Pytorch's ConvTranspose2d supported in DPU but accuracy degrades significantly. Is it expected?
Opened this issue · 0 comments
abdulazizm commented
Faced issue during compilation using Apache TVM flow. Below code helps to have conv2d_transpose onto mod['tvmgen_default_vitis_ai_main_0'] but accuracy degrades significantly (maybe due to unexpected weight transpose happens during desired layout transformation). It works great if we don't have conv2d_transpose in desired layout dict or force to have 'NCHW' layout
desired_layouts = {
'nn.conv2d': ['NHWC', 'default'],
'image.resize2d': ['NHWC'],
'nn.max_pool2d':['NHWC'],
'nn.conv2d_transpose': ['NCHW','default'],
}
Tried 'nn.conv2d_transpose': ['NCHW','OIHW']
& 'nn.conv2d_transpose': ['NCHW','HWIO']
but no luck. And not sure if it is a known bug