How to save and load the transfer-learned Inception V3 model?

vision
1

I want to save the transfer-learned inception V3 model. I couldn’t able to figure out what to be saved apart from state_dict. I didn’t made any changes to convolution layers part by freezing them. I changed fully connected layer part by adding some layers and scaling to smaller number of outputs.

Now, what are the things to be stored in order to successfully load and use the model after saving it.
precisely
i want to fill these blanks
checkpoint = {
‘input_isize’: _________,
‘output_size’:102,
‘hidden_layers’: ________,
‘state_dict’: model.state_dict()}

2

model.state_dict() basically stores weights. For example in my RNN model :
odict_keys(['i2h.weight', 'h2h.weight', 'h2o.weight']). You can reach what model.state_dict() stores with model.state_dict().keys(). I think you can do like that :

first_layer_name = list(model.state_dict().keys())[0]
input_dim = model.state_dict()[first_layer_name].size()[1]
# since weights are in the form of (hid_dim x input_dim) instead of (input_dim x hid_dim)
# we use .size[1] instead of .size[0]

You can also do same thing for hidden layer dimension.