-
Notifications
You must be signed in to change notification settings - Fork 24
/
train_unsupervised.py
218 lines (201 loc) · 12 KB
/
train_unsupervised.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
#!/usr/bin/env python
# The MIT License (MIT)
# Copyright (c) 2020 Massimiliano Patacchiola
# Paper: "Self-Supervised Relational Reasoning for Representation Learning", M. Patacchiola & A. Storkey, NeurIPS 2020
# GitHub: https://github.com/mpatacchiola/self-supervised-relational-reasoning
# Manage self-supervised training on different datasets/backbones/methods.
# Example command:
#
# python train_unsupervised.py --dataset="cifar10" --method="relationnet" --backbone="conv4" --seed=3 --data_size=64 --K=32 --gpu=0 --epochs=200
import os
import argparse
parser = argparse.ArgumentParser(description="Training script for the unsupervised phase via self-supervision")
parser.add_argument("--seed", default=-1, type=int, help="Seed for Numpy and PyTorch. Default: -1 (None)")
parser.add_argument("--epoch_start", default=0, type=int, help="Epoch to start learning from, used when resuming")
parser.add_argument("--epochs", default=200, type=int, help="Total number of epochs")
parser.add_argument("--dataset", default="cifar10", help="Dataset: cifar10|100, supercifar100, tiny, slim, stl10")
parser.add_argument("--backbone", default="conv4", help="Backbone: conv4, resnet|8|32|34|56")
parser.add_argument("--method", default="relationnet", help="Model: standard, randomweights, relationnet, rotationnet, deepinfomax, simclr")
parser.add_argument("--data_size", default=128, type=int, help="Size of the mini-batch")
parser.add_argument("--K", default=32, type=int, help="Total number of augmentations (K), sed only in RelationNet")
parser.add_argument("--aggregation", default="cat", help="Aggregation function used in RelationNet: sum, mean, max, cat")
parser.add_argument("--id", default="", help="Additional string appended when saving the checkpoints")
parser.add_argument("--checkpoint", default="", help="location of a checkpoint file, used to resume training")
parser.add_argument("--num_workers", default=8, type=int, help="Number of torchvision workers used to load data (default: 8)")
parser.add_argument("--gpu", default="0", type=str, help="GPU id in case of multiple GPUs")
args = parser.parse_args()
if(args.id!=""):
header = str(args.method)+ "_" + str(args.id) + "_" + str(args.dataset) + "_" + str(args.backbone) + "_seed_" + str(args.seed)
else:
header = str(args.method) + "_" + str(args.dataset) + "_" + str(args.backbone) + "_seed_" + str(args.seed)
os.environ["CUDA_VISIBLE_DEVICES"] = args.gpu
import torch
import torch.optim
import torchvision.datasets as dset
import torchvision.transforms as transforms
import numpy as np
import random
if(args.seed>=0):
random.seed(args.seed)
np.random.seed(args.seed)
torch.manual_seed(args.seed)
torch.cuda.manual_seed(args.seed)
torch.cuda.manual_seed_all(args.seed)
torch.backends.cudnn.deterministic = True
torch.backends.cudnn.benchmark = False
print("[INFO] Setting SEED: " + str(args.seed))
else:
print("[INFO] Setting SEED: None")
if(torch.cuda.is_available() == False): print("[WARNING] CUDA is not available.")
if(args.backbone=="conv4"):
from backbones.conv4 import Conv4
feature_extractor = Conv4(flatten=True)
elif(args.backbone=="resnet8"):
from backbones.resnet_small import ResNet, BasicBlock
feature_extractor = ResNet(BasicBlock, [1, 1, 1], channels=[16, 32, 64], flatten=True)
elif(args.backbone=="resnet32"):
from backbones.resnet_small import ResNet, BasicBlock
feature_extractor = ResNet(BasicBlock, [5, 5, 5], channels=[16, 32, 64], flatten=True)
elif(args.backbone=="resnet56"):
from backbones.resnet_small import ResNet, BasicBlock
feature_extractor = ResNet(BasicBlock, [9, 9, 9], channels=[16, 32, 64], flatten=True)
elif(args.backbone=="resnet34"):
from backbones.resnet_large import ResNet, BasicBlock
feature_extractor = ResNet(BasicBlock, layers=[3, 4, 6, 3],zero_init_residual=False,
groups=1, width_per_group=64, replace_stride_with_dilation=None,
norm_layer=None)
else:
raise RuntimeError("[ERROR] the backbone " + str(args.backbone) + " is not supported.")
tot_params = sum(p.numel() for p in feature_extractor.parameters() if p.requires_grad)
print("[INFO]", str(str(args.backbone)), "loaded in memory.")
print("[INFO] Feature size:", str(feature_extractor.feature_size))
print("[INFO] Feature extractor TOT trainable params: " + str(tot_params))
print("[INFO] Found " + str(torch.cuda.device_count()) + " GPU(s) available.")
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
print("[INFO] Device type: " + str(device))
# The datamanager is a separate class that returns
# appropriate data loaders and information based
# on the method and dataset selected.
from datamanager import DataManager
manager = DataManager(args.seed)
num_classes = manager.get_num_classes(args.dataset)
if(args.method=="relationnet"):
from methods.relationnet import Model
model = Model(feature_extractor, device, aggregation=args.aggregation)
print("[INFO][RelationNet] TOT augmentations (K): " + str(args.K))
print("[INFO][RelationNet] Aggregation function: " + str(args.aggregation))
train_transform = manager.get_train_transforms(args.method, args.dataset)
train_loader, _ = manager.get_train_loader(dataset=args.dataset,
data_type="multi",
data_size=args.data_size,
train_transform=train_transform,
repeat_augmentations=args.K,
num_workers=args.num_workers,
drop_last=False)
elif(args.method=="standard"):
from methods.standard import StandardModel
model = StandardModel(feature_extractor, num_classes, tot_epochs=args.epochs)
if(args.dataset=="stl10"):
train_transform = manager.get_train_transforms("finetune", args.dataset)
else:
train_transform = manager.get_train_transforms("standard", args.dataset)
test_loader = manager.get_test_loader(args.dataset, data_size=args.data_size, num_workers=args.num_workers)
train_loader, _ = manager.get_train_loader(dataset=args.dataset,
data_type="single",
data_size=args.data_size,
train_transform=train_transform,
repeat_augmentations=None,
num_workers=args.num_workers,
drop_last=False)
elif(args.method=="rotationnet"):
from methods.rotationnet import Model
model = Model(feature_extractor)
train_transform = manager.get_train_transforms(args.method, args.dataset)
if(args.dataset=="stl10"): data_type="unsupervised"
else: data_type="single"
train_loader, _ = manager.get_train_loader(dataset=args.dataset,
data_type=data_type,
data_size=args.data_size,
train_transform=train_transform,
repeat_augmentations=None,
num_workers=args.num_workers,
drop_last=False)
elif(args.method=="randomweights"):
if not os.path.exists("./checkpoint/"+str(args.method)+"/"+str(args.dataset)):
os.makedirs("./checkpoint/"+str(args.method)+"/"+str(args.dataset))
feature_extractor_state_dict = feature_extractor.state_dict()
checkpoint_path = "./checkpoint/"+str(args.method)+"/"+str(args.dataset)+"/"+header+".tar"
print("Saving in:", checkpoint_path)
torch.save({"backbone": feature_extractor_state_dict}, checkpoint_path)
import sys
sys.exit()
elif(args.method=="deepinfomax"):
from methods.deepinfomax import DIM
model = DIM(feature_extractor, alpha=0.0, beta=1.0, gamma=0.1)
train_transform = manager.get_train_transforms(args.method, args.dataset)
if(args.dataset=="stl10"): data_type="unsupervised"
else: data_type="single"
train_loader, _ = manager.get_train_loader(dataset=args.dataset,
data_type=data_type,
data_size=args.data_size,
train_transform=train_transform,
repeat_augmentations=None,
num_workers=args.num_workers,
drop_last=True)
elif(args.method=="simclr"):
from methods.simclr import Model
model = Model(feature_extractor)
train_transform = manager.get_train_transforms(args.method, args.dataset)
train_loader, _ = manager.get_train_loader(dataset=args.dataset,
data_type="multi",
data_size=args.data_size,
train_transform=train_transform,
repeat_augmentations=2,
num_workers=args.num_workers,
drop_last=False)
elif(args.method=="deepcluster"):
from methods.deepcluster import Model
train_transform = manager.get_train_transforms(args.method, args.dataset)
model = Model(feature_extractor, batch_size=args.data_size, num_clusters=num_classes*10, train_transform=train_transform)
if(args.dataset=="stl10"): data_type="unsupervised"
else: data_type="single"
_, train_set = manager.get_train_loader(dataset=args.dataset,
data_type=data_type,
data_size=args.data_size,
train_transform=train_transform,
repeat_augmentations=None,
num_workers=args.num_workers,
drop_last=False)
# Note: for DeepCluster we take the train-set but here for convinience
# we rename it train_loader to avoid overhead in the training loop.
train_loader = train_set
else:
raise RuntimeError("[ERROR] the model " + str(args.method) + " is not supported.")
model.to(device)
#model = model.to(device)
#model = torch.nn.DataParallel(model).to(device)
# NOTE: the checkpoint must be loaded AFTER
# the model has been allocated into the device.
if(args.checkpoint!=""):
print("Loading checkpoint: " + str(args.checkpoint))
model.load(args.checkpoint)
print("Loading checkpoint: Done!")
def main():
if not os.path.exists("./checkpoint/"+str(args.method)+"/"+str(args.dataset)):
os.makedirs("./checkpoint/"+str(args.method)+"/"+str(args.dataset))
log_file = "./checkpoint/"+str(args.method)+"/"+str(args.dataset)+"/log_"+header+".cvs"
with open(log_file, "w") as myfile: myfile.write("epoch,loss,score" + "\n") # create a new log file (it destroys the previous one)
for epoch in range(args.epoch_start, args.epochs):
loss_train, accuracy_train = model.train(epoch, train_loader) #<-- Each model must have a "train" method
with open(log_file, "a") as myfile:
myfile.write(str(epoch)+","+str(loss_train)+","+str(accuracy_train)+"\n")
if(epoch in [int(args.epochs*0.25)-1, int(args.epochs*0.5)-1, int(args.epochs*0.75)-1, args.epochs-1]):
checkpoint_path = "./checkpoint/"+str(args.method)+"/"+str(args.dataset)+"/"+header+"_epoch_"+ str(epoch+1)+".tar"
print("[INFO] Saving in:", checkpoint_path)
model.save(checkpoint_path)
if(args.method=="standard"):
#For the standard supervised method, it estimates now the test accuracy
loss_test, accuracy_test = model.test(test_loader)
print("Test loss: " + str(loss_test) )
print("Test accuracy: " + str(accuracy_test) + "%")
if __name__== "__main__": main()