Ping Chao
Ping Chao
Hi, this repo is very helpful for converting the model quickly: https://github.com/NVIDIA-AI-IOT/torch2trt
Hi, Please refer to https://github.com/sovrasov/flops-counter.pytorch for flops counting. For CIO, you can simply add a global counter in the forward pass of ConvLayer module for both input and output tensors....
Hello, thank you very much for the report. It seems like there is an accidentally deleted line for converting RGB->BGR in the last pull request. I have fixed it, please...
Hi, thank you for the feedback. They are equivalent in math while HarDBlock_v2 reduces the use of concat such that it is a little bit faster. It first decomposes convolutions...
Hello, it is mIoU stands for mean IoU
Hi, please clone the repo with --recursive. You can also clone https://github.com/MIPT-Oulu/pytorch_bn_fusion.git under FCHarDNet/
Hi, the mean/stddev is aligned with imageNet since the backbone model was pretrained with imageNet. You can still change the normalization method whatever you want. The training procedure can recover...
We use native pytorch 1.0.1 with cuda 9.2 for the speed test. You should be able to reproduce a similar speed on 0.4.0 or versions above 1.2.0 with cuda 10.1....
We actually didn't fully train the petite model with ImageNet. The training was early stopped at around 100th epoch out of 150 epochs, and unfortunately, we didn't record the accuracy.
Hi, yes we trained it with ImageNet for the first 100 epochs out of a totally 150 epochs of cosine learning rate schedule.