/
train_model.py
348 lines (301 loc) · 14.1 KB
/
train_model.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
340
341
342
343
344
345
346
347
348
'''
--- I M P O R T S T A T E M E N T S ---
'''
import os
import datetime
import logging
import coloredlogs, logging
coloredlogs.install()
import math
import torch
import torchvision
import torch.backends.cudnn as cudnn
import torch.distributed as dist
from data import iterator_factory
from train import metric
from train.model import model
from train.lr_scheduler import MultiFactorScheduler
import apex
from apex import amp
from apex.parallel import DistributedDataParallel as DDP
'''
--- S T A R T O F T R A I N _ M O D E L ---
[About]
Main function that trains the specified model on a chosen dataset.
[Args]
- sym_net:
- name: Sting for the dataset name.
- model_prefix:
- input_conf:
- clip_length: Integer, for the number of frames to be used. Defaults to 32
- clip_size: Integer, specifying the spatial dimensionality of the frames. Defaults to 224
- train_frame_interval: Integer or List for the sampling interval to be used for sampling training set frames.
Defaults to 2.
- val_frame_interval: Integer or List for the sampling interval to be used for sampling validation set frames.
Defaults to 2.
- resume_epoch: Integer to be used in the case that training is to re-start from a specific epoch. If using -1
the training will start as normal from epoch 0. Defaults to -1.
- batch_size: Integer for the batch size to be used .In the case of cycles, the batch size is the base batch
size to be used by the scheduler based on which the new batches will be calculated. Defaults to 16.
- save_frequency: Integer for the frequency in which the model is saved. It is recommended to set the `save_frequency`
to 1 and delete any epochs after training in order to ensure that you obtain the best possible results. Defaults to 1.
- lr_base: Float for the base learning rate used. This is the lr to be used for the first n steps/epochs. If cycles are
enabled, the learning rate will degrade by .5 at each step. If `resume_epoch` is not -1, the initial learning rate will change accordingly, based on the `lr_base`. Defaults to 0.1.
- lr_factor: Float for the learning rate decay factor at each of the `lr_steps`, Usually the best practice is to use a
0.1 reduction, Defaults to 0.1.
- lr_steps: List of Integers for the steps to decrease the learning rate. Defaults to [50,100,150]
- long_cycles: Boolean to specify if long cycles are to be used. Defaults to True
- short_cycles: Boolean to specify if short cycles are to be used. Defaults to True.
- end_epoch: Integer for the last epoch while training. Defaults to 300.
- pretrained_3d: String for path of pre-trained model. If None is used, the model is trained from scratch. Defaults to
None.
- fine_tune: Bolean for the case of fine-tuning. This is especially useful for decreasing the learning rate of the
convolution weights while maintaining the learning rate for the classifier (and thus training mostly the neurones
that are responible for class predictions). Defaults to False.
- dataset_location: String for the complete filepath of the dataset. This is the parent path. Along that path an
additional folder 'labels' should exist. Defaults to 'Kinetics'.
- net_name: String for the network name in order to be ported. Defaults to 'r3d_50'.
- gpus: Integer for the number of GPUs to be used. Defaults to 4.
[Returns]
None
'''
def train_model(sym_net, name, model_prefix, input_conf,
clip_length=32, clip_size=224, train_frame_interval=2, val_frame_interval=2,
resume_epoch=-1, batch_size=16, save_frequency=1,
lr_base=0.01, lr_factor=0.1, lr_steps=[50,100,150],
enable_long_cycles=True, enable_short_cycles=True, end_epoch=300,
pretrained_3d=None, fine_tune=False, dataset_location='Kinetics', net_name='r3d_50', gpus=4,
**kwargs):
assert torch.cuda.is_available(), "Only support CUDA devices."
# Make results directory for .csv files if it does not exist
results_path = str('./results/'+str(name)+'/'+str(net_name))
if not os.path.exists(results_path):
os.makedirs(results_path)
# data iterator - randomisation based on date and time values
iter_seed = torch.initial_seed() + 100 + max(0, resume_epoch) * 100
now = datetime.datetime.now()
iter_seed += now.year + now.month + now.day + now.hour + now.minute + now.second
# Get parent location
# - `data` folder should include all the dataset examples.
# - `labels` folder should inclde all labels in .csv format.
# We use a global label formating - you can have a look at the link in the `README.md` to download the files.
data_location = dataset_location.split('/data/')[0]
clip_length = int(clip_length)
clip_size = int(clip_size)
train_loaders = {}
# Create custom loaders for train and validation
train_data, eval_loader, train_length = iterator_factory.create(
name=name,
batch_size=batch_size,
return_len=True,
clip_length=clip_length,
clip_size=clip_size,
val_clip_length=clip_length,
val_clip_size=clip_size,
train_interval=train_frame_interval,
val_interval=val_frame_interval,
mean=input_conf['mean'],
std=input_conf['std'],
seed=iter_seed,
data_root=data_location)
# Create model
net = model(net=sym_net,
criterion=torch.nn.CrossEntropyLoss().cuda(),
model_prefix=model_prefix,
step_callback_freq=1,
save_checkpoint_freq=save_frequency,
opt_batch_size=batch_size, # optional
)
net.net.cuda()
# Parameter LR configuration for optimiser
# Base layers are based on the layers as loaded to the model
param_base_layers = []
base_layers_mult = 1.0
# New layers are based on fine-tuning
param_new_layers = []
new_layers_mult = 1.0
name_base_layers = []
param_transpose_layers = []
transpose_layers_mult = 1.0
param_rec_layers = []
rec_layers_mult = 1.0
# Iterate over all parameters
for name, param in net.net.named_parameters():
if fine_tune:
if 'transpose' in name.lower():
param_transpose_layers.append(param)
transpose_layers_mult = .2
elif name.lower().startswith('classifier'):
new_layers_mult = .1
param_new_layers.append(param)
elif ('lstm' or 'gru') in name.lower():
param_transpose_layers.append(param)
rec_layers_mult = .5
else:
param_base_layers.append(param)
base_layers_mult = .6
name_base_layers.append(name)
else:
if 'transpose' in name.lower():
param_transpose_layers.append(param)
transpose_layers_mult = .8
elif ('lstm' or 'gru') in name.lower():
param_transpose_layers.append(param)
rec_layers_mult = 1.
else:
param_new_layers.append(param)
# User feedback
if name_base_layers:
out = "[\'" + '\', \''.join(name_base_layers) + "\']"
logging.info("Optimiser:: >> recuding the learning rate of {} params: {}".format(len(name_base_layers),
out if len(out) < 300 else out[0:150] + " ... " + out[-150:]))
optimiser = torch.optim.SGD([
{'params': param_base_layers, 'lr_mult': base_layers_mult},
{'params': param_new_layers, 'lr_mult': new_layers_mult},
{'params': param_rec_layers, 'lr_mult': rec_layers_mult},
{'params': param_transpose_layers, 'lr_mult': transpose_layers_mult},],
lr=lr_base,
momentum=0.9,
weight_decay=0.0001,
nesterov=True
)
# Use Apex for Mixed precidion - Note: Please comment out any apex code on `train_model.py` and `model.py`
# in case you wish to switch back to standard float32. "O0" opt_level still has some bugs when also using DataParallel
net.net, optimiser = amp.initialize(net.net, optimiser, opt_level="O1")
# Create DataParallel wrapper
net.net = torch.nn.DataParallel(net.net, device_ids=[i for i in range(int(gpus))])
# load params from pretrained 3d network
if pretrained_3d:
assert os.path.exists(pretrained_3d), "cannot locate: `{}'".format(pretrained_3d)
logging.info("Initialiser:: loading model states from: `{}'".format(pretrained_3d))
checkpoint = torch.load(pretrained_3d)
net.load_state(checkpoint['state_dict'], strict=False)
num_steps = train_length // batch_size
# Long Cycle steps
if (enable_long_cycles):
count = 0
index = 0
iter_sizes = [8, 4, 2, 1]
initial_num = num_steps
# Expected to find the number of batches that fit exactly to the number of iterations.
# So the sum of the floowing batch sizes should be less or equal to the number of batches left.
while sum(iter_sizes[index:]) <= num_steps:
# Case 1: 8 x B
if iter_sizes[index] == 8:
count += 1
index = 1
num_steps -= 8
# Case 2: 4 x B
elif iter_sizes[index] == 4:
count += 1
index = 2
num_steps -= 4
# Case 3: 2 x B
elif iter_sizes[index] == 2:
count += 1
index = 3
num_steps -= 2
# Base case
elif iter_sizes[index] == 1:
count += 1
index = 0
num_steps -= 1
print ("New number of batches per epoch is {:d} being equivalent to {:1.3f} of original number of batches with Long cycles".format(count,float(count)/float(initial_num)))
num_steps = count
# Short Cycle steps
if (enable_short_cycles):
# Iterate for *every* batch
i = 0
while i <= num_steps:
m = i%3
# Case 1: Base case
if (m==0):
num_steps -= 1
# Case 2: b = 2 x B
if (m==1):
num_steps -= 2
# Case 3: b = 4 x B
else:
num_steps -= 4
i += 1
# Update new number of batches
print ("New number of batches per epoch is {:d} being equivalent to {:1.3f} of original number of batches with Short cycles".format(i,float(i)/float(initial_num)))
num_steps = i
# Split the batch number to four for every change in the long cycles
long_steps = None
if (enable_long_cycles):
step = num_steps//4
long_steps = list(range(num_steps))[0::step]
num_steps = long_steps[-1]
# Create full list of long steps (for all batches)
for epoch in range(1,end_epoch):
end = long_steps[-1]
long_steps = long_steps + [x.__add__(end) for x in long_steps[-4:]]
# Fool-proofing
if (long_steps[0]==0):
long_steps[0]=1
# resume training: model and optimiser - (account of various batch sizes)
if resume_epoch < 0:
epoch_start = 0
step_counter = 0
else:
# Try to load previous state dict in case `pretrained_3d` is None
if not pretrained_3d:
try:
net.load_checkpoint(epoch=resume_epoch, optimizer=optimiser)
except Exception:
logging.warning('Initialiser:: No previous checkpoint found in the directory! You can specify the path explicitly with `pretrained_3d` argument.')
epoch_start = resume_epoch
step_counter = epoch_start * num_steps
# Step dictionary creation
iteration_steps = {'long_0':[],'long_1':[],'long_2':[],'long_3':[],'short_0':[],'short_1':[],'short_2':[]}
#Populate dictionary
for batch_i in range(0,num_steps):
# Long cycle cases
if batch_i>=0 and batch_i<num_steps//4:
iteration_steps['long_0'].append(batch_i)
elif batch_i>=num_steps//4 and batch_i<num_steps//2:
iteration_steps['long_1'].append(batch_i)
elif batch_i>=num_steps//2 and batch_i<(3*num_steps)//4:
iteration_steps['long_2'].append(batch_i)
else:
iteration_steps['long_3'].append(batch_i)
# Short cases
if (batch_i%3==0):
iteration_steps['short_0'].append(batch_i)
elif (batch_i%3==1):
iteration_steps['short_1'].append(batch_i)
else:
iteration_steps['short_2'].append(batch_i)
# set learning rate scheduler
lr_scheduler = MultiFactorScheduler(base_lr=lr_base,
steps=[x*num_steps for x in lr_steps],
iterations_per_epoch=num_steps,
iteration_steps=iteration_steps,
factor=lr_factor,
step_counter=step_counter)
# define evaluation metric
metrics = metric.MetricList(metric.Loss(name="loss-ce"),
metric.Accuracy(name="top1", topk=1),
metric.Accuracy(name="top5", topk=5),
metric.BatchSize(name="batch_size"),
metric.LearningRate(name="lr"))
# enable cudnn tune
cudnn.benchmark = True
# Main training happens here
net.fit(train_iter=train_data,
eval_iter=eval_loader,
batch_shape=(int(batch_size),int(clip_length),int(clip_size),int(clip_size)),
workers=8,
no_cycles=(not(enable_long_cycles) and not(enable_short_cycles)),
optimiser=optimiser,
long_short_steps_dir=iteration_steps,
lr_scheduler=lr_scheduler,
metrics=metrics,
iter_per_epoch=num_steps,
epoch_start=epoch_start,
epoch_end=end_epoch,
directory=results_path)
'''
--- E N D O F T R A I N _ M O D E L F U N C T I O N ---
'''