[FRONTEND][TF] conv2d_transpose 'SAME' support kernel more than 1x1

[optima2005]

This is to support 'SAME' padding for conv3d_transpose for kernel more than 1x1

Discussions:
https://discuss.tvm.ai/t/why-we-only-support-kernel-1-1-for-tf-conv2d-transpose-same/4957

@yongwww @apivovarov

[apivovarov]

Can you also add test case for

kernel 2x2, strides 2x2, SAME
kernel 3x3, strides 2x2, SAME

E.g. if input is 5x5 then valid outputs are 9x9 or 10x10 (you can use one or another in the output_shape tensor) regardless of the kernel size.

[optima2005]

The strides in transpose convolution is used to dilate the input, see this. If the input is dilated, there would be no way to get 'SAME' size of output by only padding.

[optima2005]

ok. I can get your point. 'SAME' means the size of the enlarged input size(by dilation).

[apivovarov]

TF code for kernel 2x2, strides 2x2 and padding SAME is

dshape=(1,5,5,4)
#hwoi
kshape=(2,2,2,4)
oshape=(1,9,9,2)
# or 
oshape=(1,10,10,2)
dtype='float32'
with tf.Session() as sess:
    x = tf.placeholder(shape=dshape, dtype=dtype)
    w = tf.placeholder(shape=kshape, dtype=dtype)
    dc = tf.nn.conv2d_transpose(x, w, output_shape=oshape, strides=(1,2,2,1), padding='SAME')

    res_dc = sess.run(dc, feed_dict={x: data, w:weight})

[apivovarov]

minor formatting comment. I think we do not need spaces after ( and before )

[optima2005]

I will fix it.

[apivovarov]

Formulas are very similar except of using

2 * param->padding[0]  vs  param->padding[0] + param->padding[2]
and 
2 * param->padding[1]   vs   param->padding[1] + param->padding[3]

Can we calculate correct padding and then use the same formulas with calculated padding for both cases?

[optima2005]

The additional condition is to handle the head padding and tail padding are diffenct, for even kernel in this case. I didn't quite understand your point. Please clearify.

[apivovarov]

What about the following?

  int pad_h, pad_w;
  if (param->padding.size() == 2) {
    pad_h = 2 * param->padding[0];
    pad_w = 2 * param->padding[1];
  } else if (param->padding.size() == 4) {
    pad_h = param->padding[0] + param->padding[2];
    pad_w = param->padding[1] + param->padding[3];
  } else {
    CHECK_EQ(param->padding.size(), 4);
  }
  oshape.Set(2, (param->strides[0] * (dshape_nchw[2] - 1) + dilated_ksize_y -
                 pad_h + param->output_padding[0]));
  oshape.Set(3, (param->strides[1] * (dshape_nchw[3] - 1) + dilated_ksize_x -
                 pad_w + param->output_padding[1]));

[optima2005]

This is better. Thanks!

[FrozenGene]

@optima2005 do you mind taking up this RFC? #2682 Because I see you start to support 4D padding in Conv2dTranspose. I think the work should be similar with convolution, which could make us avoid pad operator and get potential performance.

[optima2005]

@FrozenGene let me do some evaluations first and then I will get back to you if I think I could do it. Actually I am still in the progess to understand those ops' implementation, especially those optimazation approach.

[optima2005]

@apivovarov I added the test cases per as your proposal and tried to pass them. But one condition (10x10 in your example) still not working in cuda target. I haven't found the root cause. I am wondering whether you could help.

[optima2005]

I got this error on cuda target with test case:

        _test_convolution('conv_transpose', [4, 32, 8, 8], [3, 3, 176, 32], [1, 1], [2, 2], 'SAME',
                          'NCHW', [4, 176, 16, 16])

incubator-tvm/src/pass/loop_partition.cc:544: Cannot prove: ((((floordiv(((((floordiv((floormod(dh, 2) + 7), 2)16) + 16)(floordiv((floormod(dw, 2) + 7), 2) + 1)) + 63), 64) - 1) - ((((floordiv((floormod(dh, 2) + 7), 2)16) + 16)(floordiv((floormod(dw, 2) + 7), 2) + 1)) - (floordiv(((((floordiv((floormod(dh, 2) + 7), 2)16) + 16)(floordiv((floormod(dw, 2) + 7), 2) + 1)) + 63), 64)*63))) + 1) >= 0), when generating the post doubt loop

Is it relative to issue #4470? @apivovarov

[optima2005]

This can be confirmed that after I add exclusion for this specific case in workaround #4472, that test case can pass.

[tqchen]

please comment about the status of the PR, @FrozenGene

[FrozenGene]

How about this?

[optima2005]

The case failed due to the "Cannot prove" error when compiling the module. see #4470. It seems there is no fix for it yet.

[yongwww]

What is the reason why it doen't work for cuda

[optima2005]

The case failed due to the "Cannot prove" error when compiling the module. see #4470. It seems there is no fix for it yet.

[optima2005]

For stride '2x2', I tried kernel size up to '7x7', all has the "Cannot prove" compiling error for cuda target. So I added the walkaround. for other strides > '2x2', I didn't find this error.
@yongwww @FrozenGene @apivovarov Would you please take a look again? Thanks!

[yongwww]

LGTM

[optima2005]

@yzhliu, would you mind to take a look? It is the fix to 'SAME' mode of #4300
Thanks!

[tqchen]

tHANKS @optima2005 @yongwww @FrozenGene @apivovarov