Batchsize's Change Causes Errors

1

I change the batchsize from 1 to 20, some errors caused.
I guess the reason is that the batchsize cannot be divisible by the size of dataloader, how can I fix it?

(258HLE0KS5WE9F}94A}AEP
(258HLE0KS5WE9F}94A}AEP1792×258 22.2 KB

Related code in train.py and models.py:

for epoch in range(opt.epoch, opt.n_epochs):
    for i, batch in enumerate(dataloader):


        # Set model input

        real_A = Variable(batch['A'].type(Tensor))
        real_B = Variable(batch['B'].type(Tensor))
        label_AB = Variable(batch['label_AB'].type(Tensor))
        label_BA = Variable(batch['label_BA'].type(Tensor))
        label = batch['label_name']

        real_A_features = vgg16(real_A)
        real_B_features = vgg16(real_B)

        realA_mean_std = mean_std(real_A_features)
        realB_mean_std = mean_std(real_B_features)

        weights_AB = metanet(realB_mean_std,label_AB)   #line206 in train.py
        weights_BA = metanet(realA_mean_std,label_BA)

def forward(self, mean_std_features,label):
        #mean_std_features = mean_std_features.view(-1,mean_std_features.shape[0] * mean_std_features.shape[1])
        #label_vector = torch.mean(label,dim=2).view(1,-1)
        index0 = torch.cuda.FloatTensor([torch.mean(label[0][0])])
        index1 = torch.cuda.FloatTensor([torch.mean(label[0][1])])
        index2 = torch.cuda.FloatTensor([torch.mean(label[0][2])])
        index3 = torch.cuda.FloatTensor([torch.mean(label[0][3])])
        #index4 = torch.cuda.FloatTensor([torch.mean(label[0][4])])
        label_vector = torch.cat((index0,index1,index2,index3)).view(1,-1)
        hidden = F.relu(self.hidden(torch.cat([mean_std_features,label_vector],dim=1)))   #line201 in models.py
        filters = {}
        for name, i in self.fc_dict.items():
            fc = getattr(self, 'fc{}'.format(i + 1))
            filters[name] = fc(hidden[:, i * 128:(i + 1) * 128])
        return filters
2

The error occurs at

torch.cat([mean_std_features,label_vector],dim=1))

From the error trace, I would guess the shapes are (batch_size,something) for label vector and (1,something) for mean_std_features. This implies that the the mean_std() is agnostic to changes in batch size. You would have to change the function to accomodate for different batch sizes. First, do verify this by printing out the shapes of the variables in question. Once you confirm, try tweaking the mean_std() to accmodate different batch sizes.