Quick question on inference from loading the model dictionary

1

Hello folks,

I have a small question to ask. I am currently, running image recognition on CIFAR-10 dataset. I have trained the model and have saved the model’s state_dict. The accuracy i am getting is 92.2 percent. This is trained without the bias and using batch normalization. The architecture followed is the VGG architecture.

However, when i load the dictionary of the trained model separately and run the following code my accuracies are dropping to 45 percent. Not use why there is a drop by 50 percent.

The code is as follows:

correct=0
total=0
for images,labels in valid_loader:
        images=images.to(device)
        labels=labels.to(device)
        op=F.conv2d(images,model['layer1.0.weight'],padding=1)
        op=obj1(op)
        op=F.relu(op)
        op=F.conv2d(op,model['layer1.3.weight'],padding=1)
        op=obj1(op)
        op=F.relu(op)
        op=obj(op)
        op=F.conv2d(op,model['layer2.0.weight'],padding=1)
        op=obj2(op)
        op=F.relu(op)
        op=F.conv2d(op,model['layer2.3.weight'],padding=1)
        op=obj2(op)
        op=F.relu(op)
        op=obj(op)
        op=F.conv2d(op,model['layer3.0.weight'],padding=1)
        op=obj3(op)
        op=F.relu(op)
        op=F.conv2d(op,model['layer3.3.weight'],padding=1)
        op=obj3(op)
        op=F.relu(op)
        op=obj(op)
        op=op.reshape(op.size(0),-1)
        op=F.dropout(op)
        op=F.linear(op,model['layer4.1.weight'])
        op=F.relu(op)
        op=F.dropout(op)
        op=F.linear(op,model['layer4.4.weight'])
        _,predicted=torch.max(op,1)
        total+=labels.size(0)
        correct+=(predicted==labels).sum().item()
        
print('Validation Accuracy of the model on the images:{}%'.format((correct/total)*100))

obj1,obj2,obj3 are instances of class involving batch normalization and obj is an instance of another class performing maxpooling (2 by 2 filter size with stride 2)

2

While validating your model you would usually disable dropout, so you could just remove F.dropout from your model or disable it with op = F.dropout(op, False).

Did you get the 92% accuracy for the training set?
Did you validate the model during training before storing and reloading?

3

Yes i validated before storing and reloading. Cool will try it out now

4

I did try it and the accuracy is 45 percent.
with torch.no_grad():
correct=0
total=0
for images,labels in valid_loader:
images=images.to(device)
labels=labels.to(device)
op=F.conv2d(images,model[‘layer1.0.weight’],padding=1)
op=z1(op)
op=F.conv2d(op,model[‘layer1.3.weight’],padding=1)
op=z1(op)
op=zm(op)
op=F.conv2d(op,model[‘layer2.0.weight’],padding=1)
op=z2(op)
op=F.conv2d(op,model[‘layer2.3.weight’],padding=1)
op=z2(op)
op=zm(op)
op=F.conv2d(op,model[‘layer3.0.weight’],padding=1)
op=z3(op)
op=F.conv2d(op,model[‘layer3.3.weight’],padding=1)
op=z3(op)
op=zm(op)
op=op.reshape(op.size(0),-1)
op=F.dropout(op,False)
op=F.linear(op,model[‘layer4.1.weight’])
op=F.relu(op)
op=F.dropout(op,False)
op=F.linear(op,model[‘layer4.4.weight’])
op=F.softmax(op,dim=1)
_,predicted=torch.max(op,1)
total+=labels.size(0)
correct+=(predicted==labels).sum().item()
print(‘The validation accuracy is {}%’.format((correct/total)*100))

Lets say the trained model is model=ConvNet.to(device)

Instead of storing and loading the state_dict is there a direct way of storing the model instance itself so that I could do the following:
with torch.no_grad():
correct=0
total=0
for images,labels in valid_loader:
images=images.to(device)
labels=labels.to(device)
op=model(images) #loaded from a file
_,predicted=torch.max(op,1)
total+=labels.size(0)
correct+=(predicted==labels).sum().item()
print(‘The validation accuracy is {}%’.format((correct/total)*100))