Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Jump to bottom

MixUp + Cutmix implementation is (badly!) incorrect #13490

Open
dibyaghosh opened this issue Dec 11, 2024 · 3 comments · May be fixed by #13525
Open

MixUp + Cutmix implementation is (badly!) incorrect #13490

dibyaghosh opened this issue Dec 11, 2024 · 3 comments · May be fixed by #13525
Assignees
@LakshmiKalaKadali
Labels
models:official models that come under official repository stale stat:awaiting response Waiting on input from the contributor

Comments

@dibyaghosh
Copy link

dibyaghosh commented Dec 11, 2024 •
edited
Loading

The code for sampling from a beta distribution in tfm.vision.augment.MixupAndCutmix._sample_from_beta is incorrect:

It should be

  @staticmethod
  def _sample_from_beta(alpha, beta, shape):
    sample_alpha = tf.random.gamma(shape, alpha, beta=1.0) # not: tf.random.gamma(shape, 1.0, beta=alpha)
    sample_beta = tf.random.gamma(shape, beta, beta=1.0) # not: tf.random.gamma(shape, 1.0, beta=beta)
    return sample_alpha / (sample_alpha + sample_beta)

https://github.com/tensorflow/models/blob/8c5c79c27c4bb2843308bf8623daa24f4f3cbe62/official/vision/ops/augment.py#L2712C1-L2713C1

This results in very bad mixup samples:

image

# code for plot
import tensorflow_models as tfm
import tensorflow as tf
import numpy as np
import matplotlib.pyplot as plt
a = tf.range(1000)
x = tfm.vision.augment.MixupAndCutmix._sample_from_beta(0.2, 0.2, tf.shape( tf.range(1000))).numpy()
sns.kdeplot(x, clip=(0, 1))
x = np.random.beta(0.2, 0.2, 100000)
sns.kdeplot(x, clip=(0, 1))
plt.show()
@dibyaghosh dibyaghosh mentioned this issue Dec 11, 2024
Closed
8 tasks
@LakshmiKalaKadali LakshmiKalaKadali self-assigned this Dec 11, 2024
@laxmareddyp laxmareddyp added the models:official models that come under official repository label Dec 11, 2024
@LakshmiKalaKadali
Copy link
Collaborator

Hi @dibyaghosh,

Thanks for bringing out the issue and raising PR. I have reproduced the issue and observed the deviation between beta distribution in tfm.vision.augment.MixupAndCutmix._sample_from_beta and np.random.beta(0.2, 0.2, 100000) . Here is the [gist].(https://colab.sandbox.google.com/gist/LakshmiKalaKadali/06533824610d6e85ea4aa3c6399819e6/tf_model_13490.ipynb). The observations are attached:
image
image

Please track the PR raised for this issue as well.

Thank You

@LakshmiKalaKadali LakshmiKalaKadali added the stat:awaiting response Waiting on input from the contributor label Dec 12, 2024
@github-actions GitHub Actions
Copy link

This issue has been marked stale because it has no recent activity since 7 days. It will be closed if no further activity occurs. Thank you.

@github-actions github-actions bot added the stale label Dec 20, 2024
@github-actions GitHub Actions
Copy link

This issue was closed due to lack of activity after being marked stale for past 7 days.

LakshmiKalaKadali added a commit that referenced this issue Jan 30, 2025
@LakshmiKalaKadali
Update beta sampling code in augment.py
bdbcbaa 
The function `_sample_from_beta(alpha, beta, shape)` in `MixupAndCutmix` class, is not having the same functionality as `numpy.random.beta`. So `tfm.vision.augment.MixupAndCutmix._sample_from_beta(0.2, 0.2, tf.shape( tf.range(10000))).numpy()` is also deviating as well. So suggesting the fix keeping `alpha=alpha, beta=1.0` in  `_sample_from_beta`. The reproduced [gist](https://colab.sandbox.google.com/gist/LakshmiKalaKadali/06533824610d6e85ea4aa3c6399819e6/tf_model_13490.ipynb#scrollTo=zSlE-3YDjL91) also attached. 

This PR closes [#13490](#13490)

Thank You
@LakshmiKalaKadali LakshmiKalaKadali linked a pull request Jan 30, 2025 that will close this issue
Open
15 tasks
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Assignees

@LakshmiKalaKadali LakshmiKalaKadali

Labels
models:official models that come under official repository stale stat:awaiting response Waiting on input from the contributor
Projects
None yet
Milestone
No milestone
3 participants
@dibyaghosh @laxmareddyp @LakshmiKalaKadali