MCPcopy Create free account
hub / github.com/togethercomputer/OpenChatKit / test_loop

Function test_loop

training/dist_clm_train.py:21–69  ·  view source on GitHub ↗
(args, pipe, device, test_data_loader)

Source from the content-addressed store, hash-verified

19
20
21def test_loop(args, pipe, device, test_data_loader):
22
23 if test_data_loader is None:
24 return
25
26 print('testing starts.....')
27
28 pipe.model.eval()
29
30 if get_pipeline_parallel_rank() == args.pipeline_group_size - 1:
31
32 def _lm_pred_func(x, y):
33 loss_fct = torch.nn.CrossEntropyLoss(reduction='none')
34 logits = x[:, :-1, :].contiguous().float()
35 labels = y[:, 1:].contiguous()
36 loss = loss_fct(logits.transpose(-1, -2), labels).mean(1).detach().cpu()
37 return loss
38
39 loss_list = []
40 for i, data in enumerate(test_data_loader):
41
42 if args.evaluation_num_batch is not None and i >= args.evaluation_num_batch:
43 break
44
45 input_ids = data['input_ids'].to(device)
46 labels = input_ids.clone()
47 pipe.infer_iter(input_ids, labels, output_=loss_list, pred_func=_lm_pred_func)
48
49 loss = torch.tensor(loss_list).mean()
50 ppls = torch.exp(loss)
51 metric = {"valid.perplexity": ppls.item(), "valid.loss": loss.item()}
52
53 print(metric)
54 train_log(
55 metric,
56 step=pipe.global_step,
57 )
58
59 else:
60 for i, data in enumerate(test_data_loader):
61
62 if args.evaluation_num_batch is not None and i >= args.evaluation_num_batch:
63 break
64
65 input_ids = data['input_ids'].to(device)
66 labels = input_ids.clone()
67 current_iter_time = pipe.infer_iter(input_ids, labels)
68
69 pipe.model.train()
70
71
72

Callers 1

train_loopFunction · 0.70

Calls 3

train_logFunction · 0.85
infer_iterMethod · 0.80

Tested by

no test coverage detected