Loading model, test accuracy drops

I have a PyTorch model that has test accuracy of about 97%. I save it using torch.save(my_model.state_dict(), PATH) , but whenever I try to reload it using my_model.load_state_dict(torch.load(PATH)) and test it on the same data using test_fn(my_model) my test accuracy goes down to about 0.06%. The same thing happens if I use test_fn(my_model.eval()) . Is there an extra step I need to take?

1 Like

Could you post a small executable code snippet, so that we can have a look?

my_model = GraphConv(w2i, p2i, l2i, r2i, s2i, words, pos, lems, 512, 512, 3) ## Initialise model & params
loss_function = nn.NLLLoss()
optimizer = optim.Adam(my_model.parameters(), lr=0.001)

for epoch in range(15):
    ...   ### Apply training steps
    print(test_fn(my_model))  ### Will be over 95%
    torch.save(my_model.state_dict(), PATH)

my_model2 = GraphConv(w2i, p2i, l2i, r2i, s2i, words, pos, lems, 512, 512, 3) ## Initialise new model
print(test_fn(my_model2))  ### Is about 0.06%

Please help me
I have similar problem but I cant find what’s wrong
Thank you

I faced similar issue trying to save using .pth. When I saved as .pt it worked fine.
Try this:

Saving the model using


Load the model using

my_model2 = GraphConv(w2i, p2i, l2i, r2i, s2i, words, pos, lems, 512, 512, 3)

Thank you!:slight_smile: @avinash_m

I have one more question regarding this problem.
May be it is overfitting but my training loss drop well but my validation loss keep same value.
Do you have any advise for me?

Try adding Dropouts. If layers are Convolution layers add BatchNormalization.

I already have Dropout, data augmentation.
You mean use dropout several times?

my dropout layer looks like below.

self.dropout = torch.nn.Dropout3d(dropout_prob) # 0.3

I’m not sure using BN because my batch size = 1. Even that now I’m using it after every conv layer.(I use model i3d())

Due to my gpu capability, it can only deal batch size 1.

After loading your model, you need to run model.eval() prior to testing to set dropout and batch normalization layers to evaluation mode. See *this for more details.

See my answer above.

If your dataset is small then try using any pretrained models depending upon your problem. For a small dataset, stacking up many layers doesn’t help much.

Thanks But I already use model.eval for validation and test.

my data is 3d imaging…! and I think I have enough data…
I try to solve 3d detection problem without using r-cnn or Yolo or common detection model.
I’m not sure using only cnn for detection problem will work… But it’s what i’m doing…

Could you give me some advice?