[Bug] Conv3Dtranspose default kernel layout should be IODHW

[rebel-jangys]

Hi, I found some bugs in TVM relay type relation of conv3dtranspose.

tvm/src/relay/op/nn/convolution.cc

Line 597 in fe3fa9d

static const Layout kOIDHW("OIDHW");

I believe the above line should be set to the IODHW layout, since the below conv3dtranspose kernel is made with the IODHW layout.

tvm/src/relay/op/nn/convolution.cc

Lines 629 to 630 in fe3fa9d

	Array<IndexExpr> wshape({dshape_ncdhw[1], indexdiv(param->channels, param->groups),
	param->kernel_size[0], param->kernel_size[1], param->kernel_size[2]});

Additionally, this line should use 'indexdiv(param->channels, param->groups)' like Conv2dTransposeRel instead of dividing the input channels by the number of groups.

tvm/src/relay/op/nn/convolution.cc

Line 663 in fe3fa9d

ICHECK(reporter->AssertEQ(indexdiv(dshape_ncdhw[1], param->groups), wshape[0]));

The weight of this test is treated as 'OIDHW' layout(the default kernel layout for conv3dtranspose is "OIDHW") despite having 10 input channels and 12 output channels.

tvm/tests/python/relay/test_op_level2.py

Lines 785 to 787 in fe3fa9d

	x = relay.var("x", relay.TensorType((n, c, d, h, w), "int8"))
	w = relay.var("w", relay.TensorType((10, 12, 3, 3, 3), "int8"))
	y = relay.nn.conv3d_transpose(x, w, out_dtype="int32")

Correct me if I'm wrong. Once again, thanks for your wonderful framework!

[masahi]

I believe the above line should be set to the IODHW

You are right.

Additionally, this line should use 'indexdiv(param->channels, param->groups)'

I think this line tries to compare the input channels. So the correct fix would be to remove the indexdiv.

We should also fix the followings:
https://github.com/apache/tvm/blob/main/include/tvm/relay/attrs/nn.h#L591
https://github.com/apache/tvm/blob/main/include/tvm/relay/attrs/nn.h#L433

tvm/python/tvm/relay/op/nn/nn.py

Line 439 in fd5915a

kernel_layout="OIDHW",

Can you send a PR to fix them?

[rebel-jangys]

@masahi Sure, I will send you a PR