QAT with 100k images, for W4A32， use first 1k images from validation data of evanarlian/imagenet_1k_resized_256
Top-1 accuracy: 78.3 %
versus full precision
Top-1 accuracy: 88.80%, Top-5 accuracy: 97.60%
training data and validation data are from huggingface dataset evanarlian/imagenet_1k_resized_256.