I have the following code:
x = cv2.imread('anormal_9979_AVsp2000_ciclo6.png')
transform = T.Compose([T.ToTensor(), T.Resize((64, 64)), T.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])]);
#transform_img = T.ToTensor()
x_t =transform(x)
batch_t = torch.unsqueeze(x_t,0).double()
torch.nn.CrossEntropyLoss()
resnet = torchvision.models.wide_resnet101_2(pretrained=False);
resnet.fc = torch.nn.Linear(resnet.fc.in_features, 2);
resnet.load_state_dict(torch.load('wide_resnet101_2.pt', map_location=torch.device('cpu')));
resnet.eval();
net = resnet(batch_t.float())
predictions = torch.argmax(net, axis= 1);
print(predictions);
And when it prints the prediction, the prediction is always wrong. However, when I trained the model and save it, the test_acc is aprox 95%.
In addition, this resnet is trained for 2 classes([0] for Normal and [1] for Anormal).
I tried to modify the grad_fn, but i can't apaprently. Also, at the moment i've seen that you can also save the weights of the pretrained model, but does not the model.pt has inheretly the final weights?
First parameter of nn.Module.load_state_dict is a state_dict, a dictionary of layer-parameter pairs.
This call does not load the .pt file and copy it into the model, but instead just passes the string as a state_dict.
resnet.load_state_dict('wide_resnet101_2.pt')
Correct way is to first read the weights using torch.load, and then copy them into the model.
state_dict = torch.load('wide_resnet101_2.pt')
resnet.load_state_dict(state_dict)
Torch documentation provides a neat beginner-friendly overview on the matter.
Although, this also depends on how you saved the model.
torch.save(model, PATH) # saves the entire model
model = torch.load(PATH) # loads the entire model with weights
torch.save(model.state_dict(), PATH) # saves state_dict
# now loading the state_dict first is needed, after which it can
# be copied into model
model.load_state_dict(torch.load(PATH))