Transfer-Learning/train.py at master · savan77/Transfer-Learning · GitHub

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
import torch
import torch.nn as nn
import torch.nn.functional as F
import torch.optim as optim
from torchvision import datasets, transforms
import argparse
from model import ResNetFashion
from data_loader import FashionDataset
from parameters import Args
import os
from torch.utils.tensorboard import SummaryWriter


def save_model_fn(epoch, model, optimizer, name, label_map):


	#save the model
	state = {
	    'epoch': epoch,
	    'state_dict': model.state_dict(),
	    'optimizer': optimizer.state_dict(),
	    'label_map':label_map,
	}

	torch.save(state, name)
	print("Model- {} saved successfully".format(name))


def get_class_weights(label_map, class_distribution, scheme):
	print("Calculating weights for classes...")
	# print(class_distribution)
	if scheme == 1:
		weights = [1/class_distribution[i] for i in label_map]
	elif scheme == 2:
		max_dist = max(class_distribution.values())
		weights = [max_dist/class_distribution[i] for i in label_map]
	# print(weights)
	return weights

def oversample(data):
	indices = list(range(len(data)))

	new_map = {value:key for (key,value) in data.label_map.items()}
	# print(data.class_distribution)
	# print([data.class_distribution[new_map[data[i][1]]] for i in indices][:10])
	weights = [1.0/data.class_distribution[new_map[data[i][1]]] for i in indices]
	# print(weights[:10])
	return torch.utils.data.sampler.WeightedRandomSampler(weights, len(weights))

def train_model():

	transforms_ = transforms.Compose([transforms.Resize((224,224)), transforms.ToTensor()])
	data_cls = FashionDataset(Args.data_dir, Args.train_csv, transform=transforms_)
	if Args.oversample:
		print("Oversampling minority classes, this might take a while...")
		sampler = oversample(data_cls)
		dataset_loader = torch.utils.data.DataLoader(dataset=data_cls, batch_size=Args.batch_size, sampler=sampler)
	else:
		dataset_loader = torch.utils.data.DataLoader(dataset=data_cls, batch_size=Args.batch_size, shuffle=True)
	cuda = torch.cuda.is_available()
	model = ResNetFashion(base=Args.base_model, num_classes=Args.num_classes)


	if cuda:
		model.cuda()


	if Args.optimizer == "sgd":
		optimizer = optim.SGD(model.parameters(), lr=Args.learning_rate, momentum=Args.momentum)
	elif Args.optimizer == "adam":
		optimizer = optim.Adam(model.parameters(), lr=Args.learning_rate)

	if Args.checkpoint != None:
		print("restoring from checkpoint ", Args.checkpoint)
		cp = torch.load(Args.checkpoint)
		# for n, v in cp['state_dict'].items():
			# print(n)
		if Args.diff_num_class:
			print("Restoring weights except last layer")
			model_dict = model.state_dict()
			pretrained_dict = cp['state_dict']
			# remove last two elements (fc weight and fc bias)
			# filtered_dict = {k: v for k, v in pretrained_dict.items() if "fc" not in k}
			pretrained_dict.popitem()
			pretrained_dict.popitem()
			model_dict.update(pretrained_dict)
			model.load_state_dict(model_dict)
		else:
			print("Restoring weights for all layers")
			model.load_state_dict(cp['state_dict'])
			optimizer.load_state_dict(cp['optimizer'])
		# if Args.num_class_finetune != None:
		# 	num_features = model.fc.in_features
		# 	model.fc = nn.Linear(num_features, Args.num_class_finetune).cuda()


	if Args.weighted_loss and not Args.oversample:
		weights = get_class_weights(data_cls.label_map, data_cls.class_distribution, Args.weighted_loss_scheme)
		class_weights = torch.FloatTensor(weights)
		# if cuda:
			# class_weight.to('cuda')
		criterion = nn.CrossEntropyLoss(weight=class_weights.cuda())
	else:
		criterion = nn.CrossEntropyLoss()

	writer = SummaryWriter(os.path.join(Args.output_dir, Args.name))


	for epoch in range(Args.epochs):
		total_loss = 0.0
		correct = 0
		for i, (data, target) in enumerate(dataset_loader):

			if cuda:
				data, target = data.cuda(), target.cuda()
			optimizer.zero_grad()
			output = model(data)

			loss = criterion(output, target)
			loss.backward()
			optimizer.step()
			total_loss += loss.item()
			output_idx = output.argmax(dim=1, keepdim=True)
			correct += output_idx.eq(target.view_as(output_idx)).sum().item()
			print("Epoch: {} Minibatch:{} Loss: {}".format(epoch, i, loss.item()))
			# if i % Args.avg_loss_batch == 0 and i != 0:  #print loss after every 100 mini batches
			# 	print("Avg loss over last {} batches: {}".format(Args.avg_loss_batch, total_loss/Args.avg_loss_batch))
			# 	total_loss= 0.0
		avg_loss = total_loss/len(data_cls)
		accuracy = correct / len(data_cls)
		print("Average Loss Over {} Epochs: {}".format(epoch, avg_loss))
		print("After Epoch: {}  Accuracy: {}".format(epoch, accuracy))
		if Args.tensorboard:
			writer.add_scalar("Loss/train", avg_loss, epoch)
			writer.add_scalar("Accuracy/train", accuracy)

		if Args.save_model!= None and epoch % Args.save_model==0:
			name = os.path.join(Args.output_dir, Args.name, "fashion-"+str(epoch)+".pth")
			save_model_fn(epoch, model, optimizer, name, data_cls.label_map)


if __name__ == "__main__":
	train_model()