ziplab/saq

This is the official PyTorch implementation for "Sharpness-aware Quantization for Deep Neural Networks".

saq's Issues

Quantize_first_last_layer

Hi! I noticed that in your code, you set bits_weights=8 and bits_activations=32 for first layer as default, it's not what is claimed in your paper " For the first and last layers of all quantized models, we quantize both weights and activations to 8-bit. " And I see an accuracy drop if I adjust the bits_activations to 8 for the first layer, could u please explain what is the reason? Thanks!

How to train 4 bit mobilenent v2 to 72.0 (in the paper)?

Great job! In this paper, the top1 of 4bit mobilenetv2 is 72.0 and this is beyond SOTA. Is there a way that can reproduce this result?

代码问题请教

大佬，您的代码写的太棒了！我有个小小的疑问，希望大佬能够指点一下~我根据大佬的代码把qpreresnet20改为了qresnet20, 然后使用8位量化，发现精度只有90.93，我之前做PTQ, 8位resnet20的精度也能够达到92-93,请问大佬我可能是哪里出现了问题呢？下面粘上我训练的脚本

代码问题请求帮助

你好，带佬的代码写的很好，有部分代码不太懂，想请教一下，
parser.add_argument( "--arch_bits", type=lambda s: [float(item) for item in s.split(",")] if len(s) != 0 else "", default=" ", help="bits configuration of each layer",

if len(args.arch_bits) != 0: if args.wa_same_bit: set_wae_bits(model, args.arch_bits) elif args.search_w_bit: set_w_bits(model, args.arch_bits) else: set_bits(model, args.arch_bits) show_bits(model) logger.info("Set arch bits to: {}".format(args.arch_bits)) logger.info(model)
这个arch_bits主要是做什么的呢，卡在这里有段时间了