Keras - understanding ImageDataGenerator dimensions

I'm trying to implement custom object detection by taking a trained YOLOv2 model in Keras, removing the last layer and retraining it with new data. I'm confused about how to feed the data to Keras, though. I have annotated a bunch of pictures with bounding boxes using the YOLO annotation, and put them in two separate folders (images where the .jpgs reside and annots where the .txt annotations are).

I also removed the last layer from the model and added a custom one (I'm trying to predict bounding boxes for 2 classes).

I'm trying to pass my data with an ImageDataGenerator, as my dataset is quite small.

I have the following input objects:

np.shape(train_images) # this contains RGB data from 79 pictures
(79, 1, 608, 608, 3)
np.shape(train_y)
(79,)

I'm trying to pass these to the ImageDataGenerator, but I get an error:

train_datagen = ImageDataGenerator(
rotation_range=20,
width_shift_range=0.2,
height_shift_range=0.2,
horizontal_flip=True,
fill_mode='nearest')

train_generator = train_datagen.flow(
train_images,
train_y)

ValueError: x (images tensor) and y (labels) should have the same length. Found: x.shape = (1, 608, 608, 3), y.shape = (79,)

I don't understand what the problem is. Somehow the first dimension of my images data is completely gone and thus does not match... What's wrong with it?

I think your train_images array should have shape (79, 698, 608, 3). The generator works through each of the first dimensions of those arrays, so is passing a batch of 4d numpy arrays, instead of a batch of 3d numpy arrays.

You can try seeing if that helps, using numpy.squeeze(), like this:

In [1]: import numpy as np

In [2]: a = np.random.randint(0, 10, (2, 1, 10, 10, 3))

In [3]: np.squeeze(a, axis=1).shape
Out[3]: (2, 10, 10, 3)

So be sure to set the axis=1 argument. Then just pass the updated array as you were doing.

• Thanks, I found a line in the image prep code where an additional dimension was added with the comment "adding batch dimension", I'm not sure what it meant but removing it meant the images come in the shape of (79, 608, 608, 3) and that was fine. I'm still having issues with my target values, but I'll create another question for that. – lte__ Jul 4 '19 at 14:27