-
Notifications
You must be signed in to change notification settings - Fork 0
/
train.py
258 lines (243 loc) · 14.3 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
from __future__ import print_function
from __future__ import division
import subprocess
import numpy as np
import json, os, sys, random, pickle
import torchvision.datasets as dset
import os
from PIL import Image
import urllib
from collections import OrderedDict
import torchvision.datasets as dset
import urllib
from collections import OrderedDict
import torch.nn as nn
import torch.optim as optim
import time
import copy
import torch
print("Torch version:", torch.__version__)
torch.multiprocessing.set_sharing_strategy('file_system')
import argparse
from model_init import *
from data_loader import *
from cosine_analysis.cosine_exp import *
from models_def.pytorch_models import *
from models_def.clip_model import *
from setup import *
from utils import *
import yaml
from pytorch_lightning import Trainer, seed_everything
from pytorch_lightning.loggers import TensorBoardLogger
def lightning_setup(args):
"""Loads a finetuned model into memory and returns it for bias analysis
Returns:
model_ft: finetuned model from args.trial_path
"""
model_ft = None
models_implemented = ['moco_resnet50', 'simclr_resnet50', 'alexnet', 'vgg', 'densenet', 'fasterrcnn', 'retinanet', 'googlenet', 'resnet18', 'resnet34', 'resnet50',
'resnet101', 'resnet152', 'resnext50_32x4d', 'resnext101_32x8d', 'wide_resnet50_2', 'wide_resnet101_2', 'bit_resnet50', 'virtex_resnet50']
if args.model_name == 'clip':
model_setup = CLIP_model(args, args.trial_path)
model_ft, _, _ = model_setup.setup_model()
checkpoint = torch.load(args.trial_path+'/model/model.pt')
model_ft.load_state_dict(checkpoint['model_state_dict'])
model_ft.eval()
elif args.model_name in models_implemented:
model_setup = LitPytorchModels(args, args.trial_path)
base = args.trial_path + '/model/'+args.model_name +'/' +'version_0/checkpoints/'
ckpt = os.listdir(base)[0]
checkpoint = torch.load(base+ckpt)
checkpoint['state_dict'] = dict(checkpoint['state_dict'])
state_dict_mod = {}
for i in checkpoint['state_dict']:
state_dict_mod[i[6:]] = checkpoint['state_dict'][i]
model_setup.model.load_state_dict(state_dict_mod)
model_ft = model_setup.model.eval()
else:
print("Model not implemented")
return model_ft
def lightning_train(args, dataloaders: dict, model_path: str, resume_training: bool = False):
"""Finetunes a model and saves the metadata in model_path
Args:
dataloaders: A dictionary with train and val pytorch dataloader objects
model_path: Path to save the model's metadata for a trial
resume_training: If true, uses saved checkpoint in model_path to resume training
Returns:
model_ft: finetuned model in eval mode
"""
seed_everything(args.seed, workers=True)
model_setup = None
model_ft = None
criterion = None
optimizer = None
models_implemented = ['moco_resnet50', 'simclr_resnet50', 'alexnet', 'vgg', 'densenet', 'fasterrcnn', 'retinanet', 'googlenet', 'resnet18', 'resnet34', 'resnet50',
'resnet101', 'resnet152', 'resnext50_32x4d', 'resnext101_32x8d', 'wide_resnet50_2', 'wide_resnet101_2', 'bit_resnet50', 'virtex_resnet50']
if resume_training == True:
if args.model_name == 'clip':
model_setup = CLIP_model(args, model_path)
model_ft, criterion, optimizer = model_setup.setup_model()
checkpoint = torch.load(args.checkpoint+'/model/model.pt')
model_ft.load_state_dict(checkpoint['model_state_dict'])
optimizer.load_state_dict(checkpoint['optimizer_state_dict'])
model_ft, _, _ = model_setup.train_model(args.dataset, model_ft, dataloaders, criterion, optimizer, num_epochs=args.epochs)
model_ft.eval()
torch.save(model_ft.state_dict(), model_path+'/model/model_final.pt')
elif args.model_name in models_implemented:
model_setup = LitPytorchModels(args, model_path)
logger = TensorBoardLogger(model_path+'/model', name=args.model_name)
trainer = Trainer(gpus=1, default_root_dir=model_path+'/model', logger=logger, callbacks=[], max_epochs=args.epochs, resume_from_checkpoint=args.checkpoint)
trainer.fit(model_setup, dataloaders['train'], dataloaders['val'])
trainer.validate(dataloaders=dataloaders['val'])
base = model_path + '/model/'+args.model_name +'/' +'version_0/checkpoints/'
ckpt = os.listdir(base)[0]
checkpoint = torch.load(base+ckpt)
checkpoint['state_dict'] = dict(checkpoint['state_dict'])
state_dict_mod = dict()
for i in checkpoint['state_dict']:
state_dict_mod[i[6:]] = checkpoint['state_dict'][i]
model_setup.model.load_state_dict(state_dict_mod)
model_ft = model_setup.model.eval()
else:
print("Model not implemented")
else:
if args.model_name == 'clip':
model_setup = CLIP_model(args, model_path)
model_ft, criterion, optimizer = model_setup.setup_model()
model_ft, _, _ = model_setup.train_model(args.dataset, model_ft, dataloaders, criterion, optimizer, num_epochs=args.epochs)
model_ft.eval()
elif args.model_name in models_implemented:
# call model's Lightning module trainer, train and return model_ft
model_setup = LitPytorchModels(args, model_path)
logger = TensorBoardLogger(model_path+'/model', name=args.model_name)
trainer = Trainer(gpus=args.num_gpus, default_root_dir=model_path+'/model', logger=logger, callbacks=[], max_epochs=args.epochs)
trainer.fit(model_setup, dataloaders['train'], dataloaders['val'])
trainer.validate(dataloaders=dataloaders['val'])
base = model_path + '/model/'+args.model_name +'/' +'version_0/checkpoints/'
ckpt = os.listdir(base)[0]
path = base+ckpt
checkpoint = torch.load(path)
checkpoint['state_dict'] = dict(checkpoint['state_dict'])
state_dict_mod = dict()
for i in checkpoint['state_dict']:
state_dict_mod[i[6:]] = checkpoint['state_dict'][i]
model_setup.model.load_state_dict(state_dict_mod)
model_ft = model_setup.model.eval()
else:
print("Model not implemented")
return model_ft
def extract_features(args, model_path: str, only_pretrained: bool, model_ft=None):
"""Generates features for a model and saves them in model_path
Args:
model_path: Path to model trial from which to extract features
only_pretraied: When False, returns features generated from the pretrained and finetuned model
model_ft: If only_pretrained=False, model_ft is the finetuned version of the model and used to
extract features on the finetuned model
Returns:
features: A dictionary mapping class name (specified in config file of analysis_set) to a tensor
of features of size N x d where N is the number of examples in a class and is the dimension
size of the model
"""
if args.model_name == 'clip':
feature_extractor = ClipViTFeatureExtractor(args, model_path)
else:
feature_extractor = PytorchFeatureExtractor(args, model_path)
if only_pretrained == True:
features = feature_extractor.extract_features(args, only_pretrained=True)
else:
features = feature_extractor.extract_features(args, model_ft=model_ft, only_pretrained=False)
return features
def main():
parser = argparse.ArgumentParser()
parser.add_argument('--num_classes', type=int,
help='number of classes in dataset', default=80)
parser.add_argument('--batch_size', type=int,
help='batch size for training', default=32)
parser.add_argument('--epochs', type=int,
help='number of epochs to train for', default=15)
parser.add_argument('--lr', type=float,
help='learning rate', default=0.01)
parser.add_argument('--lr_scheduler',
help='learning rate scheduler: cosine or reduce', default='none')
parser.add_argument('--momentum', type=float,
help='momentum value for sgd optimizer', default=0.9)
parser.add_argument('--feature_extract',
help='When false, finetune the whole model, when True, only update the reshaped layer parameters', action='store_true')
parser.add_argument('--optimizer', type=str,
help='optimizer for training, sgd or adam', default='sgd')
parser.add_argument('--model_name', type=str,
help='model type: clip, resnet18, resnet34, resnet50, resnet101, resnet152, resnext50_32x4d, resnext101_32x8d, wide_resnet50_2, wide_resnet101_2, alexnet, vgg, densenet, googlenet', default='resnet18')
parser.add_argument('--dataset', type=str,
help='Options: coco and openimages', default='coco')
parser.add_argument('--dataset_path', type=str,
help='folder path to training dataset', default="/localtmp/data/coco2017/coco_dataset/")
parser.add_argument('--analysis_set_path', type=str,
help='folder path to analysis dataset', default="/localtmp/data/coco2017/coco_dataset/")
parser.add_argument('--bias_analysis',
help='If True, performs cosine self similarity experiments', action='store_true')
parser.add_argument('--load_features',
help='If model has been trained, and want to use saved features for bias analysis', action='store_true')
parser.add_argument('--pretrained_features',
help='Extract features from pretrained model without finetuning', action='store_true')
parser.add_argument('--resume_training',
help='Resume training from a saved checkpoint', action='store_true')
parser.add_argument('--config_file', type=str,
help='path to config file', default="config/coco_ini")
parser.add_argument('--finetune',
help='finetune or train model from scratch', action='store_true')
parser.add_argument('--multiple_trials',
help='plot bias analysis across multiple trials', action='store_true')
parser.add_argument('--extract_cross_analysis_features',
help='extract features for an analysis set different than data the model is trained on', action='store_true')
parser.add_argument('--trial_path', type=str,
help='path to training run', default='experiments/coco/bit_resnet50/2022-01-15 14:18:21')
parser.add_argument('--checkpoint', type=str,
help='path to training run', default='none')
parser.add_argument('--analysis_set', type=str,
help='which dataset to perform bias analysis on', default='coco')
parser.add_argument('--seed', type=int,
help='random seed', default=1234)
parser.add_argument('--num_gpus', type=int,
help='random seed', default=1)
args = parser.parse_args()
torch.backends.cudnn.benchmark = True
if args.bias_analysis == True and args.load_features == True and not args.pretrained_features and not args.extract_cross_analysis_features:
# Don't finetune and just use saved features for full bias analysis with finetuned and pretrained features
print("Using saved features for full bias analysis")
if args.multiple_trials == True:
run_experiment(args.model_name, args.trial_path, args.dataset, args.analysis_set, args.config_file, features=None, only_pretrained = False, multiple_trials=args.multiple_trials)
else:
features = load_features(args.trial_path, analysis_set=args.analysis_set, only_pretrained=False)
run_experiment(args.model_name, args.trial_path, args.dataset, args.analysis_set, args.config_file, features=features, only_pretrained = False, multiple_trials=args.multiple_trials)
elif args.extract_cross_analysis_features == True:
# Extracts pretrained and finetuned features for a model on an analysis set regardless of what dataset the model was trained on
print("Extracting features on pretrained and finetuned model for dataset: "+args.analysis_set)
model_ft = lightning_setup(args)
features = extract_features(args, args.trial_path, only_pretrained=False, model_ft=model_ft)
if args.bias_analysis == True:
run_experiment(args.model_name, args.trial_path, args.dataset, args.analysis_set, args.config_file, features=features, only_pretrained = False, multiple_trials=args.multiple_trials)
elif args.pretrained_features == True:
# Only extract pretrained features and perform bias analysis on pretrained features
print("Extracting features on pretrained model for dataset: "+ args.analysis_set)
features = extract_features(args, args.trial_path, only_pretrained=True, model_ft=None)
if args.bias_analysis == True:
run_experiment(args.model_name, args.trial_path, args.dataset, args.analysis_set, args.config_file, features=features, only_pretrained=True, multiple_trials=args.multiple_trials)
else:
# Finetune, train the model from scratch or resume training, extract both pretrained and finetuned features and run bias analysis on them
datestring = datetime.datetime.now().strftime("%Y-%m-%d %H:%M:%S")
if args.finetune:
os.mkdir('experiments/'+ args.dataset + '/' + args.model_name+'/'+datestring)
model_path = 'experiments/'+ args.dataset + '/' + args.model_name+'/'+datestring
setup_dirs(model_path)
else:
os.mkdir('experiments/'+ args.dataset + '/' +args.model_name+'/model_scratch'+'/'+datestring)
model_path = 'experiments/'+ args.dataset + '/' +args.model_name+'/model_scratch'+'/'+datestring
setup_dirs(model_path, from_scratch=True)
dataloaders_dict = setup_dataset(args)
model_ft = lightning_train(args, dataloaders_dict, model_path, resume_training=args.resume_training)
features = extract_features(args, model_path, only_pretrained=False, model_ft=model_ft)
if args.bias_analysis == True:
# Run full bias experiment on finetuned and pretrained features
run_experiment(args.model_name, model_path, args.dataset, args.analysis_set, args.config_file, features=features, only_pretrained = False, multiple_trials=args.multiple_trials)
if __name__ == '__main__':
main()