PyTorch Tutorial 17 - Saving and Loading Models

[@khadijakaimous3408]

thank you so much for these series, well explained

[@user-or7ji5hv8y]

This is the best pytorch series I have come across. examples get right to the gist of lesson.

[@patloeber]

thank you!

[@npomfret]

There are a couple of mistakes at around 14:30. You are using the state from the 'checkpoint' variable, where I think you should be using state from the 'loaded_checkpoint' variable. In general, I think this part of the tutorial would be clearer if you'd used two files, one for saving and one for loading. Also, a better IDE would highlight unused variables which might have helped avoid this mistake.

[@geoffreyexoo]

Mr. Python Engineer does a good job, moves along at a reasonable speed.

[@patloeber]

Thanks, glad you like it

[@christinagiannoula3106]

In 15min, lines 32 and 32 should change to "model.load_state_dict(loaded_checkpoint["model_state"])" and "optimizer.load_state_dict(checkpoint["optim_state"]), respectively.

[@patloeber]

Oh you are correct! Thanks for the hint!

[@baotography7829]

Hi, Why not line 33 is: optimizer.load_state_dict(loaded_checkpoint["optim_state"])?

[@unknown3158]

@@baotography7829 Line 33 should be as you say.

[@sarahjamal86]

Like always beautiful video, I am giving your videos to all the beginners out there :-)

[@patloeber]

Thanks so much :)

[@amankushwaha8927]

Very helpful video. Thanks

[@chiruk2835]

Excellent and very useful, I finished all 17 tutorials, theank you

[@amerel-samman9929]

Every pytorch video: "and you can ignore this warning right here" :P I am going to be first to fall into that trap. Great series! Thank you so much.

[@ubaidghante8604]

Great video, thanks man 👍🏻

[@Footballistaas]

hello, any advice how I can prep for interview for this?

[@prajganesh]

Great series! it will be great if you can add more pytorch videos :)

[@inesylla9706]

thank you for the clear lesson! Really useful!

[@patloeber]

glad you like it!

[@longdang7791]

Hi, thank you for showing all the methods. I have a federated learning use-case where I train a single model using 32 clients. Currently, I have CUDA Out of memory issue. I want to know how each of the methods presented here affect the memory allocation and reservation on a single GPU. Any comment/idea is appreciate. Many thanks.

[@MuhammadHussain-ws1xs]

Amazing video, can i confirm that by using this method we can do fine tuning, for example train our model on one dataset, then save and load the model (with the weights), we can then train on a similar dataset but using the loaded model with the saved weights, right?

[@iNTERnazionaleNotizia589]

Hi I want to ask, Is it possible to load a Pretrained model from other Dataset (such as BTCV or BraTS2021) , and use it for inference to a different dataset (such as ISLES 2022)?

[@cyukore1796]

very clear tutorial

[@patloeber]

thanks!

[@bijjalanaganithin3798]

Thank You so much for this tutorial series

[@patloeber]

Glad you like it:)

[@nguyenduydatnguyenduydat2640]

Thank you so much for making the wonderful tutorial! Would you like to correct again the part load checkpoint. I thought the command in line 32 33 should be model.load_state_dict(loaded_checkpoint['model_state']) and optimizer.load_state_dict(loaded_checkpoint['optim_state'])

[@ahmedgdoura218]

Hi, in 8:07 why we see that requires_grad=True even when you set your model to .eval() mode thanx!

[@patloeber]

Very good observation, sorry if I miscommunicated that! model.eval() will not deactivate the requires_grad! See here: discuss.pytorch.org/t/model-eval-vs-with-torch-no-grad/19615

[@GurpreetSingh-si2gh]

Thank you very much.🙏

[@shivanshchakrawarti9289]

How do you save losses and accuracy in a csv file? after each epoch?

[@erictian8075]

Thanks for the great video! I have a question at 14:56, when you print the loaded optimizer.state_dict(), the lr is loaded correctly, but the 'params' are not the same. What do 'params' of the optimizer mean? If I have saved checkpoint at epoch 100 and want to resume training by loading checkpoint, does the inconformity of params influence my training?

[@qhliucy]

'params' are always the memory addresses of params which you used to init the optimizer and will not change when you call optimizer.load_state_dict(). The memory addresses saved in checkpoint are meaningless unless you save and then load the checkpoint in the same process. In the video, the new optimizer is inited with the new model's parameters( optimizer = torch.optim.SGD(model.parameters(), lr=0) ).

[@m3dXX]

Great series, thanks for your work. Could you cover topic of NN for non-linear regression tasks with PyTorch?

[@patloeber]

I will consider this :)

[@Sparklerated]

could you make video about semantic segmentation with ensemble model with different encoders on pytorch, thank you!

[@tehreemsyed8621]

Nice Explanation!! I trained my model on GPU and saved but when loaded my .pt file using CPU, got an error? RuntimeError: Expected all tensors to be on the same device, but found at least two devices, cuda:0, and CPU! What should I do in this case? Thank you

[@patloeber]

did you use map_location? device = torch.device('cpu') model = Model(*args, **kwargs) model.load_state_dict(torch.load(PATH, map_location=device)) Also, you must push all your tensors and the model to the same device (either to cpu or gpu)

[@krishj8011]

Awesome Tutorial..👍👍👍

[@patloeber]

Thanks :)

[@mywayluna1715]

Thank you so much for your lecture! I tried to save all my testing model result with nii(image) format. Could you give me a tip for me? I saved and loaded model to test. I did the image segmentation, my goal is to get a predicted segmentation images.

[@annemahoro9966]

I need your help on how I can resolve below issues: RuntimeError: Attempting to deserialize object on a CUDA device but torch.cuda.is_available() is False. If you are running on a CPU-only machine, please use torch.load with map_location=torch.device('cpu') to map your storages to the CPU.

[@mohammadusamah819]

Hi I love your videos and it has been very much help for me. I have a question that how can we use pytorch custom models on different files

[@omerbinali9472]

How can we convert saved .pt model to .pb ?

[@newajsorifnishad9169]

Why my model run again after load. Pth file? Can u plz help me

[@teetanrobotics5363]

amazing video

[@innotech4079]

how to load *.pt?

[@patloeber]

model = torch.load(filename.pt) or model.load_state_dict(torch.load(filename.pt))

[@Dr.Bimechi]

tnx

Hey man i have been trying to reach you out via email. I need some help with a project that i am working on. lmk if you can help me out

[@ewunate]

Tanx

[@ashwiniaditya8445]

Your voice is irritating. Giving knowledge is one thing but that voice is distracting me from concentrating on the topic (looks like you are putting on an accent). And what is with ending each sentence like you are asking a question?

[@lakeguy65616]

HELP! I am confused. I have created and trained a model successfully. I have saved the model using the lazy method (torch.save(model, MODEL_NAME)) ** model_name is the file name and full path. Now, in a separate file(forecast), I want to load the trained model. So in this new forecast file, I've loaded the model from disk "model = torch.load("C:\\Users\\user\\Desktop\\datasets\ ain\\model.pth") " as per the lazy method. it generates an error message AttributeError: Can't get attribute 'NeuralNet' on

[@longdang7791]

Do not use the lazy method. You should use the recommended ones to avoid this situation. After training is done, it is better to save the trained model and load it and do an evaluation on test data to confirm your model's prediction performance.