-
Notifications
You must be signed in to change notification settings - Fork 0
/
train_digit_classifier.py
166 lines (131 loc) · 6.45 KB
/
train_digit_classifier.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
from sklearn.metrics import balanced_accuracy_score, accuracy_score
import wandb
from constants import *
from torch.utils.data import random_split, DataLoader
import torchvision.datasets as datasets
import torchvision
from ex3.AutoEncoder import Encoder
from ex3.config_loader import load_config
from digit_classifier import DigitClassifier
train_dataloader = None
test_dataloader = None
reload_model = False
def get_config(max_fc_layer_num):
sweep_config = {}
sweep_config['method'] = 'bayes'
sweep_config['metric'] = {'name': 'balanced_accuracy', 'goal': 'maximize'}
sweep_config['name'] = f"DigitClassifier_{get_time()}"
param_dict = {
'model_name': {'value': 'DigitClassifier'},
'model_type': {'value': 2},
"fc_layers_num": {'distribution': 'int_uniform', 'min': 1, 'max': max_fc_layer_num},
'grad': {'values': [0, 1]},
'optimizer': {'values': [0, 1]},
'activation': {'values': [0, 1, 2]},
'encoder_index': {'values': list(range(len(models)))}
}
for i in range(max_fc_layer_num):
param_dict[f"fc_out_features_{i}"] = {'distribution': 'int_uniform', 'min': 3, 'max': 15}
sweep_config['parameters'] = param_dict
sweep_config['parameters'].update(
{
# "learning_rate": {'distribution': 'uniform', 'min': 0.001, 'max': 0.003},
"learning_rate": {'values': [0.0001, 0.001, 0.01, 0.1, 0.005, 0.0005, 0.05]},
'epoch_num': {'distribution': 'int_uniform', 'min': 10, 'max': 700}, })
# 'epoch_num': {'value': 700}})
return sweep_config
def collect_batch(batch):
def one_hot(num: int):
out = tr.zeros(10)
out[num] = 1
return out
pictures = tr.stack([padding(item[0]) for item in batch])
labels = tr.stack([one_hot(item[1]) for item in batch])
return pictures, labels
def load_data(train_batch_size, test_batch_size):
mnist_train_set = datasets.MNIST(root='./data', train=True, download=True,
transform=torchvision.transforms.Compose(
[torchvision.transforms.ToTensor()]))
mnist_train_set, _ = tr.utils.data.random_split(mnist_train_set, [100, len(mnist_train_set) - 100])
mnist_test_set = datasets.MNIST(root='./data', train=False, download=True,
transform=torchvision.transforms.Compose(
[torchvision.transforms.ToTensor()]))
train_dataloader1 = DataLoader(mnist_train_set, batch_size=train_batch_size, shuffle=True,
collate_fn=collect_batch)
test_dataloader1 = DataLoader(mnist_test_set, batch_size=test_batch_size, shuffle=True,
collate_fn=collect_batch)
return train_dataloader1, test_dataloader1
def train(config=None):
with wandb.init(config=config):
config = wandb.config
model_path, config_path = models[config["encoder_index"]]
AE_loaded = tr.load(model_path)
encoder_config = load_config(config_path)
custom_encoder = Encoder((32, 32), encoder_config["latent_dim"], encoder_config)
custom_encoder.load_state_dict(AE_loaded['encoder'])
wandb.log({"latent_dim": encoder_config["latent_dim"]})
digit_classifier = DigitClassifier(custom_encoder, encoder_config["latent_dim"], config)
wandb.watch(digit_classifier)
criterion = nn.CrossEntropyLoss()
optimizer_dict = {0: tr.optim.SGD, 1: tr.optim.Adam}
optimizer = optimizer_dict[config["optimizer"]](digit_classifier.parameters(), lr=config[
'learning_rate'])
train_loss = 1.0
test_loss = 1.0
test_interval = 3
all_itr = 0
num_epochs = config['epoch_num']
for epoch in range(num_epochs):
itr = 0
for pictures, labels in train_dataloader: # getting training batches
itr += 1
all_itr += 1
if (itr + 1) % test_interval == 0:
test_iter = True
pictures, labels = next(iter(test_dataloader))
else:
test_iter = False
# _______________________________________TRAINING:____________________________________________
output = digit_classifier(pictures)
loss = criterion(output, labels)
if not test_iter:
optimizer.zero_grad()
loss.backward()
optimizer.step()
if test_iter:
test_loss = 0.8 * float(loss.detach()) + 0.2 * test_loss
else:
train_loss = 0.9 * float(loss.detach()) + 0.1 * train_loss
if test_iter:
y_pred = tr.argmax(output, dim=1).cpu().detach().numpy()
y_true = tr.argmax(labels, dim=1).cpu().detach().numpy()
balanced_accuracy = balanced_accuracy_score(y_true, y_pred)
accuracy = accuracy_score(y_true, y_pred, normalize=True)
wandb.log({"epoch": epoch,
"test_loss": test_loss,
"train_loss": train_loss,
"accuracy": accuracy,
"balanced_accuracy": balanced_accuracy},
step=all_itr)
print(
f"Epoch [{epoch + 1}/{num_epochs}], "
f"Step [{itr + 1}/{len(train_dataloader)}], "
f"Train Loss: {train_loss:.4f}, "
f"Test Loss: {test_loss:.4f}"
)
# saving the model
if not reload_model:
tr.save({"encoder": custom_encoder.state_dict(),
"digit_classifier": digit_classifier.state_dict()},
f"models/digit_classifier"
f"/{config._settings.run_id}_digit_classifier.pth")
# if itr > len(train_dataloader) // 2 and test_loss > 0.05:
# wandb.finish()
def main():
global train_dataloader
global test_dataloader
train_dataloader, test_dataloader = load_data(10, 200)
sweep_id = wandb.sweep(get_config(7), project="ex3_digit_classification_test1", entity="malik-noam-idl")
wandb.agent(sweep_id, train, count=1000)
if __name__ == '__main__':
main()