@@ -1129,9 +1129,9 @@ def forward(self, input):
11291129print (acc_train )
11301130
11311131#%%
1132- # now we are ready to see how finte -tuning works in Pytorch
1133- # there are several models in github models repository that we can use
1134- # lets choose one but before lets see what we have at our disposal
1132+ # OK we are ready to see how fine -tuning works in Pytorch
1133+ # there are several models in models repository that we can use
1134+ # lets choose one but before that lets see what we have at our disposal
11351135from torchvision import models
11361136import torch .nn as nn
11371137
@@ -1142,19 +1142,23 @@ def forward(self, input):
11421142# lets print the model
11431143print (f'\n ORIGINAL MODEL : \n { resnet18 } \n ' )
11441144
1145- # by looking at the architecure, we notice
1145+ # by looking at the architecure, we notice :
11461146# (fc): Linear(in_features=512, out_features=1000, bias=True)
1147- # this means in order to make retrain this network for our usecase
1148- # lets train this for cifar10 which has 10 classes.
1147+ # In order to retrain this network for our usecase
1148+ # we need to alter this layer. this was trained on imagenet
1149+ # which had 1000 classes. lets train this for cifar10 which
1150+ # has 10 classes. all we need to do is just defining a new
1151+ # fully connected (fc) layer and assigning it back to
1152+ # resnet18.fc attribute!
11491153resnet18 .fc = nn .Linear (512 , 10 )
1150- # instead of hardcoding the 512 which we saw from the printed version of
1151- # our model. we can simply use the in_features attribute of the fc layer!
1152- # and write :
1154+ # instead of hardcoding the 512 which we saw by looking at the
1155+ # printed version of our model, we can simply use the
1156+ # 'in_features' attribute of the fc layer! and write :
11531157# resnet18.fc = nn.Linear(resnet18.fc.in_features, 10)
11541158
11551159print (f'\n NEW MODEL(after adding the new fc layer): \n { resnet18 } )
1156- # now before we dive in to train our net we should frst
1157- # freeze all layers but this new one, and train for several epochs,
1160+ # now before we dive in to train our network we should first
1161+ # freeze all layers except this new one, and train for several epochs,
11581162# so that it converges to a reasonable set of weights
11591163# then we unfreeze all previous layers and train the whole net
11601164# altogether again.
0 commit comments