keras icon indicating copy to clipboard operation
keras copied to clipboard

Why my code get better f1_score on older versions

Open DSOTM-pf opened this issue 3 years ago • 3 comments
trafficstars

System information.

  • Have I written custom code (as opposed to using a stock example script provided in Keras): No
  • OS Platform and Distribution: Linux Ubuntu 16.04
  • TensorFlow installed from binary
  • TensorFlow version (use command below): 2.10.0 , 2.7.0,
  • Python version: 3.7.10
  • GPU model and memory: TITAN V

Describe the problem. Hi, I adopt different Tensorflow and Keras versions with my code. But it returns a better f_1 Score on older versions.

To avoid the effect of random seeds, I try to set the random seed using this approach. https://github.com/keras-team/keras/issues/2743#issuecomment-904169662

import random
random.seed(2022) 
np.random.seed(2022)  
tf.random.set_seed(2022) 

I also added shuffle = False in model.fit()

Describe the current behavior. The reproduction results are as follows:

Tensorflow Keras F1 Score
2.10.0 2.10.0 0.5369807497467073
2.7.0 2.7.0 0.29846938775510207
2.4.0 2.4.0 0.7376928728875827

Describe the expected behavior. Why the new version has worse results? The latest version of the score should be better.

Standalone code to reproduce the issue. I post the code on colad. click me to reproduce

click me to download dataset

DSOTM-pf avatar Nov 04 '22 08:11 DSOTM-pf

https://github.com/tensorflow/tensorflow/issues/58431

DSOTM-pf avatar Nov 04 '22 08:11 DSOTM-pf

@DSOTM-pf, I tried to execute the mentioned code with the dataset attached and was able to fetch the better f1-score than the older versions. Kindly find the gist of it here and let us know if anything required here.

tilakrayal avatar Nov 07 '22 09:11 tilakrayal

Hi, I noticed that your execution achieved a 1.0 F1 score. But when I execute this code with tf-nightly, it returns 0.5369807497467073. I only uploaded the dataset on colab and modified the folder to '***.csv'. May I know what version of TensorFlow you are using when running, and whether the lower version also has a 1.0 f1 score?

DSOTM-pf avatar Nov 07 '22 12:11 DSOTM-pf

@DSOTM-pf, I tried to execute the given code again on tf-nightly and I was still able to get the same f1 1.0. The tf-nightly version which I am using is 2.12.0-dev20221110. Kindly find the gist of it here. Thank you!

tilakrayal avatar Nov 11 '22 07:11 tilakrayal

@tilakrayal I achieve the same @DSOTM-pf f1_score: 0.5369807497467073 on your colab gist with today tf-nightly-2.12.0.dev20221111.

bhack avatar Nov 12 '22 00:11 bhack