poster-var/valid_encoder.py at main · lg2578/poster-var · GitHub

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
import torch
import warnings
from tqdm import tqdm
from utils.visualization import plot_confusion_matrix
from loss_encoder import compute_loss_em
from utils.metrics import Metrics
import swanlab
warnings.simplefilter("ignore")
@torch.no_grad()
def validloop(wandb,args, model,epoch, valid_loader,embeddings):
    torch.cuda.empty_cache()
    #记录每个mini batch的数据，用于计算一个epoch的数据，ids_epoch可以追踪错的最多的ids
    metrics=Metrics(args.num_classes,args.valid_metrics,args.device)
    batchs=len(valid_loader)
    loss_epoch=0

    model.eval()
    valid_pbar=tqdm(valid_loader,desc=f'----- {epoch},valid')
    for batch_id, (idxs, datas, labels,labels_em,image_path) in enumerate(valid_pbar):
        batch_size = labels.size(0)
        labels=labels.to(args.device)
        labels_em=labels_em.to(args.device)
        idxs=idxs.to(args.device)
        """ forward and calculate loss  """
        datas = datas.to(args.device)

        preds_em = model(datas)
        idxs=idxs.to(args.device)
        if epoch<args.debug_after['valid']:
            idxs=None
        loss_batch,pred_labels=compute_loss_em(args.loss_name,preds_em,embeddings,idxs,labels,datas,labels_em)
        loss_epoch+=loss_batch
        valid_pbar.set_postfix(loss=loss_batch.item())
        metrics.add_batch(pred_labels,labels,preds_em,labels_em)
    #logs
    return_metrics=metrics.compute_and_reset()
    loss_epoch=loss_epoch/batchs
    return_metrics['loss_epoch/valid']=loss_epoch
    return_metrics['epoch']=epoch
    wandb.log(return_metrics)
    swanlab.log(return_metrics)
    metric_name=f'{args.metric_name}/valid'
    return return_metrics[metric_name]