Hello guys,
\nAs I have noticed in some discussions like 1 2 3 and Kernels like 4 that SnapMix Augmentation has a nice improvement in the performance, so I have decided to apply it with Tensorflow, as all the implementations I have found were in Pytorch.
\nYou can find the notebook here
I wrote 2 versions of it
\ndef transform(image, label) function and use it like other augmentations there.I will share a notebook that demonstrates using the TPU version so soon.
\nOne problem I faced with the TPU version, is that when apply dataset.map(snapmix), it runs in \"disabled eager execution mode\" and [conv_outputs, predictions] = get_output(x) raise an error, as the map() method runs in \"graph mode\" and then I can't return Numpy values.
\nI tried deifferent solutions like using tf.py_function(func=snapmix, inp=[image, label, model, ...], Tout=[tf.float32,tf.float32]) but couldn't reach to a solution that work.
\nSo if you can edit the TPU version and make it run, that would be interesting😃.
Hello guys,
\nAs I have noticed in some discussions like 1 2 3 and Kernels like 4 that SnapMix Augmentation has a nice improvement in the performance, so I have decided to apply it with Tensorflow, as all the implementations I have found were in Pytorch.
\nYou can find the notebook here
I wrote 2 versions of it
\ndef transform(image, label) function and use it like other augmentations there.I will share a notebook that demonstrates using the TPU version so soon.
\nOne problem I faced with the TPU version, is that when apply dataset.map(snapmix), it runs in \"disabled eager execution mode\" and [conv_outputs, predictions] = get_output(x) raise an error, as the map() method runs in \"graph mode\" and then I can't return Numpy values.
\nI tried deifferent solutions like using tf.py_function(func=snapmix, inp=[image, label, model, ...], Tout=[tf.float32,tf.float32]) but couldn't reach to a solution that work.
\nSo if you can edit the TPU version and make it run, that would be interesting😃.
Note: when I started (3 months ago) building the snapmix augmentation , I have not found any tensorflow implementation of it, all were in pytorch, so I decided to build it, and finished it early but I was busy to reorganize the notebook and share it, so while I was writing this discussion I found that someone was working (nearly the same period) on SnapMix too and have uploaded it a weeks ago in github, so feel free to check it
", "rawMarkdown": "**Note**: when I started (3 months ago) building the snapmix augmentation , I have not found any tensorflow implementation of it, all were in pytorch, so I decided to build it, and finished it early but I was busy to reorganize the notebook and share it, so while I was writing this discussion I found that someone was working (nearly the same period) on SnapMix too and have uploaded it a weeks ago in github, so feel free to check [it](https://github.com/wangermeng2021/SnapMix-tensorflow2)", "votes": null } ], "comments": [ { "id": 1213915, "author_name": "omarzaghlol", "author_url": "", "post_date": "02/22/2021 12:49:33", "content": "Note: when I started (3 months ago) building the snapmix augmentation , I have not found any tensorflow implementation of it, all were in pytorch, so I decided to build it, and finished it early but I was busy to reorganize the notebook and share it, so while I was writing this discussion I found that someone was working (nearly the same period) on SnapMix too and have uploaded it a weeks ago in github, so feel free to check it
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "1213907": "Hello guys,\nAs I have noticed in some discussions like [1](https://www.kaggle.com/c/cassava-leaf-disease-classification/discussion/204631) [2](https://www.kaggle.com/c/cassava-leaf-disease-classification/discussion/202627) [3](https://www.kaggle.com/c/cassava-leaf-disease-classification/discussion/205630) and Kernels like [4](https://www.kaggle.com/shaolihuang/training-with-snapmix/notebook) that **SnapMix** Augmentation has a nice improvement in the performance, so I have decided to apply it with Tensorflow, as all the implementations I have [found](https://github.com/Shaoli-Huang/SnapMix) were in Pytorch.\nYou can find the notebook [here](https://colab.research.google.com/drive/14_Mhw3ATNyrXvwhDK8PdLSGbOgvJ9qzx?usp=sharing)\n\nI wrote 2 versions of it\n1. **GPU** version: works with the normal pipeline of *loading images* then use augmentations (like Keras ones)\n2. **GPU/TPU** version: works with **TFRecords** *data type*, and I built it to be *easy to use* like the **CutMix/MixUp** augmentations in this [kernel](https://www.kaggle.com/itsuki9180/efficientnet-and-cutmixup-with-tpu-train-phase#CutMix-and-MixUp-Augmentation), so you can simply add it in the `def transform(image, label)` function and use it like other augmentations there.\n\nI will share a notebook that demonstrates using the TPU version so soon.\n\n---\nOne problem I faced with the **TPU** version, is that when apply `dataset.map(snapmix)`, it runs in \"*disabled eager execution mode*\" and `[conv_outputs, predictions] = get_output(x)` raise an *error*, as the `map()` method runs in \"*graph mode*\" and then I can't return **Numpy** values.\nI tried deifferent solutions like using `tf.py_function(func=snapmix, inp=[image, label, model, ...], Tout=[tf.float32,tf.float32])` but couldn't reach to a solution that work.\nSo if you can edit the **TPU** version and make it run, that would be interesting😃.", "1213915": "**Note**: when I started (3 months ago) building the snapmix augmentation , I have not found any tensorflow implementation of it, all were in pytorch, so I decided to build it, and finished it early but I was busy to reorganize the notebook and share it, so while I was writing this discussion I found that someone was working (nearly the same period) on SnapMix too and have uploaded it a weeks ago in github, so feel free to check [it](https://github.com/wangermeng2021/SnapMix-tensorflow2)" }, "source": "meta" }