
model.eval()
total_loss = 0.0
total,correct = 0,0
for i,data in enumerate(dataloader):
inputs, labels = data
inputs, labels = inputs.to(device),labels.to(device)
labels = labels.long()
#Perfoming eval
outputs = model(inputs)
batch_loss = loss_fn(outputs,labels)
#Adding loss
total_loss += batch_loss.cpu().detach().float()
#Calculating correct predictions for accuracy
_, preds = torch.max(outputs.data, 1)
total += labels.size(0)
#print(preds,labels)
correct += (preds==labels).sum().item()
acc = 100 * correct / total
typ = 'Val'
if test:
typ = 'Test'
#Write the information about current iteration
sys.stdout.write("\r%s: Iteration %i/%i | Loss: %0.3f | Acc: %0.2f" % (typ,i+1,len(dataloader),(total_loss/(i+1)),acc))
del inputs,labels,outputs
torch.cuda.empty_cache()
model.train()
def train_epoch(model=None,loss_fn=None,trainloader=None,optimiser=None,scheduler=None,device='cpu',epoch=0,chkpt_dir=None):
"""
Used to train a epoch
Attributes
----------
model:
The model to train
loss_fn:
The loss function
trainloader:
The batched dataloader for train dataset
optimiser:
The optimiser to use
scheduler:
The learning rate scheduler to use
device:
The device to use for training, cpu or gpu (default cpu)
epoch:
The current epoch
chkpt_dir:
The directory where to store checkpoint files
Returns
-------
(float) -> Epoch loss
"""
if model is None:
raise AttributeError("Model cannot be empty")
if loss_fn is None:
raise AttributeError("loss_fn cannot be empty")
if trainloader is None:
raise AttributeError("trainloader cannot be empty")
if optimiser is None:
raise AttributeError("optimiser cannot be empty")
total_loss = 0.0
total,correct = 0,0
for i,data in enumerate(trainloader):
start = time.time()
inputs, labels = data
inputs, labels = inputs.to(device),labels.to(device)
labels = labels.long()
#Compute and backpropagate loss
outputs = model(inputs)
loss = loss_fn(outputs, labels)
total_loss += loss.item()
loss.backward()
optimiser.step()
optimiser.zero_grad()
#Calculating correct predictions for accuracy
_, preds = torch.max(outputs.data, 1)
total += labels.size(0)
correct += (preds==labels).sum().item()
acc = 100 * correct / total
sys.stdout.write("\rTrain: Iteration %i/%i | Loss: %0.3f | Time Taken: %0.3f | Acc: %0.2f" % (i+1,len(trainloader),(total_loss/(i+1)),(time.time()-start),acc))
del inputs,labels,outputs
torch.cuda.empty_cache()
if scheduler is not None:
scheduler.step()
if chkpt_dir is not None:
#Checkpoint the model
path = Path(chkpt_dir).joinpath(epoch+'_saved_model.pth')
torch.save({'weights':model.state_dict(),'optimizer_state_dict':optimiser.state_dict(),'epoch':epoch},str(path))
sys.stdout.write('\n')
return loss.item()
def train_model(model=None,loss_fn=None,trainloader=None,valloader=None,epochs=1,batch_size=16,optimiser=None,scheduler=None,device='cpu',chkpt=False):
"""
Used to train the specified models
Attributes
----------
model:
The model to train
loss_fn:
The loss function
trainloader:
The batched dataloader for train dataset
valloader:
The batched dataloader for validation dataset
epochs:
The number of epochs to run (default 1)
batch_size:
The batch size to use during training (default 16)
optimiser:
The optimiser to use
scheduler:
The learning rate scheduler to use
device:
The device to use for training, cpu or gpu (default cpu)
"""
if model is None:
raise AttributeError("Model cannot be empty")
if loss_fn is None:
raise AttributeError("loss_fn cannot be empty")
if trainloader is None:
raise AttributeError("trainloader cannot be empty")
if optimiser is None:
raise AttributeError("optimiser cannot be empty")
if valloader is None:
raise AttributeError("valloader cannot be empty")
print("Starting the training\n----------------------")
start = time.time()
loss_epoch_arr = []
n_iters = np.ceil(len(trainloader))
#Perform an inital val evaluation
val_evaluation(valloader,model,loss_fn,device=device)
training_folder = None
if chkpt:
training_folder = 'training'+str(time.time())
Path(training_folder).mkdir(parents=True,exists_ok=True)
# The training loop
for epoch in range(epochs):
print('\nEpoch %i/%i\n-------------' % (epoch+1,epochs))
epoch_loss = train_epoch(model=model,loss_fn=loss_fn,trainloader=trainloader,optimiser=optimiser,scheduler=scheduler,device=device,epoch=epoch,chkpt_dir=training_folder)
loss_epoch_arr.append(epoch_loss)
val_evaluation(valloader,model,loss_fn,device=device)
#Save the weights such that training can be resumed
torch.save({'weights':model.state_dict(),'optimizer_state_dict':optimiser.state_dict()},'saved_model_to_resume.pth')
#Save the model for eval
torch.save(model.state_dict(),'saved_model_final.pth')
print('\nTraining finished in',timedelta(seconds=time.time()-start))
if __name__ == "__main__":
#Parse the arguments
parser = argparse.ArgumentParser(description="Create a dataset")
parser.add_argument('folder' ,type=str,nargs=1, help="folder which will be used to build the dataset")
parser.add_argument('--weight-file',type=str,nargs=1, help="The weight file for initializing the weights")
parser.add_argument('--weight-type',type=str,nargs=1, help="The weight file type")
parser.add_argument('--resume',action=argparse._StoreTrueAction,help="Resume training")
parser.add_argument('--class-list-file',type=str,nargs=1,help="The csv file containing class embeddings")
parser.add_argument('--val-size',type=float,nargs=1,help="The size of the val set")
parser.add_argument('--stratify',type=argparse._StoreTrueAction,help="Whether to use stratify during train test split")
parser.add_argument('--shuffle',type=argparse._StoreTrueAction,help="Whether to shuffle the datasets")
parser.add_argument('--batch-size',default=16,type=int,nargs=1,help="The batch size for the loader")
parser.add_argument('--lr',default=0.001,type=float,nargs=1,help="The learning rate")
parser.add_argument('--epochs',default=5,type=int,nargs=1,help="The number of epochs to run the training for")
parser.add_argument('--checkpoint',type=argparse._StoreTrueAction,help="Specify whether to checkpoint")
parser.add_argument('--ignore-weights',type=str,nargs='+',default=['fc.bias','fc.weight'],help="The weights to ignore while loading the weight file")
args = parser.parse_args()
parsed_args = vars(args)
#print(parsed_args)
#Use GPU if available
device = torch.device('cuda:0' if torch.cuda.is_available() else 'cpu')
print('Using device',device)
folder_name = str(Path(parsed_args['folder'][0]).absolute())
class_list_file = str(Path(parsed_args['class_list_file'][0]).absolute())
#Load the dataset from the specified folder
loaded_dataset = LoadDatasetFromFolder(folder_name,class_list_file)
X,y = loaded_dataset.load()
#Create the train-val split
val_size = 0.12
if parsed_args['val_size']:
val_size = parsed_args['val_size'][0]
datasets = CreateTrainValDatasets(X,y,val_size=val_size,stratify=parsed_args['stratify'])
trainset = datasets.get_trainset()
valset = datasets.get_valset()
print('Train set is of length',trainset.__len__())
print('Val set is of length',valset.__len__())
batch_size = parsed_args["batch_size"]
shuffle = parsed_args['shuffle']
#Create the batched loaders
trainloader = create_batched_loader(trainset,batch_size=batch_size,shuffle=shuffle)
valloader = create_batched_loader(valset,batch_size=batch_size,shuffle=shuffle)
#Define the model
model = resnet50(num_classes=loaded_dataset.num_classes())
#Define the loss_fn
loss_fn = nn.CrossEntropyLoss()
#Define the optimizer
lr = parsed_args['lr']
opt = optim.Adam(model.parameters(),lr=0.001)
#Define the scheduler
scheduler = optim.lr_scheduler.MultiStepLR(opt,[5,10])
resume =False
if parsed_args['resume'] is not None:
resume = True
# Try weight loading, if weight file provided.
try:
weight_file = str(Path(parsed_args['weight_file'][0]).absolute())
model = load_init_weights(weight_file,parsed_args['weight_type'][0],model,optimizer=opt,resume=resume,ignore_weights=parsed_args['ignore_weights'])
print("Weights loaded successfully")
except Exception as e:
print(e)
print("Weight file not provided,continuing without pre training")
#Move model to specified device
model = model.to(device)
epochs = parsed_args['epochs'][0]
#Start the training
train_model(model,loss_fn,trainloader,valloader,epochs=epochs,batch_size=batch_size,optimiser=opt,scheduler=scheduler,device=device,chkpt=parsed_args['checkpoint'])