How to save all test data images

[bibinwils]

Hi, I trained a UNet with input images and mask. For that I split the entire dataset to train - valid - test. I have trained the network and got the output metrics for the test dataset. There are around 100 test images. I would like to pass through the model and save the corresponding output images to a folder. How to save the test output so?

[rohitgr7]

for that you can try trainer.predict by passing the test dataloader inside predict. https://pytorch-lightning.readthedocs.io/en/stable/common/trainer.html#predict

[bibinwils]

Hi, Thanks for the reply. May I know how do I save all the images?

[rohitgr7]

https://discuss.pytorch.org/t/how-can-i-save-generate-image-in-new-separate-folder/84080/2
https://discuss.pytorch.org/t/converting-tensors-to-images/99482/6
https://discuss.pytorch.org/t/how-to-save-image/48808

[bibinwils]

Hi, I couldn't succeed. Can you help me out? I tried like
Trainer.predict(model=clf)
for i in range(100):
torchvision.utils.save_image(image_list[-1][i], “/images/generated_image%d.JPG” %i)

but not getting any results

[bibinwils]

anybody here to help?

[tchaton]

Won't this work ?

predictions = trainer.predict(model=clf)
for i in range(100):
    torchvision.utils.save_image(predictions[-1][i], “/images/generated_image%d.JPG” %i)

[bibinwils]

MisconfigurationException: Dataloader not found for Trainer.predict - it's the error coming there

[bibinwils]

import torchvision
predictions = trainer.predict(model=clf)

for i in range(32):

torchvision.utils.save_image(predictions[-1][i], 'generated_image'+'%d'+'.JPG' %i)

This is what I tried. I put 32 since my batch size is 32

[rohitgr7]

did you define a predict_dataloader? if not you need to pass one inside trainer.predict.

[bibinwils]

Thanks. I tried. Now i'm getting this error

[rohitgr7]

I don't remember what the exact dimension of the predictions is supposed to be, but you can analyze the prediction output a little bit and figure out how to access the required image output. If it doesn't work out pls share a minimal example, I'll check it :)