Significant Performance Drop When Training Sequential Model Using `tf.data.Dataset.from_generator` about tensorflow HOT 2 OPEN

@nrad,
I can see you are using tensorflow 2.13 which is little older than the latest. I tried with the tf-nightly and couldn't found any difference in the performance with tf.data.dataset.from_generator. Kindly find the gist of it here.

Performance using pd.DataFrame:
20/20 ━━━━━━━━━━━━━━━━━━━━ 1s 2ms/step - accuracy: 0.6222 - loss: 0.6594   
Performance using tf.dataset:
/usr/local/lib/python3.10/dist-packages/keras/src/layers/normalization/batch_normalization.py:143: UserWarning: Do not pass an `input_shape`/`input_dim` argument to a layer. When using Sequential models, prefer using an `Input(shape)` object as the first layer in the model instead.
  super().__init__(**kwargs)
20/20 ━━━━━━━━━━━━━━━━━━━━ 0s 2ms/step - accuracy: 0.6222 - loss: 0.6594  
[0.6589511632919312, 0.6225550174713135]

Thank you!

from tensorflow.

@tilakrayal
Thank you for the quick response!
Sorry but in the code I had posted the actual training line history = model.fit(...) was commented out!

I just tried the code again with the tf-nightly and also with v2.16.1 and unfortunately the I see the issue persisting.
Here is the corrected google colab. I also edited the original post.

from tensorflow.