Transfer Learning help me

vision
1

Hello. I want to transfer learning with pytorch but my error rate is increasing and decreasing abnormally and my accuracy rate is not increasing. Am I doing something wrong? Yield is 19201080 but due to lack of memory there are 2828 and formula 1 car images

note=The ready models and weights I use are not suitable for my dataset. what do you think

import torchvision.models as models
    
model = models.resnet152(pretrained=True) # Önceden eğitilmiş modelleri kullanmak için pretrained=True 

for param in model.parameters(): # Tüm modeli eğitmek için True
    param.requires_grad = False # Parametreleri donduruyoruz

num_ftrs = model.fc.in_features
model.fc = nn.Linear(num_ftrs, 100)
model.fc2=nn.ReLU()
model.fc3=nn.Linear(100,20)
model.fc4=nn.ReLU()
model.fc5=nn.Linear(20,10)
model.fc6=nn.ReLU()
model.fc7=nn.Linear(10,4)

error = nn.CrossEntropyLoss()
optimizer = optim.Adamax(model.parameters(), lr=0.0001)
num_epochs=20
count=0
losses = []
iterasyon=[]

for epoch in range(num_epochs):

    for i,(images,label) in enumerate (train_loader):
        
        out = model(images)
        loss = error(out, label)

        optimizer.zero_grad()
        loss.backward()
        optimizer.step()
        
        losses.append(loss.item())
        
        count+=1
        
        if count % 200 == 0:
            
            total=0
            correct=0
            correct_hata=0
            
            for image,labels in test_loader:
                
                out=model(images)
                
                pred=torch.max(out.data,1)[1]
                
                total+=len(label)
                
                correct+= (pred==labels).sum()
                correct_hata+=(pred!=labels).sum()
                
            dogruluk=100*correct /float(total)
            hata=100*correct_hata /float(total)
            
            losses.append(loss.data)
            iterasyon.append(count)
            
            
        if count % 200 == 0:
            print('Iteration: {}  Loss: {}  Accuracy: {}% Error: {}%'.format(count, loss.data, dogruluk,hata))
        ```
result:

Iteration: 200 Loss: 1.5698899030685425 Accuracy: 27.848100662231445% Error: 72.15190124511719%
Iteration: 400 Loss: 1.530423879623413 Accuracy: 26.582279205322266% Error: 73.417724609375%
Iteration: 600 Loss: 1.5104633569717407 Accuracy: 26.582279205322266% Error: 73.417724609375%
Iteration: 800 Loss: 1.3361421823501587 Accuracy: 27.848100662231445% Error: 72.15190124511719%
Iteration: 1000 Loss: 1.297659158706665 Accuracy: 20.253164291381836% Error: 79.74683380126953%
Iteration: 1200 Loss: 1.3579907417297363 Accuracy: 27.848100662231445% Error: 72.15190124511719%
Iteration: 1400 Loss: 1.1254953145980835 Accuracy: 25.316455841064453% Error: 74.68354797363281%
Iteration: 1600 Loss: 1.9074203968048096 Accuracy: 25.316455841064453% Error: 74.68354797363281%
Iteration: 1800 Loss: 1.3322985172271729 Accuracy: 27.848100662231445% Error: 72.15190124511719%
Iteration: 2000 Loss: 1.3254060745239258 Accuracy: 25.316455841064453% Error: 74.68354797363281%
Iteration: 2200 Loss: 1.1631817817687988 Accuracy: 20.253164291381836% Error: 79.74683380126953%
Iteration: 2400 Loss: 1.2752320766448975 Accuracy: 20.253164291381836% Error: 79.74683380126953%
Iteration: 2600 Loss: 1.3860232830047607 Accuracy: 27.848100662231445% Error: 72.15190124511719%
Iteration: 2800 Loss: 1.2704166173934937 Accuracy: 20.253164291381836% Error: 79.74683380126953%
Iteration: 3000 Loss: 1.081123948097229 Accuracy: 25.316455841064453% Error: 74.68354797363281%

2.accuracy :

Train Doğruluk:
Got 180 / 200 with accuracy 90.00

Test Doğruluk:
Got 43 / 79 with accuracy 54.43```

2

You are freezing all parameters of the untrained model besides the new linear layer in model.fc.
Since the model wasn’t pretrained, I would assume that your training fails as you are trying to use transfer learning to a randomly initialized model.

Also, model.fc2, model.fc3, etc. won’t be used unless you override the forward and use them explicitly. If you want to use more than a single new linear layer in model.fc, replace model.fc with an nn.Sequential container.

3

thanks @ptrblck. Thank you for your help. What I don’t understand is this. I want to use pre-trained weights and a proven model. When I say .pretrained=True, it downloads the pth file and these weights. I want to use them. but don’t I need to freeze the parameters? what happens if i don’t freeze the parameters? I don’t want to train from scratch I want to use a certain model and weights and add a few final layers to the model. i just want to train the layers i add

4

Sorry, I might have looked at another post and saw pretrained=False in the code snippet.
You are right that you are already using a pretrained model, so skip this part and check the last issue (i.e. replacing the model.fc layer with nn.Sequential).

5

thank you @ptrblck I did it as sequential as you said, but my problem is still not fixed. The ready-made models and weights I use are not suitable for my dataset. What do you think, could it be caused by this? and is my code correct

6

If you assume the pretrained weights are not suitable for your task, you could either fine-tune the entire model (so remove freezing the parameters) or train it from scratch.