[Q] Need an example to show how to generate an image. about pytorch-generative HOT 13 CLOSED

@showmin An example of how to pass the sample_fn is here.

The sample_fn should call model.sample() and do any pre/post processing to convert the model samples into images. In the example above, we first call model.sample() to generate the raw samples then binarize the output to 0 or 1 since we're generating binarized MNIST images.

For ImageGPT, I believe the sample_fn should be as straightforward as passing sample_fn=model.sample.

Would you please add a simple example in the readme?

Good point, I will update the readme with an example of how to sample from a trained model.

from pytorch-generative.

Yes, like I said above it only works for new checkpoints unfortunately. For old checkpoints you'd need to manually use model.sample.

from pytorch-generative.

@EugenHotaj Thanks. It works without any other issue.

from pytorch-generative.

@showmin I thought a bit more about the sample_fn and decided to remove it altogether because it's confusing. You should now no longer need to pass in sample_fn and can just set sample_epochs on the Trainer. The trainer will directly call model.sample().

Hopefully this is a lot easier. Let me know if you run into issues :) .

from pytorch-generative.

@EugenHotaj Thanks for the reply. That's really easier for me. Well, now I have to wait for the result after trainning 1 epoch but cannot restore the epoch that I've trained then generate the samples. It would be perfect for me if I can generate samples from trained epoch.

from pytorch-generative.

@showmin If you want to generate samples at a particular epoch, the easiest thing to do is manually call model.sample() after you've loaded the checkpoint. The samples generated in the Trainer are mostly meant to track the progress during training.

from pytorch-generative.

I try to add self.sample_one_batch() after the code:

such that I can generate samples after I restore the latest epoch. But it came out an error:

Found 2 saved checkpoints.
Restoring trainer state from checkpoint trainer_state_2.ckpt.
Falied to sample from the model: 'ImageGPT' object has no attribute '_c'

from pytorch-generative.

The issue is that we weren't saving some variables during checkpointing. I just landed 406a7be which should fix this issue for new checkpoints only.

from pytorch-generative.

Still has the same error message. Does it only work for the new trained epoch after I checkout the latest code?

from pytorch-generative.

After training a new epoch, errors here:

Found 3 saved checkpoints.
Restoring trainer state from checkpoint trainer_state_3.ckpt.
Traceback (most recent call last):
  File "train.py", line 76, in <module>
    main(args)
  File "train.py", line 44, in main
    MODEL_DICT[args.model].reproduce(args.epochs, args.batch_size, args.logdir)
  File "D:\Jupyter_Home\pytorch-generative\pytorch_generative\models\autoregressive\image_gpt.py", line 175, in reproduce
    model_trainer.interleaved_train_and_eval(n_epochs)
  File "D:\Jupyter_Home\pytorch-generative\pytorch_generative\trainer.py", line 232, in interleaved_train_and_eval
    self.restore_checkpoint()
  File "D:\Jupyter_Home\pytorch-generative\pytorch_generative\trainer.py", line 134, in restore_checkpoint
    self.model.load_state_dict(checkpoint["model"])
  File "C:\Users\w00665547\Anaconda3\envs\env_ml\lib\site-packages\torch\nn\modules\module.py", line 1671, in load_state_dict
    raise RuntimeError('Error(s) in loading state_dict for {}:\n\t{}'.format(
RuntimeError: Error(s) in loading state_dict for ImageGPT:
        Unexpected key(s) in state_dict: "_c", "_h", "_w".

from pytorch-generative.

It seems that the error was caused by putting register_buffer() in call method in base.py. I tried to put them in init method and it removed the error successfully. If it is the root cause, probably you need to assign an input image when creating the object.

from pytorch-generative.

@showmin busy week so just had a chance to look at this. You're right, the issue is that we create some dynamic buffers during forward which won't be present on a newly initialized model causing the checkpoint loading to fail. I just pushed in 449795f which should fix the issue. You should be able to load in any existing checkpoints now.

Let me know if you're still seeing issues.

from pytorch-generative.

Great! Going to close this issue out, please feel free to open again if needed.

from pytorch-generative.