Script that demonstrates the LSQ over ImageNet

Context

Now that all the elements of the LSQ Quantization Aware Training are ready, we want a script that train a real-size network (e.g. ResNet18) with LSQ nodes inserted, and that evaluate the accuracy drop when using 8 and 4 bits precisions.

Key Steps

• LSQ of a ConvNet using SGD on CIFAR-10
• Add support of Adam optimizer
• Add support of Cross-Entropy Loss
• Training of a ResNet18 using Adam on ImageNet
• LSQ of a ResNet18 using Adam on ImageNet

Results

TODO ...