Example or support for .fit_generator about tf-explain HOT 4 OPEN

Terribly sorry for not answering earlier. My guess is that the condition np.all(np.argmax(label) == 0) is not met when you use the flow_from_directory, so your validation class zero is empty and hence has shape [0]. To validate this, could you tell what y looks like?

I'm looking for the next release to support ImageDataGenerator and tf.data.Dataset, which should solve this.

from tf-explain.

I have the same problem. Is there support for the ImageDataGenerator in version 2.1.0? I haven't found it yet.

from tf-explain.

A possible solution could look like this:

valid_batches = valid_datagen.flow_from_directory(...)
(x,y) = valid_batches.next()
validation_class_zero = (np.array([
    el for el, label in zip(x,y)
    if np.all(np.argmax(label) == 0)
][0:5]), None)

But when i try to select the first conv layer in InceptionResNetV2, I get the following error:
ValueError: Input 0 of layer conv2d is incompatible with the layer: expected ndim=4, found ndim=1. Full shape received: [0]
And I'm not sure, if this is related to my workaround or something else.

from tf-explain.

Terribly sorry for not answering earlier 😄
Looks like your thought was on point, the condition np.all is aparently not always met.

(x,y) = valid_batches.next()
for el, label in zip(x,y):
    print(el.shape)
    print(label.shape)
    print(np.argmax(label))

validation_class_zero = (np.array([
    el for el, label in zip(x,y)
    if np.all(np.argmax(label) == 0)
][0:5]), None)
print(validation_class_zero)

(300, 300, 3)
(3,)
0
(300, 300, 3)
(3,)
0
(300, 300, 3)
(3,)
1
(300, 300, 3)
(3,)
1
(300, 300, 3)
(3,)
1
(300, 300, 3)
(3,)
1
(array([], dtype=float64), None)

This is probably due to the fact, that I rely on transfer learning and do not have a sufficient amount of samples.

from tf-explain.