aten._adaptive_avg_pool2d.default

[ayerofieiev-tt]

No description provided.

[swimdi]

Need to solved tenstorrent/tt-metal#12151 or the ttnn.avg_pool2d cannot support case of output_size != [1, 1]

input = torch.randn([1, 512, 7, 7]).to(torch.bfloat16)
output_size = [7, 7]
adaptive_avg_pool2d = torch.ops.aten._adaptive_avg_pool2d.default(input, output_size)
print(adaptive_avg_pool2d.shape)
# torch.Size([1, 512, 7, 7])

[swimdi]

Sorry the above comment is wrong, please ignore it

Here is the description of aten._adaptive_avg_pool2d.default:

aten._adaptive_avg_pool2d.default is not same with aten.avg_pool2d.default, but it can convert to aten.avg_pool2d.default by mapping aten._adaptive_avg_pool2d.default's output_size argument to aten.avg_pool2d.default's stride & size by

Stride = (input_size//output_size)  
Kernel size = input_size - (output_size-1)*stride  
Padding = 0

So if want to support lowering of aten._adaptive_avg_pool2d.default, then it needs either ttnn.adaptive_avg_pool2d op for direct converting or ttnn.avg_pool2d with above adjust

But seems there no ttnn.adaptive_avg_pool2d op now, and ttnn.avg_pool2d is not finished according to tenstorrent/tt-metal#12151

Luckly, there's only one input variation appeared on models

Tensor<[1, 512, 7, 7]> self = ?, List[int] output_size = [7, 7]

And this input variation actually let aten._adaptive_avg_pool2d.default do nothing, so if we check these input varation type during lowering and just remove it, then we can say aten._adaptive_avg_pool2d.default is fully supported of the current scope of models, so the supporting of actul lowering maybe can set lower priority