Peng Chen
I am trying to only quantize the last three fc layers and keep conv layers by set layer level lr to be 0. Fine-tuning only happens on FC layers. Wish...
Hi, @Zhouaojun I met some problem when re-implement your experiments as described above. The accuracy seemed not re-gain after quantization. I wonder if it is convenient for you to share...
@Zhouaojun Thanks for the comment of employment of the mask variable. However, would it be more nice if we revise the 235 line in src/caffe/solvers/sgd_solver.cpp as following: ` caffe_gpu_mul(net_params[param_id]->count(),net_params[param_id]->gpu_mask(),history_[param_id]->mutable_gpu_diff(),net_params[param_id]->mutable_gpu_diff()); `...
Again, we might move the gradient update in the Update() function in the blob.cpp, so that all kinds of solvers(except SGD) could take benefit of the increasement quantization.
@Zhouaojun Hi, Thanks for sharing the awesome code. About the log2 represent of the filter, I have a problem. In the default 5 bit filter, does 1 bit for the...
@XinDongol Any update on the thread?
Hi, I'm new in the training and not able to give authority answers. However, in my opinion, it doesn't matter whether the batch size equals 1 or some one bigger....
I only focus on mobilenetv2. Which pytorch version and python version do you employ?
Yes, pre-trained model was used. My environment is python2.7/3.6 and pytorch 0.4.1
It seems the author employ version before 0.4 when setting up the project. Many features after 0.4 are changed. such as the variable, the nogradient clause