Dumping weights from transfer learning - inception v4

[antspy]

Hi,

Is there a way to load inception_v4 weights from a custom .pb tensorflow model? The weights are different than the one from the original inception but the architechture is the same.
It looks like it should be an easy modification but I know nothing about tensorflow :)

[Cadene]

Unfortunatly I don't know a "simple" way to do it. I did it by dumping parameters from a loaded model in tensorflow. Please have a look at:
https://github.com/Cadene/tensorflow-model-zoo.torch/blob/master/inceptionv4/tensorflow_dump.py

Then you can load the pretrained parameters and test if you get the same result with the pytorch model:
https://github.com/Cadene/tensorflow-model-zoo.torch/blob/master/inceptionv4/pytorch_load.py

[antspy]

Thank you for your answer! I was referring to those files, in particular tensorflow_dump. I see you download the model from the internet, but what if you only have a .pb file you want to dump the weights of? Assuming that the structure is the same as inception v4 (or similar enough anyway)

[Cadene]

You should definitely try. Beware the tensorflow API has changed a bit. You should have a look to the nasnet directory (in this repo) which uses the new API.

[berkinimamoglu]

@Cadene thanks alot for your reply , I really need a clarification on a paper that I read related with transfer learning portion. The subject is a classification problem for numbering disease level from 0 to 5

Inside the artice there is a statement like this " An ImageNet pre-trained Resnet-50 model was used, and the Inception’s bottom layers were frozen to prevent their weights from getting updated during training process while the remaining top layers were trained with the pre-processed fundus images.
" under CNN Model section inside paper. What does this mean? Which bottom layer's do they speak? and for this problem is there any available code that u might advise to us as a reference . Thank you.

[Cadene]

@berkinimamoglu It sounds a mistake to me, because ResNet-50 doesn't have any "Inception's bottom layers". Source
In fact, an inception module has multiple branches which produces features maps that are concatenated at the end. Source

The authors of the paper are probably speaking about fine-tuning where most of the layers (starting from the beginning) are frozen. Unfortunately, they don't provide further details. What you should do is to download resnet50 with pretrained weights and try to fine tune specific layers.

import pretrainedmodels as pm
model = pm.__dict__['resnet50'](num_classes=1000, pretrained='imagenet')

nb_classes = 10 # your new dataset
model.last_linear = nn.Linear(model.last_linear.in_features, nb_classes)

# freeze all parameters
for param in model.parameters():
    param.requires_grad = False

# unfreeze the last block residual before the classif
for param in model.layer4.parameters():
    param.requires_grad = True

# unfreeze classif
for param in model.last_linear.parameters():
    param.requires_grad = True

optimizer = torch.optim.Adam(
    filter(lambda p: p.requires_grad, model.parameters()),
    3e-4)

# train the network