IntegerLookup layer performance issue when inited with vocabulary about tensorflow

[sushreebarsa]

Hi, @jamm1985!
Could you please have a look at this gist where I could reproduce this output ? Kindly confirm the results?

8.22 ms ± 363 µs per loop (mean ± std. dev. of 7 runs, 100 loops each)
509 µs ± 91.1 µs per loop (mean ± std. dev. of 7 runs, 1000 loops each)
421 µs ± 7.34 µs per loop (mean ± std. dev. of 7 runs, 1000 loops each)

Thank you!

[jamm1985]

Hi @sushreebarsa!

This is the same code snippet from the bug report. My timings are slightly different as it depends on system load, but I confirm that tf.keras.layers.IntegerLookup(vocabulary=vocab.tolist()) is an order of magnitude slower.

[jamm1985]

@sushreebarsa from your colab gist:

print(tf.__version__)
# 2.15.0

this bug also affects version 2.15.

[SuryanarayanaY]

Hi @jamm1985 ,

I have checked with tf-nightly and there seems no performance issue. Please check the attached gist and confirm.

[jamm1985]

@SuryanarayanaY

I confirm that this bug does not affect tf-nightly.