I am using tensorflow slim to load pre-trained models like vgg and resnet-50. So for vgg , tf-slim provides a way to load RGB mean values like:
from preprocessing.vgg_preprocessing import (_mean_image_subtraction,
_R_MEAN, _G_MEAN, _B_MEAN)
I couldn't find anything similar for resnets. Is it not implemented yet? Also I do know that some libraries like py-torch provide global mean values for every model. Is that the case with tf-slim too?
Actually despite preprocessing_factory says that restnetv2 uses vgg_preprocessing it isn't.
The correct preprocessing for resnetv2 is inception_preprocessing as noted in this github issue https://github.com/tensorflow/models/issues/2217
The previous answer is actually not completely correct. Have a look at https://github.com/tensorflow/models/blob/master/research/slim/preprocessing/preprocessing_factory.py
You can see that both Resnet-V1 and ResNet-V2 use VGG preprocessing
preprocessing_fn_map = {
'cifarnet': cifarnet_preprocessing,
'inception': inception_preprocessing,
'inception_v1': inception_preprocessing,
'inception_v2': inception_preprocessing,
'inception_v3': inception_preprocessing,
'inception_v4': inception_preprocessing,
'inception_resnet_v2': inception_preprocessing,
'lenet': lenet_preprocessing,
'mobilenet_v1': inception_preprocessing,
'mobilenet_v2': inception_preprocessing,
'mobilenet_v2_035': inception_preprocessing,
'mobilenet_v2_140': inception_preprocessing,
'nasnet_mobile': inception_preprocessing,
'nasnet_large': inception_preprocessing,
'pnasnet_mobile': inception_preprocessing,
'pnasnet_large': inception_preprocessing,
'resnet_v1_50': vgg_preprocessing,
'resnet_v1_101': vgg_preprocessing,
'resnet_v1_152': vgg_preprocessing,
'resnet_v1_200': vgg_preprocessing,
'resnet_v2_50': vgg_preprocessing,
'resnet_v2_101': vgg_preprocessing,
'resnet_v2_152': vgg_preprocessing,
'resnet_v2_200': vgg_preprocessing,
'vgg': vgg_preprocessing,
'vgg_a': vgg_preprocessing,
'vgg_16': vgg_preprocessing,
'vgg_19': vgg_preprocessing,
}
I have tested the Slim checkpoints of both Resnet_v1_50 and Resnet_v2_50 on ImageNet validation dataset, and the result is that Resnet_v1_50 uses VGG preprocessing, while Resnet_v2_50 uses Inception preprocessing.
The Github issue has also indicated misleading information in the source code.
In Tensorflow Slim resnet-v1 uses vgg_preprocessing. Resnet-v2 uses inception preprocessing by default, which uses a lot of color, hue, and saturation augmentation. This makes mean subtraction less relevant.
