-
Notifications
You must be signed in to change notification settings - Fork 424
/
trainer.py
78 lines (58 loc) · 2.42 KB
/
trainer.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
import numpy as np
import torch
import time
class Trainer:
def __init__(self, model, optimizer, batch_size, get_batch, loss_fn, scheduler=None, eval_fns=None):
self.model = model
self.optimizer = optimizer
self.batch_size = batch_size
self.get_batch = get_batch
self.loss_fn = loss_fn
self.scheduler = scheduler
self.eval_fns = [] if eval_fns is None else eval_fns
self.diagnostics = dict()
self.start_time = time.time()
def train_iteration(self, num_steps, iter_num=0, print_logs=False):
train_losses = []
logs = dict()
train_start = time.time()
self.model.train()
for _ in range(num_steps):
train_loss = self.train_step()
train_losses.append(train_loss)
if self.scheduler is not None:
self.scheduler.step()
logs['time/training'] = time.time() - train_start
eval_start = time.time()
self.model.eval()
for eval_fn in self.eval_fns:
outputs = eval_fn(self.model)
for k, v in outputs.items():
logs[f'evaluation/{k}'] = v
logs['time/total'] = time.time() - self.start_time
logs['time/evaluation'] = time.time() - eval_start
logs['training/train_loss_mean'] = np.mean(train_losses)
logs['training/train_loss_std'] = np.std(train_losses)
for k in self.diagnostics:
logs[k] = self.diagnostics[k]
if print_logs:
print('=' * 80)
print(f'Iteration {iter_num}')
for k, v in logs.items():
print(f'{k}: {v}')
return logs
def train_step(self):
states, actions, rewards, dones, attention_mask, returns = self.get_batch(self.batch_size)
state_target, action_target, reward_target = torch.clone(states), torch.clone(actions), torch.clone(rewards)
state_preds, action_preds, reward_preds = self.model.forward(
states, actions, rewards, masks=None, attention_mask=attention_mask, target_return=returns,
)
# note: currently indexing & masking is not fully correct
loss = self.loss_fn(
state_preds, action_preds, reward_preds,
state_target[:,1:], action_target, reward_target[:,1:],
)
self.optimizer.zero_grad()
loss.backward()
self.optimizer.step()
return loss.detach().cpu().item()