-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathtrain_vanilla_sam.py
128 lines (101 loc) · 5.02 KB
/
train_vanilla_sam.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
"""
train feature extractor with vanilla sam
"""
#######################################################
# load helper functions and libraries
import tensorflow as tf
tf.compat.v1.disable_eager_execution()
import torch
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
from torch.optim.lr_scheduler import CosineAnnealingWarmRestarts
import os
from tqdm import tqdm
import numpy as np
import argparse
from models.resnet18 import create_resnet18
from models.losses import cross_entropy_loss, prototype_loss
from models.sam import SAM
from util_function import enable_running_stats, disable_running_stats
parser = argparse.ArgumentParser('train feature extractor with sam')
parser.add_argument("--eval_size", default=300, type = int)
parser.add_argument("--job_id", default=0, type = int)
parser.add_argument("--dataset_id", default=0, type = int)
parser.add_argument("--weight_decay", default=0.0007, type = float)
args = parser.parse_args()
#######################################################
# set hyperparameters
dataset_list = "ilsvrc_2012 omniglot aircraft cu_birds dtd quickdraw fungi vgg_flower".split(' ')
trainsets = [dataset_list[args.dataset_id]]
train_dataset = dataset_list[args.dataset_id]
valsets = trainsets
hyparam_list = {"ilsvrc_2012": [64, 3e-2, 480000, 48000, 0.05],
'omniglot': [16, 3e-2, 50000, 3000, 0.01],
'aircraft':[8, 3e-2, 50000, 3000, 0.1],
'cu_birds': [16, 3e-2, 50000, 3000, 0.1],
'dtd': [32, 3e-2, 50000, 1500, 0.05],
'quickdraw': [64, 1e-2, 480000, 48000, 0.05],
'fungi': [32, 3e-2, 480000, 15000, 0.05],
'vgg_flower': [8, 3e-2, 50000, 1500, 0.1]}
batch_size, lr, iter_num, optimizer_restart, rho = hyparam_list[train_dataset]
eval_every = optimizer_restart
weight_decay = args.weight_decay
eval_size = args.eval_size
#######################################################
# data set loader
from meta_dataset import MetaDatasetEpisodeReader, MetaDatasetBatchReader
testsets = ['mnist']
batch_loader = MetaDatasetBatchReader('train', trainsets, trainsets, testsets, batch_size=batch_size)
val_loader = MetaDatasetEpisodeReader('val', trainsets, valsets, testsets)
num_train_classes = batch_loader.num_classes('train')
#######################################################
# create model
model = create_resnet18(num_train_classes).cuda()
#######################################################
# start traning
config = tf.compat.v1.ConfigProto()
config.gpu_options.allow_growth = False
session = tf.compat.v1.Session(config=config)
base_optimizer = torch.optim.SGD
optimizer = SAM(model.parameters(), base_optimizer, lr = lr, rho = rho, weight_decay = weight_decay, momentum=0.9, nesterov=False)
lr_manager = CosineAnnealingWarmRestarts(optimizer.base_optimizer, optimizer_restart)
best_val_acc = 0
for iteration in tqdm(range(iter_num)):
model.train()
################################### train ###################################
batch_data = batch_loader.get_train_batch(session)
# first forward-backward step
enable_running_stats(model)
pred = model.forward(batch_data['images'])
loss, _ = cross_entropy_loss(pred, batch_data['labels'])
loss.backward()
optimizer.first_step(zero_grad=True)
# second forward-backward step
disable_running_stats(model)
pred = model.forward(batch_data['images'])
loss, _ = cross_entropy_loss(pred, batch_data['labels'])
loss.backward()
optimizer.second_step(zero_grad=True)
with torch.no_grad():
sl_loss, sl_acc = cross_entropy_loss(pred, batch_data['labels'])
# learniing rate scheduler
lr_manager.step(iteration)
################################### validation ###################################
if (iteration + 1) % eval_every == 0:
model.eval()
total_val_acc = []
for valset in valsets:
dataset_acc = []
for j in tqdm(range(eval_size)):
episode_data = val_loader.get_validation_task(session, valset)
with torch.no_grad():
context_features = model.embed(episode_data['context_images'].to(device))
target_features = model.embed(episode_data['target_images'].to(device))
fsc_loss, fsc_acc = prototype_loss(context_features, episode_data['context_labels'],
target_features, episode_data['target_labels'], distance = 'cos')
dataset_acc.append(fsc_acc.cpu().item())
total_val_acc.append(np.mean(dataset_acc))
if np.mean(total_val_acc) > best_val_acc:
best_val_acc = np.mean(total_val_acc)
print(f"--------------------------------- best model at {iteration}, val acc {best_val_acc} ---------------------------------")
state = {'epoch': iteration + 1, 'state_dict': model.get_state_dict()}
torch.save(state, os.path.join(f"saved_model/sam/vanilla/{train_dataset}.pth"))