/
train.py
177 lines (156 loc) · 7.91 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
import argparse
import copy
import datetime
import os
import shutil
import sys
import warnings
import numpy as np
import torch.utils.data
from torch import nn
from torch.utils.tensorboard import SummaryWriter
from torchsummary import summary
from config import TRAIN_BATCH_SIZE, TEST_BATCH_SIZE
from dataset import AudioSamplePairDataset
from models import SmallEncoder, LargeEncoder
from searcher import Searcher
from utils import load_train_state, save_train_state
SAVED_MODELS_PATH = 'saved_models'
CHECKPOINT_FILE = 'checkpoint.pck'
LOGS_DIR = 'runs'
def evaluate_encoder(encoder: nn.Module, test_loader: torch.utils.data.DataLoader, loss_fn: nn.TripletMarginLoss = None,
writer: SummaryWriter = None,
epoch: int = -1):
with torch.no_grad():
encoder.eval()
searcher = Searcher.get_simple_index(encoder.embedding_dim)
embeddings_x = []
embeddings_y = []
epoch_losses = []
for step, (x, y_pos, y_neg) in enumerate(test_loader):
x, y_pos, y_neg = x.cuda(), y_pos.cuda(), y_neg.cuda()
x_enc = encoder(x)
y_pos_enc = encoder(y_pos)
y_neg_enc = encoder(y_neg)
if loss_fn:
loss_val = loss_fn(x_enc, y_pos_enc, y_neg_enc)
epoch_losses.append(loss_val.item())
embeddings_x.append(x_enc)
embeddings_y.append(y_pos_enc)
print(' Test batch {} of {}'.format(step + 1, len(test_loader)), file=sys.stderr)
embeddings_x = torch.cat(embeddings_x, dim=0)
embeddings_y = torch.cat(embeddings_y, dim=0)
searcher.add(embeddings_x)
lookup = searcher.search(embeddings_y, 100)
correct_100 = sum(y in x for y, x in enumerate(lookup[1])) / len(lookup[1])
correct_50 = sum(y in x[:50] for y, x in enumerate(lookup[1])) / len(lookup[1])
correct_10 = sum(y in x[:10] for y, x in enumerate(lookup[1])) / len(lookup[1])
correct_1 = sum(y == x[0] for y, x in enumerate(lookup[1])) / len(lookup[1])
print(f'Test loss: {np.mean(epoch_losses):.4f}')
print(
'Test accuracy:\n top1 {}\n top10 {}\n top50 {}\n top100 {}'.format(
correct_1, correct_10, correct_50, correct_100))
if writer:
writer.add_scalars('Accuracy', {
'top1': correct_1,
'top10': correct_10,
'top50': correct_50,
'top100': correct_100,
}, global_step=epoch)
writer.add_scalar('Loss/test', np.mean(epoch_losses), global_step=epoch)
if epoch == -1 or epoch % 5 == 1:
mat = torch.cat([embeddings_x[:1000], embeddings_y[:1000]], dim=0)
labels = list(range(1000)) + list(range(1000))
writer.add_embedding(mat, labels, tag='Embeddings', global_step=epoch)
return correct_1 * 100, correct_100 * 100, lookup, embeddings_x, embeddings_y
def train_encoder(train_loader, test_loader, encoder, epochs=30):
LR = 0.01
summary(model=encoder, input_size=next(iter(train_loader))[0].shape[1:], device='cuda')
optimizer = torch.optim.AdamW(encoder.parameters(), lr=LR)
scheduler = torch.optim.lr_scheduler.ExponentialLR(optimizer, gamma=0.9)
epoch = 0
best_model = copy.deepcopy(encoder)
best_score = -1
if os.path.exists(CHECKPOINT_FILE):
print('Loading checkpoint')
epoch, best_score = load_train_state(CHECKPOINT_FILE, encoder, optimizer, scheduler)
log_file = os.path.join(LOGS_DIR, encoder.__class__.__name__)
writer = SummaryWriter(log_file, purge_step=epoch, flush_secs=60)
print("Learning started")
while epoch < epochs:
epoch += 1
print(f"Epoch: {epoch}")
epoch_losses = []
encoder.train()
margin = np.sqrt(encoder.embedding_dim) / 4
loss_fn = torch.nn.TripletMarginLoss(margin=margin)
for step, (x, y_pos, y_neg) in enumerate(train_loader):
x, y_pos, y_neg = x.cuda(), y_pos.cuda(), y_neg.cuda()
x_enc = encoder(x)
y_pos_enc = encoder(y_pos)
y_neg_enc = encoder(y_neg)
loss_val = loss_fn(x_enc, y_pos_enc, y_neg_enc)
optimizer.zero_grad()
loss_val.backward()
optimizer.step()
epoch_losses.append(loss_val.item())
print(' Batch {} of {} loss: {}, lr: {}'.format(step + 1, len(train_loader),
loss_val.item(), optimizer.param_groups[0]["lr"]),
file=sys.stderr)
print(f'Train loss: {np.mean(epoch_losses):.4f}')
writer.add_scalars('Loss', {
'train': np.mean(epoch_losses),
'lr': optimizer.param_groups[0]["lr"]
}, global_step=epoch)
score = evaluate_encoder(encoder, test_loader, loss_fn, writer=writer, epoch=epoch)[0]
if score > best_score:
best_model = copy.deepcopy(encoder)
best_score = score
print('New best score')
save_train_state(epoch, encoder, optimizer, scheduler, best_score, CHECKPOINT_FILE)
scheduler.step()
if best_score < 0:
best_score = evaluate_encoder(encoder, test_loader, writer=writer)[0]
writer.close()
save_file_path = os.path.join(SAVED_MODELS_PATH, '{}.{}.{:.2f}.pck'.format(encoder.__class__.__name__,
datetime.datetime.now().isoformat(),
best_score))
log_file_path = os.path.join(LOGS_DIR, '{}.{}.{:.2f}.pck'.format(encoder.__class__.__name__,
datetime.datetime.now().isoformat(),
best_score))
os.makedirs(os.path.dirname(save_file_path), exist_ok=True)
shutil.move(CHECKPOINT_FILE, save_file_path)
shutil.move(log_file, log_file_path)
return best_model, best_score
def train(large_model=False, epochs=30):
print('Training {} encoder for {} epochs'.format('large' if large_model else 'small', epochs))
train_data = AudioSamplePairDataset(root_path='data/fma_small_samples', test=False, large=large_model)
test_data = AudioSamplePairDataset(root_path='data/fma_small_samples', test=True, large=large_model)
train_loader = torch.utils.data.DataLoader(train_data,
batch_size=TRAIN_BATCH_SIZE,
shuffle=True,
num_workers=12,
pin_memory=True,
prefetch_factor=4)
test_loader = torch.utils.data.DataLoader(test_data,
batch_size=TEST_BATCH_SIZE,
shuffle=True,
num_workers=8,
pin_memory=True,
prefetch_factor=4)
encoder = LargeEncoder() if large_model else SmallEncoder()
encoder = encoder.cuda()
train_encoder(train_loader, test_loader, encoder, epochs)
if __name__ == "__main__":
warnings.simplefilter("ignore")
# noinspection PyTypeChecker
parser = argparse.ArgumentParser(description='Train models.',
formatter_class=argparse.ArgumentDefaultsHelpFormatter)
parser.add_argument('--model', type=str, choices=['small', 'large', 'both'], default='both',
help='Which model to train')
parser.add_argument('--epochs', type=int, default=30, help='How many epochs to train for')
args = parser.parse_args()
if args.model in ['small', 'both']:
train(False, args.epochs)
if args.model in ['large', 'both']:
train(True, args.epochs)