Accessing weights and bias values of a pretrained and quantized model

1

How can we access all weights and biases for a pretrained and quantized model of a neural network?
I have downloaded the model using the following command:

model = models.quantization.resnet18(pretrained=True, quantize=True)
for param_tensor in model.state_dict():
print(’ For param_tensor is ',param_tensor)
if model.state_dict()[param_tensor].dtype==torch.qint8:
double_x=(torch.int_repr(model.state_dict()[param_tensor]).numpy())
else:
double_x=(model.state_dict()[param_tensor]).detach().numpy()

This leads to the following error:
For param_tensor is fc._packed_params.dtype
Traceback (most recent call last):
File “compression_quantize_test.py”, line 63, in
if model.state_dict()[param_tensor].dtype==torch.qint8:
AttributeError: ‘torch.dtype’ object has no attribute ‘dtype’

Can anyone let me know the correct way to access all the weights and biases?

2

You are seeing this because the quantized model saves the weights for fully connected layers using this odd _packed_params object. That is what is actually stored in the state_dict.

Try to just print the objects in the state_dict to see. And then check what is in the _packed_params object. That way you’ll know how to access weights for FC layers.