input size

[ireneb612]

my images are 288 288 but i'm using an input size of 32 32 and the training scripts works fine! So does the scripts automatically resize the image?

[TorbenSDJohansen]

The images are automatically resized - both during training, validation, and inference. The --input-size argument is the input size of the models, and not the size of the raw images (which might cause the confusion). Thus, the images are resized before they are passed on to the model.

To follow the logic on where this resizing happens/how it is created, you should look into create_loader, which is used in both train.py, validate.py, and inference.py to create the loaders. Specifically, note how this function uses create_transform to create the transformations that are applied to the images. Depending on the exact arguments you use (whether you are training or not, whether you use --no-aug), the first transformation will either be (during normal training) RandomResizedCropAndInterpolation or (during validation, inference, training with no augmentation) Resize. In both cases, this will resize your images to the size you specify using --input-size.

[ireneb612]

Thank you for your answer, I saw that part od the code but I was wondering if there was a function that just proportionally resized the image without cropping it. Do you know if that is possible? maybe using the no-aug flag? thabk you for your answer!

[TorbenSDJohansen]

No problems! You probably do not want to use the --no-aug flag, as that disables augmentation which you'd likely want (although that depends on your use case). Personally, I have had use cases where I wanted to change the initial transform, and the easiest way I found to do that was to use a modified version of create_loader. Specifically, you can modify L206 to use your own CustomResize by adding:

nb_drop = 1 if is_training and not no_aug else 2
del dataset.transform.transforms[:nb_drop]
dataset.transform.transforms.insert(
    0, CustomResize(input_size),
    )

[ireneb612]

Wow Thank you! I'll try that and let you know if all went well.