[Conv2DTransposed] Fix wrong shape check and add new TOPI module to support groups

[Lyken17]

This PR includes two bug fixs. The first one is the wrong shape check in src/relay/op/nn/convolution.h, the other is about enable groups > 1 in current relay generic as well as specializations.

Shape check

The default shape format of TVM is N x C x iH x iW for input and I x O x kH x kW for weight, a proper shape for Conv2dTransposed should be

• input: (batch, in_channels, iH, iW)
• weight: (in_channels, out_channels // groups, kH, kW)

Therefore the original channel checking.

ICHECK(reporter->AssertEQ(param->channels, wshape[1]))

is wrong. This should be wshape[0] rather than wshape[1].

Thus the original groups checking

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

is wrong. The proper comparison dimension should be wshape[1] rather than wshape[0].

Besides, the name for debug is not correct either. All logging information are using conv2d rather than conv2d_transposed, which is confusing.

Example to trigger error in current implementation

importnumpyasnpimporttvmfromtvmimportrelayfromtvmimportrelay, auto_schedulerfromtvm.relayimporttestingSEMVER='#[version = "0.0.5"]\n'defassert_graph_equal(lhs, rhs):
tvm.ir.assert_structural_equal(lhs, rhs, map_free_vars=True)
defroundtrip(expr):
x=tvm.parser.fromtext(expr.astext())
assert_graph_equal(x, expr)
# Testing Utilities for full modules.defparse_module(code):
mod=tvm.parser.parse(SEMVER+code)
roundtrip(mod)
returnmodprogram="""def @main(%input0: Tensor[(1, 32, 224, 224), float32],  %v0_0_weight: Tensor[(32, 1, 3, 3), float32]) -> Tensor[(1, 32, 224, 224), float32] { /* test comment */ %0 = nn.conv2d_transpose(%input0, %v0_0_weight, strides=[1, 1], padding=[1, 1, 1, 1], groups=32, channels=32, kernel_size=[3, 3]); %0}"""mod=parse_module(program)
print(mod)
# This module cannot be built yet. There are bugs in the topi needed to be fix as well. # target = "llvm"# lib = relay.build(mod, target=target, params=None)# print("build [fwd] pass successful")"""An error occurred during the execution of TVM.For more information, please see: https://tvm.apache.org/docs/errors.html--------------------------------------------------------------- Check failed: (false) is false: [18:09:32] /home/ligeng/Workspace/tvm/src/relay/op/nn/convolution.h:1111: ---------------------------------------------------------------An error occurred during the execution of TVM.For more information, please see: https://tvm.apache.org/docs/errors.html--------------------------------------------------------------- Check failed: (reporter->AssertEQ(indexdiv(dshape_nchw[1], param->groups), wshape[0])) is false: Conv2DTransposed: data.shape[1] // groups != weight.shape[0], data.shape= [1, 32, 224, 224] groups= 32 weight.shape= [32, 2, 3, 3]"""

Enable groups > 1 for Conv2dTranspose

This bug was originally raised in #8182 and #8799. Since the two issues are both about Conv2dTranspose, I merge them into one to better test.

In current tvm.relay.op.strategy.generic.py, the groups of Conv2dTranspose are forced to be one

@override_native_generic_func("conv2d_transpose_strategy")
def conv2d_transpose_strategy(attrs, inputs, out_type, target):
"""conv2d_transpose generic strategy"""
logger.warning("conv2d_transpose is not optimized for this platform.")
layout = attrs.data_layout
dilation = get_const_tuple(attrs.dilation)
groups = attrs.groups
assert layout == "NCHW", "only support nchw for now"
assert dilation == (1, 1), "not support dilate now"
assert groups == 1, "only support groups == 1 for now"

This PR aims to enlables groups > 1 for this OP on multiple platforms.

[junrushao]

Thank you much for the quick fix!! Would you like to include your minimal reproducible example as a unittest test case so that such regression could always be tested properly?

[Hzfengsy]

Please add a minimum test and fix the lint.

[Lyken17]

@Hzfengsy@junrushao1994 I've linted the code and added the test cases. However, current TOPI does not support groups in conv2dtranspose and I temporarilly mark the test as skip. This relies on another PR to fix #8799, will come back to this PR after that was fixed.

[junrushao]

Thank you all for the effort! @Lyken17 let's do the last mile, fixing the merge conflict and get the CI green :-)

[Lyken17]

@junrushao1994 resolved, please have a check.

[junrushao]

Looks good! Please fix the lint btw

[Lyken17]

Lint fixed. Let wait for the green CI!

[junrushao]

Thank you SO much @Lyken17@alicja-SiMa-ai@vinx13@AndrewZhaoLuo@Hzfengsy and many!