batch size > 1 (per GPU)

[varun19299]

Hi,

Is it possible to use a batch-size (per GPU) larger than 1?

The model's forward pass currently contains statements such as,

Shift-Net/basicsr/models/archs/gshift_deblur2.py

Line 750 in 816f1b2

x = x[0]

.

It seems that stage0 and stage2 should be amenable to batch processing by a simple rearrangement:
x_rearranged = rearrange(x, "b f c h w -> (b f) c h w").

How about stage_1? What modifications are needed for it to operate on a batchsize > 1?

[dasongli1]

Suppose the batch size is B and the batch size of input tensor is B, N, C, H, W = x.shape.
For frame-wise processing, x = x.view(B*N, C, H, W) -> 2D CNN processing.
For y in stage 2, y = x.view(B,N*C,H,W), y = torch.roll(y, slice_c,1), y = y.reshape(B*N, C, H, W)

[hiredd]

Suppose the batch size is B and the batch size of input tensor is B, N, C, H, W = x.shape. For frame-wise processing, x = x.view(B*N, C, H, W) -> 2D CNN processing. For y in stage 2, y = x.view(B,N*C,H,W), y = torch.roll(y, slice_c,1), y = y.reshape(B*N, C, H, W)
Can you explain it in more detail? thanks ~