What is the mean and std of imagenet dataset if using 'autoaugment'?
YongtaoGe opened this issue · 4 comments
It's the same means and stds the model was trained with on ImageNet. For most of the models from torchvision, according to pretrained_models_pytorch:
means = [0.485, 0.456, 0.406]
stds = [0.229, 0.224, 0.225]
Do you mean that the autoaugment would not change the distribution of mean and stds in original imagenet dataset? But as I know, differenet preprocessing ways could lead to different means and stds.
for model_name in __all__:
input_sizes[model_name] = [3, 224, 224]
means[model_name] = [0.485, 0.456, 0.406]
stds[model_name] = [0.229, 0.224, 0.225]
for model_name in ['inceptionv3']:
input_sizes[model_name] = [3, 299, 299]
means[model_name] = [0.5, 0.5, 0.5]
stds[model_name] = [0.5, 0.5, 0.5]
This only matters if you are using a pretrained model. Also, means and stds of the dataset are calculated from the raw images without data augmentation as far as I know. And the values can be chosen quite arbitrarily as you can see from inceptionv3.
Is it normal to use these values to normalize and then the images channels 0 and 2 are transformed to inf, but the channel 1 maintains its values ? (I can't plot the images if I normalize the data with Image Net' mean and std values)