lsq-net

lsq-net copied to clipboard

Error: Divided by 0.

3

Thanks for your the implementation! In LsqQuan. https://github.com/zhutmost/lsq-net/blob/2c24a96be06d044fa4c7d651727f4574b8d88c86/quan/quantizer/lsq.py#L54 If s_scale contains 0, there will be Divided by 0 Error, and the training loss will become nan. Maybe eps should be...

woriazzc

Distillation degrades the accuracy

1

I add distillation when training resnet18. But the Top-1 Acc degrades from 68.150 % to 67.364%。 Hyperparameters as follow: 4gpu epochs: 90 learning_rate: 0.01 momentum: 0.9 weight_decay: 0.0001 mode: step...

lippman1125

TypeError: Conv2d._conv_forward() missing 1 required positional argument: 'bias'

2

python main.py config.yaml INFO - Log file for this run: /home/itl/Documents/xrh/lsq-net/out/MyProject_20220505-162725/MyProject_20220505-162725.log INFO - TensorBoard data directory: /home/itl/Documents/xrh/lsq-net/out/MyProject_20220505-162725/tb_runs Files already downloaded and verified Files already downloaded and verified INFO - Dataset...

rahulxie

Quantized checkpoint is not smaller / faster

8

Hi, first, thanks for your implementation! It's not too hard to apply your code to my current model! However, the dumped checkpoint has the same size as the original model,...

sophia1488

How to deploy the quantized model?

1

When we have trained the quantization model, how to deploy it in cpu backend?

jiinhui

Accuray for Cifar-10

2

Hey, thanks for your work. Have you tested the accuracy for Cifar10? I tried to quantized ResNet20 into 2/2 4/4 bit, and the accuracy only 88%~90%. If possible, can you...

loujie1994

Did you improve the performance using per-channel in weight quantization?

4

Hi, Great implementation! Since per-channel weight quantization is implemented in you code, I'm wondering if there is any improvement compared to per-tensor weight quantization.

talenz

Why use fixstep lr

3

Thanks for your work! I have 2 questions: In the paper, the author use cos lr but why fixsLR used in your code? And now W3A2 acc has reported but...

liu6381810

cannot reproduce the same accuracy in the paper

9

Thanks for your great work! But I trained resnet18 with the default config, and only got the top1 acc 52, top5 acc 75

HuDi2018

Train and validation set are imbalanced

2

When using the `random_split` function provided in PyTorch to split the dataset to training and validation sets, the propotion of categories are imbalanced.

zhutmost