search

ModelTC / MQBench

Model Quantization Benchmark
Apache License 2.0
766 stars 140 forks source link

The `ch_axis` of ConvTransposeReLU2d cannot be set as 1 during the PTQ #113

Closed sherlockhz1415 closed 2 years ago

sherlockhz1415 commented 2 years ago

TL;DR: the ch_axis of ConvTransposeReLU2d cannot be set as 1 during the PTQ, which will cause errors.

During the PTQ, if we choose the per-channel mode, we are supposed to set the ch_axis of ConvTranspose as 1 otherwise we'll face errors.

According to: https://github.com/ModelTC/MQBench/blob/e2175203c8e62596e66500a720a6cb1d1fc1dacd/mqbench/nn/qat/modules/deconv.py#L21-L23 The ch_axis will be automatically converted to 1 if the input ch_axis is not -1, i.e., if we set ch_axis = 0 in configs, it will still be 1 in ConvTranspose.

While in the __init__ of ConvTransposeReLU2d: https://github.com/ModelTC/MQBench/blob/e2175203c8e62596e66500a720a6cb1d1fc1dacd/mqbench/nn/intrinsic/qat/modules/deconv_fused.py#L372-L386 The ch_axis will be overridden by the input qconfig, the above automatic conversion of ch_axis is not working.

Please correct me if I'm wrong: shall we consider deleting L385 and L386?

Thanks!

Tracin commented 2 years ago

You are correct, just delete deconv_fused.py:385-386 will work. Look forward to your MR!