Xiandong Zhao

我测试的时候也是图片会变黑，请问现在有方法可以解决吗?

Thanks for the wonderful job. I only have trained model on VOC. https://pan.baidu.com/s/1L_TdCeQpXOFCy2P2X9sK0w 提取码: m5bw And I test mAP on VOC2007 [here](https://github.com/hustzxd/Yolo-v2-pytorch/blob/master/README.md#TODO).

I added the mixed-precision experiments. https://github.com/rhhc/ZeroQ-MP However, there are relatively large accuracy gaps with the results of the paper in some cases. So far, I have done several sets of...

There is currently no support, but you can add kernel-wise LSQ as follows: ```python w_reshape = self.weight.reshape([self.weight.shape[0], -1]).transpose(0, 1) # quantize w_reshape # ... w_q = w_reshape_q.transpose(0, 1).reshape(self.weight.shape).detach() + self.weight...

The accuracy of ResNet18-w4a4 is about 0.60. I think there may be some problems in the code or the hyperparameters used in the experiments is not suitable. Unluckily, I haven't...

Thank you for pointing out the problem, I will retrain the model. By the way, I think 70% is excellent for ResNet18-w4a4. I will share it with you if I...

Thank for your question. In the paper ``` s = gradscale(s, gradScaleFactor) Function 1 gradscale(x, scale): # x: Input tensor # scale: Scale gradient by this yOut = x yGrad...

Thanks for your questions. In the w4a4 case, I don't quantize the first and last layers. In the w8a8 case, I quantize the data into INT8 and activations after ReLU...

The hyper-parameters are shared in the README.md And maybe you can try to use weight decay for weight only. ```python def add_weight_decay(model, weight_decay, skip_keys): decay, no_decay = [], [] for...

Maybe you can find a better pretrained-model here. (https://pypi.org/project/pytorchcv/)