EfficientNet-PyTorch the inference speed is much slower than original TensorFlow code

hi，I ran your sample code and then tested the inference time，but I find that the inference speed is much slower than original TensorFlow code on my computer。

Jun 06 '19 08:06 tensorflower

hi，I ran your sample code and then tested the inference time，but I find that the inference speed is much slower than original TensorFlow code on my computer。

Jun 06 '19 08:06 tensorflower

Hello,

Yes, this is expected because grouped convolutions in PyTorch are slow (the core devs are working on making them faster). See pytorch/pytorch#18631.

Jun 06 '19 16:06 lukemelas

hi, since the speed is too slow, is that can be modified the implementation from " self._depthwise_conv = Conv2dSamePadding( in_channels=oup, out_channels=oup, groups=oup, # groups makes it depthwise kernel_size=k, stride=s, bias=False)"

to " self._depthwise_conv_in = Conv2dSamePadding( in_channels=oup, out_channels=1, groups=1, # groups makes it depthwise kernel_size=1, stride=1, bias=False) self._depthwise_conv = Conv2dSamePadding( in_channels=1, out_channels=oup, groups=1, # groups makes it depthwise kernel_size=k, stride=s, bias=False) " thanks

Jun 16 '19 07:06 semchan

Now with #44 you can export to ONNX. That may help in terms of inference speed, as it actually compiles a graph.

Jun 30 '19 00:06 lukemelas

@semchan I set groups=1, and the size of the weight file changed from ～30m to ～900m. It is unacceptable.

@lukemelas The TensorRT engine file form onnx still slow than the engine file from tf model, it cost ~10ms in my task and tf model cost ~5ms. It seems the onnx graph will affect the build method of TensorRT engine file.