Activation Bits Support + qconfig kwargs for QuantizationModifier

[rahul-tuli]

import torch
from sparseml.pytorch.models import resnet50
from sparseml.pytorch.optim import QuantizationModifier


# Pretrained Model
model = resnet50(pretrained=True)

# Modifier init
modifier = QuantizationModifier(start_epoch=0, activation_bits=4)

# Modifier Application
modifier.apply(
    module=model,
)

# Book-Keeping, the model should have non infinity weights
print(model(torch.rand(1, 3, 224, 224)))

for param in model.parameters():
    print(param.data.min(), param.data.max())

[bfineran]

as discussed we need to have modifier properties to control both the weight and activation qconfig kwargs

[bfineran]

@rahul-tuli pushed the discussed fix for propagating observer quant range

[rahul-tuli]

@rahul-tuli pushed the discussed fix for propagating observer quant range

Thanks, incorporated the requested changes!

[bfineran]

looks close! left a few comments then good from my end

[bfineran]

one more comment then good to go

[KSGulin]

Good stuff

synced offline