Batchsize

[CharlesAntoineParent]

Hello, I was wondering why it's not possible to have bigger batch. Like if you just do two forward propagation in the JL phase and combine the result in the CM block ?

[kerenfu]

Hello, I was wondering why it's not possible to have bigger batch. Like if you just do two forward propagation in the JL phase and combine the result in the CM block ?

Indeed yes, you can manage this issue in the code level. By using bigger batch, the code should be modified to distinguish different samples in the batch and avoid mixing them up in the CM block. As long as the samples are distinguished properly, they can then form a new batch after the CM block and compute loss with a batch of GT. To realize this, certain modification should be made to the CM block.

[jiangyao-scu]

Since RGB images and Depth maps are first loaded as batch data and then connected in the batch before entering the network, they can be split in half in the batch dimension to distinguish different samples. The ``forward'' function of the CM module (the CMLayer in JL_DCF.py) can be modified as follows:

    def forward(self, list_x):
        b, c, h, w = list_x[0].size()
        batch_size = int(b/2)
        resl = []
        for i in range(len(list_x)):
            part1 = list_x[i][0:batch_size]
            part2 = list_x[i][batch_size:]
            sum = (part1 + part2 + (part1 * part2))
            resl.append(sum)
        return resl

In this way, you can enable bigger batch.