-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathtrain.py
217 lines (186 loc) · 10.3 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
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
from pathlib import Path
from datetime import datetime
import torch
from torchmetrics.classification import BinaryF1Score
from torcheval.metrics.functional import perplexity
from tqdm import tqdm
import hydra
from omegaconf import DictConfig, OmegaConf
import wandb
from dataset import OsuDataset, collate
from models import OsuModel, ControlModel, OsuModelNpPred
from loss import binary_focal_loss, multi_focal_loss
DEV = 'cpu'
if torch.cuda.is_available():
DEV = 'cuda'
else:
print('Warning: Model running on CPU.')
class Trainer():
def __init__(self, model, optimizer, train_loader,
valid_loader, device, experiment_config, hyperparams):
self.model = model
self.optimizer = optimizer
self.train_loader = train_loader
self.valid_loader = valid_loader
self.device = device
self.checkpoint_path = Path(experiment_config.checkpoint_path)
self.checkpoint_path.mkdir(exist_ok=True)
self.np_fl_gamma = hyperparams.np_fl_gamma
self.np_fl_weight = hyperparams.np_fl_weight
self.ns_fl_gamma = hyperparams.ns_fl_gamma
self.ns_fl_weight = hyperparams.ns_fl_weight
self.np_loss_multiplier = 7
self.start_epoch = 0
self.f1 = BinaryF1Score().to(self.device)
def load_checkpoint(self, fn):
checkpoint = torch.load(fn, map_location=self.device)
self.start_epoch = checkpoint['epoch'] + 1
self.model.load_state_dict(checkpoint['model_state_dict'])
self.optimizer.load_state_dict(checkpoint['optimizer_state_dict'])
for state in self.optimizer.state.values():
for k, v in state.items():
if (torch.is_tensor(v)):
state[k] = v.to(self.device)
def save_checkpoint(self, epoch, fn):
checkpoint = {
'epoch': epoch,
'model_state_dict': self.model.state_dict(),
'optimizer_state_dict': self.optimizer.state_dict()
}
torch.save(checkpoint, fn)
def run(self, experiment_config, hyperparams):
if experiment_config.log_to_wandb:
wandb.init(project='AutoOsu',
name=experiment_config.run_name, config=hyperparams)
self.model.to(self.device)
for epoch in tqdm(range(self.start_epoch, self.start_epoch + hyperparams.num_epochs)):
if experiment_config.train:
self.model.train()
for batch in tqdm(self.train_loader, leave=False):
specs, beat_phases, beat_nums, difficulties, onsets, actions = batch
specs = specs.to(self.device)
beat_phases = beat_phases.to(self.device)
beat_nums = beat_nums.to(self.device)
difficulties = difficulties.to(self.device)
onsets = onsets.to(self.device)
actions_gt = actions[:, 1:].to(self.device)
actions_shifted = actions[:, :-1].to(self.device)
np_pred, ns_logit = self.model(
specs, beat_phases, beat_nums, difficulties, actions_shifted)
np_pred = torch.reshape(np_pred, [-1])
np_label = torch.reshape(onsets, [-1])
ns_pred = torch.reshape(
ns_logit, [-1, ns_logit.shape[-1]]).softmax(dim=-1)
ns_label = torch.reshape(actions_gt, [-1])
np_loss = binary_focal_loss(
np_label, np_pred, self.np_fl_gamma, self.np_fl_weight)
ns_loss = multi_focal_loss(
ns_label, ns_pred, self.ns_fl_gamma, self.ns_fl_weight)
batch_loss = np_loss * self.np_loss_multiplier + ns_loss
self.optimizer.zero_grad()
batch_loss.backward()
self.optimizer.step()
ns_acc = (ns_pred.argmax(dim=-1) ==
ns_label).float().mean()
if experiment_config.log_to_wandb:
wandb.log({'train_np_loss': np_loss.item(),
'train_ns_loss': ns_loss.item(),
'train_loss': batch_loss.item(),
'train_acc': ns_acc.item(),
'train_np_f1': self.f1(np_pred, np_label.int()).item()})
if experiment_config.validate:
self.model.eval()
with torch.inference_mode():
valid_np_loss_sum = 0
valid_ns_loss_sum = 0
valid_loss_sum = 0
valid_np_f1_sum = 0
valid_ns_acc_sum = 0
valid_ns_onset_acc_sum = 0
valid_ns_ppl_sum = 0
for batch in tqdm(self.valid_loader, leave=False):
specs, beat_phases, beat_nums, difficulties, onsets, actions = batch
specs = specs.to(self.device)
beat_phases = beat_phases.to(self.device)
beat_nums = beat_nums.to(self.device)
difficulties = difficulties.to(self.device)
onsets = onsets.to(self.device)
actions_gt = actions[:, 1:].to(self.device)
actions_shifted = actions[:, :-1].to(self.device)
np_pred, ns_logit = self.model(
specs, beat_phases, beat_nums, difficulties, actions_shifted)
np_pred = torch.reshape(np_pred, [-1])
np_label = torch.reshape(onsets, [-1])
ns_pred = torch.reshape(
ns_logit, [-1, ns_logit.shape[-1]]).softmax(dim=-1)
ns_label = torch.reshape(actions_gt, [-1])
np_loss = binary_focal_loss(
np_label, np_pred, self.np_fl_gamma, self.np_fl_weight) * self.np_loss_multiplier
ns_loss = multi_focal_loss(
ns_label, ns_pred, self.ns_fl_gamma, self.ns_fl_weight)
batch_loss = np_loss + ns_loss
valid_np_loss_sum += np_loss.item()
valid_ns_loss_sum += ns_loss.item()
valid_loss_sum += batch_loss.item()
ns_acc = (ns_pred.argmax(dim=-1) ==
ns_label).float().mean()
is_onset = onsets.flatten() == 1
valid_ns_onset_acc_sum += (ns_pred[is_onset].argmax(
dim=-1) == ns_label[is_onset]).float().mean()
valid_ns_acc_sum += ns_acc.item()
valid_np_f1_sum += self.f1(np_pred,
np_label.int()).item()
valid_ns_ppl_sum += perplexity(ns_logit, actions_gt, ignore_index=0).item()
if experiment_config.log_to_wandb:
wandb.log({'valid_np_loss': valid_np_loss_sum / len(self.valid_loader),
'valid_ns_loss': valid_ns_loss_sum / len(self.valid_loader),
'valid_loss': valid_loss_sum / len(self.valid_loader),
'valid_acc': valid_ns_acc_sum / len(self.valid_loader),
'valid_ns_onset_acc': valid_ns_onset_acc_sum / len(self.valid_loader),
'valid_np_f1': valid_np_f1_sum / len(self.valid_loader),
'valid_ns_ppl': valid_ns_ppl_sum / len(self.valid_loader)})
if experiment_config.print_metrics:
print(f'Model: {experiment_config.model}')
print(f'valid_np_loss: {valid_np_loss_sum / len(self.valid_loader)}')
print(f'valid_ns_loss: {valid_ns_loss_sum / len(self.valid_loader)}')
print(f'valid_loss: {valid_loss_sum / len(self.valid_loader)}')
print(f'valid_acc: {valid_ns_acc_sum / len(self.valid_loader)}')
print(f'valid_ns_onset_acc: {valid_ns_onset_acc_sum / len(self.valid_loader)}')
print(f'valid_np_f1: {valid_np_f1_sum / len(self.valid_loader)}')
print(f'valid_ns_ppl: {valid_ns_ppl_sum / len(self.valid_loader)}')
if experiment_config.save_checkpoint:
time = datetime.now().strftime('%m-%d-%H-%M-%S')
checkpoint_path = Path(
self.checkpoint_path / f'{experiment_config.model}-{time}-epoch{epoch}.pt')
self.save_checkpoint(epoch, checkpoint_path)
if experiment_config.log_to_wandb:
wandb.finish()
@hydra.main(version_base=None, config_path='.', config_name='train_config')
def main(config: DictConfig):
experiment_config = config.experiment_config
hyperparams = config.hyperparams
base_set = OsuDataset(Path(experiment_config.beatmap_dir),
Path(experiment_config.audio_dir))
generator = torch.Generator().manual_seed(experiment_config.random_seed)
train_set, valid_set = torch.utils.data.random_split(
base_set, [0.95, 0.05], generator)
train_loader = torch.utils.data.DataLoader(
train_set, batch_size=hyperparams.batch_size, shuffle=True, generator=generator, collate_fn=collate, drop_last=False)
valid_loader = torch.utils.data.DataLoader(
valid_set, batch_size=12, shuffle=False, collate_fn=collate, drop_last=False)
if experiment_config.model == 'default':
model = OsuModel(hyperparams)
elif experiment_config.model == 'control':
model = ControlModel(hyperparams)
elif experiment_config.model == 'np_pred':
model = OsuModelNpPred(hyperparams)
print("Using NP Pred Model...")
optimizer = torch.optim.Adam(
model.parameters(), lr=hyperparams.learning_rate)
trainer = Trainer(model, optimizer, train_loader, valid_loader,
DEV, experiment_config, hyperparams)
if experiment_config.load_checkpoint is not None:
trainer.load_checkpoint(Path(experiment_config.load_checkpoint))
trainer.run(experiment_config, hyperparams)
if __name__ == '__main__':
main()