analogdevicesinc/ai8x-training

KeyError conv1.shift_qunatile not found

Opened this issue · 5 comments

flaviose commented

Dear All,

when following the instructions in the github page for training and quantizing the mnist model, when evaluating the model after quzantizing it, the error KeyError conv1.shift_qunatile not found.

ermanok commented

Hello,

Could you please give more details about the steps you followed and the command you run that causes the error.

Thanks.

flaviose commented

Hi, sorry for the late reply , I just followed the instructions on the GitHub exactly using the script provided by the repository instead of me crafting one. The steps that I followed are exactly the ones in the GitHub where mnist is used as an example

(ai8x-synthesis) $ scripts/quantize_mnist.sh runs as it should but
(ai8x-training) $ scripts/evaluate_mnist.sh throws the error regarding shift_quantile.

flaviose commented

I trained a model myself When evaluating the resulting models using train.py whenever I use load weights from and a model that has qat_, for example qat_best.pth.tar in the name , I get a state Dict missing keys problem. This happens if I have a batchnorm layer like in my case the first layer FusedConv2dBNRelu, I don't how the training can generate the qat model but can't load the weights It created, and instead throws the error the loaded checkpoint is missing 4 state keys. The keys missing are 'conv1.bn.weight, cond1.bn.bias, cond1.bn.running_mean, conv1.bn.running_var'

asyatrhl commented

Hi, this issue has been resolved in the latest commit. Could you please fetch the ai8x-training repository?

github-actions commented

This issue has been marked stale because it has been open for over 30 days with no activity. It will be closed automatically in 10 days unless a comment is added or the "Stale" label is removed.