Add group_conv2d_transpose_nchw to support groups argument

[Lyken17]

As mentioned in #8182, TranposedConv2d is an important operator in GAN related applications and groups should also be supported for this operator. This PR was set to implement this missing feature.

However, there are parts I am still confusing

• Which names should be used for the topi function?
  In topi.nn.conv2d, there are data, kernel, strides, padding, out_dtype, output_padding as well as Input, Filter, strides, padding, out_dtype, output_padding, which one should be used for newly added function?
• Should we add a new function, or extend existing ones?
  In most DL frameoworks (e.g., PyTorch), conv2d is a unified function with support of various arguments such as padding, dilation, groups. But in topi, conv2d (w/o groups) and group_conv2d are two differnet function. While I understand this might be important to backward compability, I would recommend to merge these functions for simplicty.

Please comment if you have any thoughts. I will prepare unit tests after the discussion.

[vinx13]

Merging these functions sounds good to me. Note that we have unified operators in Relay level, it might be fine to lower to different topi operators if this is simpler

[Lyken17]

@vinx13 What is relationship between topi.nn and relay.nn?

[vinx13]

relay.nn is graph level operator, topi.nn is the platform-aware implementation of each operator. There is a lowering rule registered here

tvm/python/tvm/relay/op/strategy/generic.py

Line 249 in 596ff5c

defconv2d_strategy(attrs, inputs, out_type, target):

[vinx13]

@Lyken17 could you add a test case?

[Lyken17]

As mentioned in #8182, TranposedConv2d is an important operator in GAN related applications and groups should also be supported for this operator. This PR was set to implement this missing feature.

However, there are parts I am still confusing

• Which names should be used for the topi function?
  In topi.nn.conv2d, there are data, kernel, strides, padding, out_dtype, output_padding as well as Input, Filter, strides, padding, out_dtype, output_padding, which one should be used for newly added function?
• Should we add a new function, or extend existing ones?
  In most DL frameoworks (e.g., PyTorch), conv2d is a unified function with support of various arguments such as padding, dilation, groups. But in topi, conv2d (w/o groups) and group_conv2d are two differnet function. While I understand this might be important to backward compability, I would recommend to merge these functions for simplicty.

Please comment if you have any thoughts. I will prepare unit tests after the discussion.

[vinx13]

Merging these functions sounds good to me. Note that we have unified operators in Relay level, it might be fine to lower to different topi operators if this is simpler

[Lyken17]

@vinx13 What is relationship between topi.nn and relay.nn?

[vinx13]

relay.nn is graph level operator, topi.nn is the platform-aware implementation of each operator. There is a lowering rule registered here

tvm/python/tvm/relay/op/strategy/generic.py

Line 249 in 596ff5c

defconv2d_strategy(attrs, inputs, out_type, target):

[vinx13]

@Lyken17 could you add a test case?

[Lyken17]

As mentioned in #8182, TranposedConv2d is an important operator in GAN related applications and groups should also be supported for this operator. This PR was set to implement this missing feature.

However, there are parts I am still confusing

• Which names should be used for the topi function?
  In topi.nn.conv2d, there are data, kernel, strides, padding, out_dtype, output_padding as well as Input, Filter, strides, padding, out_dtype, output_padding, which one should be used for newly added function?
• Should we add a new function, or extend existing ones?
  In most DL frameoworks (e.g., PyTorch), conv2d is a unified function with support of various arguments such as padding, dilation, groups. But in topi, conv2d (w/o groups) and group_conv2d are two differnet function. While I understand this might be important to backward compability, I would recommend to merge these functions for simplicty.

Please comment if you have any thoughts. I will prepare unit tests after the discussion.

[vinx13]

Merging these functions sounds good to me. Note that we have unified operators in Relay level, it might be fine to lower to different topi operators if this is simpler

[Lyken17]

@vinx13 What is relationship between topi.nn and relay.nn?

[vinx13]

relay.nn is graph level operator, topi.nn is the platform-aware implementation of each operator. There is a lowering rule registered here

tvm/python/tvm/relay/op/strategy/generic.py

Line 249 in 596ff5c

defconv2d_strategy(attrs, inputs, out_type, target):

[vinx13]

@Lyken17 could you add a test case?

[Lyken17]

As mentioned in #8182, TranposedConv2d is an important operator in GAN related applications and groups should also be supported for this operator. This PR was set to implement this missing feature.

However, there are parts I am still confusing

• Which names should be used for the topi function?
  In topi.nn.conv2d, there are data, kernel, strides, padding, out_dtype, output_padding as well as Input, Filter, strides, padding, out_dtype, output_padding, which one should be used for newly added function?
• Should we add a new function, or extend existing ones?
  In most DL frameoworks (e.g., PyTorch), conv2d is a unified function with support of various arguments such as padding, dilation, groups. But in topi, conv2d (w/o groups) and group_conv2d are two differnet function. While I understand this might be important to backward compability, I would recommend to merge these functions for simplicty.

Please comment if you have any thoughts. I will prepare unit tests after the discussion.

[vinx13]

Merging these functions sounds good to me. Note that we have unified operators in Relay level, it might be fine to lower to different topi operators if this is simpler

[Lyken17]

@vinx13 What is relationship between topi.nn and relay.nn?

[vinx13]

relay.nn is graph level operator, topi.nn is the platform-aware implementation of each operator. There is a lowering rule registered here

tvm/python/tvm/relay/op/strategy/generic.py

Line 249 in 596ff5c

defconv2d_strategy(attrs, inputs, out_type, target):

[vinx13]

@Lyken17 could you add a test case?

[Lyken17]

As mentioned in #8182, TranposedConv2d is an important operator in GAN related applications and groups should also be supported for this operator. This PR was set to implement this missing feature.

However, there are parts I am still confusing

• Which names should be used for the topi function?
  In topi.nn.conv2d, there are data, kernel, strides, padding, out_dtype, output_padding as well as Input, Filter, strides, padding, out_dtype, output_padding, which one should be used for newly added function?
• Should we add a new function, or extend existing ones?
  In most DL frameoworks (e.g., PyTorch), conv2d is a unified function with support of various arguments such as padding, dilation, groups. But in topi, conv2d (w/o groups) and group_conv2d are two differnet function. While I understand this might be important to backward compability, I would recommend to merge these functions for simplicty.

Please comment if you have any thoughts. I will prepare unit tests after the discussion.

[vinx13]

Merging these functions sounds good to me. Note that we have unified operators in Relay level, it might be fine to lower to different topi operators if this is simpler

[Lyken17]

@vinx13 What is relationship between topi.nn and relay.nn?

[vinx13]

relay.nn is graph level operator, topi.nn is the platform-aware implementation of each operator. There is a lowering rule registered here

tvm/python/tvm/relay/op/strategy/generic.py

Line 249 in 596ff5c

defconv2d_strategy(attrs, inputs, out_type, target):

[vinx13]

@Lyken17 could you add a test case?

[Lyken17]

As mentioned in #8182, TranposedConv2d is an important operator in GAN related applications and groups should also be supported for this operator. This PR was set to implement this missing feature.

However, there are parts I am still confusing

• Which names should be used for the topi function?
  In topi.nn.conv2d, there are data, kernel, strides, padding, out_dtype, output_padding as well as Input, Filter, strides, padding, out_dtype, output_padding, which one should be used for newly added function?
• Should we add a new function, or extend existing ones?
  In most DL frameoworks (e.g., PyTorch), conv2d is a unified function with support of various arguments such as padding, dilation, groups. But in topi, conv2d (w/o groups) and group_conv2d are two differnet function. While I understand this might be important to backward compability, I would recommend to merge these functions for simplicty.

Please comment if you have any thoughts. I will prepare unit tests after the discussion.

[vinx13]

Merging these functions sounds good to me. Note that we have unified operators in Relay level, it might be fine to lower to different topi operators if this is simpler

[Lyken17]

@vinx13 What is relationship between topi.nn and relay.nn?

[vinx13]

relay.nn is graph level operator, topi.nn is the platform-aware implementation of each operator. There is a lowering rule registered here

tvm/python/tvm/relay/op/strategy/generic.py

Line 249 in 596ff5c

defconv2d_strategy(attrs, inputs, out_type, target):

[vinx13]

@Lyken17 could you add a test case?

[Lyken17]

As mentioned in #8182, TranposedConv2d is an important operator in GAN related applications and groups should also be supported for this operator. This PR was set to implement this missing feature.

However, there are parts I am still confusing

• Which names should be used for the topi function?
  In topi.nn.conv2d, there are data, kernel, strides, padding, out_dtype, output_padding as well as Input, Filter, strides, padding, out_dtype, output_padding, which one should be used for newly added function?
• Should we add a new function, or extend existing ones?
  In most DL frameoworks (e.g., PyTorch), conv2d is a unified function with support of various arguments such as padding, dilation, groups. But in topi, conv2d (w/o groups) and group_conv2d are two differnet function. While I understand this might be important to backward compability, I would recommend to merge these functions for simplicty.

Please comment if you have any thoughts. I will prepare unit tests after the discussion.

[vinx13]

Merging these functions sounds good to me. Note that we have unified operators in Relay level, it might be fine to lower to different topi operators if this is simpler

[Lyken17]

@vinx13 What is relationship between topi.nn and relay.nn?

[vinx13]

relay.nn is graph level operator, topi.nn is the platform-aware implementation of each operator. There is a lowering rule registered here

tvm/python/tvm/relay/op/strategy/generic.py

Line 249 in 596ff5c

defconv2d_strategy(attrs, inputs, out_type, target):

[vinx13]

@Lyken17 could you add a test case?