decathlon / decavision Goto Github PK

To save in Colab, we need to add some lines if we train with TPU because it usually requires to be saved to a Bucket. To overcome the use of a Bucket, add the SaveOptions as follows to the fit() function:

        if export_model:
            save_locally = tf.saved_model.SaveOptions(experimental_io_device='/job:localhost')
            self.model.save(export_model, options=save_locally)
            print('Model exported')

When using tensorflow higher than 2.9.1, B0, B5, B7 model can have this error when fitting

TypeError: Unable to serialize [2.0896919 2.1128857 2.1081853] to JSON. Unrecognized type <class 'tensorflow.python.framework.ops.EagerTensor'>.

So in the requirements.txt, we need to put tensorflow == 2.9.1

1- Currently, I believe the images are not scaled to a [0-1] range when we run the tfrecords generation followed by model training. This is OK for efficientnet models, because preprocessing is done by the model itself. That said, other models, like inception expect a scaled input.

2- I might we wrong, but I believe the scale variable, used for data augmentation and created here, should be the opposite. It should be 1 for models of the efficientnet family (because the model expect a tensor of [0-255] floats), while it should be 255 for the other models.

By default the validation batch size is the same as the training batch size, but sometimes there are not many validation images and this gives an error when the batch size is larger than the number of validation images. I think we should add a check to make sure this does not happen.

Here is what I suggest. Add an extra value in the __init__ for the val_batch_size. If the batch size is smaller than the size of the validation dataset, then this value is the batch size. Otherwise we take something like half the size of the dataset. This will require to add the batch size as an argument in the _get_dataset method and use the correct one when we get a dataset.

Found this bug when trying to unzip 26 000 images from gdrive. The solution I found was to unzip with a command line in the Colab terminal and it did the trick.

!unzip gdrive/MyDrive/image_dataset.zip > /dev/null

Maybe updating to PyDrive2 might solve the issue, but i am not shure !

Resnet is supported by decavision according to the docstring of the ImageClassifier class and has seen in the code here.

However, Resnet seems to be missing from the dict of input_dims

Colab now uses tensorflow 2.5 by default so using 2.4 causes problems when working with TPUs. For example saving a model to .pb format on google storage doesn't work even if it worked in the past. Switching to tensorflow 2.5 seems to fix the issue.