ma7555 / kerasgen Goto Github PK

Describe the bug
Importing kerasgen as from kerasgen import balanced_image_dataset fails with ImportError on the latest version of tensorflow (2.9.1).

To Reproduce
Steps to reproduce the behavior:

1. mkdir /tmp/repro && cd /tmp/repro
2. python -m virtualenv .venv && source .venv/bin/activate
3. python -m pip install tensorflow
4. python -c 'from kerasgen import balanced_image_dataset'

Observed

Traceback (most recent call last):
  File "<string>", line 1, in <module>
  File "/tmp/repro/.venv/lib/python3.8/site-packages/kerasgen/balanced_image_dataset.py", line 8, in <module>
    from keras.preprocessing import dataset_utils
ImportError: cannot import name 'dataset_utils' from 'keras.preprocessing' (/home/azureuser/asru/tmp/.venv/lib/python3.8/site-packages/keras/preprocessing/__init__.py)

Expected behavior
kerasgen imports module successfully.

Is it possible to pass a preprocessing function to the batch generator? I want to normalize images during the training

Hi! After training with balanced_image_dataset_from_directory , I have tried to test the validation dataset that your app made.
What I get from it is the following:

Found 161866 files belonging to 9198 classes.
Using 129493 files for training.
Found 161866 files belonging to 9198 classes.
Using 32373 files for validation.

I thought that every class would be used in validation database, but what I get using:
x = np.concatenate([x for x, y in val_ds], axis=0)
y = np.concatenate([y for x, y in val_ds], axis=0)

The variable y has just 4500 unique classes. Where are the ones missing of those 9198 classes? It's not about not having enough photos because the class 3 (for example, one missing in validation dataset), it has more than 10 photos. The parameters of my balanced_image_dataset are these ones:

val_ds = balanced_image_dataset_from_directory(
    directory, num_classes_per_batch=64,
    num_images_per_class=4, image_size=(160, 160),
    seed=6, validation_split=0.2, subset='validation',
    safe_triplet=True)