Meaning of transform_input parameter in inceptionV3

vision
1

Hi, I’m trying to modify the output of inceptionV3 by creating new class that derives from InceptionV3.

My Question is regarding the parameter transform_input in:
torchvision/models/inception.py

If I’m using the pretrained version and not passing this parameter this is set to True, does this mean that I should not do any preprocesing to the images that I feed the network?
Currently I’m using a composed Transform that I apply to the images, most of it is for data augmentation but I also have normalization there:

deep_drive_training_trafo = transforms.Compose([
    transforms.Resize((299, 299)),
    transforms.RandomResizedCrop(size=299, scale=(0.333, 1.0), ratio=(0.75, 1.3333333333333333)),
    transforms.RandomHorizontalFlip(),
    transforms.ColorJitter(brightness=0.6, contrast=0.5, saturation=0.8),
    transforms.ToTensor(),
    transforms.Normalize(mean=[0.485, 0.456, 0.406],  # here we need 3 values, 1 for each chanel
                         std=[0.229, 0.224, 0.225]),
    transforms.RandomErasing(p=0.3, scale=(0.033, 0.2), ratio=(0.3, 3.3), value='random', inplace=False)
])

Is this correct or by not passing transform_input to InceptionV3 the model is applying another transform on top of that? If it is true which should I use?

2

If you set transform_input=True, this line will call into this method.

However, by default this argument is set to False.

3

Pretrained model has transform_input as True by default. I encountered it when there was mismatch in output when loading the statedict separately. You may need to specify in documentations clearly.