itayhubara / calibtip Goto Github PK

[0] index into the first batch.

isn't sequential adaquant supposed to update the input cache of all batches to the quantized values?

Hi!
Thanks for your work! In Table 2, you showed the results of w4a4 configuration for ResNet models. But you omitted the result of MobileNet-V2 at that table. From figure 3, it's hard to know the exact value of quantized MobileNet-V2. May I know w4a4 quantized MobileNet-V2, please?

when i use advanced_pipeline.sh, and i achieve bugs at line 3 "sh scripts/integer-programing.sh resnet resnet50 4 4 8 8 50 loss True",
which is 'No such file or directory: 'results/resnet50_w8a8.adaquant/IP_resnet50_loss.txt'.
How can I fix it? Thanks for response~

Hello. I was trying to running the code as per instructed but got this error while running the code. Is there anything I am missing to add during the run? Please look into the errors.

main.py: error: invalid checkpoint: results/resnet50_w8a8/resnet.absorb_bn.measure_perC

error: argument -e/--evaluate: expected one argument

Good afternoon,

I found your paper very interesting and wanted to try out your code. I have a several questions I would be grateful if you could answer:

1. In your paper you use a 71.97% benchmark for FP32 ResNet18 model. Is this your own trained checkpoint? The torchvision model zoo provides 69.76 top-1 accuracy.
2. I am trying to measure accuracy of a quantized model(both weights and activations). The model saved after, for example, bn_tuning, contains weights that have attributes: quantize_weight.running_range, quantize_weight.running_zero_point, quantize_input.running_range, quantize_input.running_zero_point, beta, gamma, equ_scale, num_measured.

• For activations, naturally, I placed forward hooks that quantize inputs to layers using quantize_input.running_zero_point and quantize_input.running_range. However, there is no range and zero point for quantizing inputs to residual connection(where residuals are summed with outputs from the previous layer). Did you not quantize those inputs when measuring accuracy?
• Do I understand correctly, that the output model already contains quantized weights or do I have to quantize them myself using quantize_weight.running_range, quantize_weight.running_zero_point, gamma and beta attributes?
• Maybe you already have a script for measuring accuracy of a quantized model and I just missed it?
• What is the purpose of equ_scale attribute? In all of my tests on ResNet18 model they only contained ones(tensor of 1s).

Hi there!

In your paper https://arxiv.org/abs/2006.10518 there is a parameter V which was added to the weights of some layer in formula (2) or (3) and which should be optimized to minimize MSE between original and quantized outputs.

However, I can't find this offset parameter in your current repo. Is there something I have missed or you have decided to remove this parameter?

Hello and thank you for an interesting paper!
I have a question concerning the optimization of the quantization step size. In section D.1 of the paper, you mentioned the usage of Adam optimizer for solving the AdaQuant optimization problem, but I have failed to find a description of the methodology used for backpropagating gradients through formally nondifferentiable rounding operation. Is it just Straight Through Estimator?
Thank you for your time and consideration.

What data are you use in /media/drive/Datasets/imagenet/calib?

This implies to calculate MSE between relu(conv_out) for conv1 and conv2 layers

But in ResNet architecture, conv2 is not followed by a direct Relu. Instead it follows by a residual addition, then Relu.

How was this difference justified?