Accuracy decrease after load the saved model

vision
1

Hi. Can you tell me why accuracy of my model decreases when I save and load it to predict. My accuracy after training is about 88%

I use resnet model of pytorch

model_ft = models.resnet34(pretrained=True)
num_ftrs = model_ft.fc.in_features

# replace the last fc layer with an untrained one (requires grad by default)
model_ft.fc = nn.Linear(num_ftrs, 216)
model_ft = model_ft.to(device)

# uncomment this block for half precision model
# """
model_ft = model_ft.half()


for layer in model_ft.modules():
    if isinstance(layer, nn.BatchNorm2d):
        layer.float()
# """
criterion = nn.CrossEntropyLoss()
optimizer = optim.SGD(model_ft.parameters(), lr=0.01, momentum=0.9)

lrscheduler = optim.lr_scheduler.ReduceLROnPlateau(optimizer, mode='max', patience=3, threshold = 0.9)

I use this code to save the model

PATH = './car_classifier.pth'
torch.save(model_ft.state_dict(), PATH)

then I load it with

image
image713×172 5.99 KB

Thanks in advanced!

2

My guess is here model accuracy was depend on the batches that go through the testing phase so instead of performing once test for 5 time and average it then save the model and then redo it

3

Sorry I don’t really understand what you want me to do. Can you make it more clearly

4

I use this code to transform the input images.
image

5

How large is the decrease in accuracy?
Are you using the same data for both use cases or are you loading another dataset for the validation use case?

6

I trained the model on jupyter note book and get 88% validation accuracy. Then I use it to predict some random images and it go well. After that I save and load it to my web app to predict. But none of them is correct. Is that because of the way I load my model to the web app?

7

can you please help me with it

8

Hi
I Have the same problem with ResNet50?
my validation and train accuracy decreased 50% when I loaded the saved model with load_state_dict() function !!!
However when I save training results everything was OK