{ "id": 344301, "title": "GCViT: Global Context Vision Transformer", "url": "/competitions/rsna-2022-cervical-spine-fracture-detection/discussion/344301", "author_name": "Awsaf", "post_date": "2022-08-14T18:50:34.025000", "votes": 13, "comment_count": 4, "views": 0, "content": "

\n

Hello everyone,
\nAs you may already know, NVIDIA has recently published (20 June 2022) its paper, GCViT: Global Context Vision Transformer which outperforms ConvNeXt and SwinTransformer . As this competition is a Classification Problem (give or take), I think GCViT can be useful here. You are welcome to explore this model for this competition.

\n
\n

There was some issue with their weights and they recently released their final version of ImageNet pretrained weights which brought over ~1% improvement over older weights.

\n
\n

To ease things up, I've implemented this model using TensorFlow and created an open-source library gcvit-tf. I've also made a notebook to help you get started with GCViT,.

\n

Resources:

\n\n

Features of gcvit-tf:

\n\n

Supported Models

\n

The official codebase had some issue which has been fixed recently (27 July 2022). Here's the result of ported weights on ImageNetV2-Test data,

\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n
ModelAcc@1Acc@5#Params
GCViT-XXTiny668712M
GCViT-XTiny698820M
GCViT-Tiny719028M
GCViT-Small729051M
GCViT-Base739190M
\n

Usage

\n

Install Library

\n
pip install gcvit\n
\n

Load model using the following codes,

\n
from gcvit import GCViTTiny\nmodel = GCViTTiny(pretrain=True)\n
\n

For feature extraction:

\n
model.reset_classifier(num_classes=0, head_act=None)\nfeature = model(img)\nprint(feature.shape)\n
\n

Feature:

\n
(None, 512)\n
\n

For feature map:

\n
feature = model.forward_features(img)\nprint(feature.shape)\n
\n

Feature map:

\n
(None, 7, 7, 512)\n
", "messages": [ { "id": 1898717, "postDate": "2022-08-14T18:50:34.027Z", "content": "

\n

Hello everyone,
\nAs you may already know, NVIDIA has recently published (20 June 2022) its paper, GCViT: Global Context Vision Transformer which outperforms ConvNeXt and SwinTransformer . As this competition is a Classification Problem (give or take), I think GCViT can be useful here. You are welcome to explore this model for this competition.

\n
\n

There was some issue with their weights and they recently released their final version of ImageNet pretrained weights which brought over ~1% improvement over older weights.

\n
\n

To ease things up, I've implemented this model using TensorFlow and created an open-source library gcvit-tf. I've also made a notebook to help you get started with GCViT,.

\n

Resources:

\n\n

Features of gcvit-tf:

\n\n

Supported Models

\n

The official codebase had some issue which has been fixed recently (27 July 2022). Here's the result of ported weights on ImageNetV2-Test data,

\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n
ModelAcc@1Acc@5#Params
GCViT-XXTiny668712M
GCViT-XTiny698820M
GCViT-Tiny719028M
GCViT-Small729051M
GCViT-Base739190M
\n

Usage

\n

Install Library

\n
pip install gcvit\n
\n

Load model using the following codes,

\n
from gcvit import GCViTTiny\nmodel = GCViTTiny(pretrain=True)\n
\n

For feature extraction:

\n
model.reset_classifier(num_classes=0, head_act=None)\nfeature = model(img)\nprint(feature.shape)\n
\n

Feature:

\n
(None, 512)\n
\n

For feature map:

\n
feature = model.forward_features(img)\nprint(feature.shape)\n
\n

Feature map:

\n
(None, 7, 7, 512)\n
", "rawMarkdown": "\n\nHello everyone,\nAs you may already know, NVIDIA has recently published (20 June 2022) its paper, **GCViT: Global Context Vision Transformer** which outperforms **ConvNeXt** and **SwinTransformer** . As this competition is a Classification Problem (give or take), I think **GCViT** can be useful here. You are welcome to explore this model for this competition.\n\n> There was some issue with their weights and they recently released their **final version of ImageNet pretrained weights** which brought over `~1%` improvement over older weights.\n\nTo ease things up, I've implemented this model using **TensorFlow** and created an open-source library [gcvit-tf](https://github.com/awsaf49/gcvit-tf). I've also made a notebook to help you get started with **GCViT**,. \n\n## Resources:\n* **Notebook:** [GCViT: Global Context Vision Transformer](https://www.kaggle.com/code/awsaf49/gcvit-global-context-vision-transformer)\n * Refer to [Reusability](https://www.kaggle.com/code/awsaf49/gcvit-global-context-vision-transformer#6.-Reusability) section if are interested in using this model for some other tasks (classification or feature_extraction) on another dataset.\n* **GitHub Code:** [gcvit-tf](https://github.com/awsaf49/gcvit-tf)\n * Feel free to send **Pull-Request** if you are interested in contributing to this project.\n\n## Features of [gcvit-tf](library):\n* This library loads **ImageNet** weights from the official repo.\n* Also, it has `timm` like features such as`forward_features`, `forward_head`, and `reset_classifier` which might come in handy.\n* It can be used in both **GPU** and **TPU**.\n\n## Supported Models\nThe official codebase had some issue which has been fixed recently (27 July 2022). Here's the result of ported weights on **ImageNetV2-Test** data,\n\n| Model | Acc@1 | Acc@5 | #Params |\n|--------------|-------|-------|---------|\n| GCViT-XXTiny | 66 | 87 | 12M |\n| GCViT-XTiny | 69 | 88 | 20M |\n| GCViT-Tiny | 71 | 90 | 28M |\n| GCViT-Small | 72 | 90 | 51M |\n| GCViT-Base | 73 | 91 | 90M |\n\n## Usage\nInstall Library\n```shell\npip install gcvit\n```\nLoad model using the following codes,\n```py\nfrom gcvit import GCViTTiny\nmodel = GCViTTiny(pretrain=True)\n```\n\nFor feature extraction:\n```py\nmodel.reset_classifier(num_classes=0, head_act=None)\nfeature = model(img)\nprint(feature.shape)\n```\nFeature:\n```py\n(None, 512)\n```\nFor feature map:\n```py\nfeature = model.forward_features(img)\nprint(feature.shape)\n```\nFeature map:\n```py\n(None, 7, 7, 512)\n```\n\n", "votes": 13 }, { "id": 1904577, "postDate": "2022-08-18T09:56:29.127Z", "content": "

Repeat. Same discussion thread, at a same on going competition.

\n

4 days ago and 22 days ago

\n

\"image\"

\n

\"image\"

", "rawMarkdown": "Repeat. Same discussion thread, at a same on going competition.\n\n4 days ago and 22 days ago\n\n![image](https://user-images.githubusercontent.com/45315076/185367129-4e4eaf21-59e3-4b7e-a91f-786de2144b30.png)\n\n![image](https://user-images.githubusercontent.com/45315076/185367451-99916fc8-8bc6-4ffd-91f0-45bcad57a752.png)", "replies": [ { "id": 1908134, "postDate": "2022-08-21T11:46:04.800Z", "content": "

The title same but the purpose is different. I've mentioned the purpose in this post,

\n
\n

There was some issue with their weights and they recently released their final version of ImageNet pretrained weights which brought over ~1% improvement over older weights.

\n
", "rawMarkdown": "The title same but the purpose is different. I've mentioned the purpose in this post,\n> There was some issue with their weights and they recently released their final version of ImageNet pretrained weights which brought over ~1% improvement over older weights." }, { "id": 1908140, "postDate": "2022-08-21T11:50:07.857Z", "content": "

Another purpose is to make people in this competition aware of GCViT. People often publish similar publish similar notebook just changing the dataset for a new competition so I don't see any problem here.

", "rawMarkdown": "Another purpose is to make people in this competition aware of **GCViT**. People often publish similar publish similar notebook just changing the dataset for a new competition so I don't see any problem here." } ] }, { "id": 1900952, "postDate": "2022-08-16T11:28:33.240Z", "content": "

It is incredible that each time a new paper comes out we can just find it on your github.
\nYour work is impressive as hell.

\n

As for the model itself: Those performance plots of the models are just SICK! They outperform everyone.

\n

The Devastator.

", "rawMarkdown": "It is incredible that each time a new paper comes out we can just find it on your github. \n**Your work is impressive as hell.**\n\nAs for the model itself: Those performance plots of the models are just SICK! They outperform everyone.\n\n\nThe Devastator.\n" } ], "comments": [ { "id": 1904577, "author_name": "Simon Alerdic", "author_url": "", "post_date": "2022-08-18T09:56:29.127000", "content": "

Repeat. Same discussion thread, at a same on going competition.

\n

4 days ago and 22 days ago

\n

\"image\"

\n

\"image\"

", "votes": 0, "replies": [ { "id": 1908134, "author_name": "Awsaf", "author_url": "", "post_date": "2022-08-21T11:46:04.800000", "content": "

The title same but the purpose is different. I've mentioned the purpose in this post,

\n
\n

There was some issue with their weights and they recently released their final version of ImageNet pretrained weights which brought over ~1% improvement over older weights.

\n
", "votes": 0, "replies": [] }, { "id": 1908140, "author_name": "Awsaf", "author_url": "", "post_date": "2022-08-21T11:50:07.857000", "content": "

Another purpose is to make people in this competition aware of GCViT. People often publish similar publish similar notebook just changing the dataset for a new competition so I don't see any problem here.

", "votes": 0, "replies": [] } ] }, { "id": 1900952, "author_name": "The Devastator", "author_url": "", "post_date": "2022-08-16T11:28:33.240000", "content": "

It is incredible that each time a new paper comes out we can just find it on your github.
\nYour work is impressive as hell.

\n

As for the model itself: Those performance plots of the models are just SICK! They outperform everyone.

\n

The Devastator.

", "votes": 0, "replies": [] } ], "raw_markdown_by_id": { "1898717": "\n\nHello everyone,\nAs you may already know, NVIDIA has recently published (20 June 2022) its paper, **GCViT: Global Context Vision Transformer** which outperforms **ConvNeXt** and **SwinTransformer** . As this competition is a Classification Problem (give or take), I think **GCViT** can be useful here. You are welcome to explore this model for this competition.\n\n> There was some issue with their weights and they recently released their **final version of ImageNet pretrained weights** which brought over `~1%` improvement over older weights.\n\nTo ease things up, I've implemented this model using **TensorFlow** and created an open-source library [gcvit-tf](https://github.com/awsaf49/gcvit-tf). I've also made a notebook to help you get started with **GCViT**,. \n\n## Resources:\n* **Notebook:** [GCViT: Global Context Vision Transformer](https://www.kaggle.com/code/awsaf49/gcvit-global-context-vision-transformer)\n * Refer to [Reusability](https://www.kaggle.com/code/awsaf49/gcvit-global-context-vision-transformer#6.-Reusability) section if are interested in using this model for some other tasks (classification or feature_extraction) on another dataset.\n* **GitHub Code:** [gcvit-tf](https://github.com/awsaf49/gcvit-tf)\n * Feel free to send **Pull-Request** if you are interested in contributing to this project.\n\n## Features of [gcvit-tf](library):\n* This library loads **ImageNet** weights from the official repo.\n* Also, it has `timm` like features such as`forward_features`, `forward_head`, and `reset_classifier` which might come in handy.\n* It can be used in both **GPU** and **TPU**.\n\n## Supported Models\nThe official codebase had some issue which has been fixed recently (27 July 2022). Here's the result of ported weights on **ImageNetV2-Test** data,\n\n| Model | Acc@1 | Acc@5 | #Params |\n|--------------|-------|-------|---------|\n| GCViT-XXTiny | 66 | 87 | 12M |\n| GCViT-XTiny | 69 | 88 | 20M |\n| GCViT-Tiny | 71 | 90 | 28M |\n| GCViT-Small | 72 | 90 | 51M |\n| GCViT-Base | 73 | 91 | 90M |\n\n## Usage\nInstall Library\n```shell\npip install gcvit\n```\nLoad model using the following codes,\n```py\nfrom gcvit import GCViTTiny\nmodel = GCViTTiny(pretrain=True)\n```\n\nFor feature extraction:\n```py\nmodel.reset_classifier(num_classes=0, head_act=None)\nfeature = model(img)\nprint(feature.shape)\n```\nFeature:\n```py\n(None, 512)\n```\nFor feature map:\n```py\nfeature = model.forward_features(img)\nprint(feature.shape)\n```\nFeature map:\n```py\n(None, 7, 7, 512)\n```\n\n", "1904577": "Repeat. Same discussion thread, at a same on going competition.\n\n4 days ago and 22 days ago\n\n![image](https://user-images.githubusercontent.com/45315076/185367129-4e4eaf21-59e3-4b7e-a91f-786de2144b30.png)\n\n![image](https://user-images.githubusercontent.com/45315076/185367451-99916fc8-8bc6-4ffd-91f0-45bcad57a752.png)", "1900952": "It is incredible that each time a new paper comes out we can just find it on your github. \n**Your work is impressive as hell.**\n\nAs for the model itself: Those performance plots of the models are just SICK! They outperform everyone.\n\n\nThe Devastator.\n" } }