mrusci / training-mixed-precision-quantized-networks Goto Github PK

Hi,
The asymmetric quantization function used in the code is not clamping the data tensor for which the values exceed the specified bit width. For 4-bit quantization, the deployment model contains values (weights) that exceed the 4-bit range.
Is there any problem if I clamp the data tensor before converting it back to floating-point on the accuracy?
If it is, then what could be the solution?

Thanks

Hi
Firstly, I appreciate you for this project. It seems you haven't updated your code since four years ago. Some errors occur for

python3 main_binary.py -a mobilenet --mobilenet_width 1.0 --mobilenet_input 224 --save Imagenet/mobilenet_224_1.0_w8a8 --dataset imagenet --type_quant 'PerLayerAsymPACT' --weight_bits 8 --activ_bits 8 --activ_type learned --gpus 0,1,2,3 -j 8 --epochs 12 -b 128 --save_check --quantizer --batch_fold_delay 1 --batch_fold_type folding_weights

I solved some of them by using

• transforms.Resize instead of transforms.Scale
• reader = tf.train.load_checkpoint(tensorflow_model) instead of reader = pywrap_tensorflow.NewCheckpointReader(tensorflow_model)
• correct_k = correct[:k].reshape(-1).float().sum(0) instead of correct_k = correct[:k].view(-1).float().sum(0)

I got stucked at this error:

Traceback (most recent call last):
  File "/content/training-mixed-precision-quantized-networks/main_binary.py", line 456, in <module>
    main()
  File "/content/training-mixed-precision-quantized-networks/main_binary.py", line 286, in main
    train_loss, train_prec1, train_prec5 = train(
  File "/content/training-mixed-precision-quantized-networks/main_binary.py", line 443, in train
    return forward(data_loader, model, criterion, epoch,
  File "/content/training-mixed-precision-quantized-networks/main_binary.py", line 411, in forward
    quantizer.backprop_quant_gradients()    
  File "/content/training-mixed-precision-quantized-networks/quantization/quantop.py", line 1037, in backprop_quant_gradients
    w_max_param.grad.data = grad_a_max
AttributeError: 'NoneType' object has no attribute 'data'

Can you take a look at it?