请问有支持别的batch size的代码吗

[guohanp]

请问有支持别的batch size的代码吗

[ArtoriaKawaii]

train.py and test.py each have the param --bsize which can be set, not sure this is what you're looking for?

[guohanp]

哦我看到代码里对比损失这写的只能是1的batch

[guohanp]

[guohanp]

这个代码里写的只能1的batch size啊

[pokaaa]

可以尝试一下这个新的函数： ` def get_contrastive_loss(self, style, mask, pred,style_comparison):

    style_feats = self.encode_with_small_intermediate(style)
    fine_feats = self.encode_with_small_intermediate(pred)

    loss_contra=[]
    for j in range(style.shape[0]):
        style_comparison_cur=style_comparison[j]
        style_comparison_feats=[]
        for i in range(self.compare_num):
            style_comparison_feats.append(self.encode_with_small_intermediate(style_comparison_cur[i:i+1]))
        style_comparison_feats=torch.cat(style_comparison_feats,dim=0)  
        cur_loss_contra=self.calc_contrastive_loss(style_feats[j].unsqueeze(0),style_comparison_feats,fine_feats[j].unsqueeze(0),mask[j].unsqueeze(0))
        loss_contra.append(cur_loss_contra)

    loss_contra=torch.mean(torch.stack(loss_contra,dim=0))

    return loss_contra`

感谢您的提醒，我们已经在库里进行了更新，如果使用中有任何问题欢迎再进行反馈，我们会及时跟进。