CLIP Variants

The CLIP model was developed by researchers at OpenAI to learn about what contributes to robustness in computer vision tasks. The model was also developed to test the ability of models to generalize to arbitrary image classification tasks in a zero-shot manner. It was not developed for general model deployment - to deploy models like CLIP, researchers will first need to carefully study their capabilities in relation to the specific context they’re being deployed within.

See the original CLIP Model Card for more details on limitations and biases.

This repository holds OpenAI's CLIP models converted into many other variants, see below for more details.

Disclaimer & License

I haven't done many tests on these conversions. I've briefly tried the float16 versions, which seem very similar to the original float32, however the similarity seems to drop more with the qint8/quint8 versions as expected. I couldn't try qint8 as it seemed unsupported for some operations, but I'm including it for completeness. From a brief test the quint8 version seemed to work fine.

The license for the conversion code is MIT, the license for the models is the same as the original license for the OpenAI models (🤷‍♂️). I have no affiliation with OpenAI.

Acknowledgements

Example

See example.py

❯ python .\example.py
Loading visual model: models/clip-vit-base-patch32-visual-float16.onnx
Visual inference ready, input size 224, type tensor(float16)
Images shape: (2, 3, 224, 224)
Embeddings shape: (2, 512)

Loading textual model: models/clip-vit-base-patch32-textual-float16.onnx
Textual inference ready, input size 77, type tensor(int32)
Texts shape: (14, 77)
Embeddings shape: (14, 512)

flowers.jpg
  similarity  bar chart    text
------------  -----------  ---------------------------------------------------------------
    0.294922  >>>>>>>>     a close up photo of a cherry blossom
    0.267578  >>>>>>>>     cherry blossom
    0.249878  >>>>>>>      flowers
    0.242554  >>>>>>>      a photo taken on a bright and sunny day
    0.228882  >>>>>>       bees
    0.222778  >>>>>>       plant
    0.216187  >>>>>>       a photo taken on a dark and cloudy day
    0.201538  >>>>>>       ruhrgebiet
    0.196655  >>>>>        processing plant
    0.192139  >>>>>        a photo taken at midnight
    0.18689   >>>>>        industry
    0.177856  >>>>>        cars
    0.176636  >>>>>        dogs and cats
    0.111267  >>>          a large industrial plant with many pipes, walkways and railings

heavy-industry.jpg
  similarity  bar chart    text
------------  -----------  ---------------------------------------------------------------
    0.336182  >>>>>>>>>>   a large industrial plant with many pipes, walkways and railings
    0.316895  >>>>>>>>>    processing plant
    0.302002  >>>>>>>>>    industry
    0.27417   >>>>>>>>     ruhrgebiet
    0.254883  >>>>>>>      plant
    0.22876   >>>>>>       a photo taken on a dark and cloudy day
    0.219482  >>>>>>       a photo taken on a bright and sunny day
    0.211304  >>>>>>       a photo taken at midnight
    0.198608  >>>>>        cars
    0.190552  >>>>>        flowers
    0.181885  >>>>>        bees
    0.180542  >>>>>        cherry blossom
    0.174438  >>>>>        dogs and cats
    0.14917   >>>>         a close up photo of a cherry blossom

Parameters

The only format supported right now is Open Neural Network Exchange (ONNX).

All the currently available OpenAI models have been converted. Some of the IDs were taken from Open AI models on Hugging Face, others were made up following the same format.

Model name	Model ID
RN50	resnet-50
RN101	resnet-101
RN50x4	resnet-50x4
RN50x16	resnet-50x16
RN50x64	resnet-50x64
RN50	resnet-50
RN50	resnet-50
RN50	resnet-50
ViT-B/16	vit-base-patch16
ViT-B/32	vit-base-patch32
ViT-L/14	vit-large-patch14
ViT-L/14@336px	vit-large-patch14-336

As CLIP is a multimodal model, the original models are split into two separate "modes", one for processing images and the other for processing text.

Mode
visual
textual

The models were converted into multiple data types as well.

Data Type
float16
qint8
quint8

Variants

Path	Model ID	Mode	Data Type	Available	Size (MB)
models/clip-resnet-50-visual-float32.onnx	resnet-50	visual	float32	✅	153
models/clip-resnet-50-visual-float16.onnx	resnet-50	visual	float16	✅	77
models/clip-resnet-50-visual-qint8.onnx	resnet-50	visual	qint8	✅	39
models/clip-resnet-50-visual-quint8.onnx	resnet-50	visual	quint8	✅	39
models/clip-resnet-50-textual-float32.onnx	resnet-50	textual	float32	✅	255
models/clip-resnet-50-textual-float16.onnx	resnet-50	textual	float16	✅	128
models/clip-resnet-50-textual-qint8.onnx	resnet-50	textual	qint8	✅	64
models/clip-resnet-50-textual-quint8.onnx	resnet-50	textual	quint8	✅	64
models/clip-resnet-101-visual-float32.onnx	resnet-101	visual	float32	✅	225
models/clip-resnet-101-visual-float16.onnx	resnet-101	visual	float16	✅	112
models/clip-resnet-101-visual-qint8.onnx	resnet-101	visual	qint8	✅	57
models/clip-resnet-101-visual-quint8.onnx	resnet-101	visual	quint8	✅	57
models/clip-resnet-101-textual-float32.onnx	resnet-101	textual	float32	✅	254
models/clip-resnet-101-textual-float16.onnx	resnet-101	textual	float16	✅	127
models/clip-resnet-101-textual-qint8.onnx	resnet-101	textual	qint8	✅	64
models/clip-resnet-101-textual-quint8.onnx	resnet-101	textual	quint8	✅	64
models/clip-resnet-50x4-visual-float32.onnx	resnet-50x4	visual	float32	✅	348
models/clip-resnet-50x4-visual-float16.onnx	resnet-50x4	visual	float16	✅	174
models/clip-resnet-50x4-visual-qint8.onnx	resnet-50x4	visual	qint8	✅	88
models/clip-resnet-50x4-visual-quint8.onnx	resnet-50x4	visual	quint8	✅	88
models/clip-resnet-50x4-textual-float32.onnx	resnet-50x4	textual	float32	✅	365
models/clip-resnet-50x4-textual-float16.onnx	resnet-50x4	textual	float16	✅	183
models/clip-resnet-50x4-textual-qint8.onnx	resnet-50x4	textual	qint8	✅	92
models/clip-resnet-50x4-textual-quint8.onnx	resnet-50x4	textual	quint8	✅	92
models/clip-resnet-50x16-visual-float32.onnx	resnet-50x16	visual	float32	✅	669
models/clip-resnet-50x16-visual-float16.onnx	resnet-50x16	visual	float16	✅	335
models/clip-resnet-50x16-visual-qint8.onnx	resnet-50x16	visual	qint8	✅	169
models/clip-resnet-50x16-visual-quint8.onnx	resnet-50x16	visual	quint8	✅	169
models/clip-resnet-50x16-textual-float32.onnx	resnet-50x16	textual	float32	✅	495
models/clip-resnet-50x16-textual-float16.onnx	resnet-50x16	textual	float16	✅	248
models/clip-resnet-50x16-textual-qint8.onnx	resnet-50x16	textual	qint8	✅	124
models/clip-resnet-50x16-textual-quint8.onnx	resnet-50x16	textual	quint8	✅	124
models/clip-resnet-50x64-visual-float32.onnx	resnet-50x64	visual	float32	✅	1681
models/clip-resnet-50x64-visual-float16.onnx	resnet-50x64	visual	float16	✅	840
models/clip-resnet-50x64-visual-qint8.onnx	resnet-50x64	visual	qint8	✅	424
models/clip-resnet-50x64-visual-quint8.onnx	resnet-50x64	visual	quint8	✅	424
models/clip-resnet-50x64-textual-float32.onnx	resnet-50x64	textual	float32	✅	812
models/clip-resnet-50x64-textual-float16.onnx	resnet-50x64	textual	float16	✅	406
models/clip-resnet-50x64-textual-qint8.onnx	resnet-50x64	textual	qint8	✅	204
models/clip-resnet-50x64-textual-quint8.onnx	resnet-50x64	textual	quint8	✅	204
models/clip-resnet-50-visual-float32.onnx	resnet-50	visual	float32	✅	153
models/clip-resnet-50-visual-float16.onnx	resnet-50	visual	float16	✅	77
models/clip-resnet-50-visual-qint8.onnx	resnet-50	visual	qint8	✅	39
models/clip-resnet-50-visual-quint8.onnx	resnet-50	visual	quint8	✅	39
models/clip-resnet-50-textual-float32.onnx	resnet-50	textual	float32	✅	255
models/clip-resnet-50-textual-float16.onnx	resnet-50	textual	float16	✅	128
models/clip-resnet-50-textual-qint8.onnx	resnet-50	textual	qint8	✅	64
models/clip-resnet-50-textual-quint8.onnx	resnet-50	textual	quint8	✅	64
models/clip-resnet-50-visual-float32.onnx	resnet-50	visual	float32	✅	153
models/clip-resnet-50-visual-float16.onnx	resnet-50	visual	float16	✅	77
models/clip-resnet-50-visual-qint8.onnx	resnet-50	visual	qint8	✅	39
models/clip-resnet-50-visual-quint8.onnx	resnet-50	visual	quint8	✅	39
models/clip-resnet-50-textual-float32.onnx	resnet-50	textual	float32	✅	255
models/clip-resnet-50-textual-float16.onnx	resnet-50	textual	float16	✅	128
models/clip-resnet-50-textual-qint8.onnx	resnet-50	textual	qint8	✅	64
models/clip-resnet-50-textual-quint8.onnx	resnet-50	textual	quint8	✅	64
models/clip-resnet-50-visual-float32.onnx	resnet-50	visual	float32	✅	153
models/clip-resnet-50-visual-float16.onnx	resnet-50	visual	float16	✅	77
models/clip-resnet-50-visual-qint8.onnx	resnet-50	visual	qint8	✅	39
models/clip-resnet-50-visual-quint8.onnx	resnet-50	visual	quint8	✅	39
models/clip-resnet-50-textual-float32.onnx	resnet-50	textual	float32	✅	255
models/clip-resnet-50-textual-float16.onnx	resnet-50	textual	float16	✅	128
models/clip-resnet-50-textual-qint8.onnx	resnet-50	textual	qint8	✅	64
models/clip-resnet-50-textual-quint8.onnx	resnet-50	textual	quint8	✅	64
models/clip-vit-base-patch16-visual-float32.onnx	vit-base-patch16	visual	float32	✅	345
models/clip-vit-base-patch16-visual-float16.onnx	vit-base-patch16	visual	float16	✅	173
models/clip-vit-base-patch16-visual-qint8.onnx	vit-base-patch16	visual	qint8	✅	87
models/clip-vit-base-patch16-visual-quint8.onnx	vit-base-patch16	visual	quint8	✅	87
models/clip-vit-base-patch16-textual-float32.onnx	vit-base-patch16	textual	float32	✅	254
models/clip-vit-base-patch16-textual-float16.onnx	vit-base-patch16	textual	float16	✅	127
models/clip-vit-base-patch16-textual-qint8.onnx	vit-base-patch16	textual	qint8	✅	64
models/clip-vit-base-patch16-textual-quint8.onnx	vit-base-patch16	textual	quint8	✅	64
models/clip-vit-base-patch32-visual-float32.onnx	vit-base-patch32	visual	float32	✅	352
models/clip-vit-base-patch32-visual-float16.onnx	vit-base-patch32	visual	float16	✅	176
models/clip-vit-base-patch32-visual-qint8.onnx	vit-base-patch32	visual	qint8	✅	89
models/clip-vit-base-patch32-visual-quint8.onnx	vit-base-patch32	visual	quint8	✅	89
models/clip-vit-base-patch32-textual-float32.onnx	vit-base-patch32	textual	float32	✅	254
models/clip-vit-base-patch32-textual-float16.onnx	vit-base-patch32	textual	float16	✅	127
models/clip-vit-base-patch32-textual-qint8.onnx	vit-base-patch32	textual	qint8	✅	64
models/clip-vit-base-patch32-textual-quint8.onnx	vit-base-patch32	textual	quint8	✅	64
models/clip-vit-large-patch14-visual-float32.onnx	vit-large-patch14	visual	float32	✅	1216
models/clip-vit-large-patch14-visual-float16.onnx	vit-large-patch14	visual	float16	✅	608
models/clip-vit-large-patch14-visual-qint8.onnx	vit-large-patch14	visual	qint8	✅	306
models/clip-vit-large-patch14-visual-quint8.onnx	vit-large-patch14	visual	quint8	✅	306
models/clip-vit-large-patch14-textual-float32.onnx	vit-large-patch14	textual	float32	✅	495
models/clip-vit-large-patch14-textual-float16.onnx	vit-large-patch14	textual	float16	✅	248
models/clip-vit-large-patch14-textual-qint8.onnx	vit-large-patch14	textual	qint8	✅	124
models/clip-vit-large-patch14-textual-quint8.onnx	vit-large-patch14	textual	quint8	✅	124
models/clip-vit-large-patch14-336-visual-float32.onnx	vit-large-patch14-336	visual	float32	✅	1217
models/clip-vit-large-patch14-336-visual-float16.onnx	vit-large-patch14-336	visual	float16	✅	609
models/clip-vit-large-patch14-336-visual-qint8.onnx	vit-large-patch14-336	visual	qint8	✅	307
models/clip-vit-large-patch14-336-visual-quint8.onnx	vit-large-patch14-336	visual	quint8	✅	307
models/clip-vit-large-patch14-336-textual-float32.onnx	vit-large-patch14-336	textual	float32	✅	495
models/clip-vit-large-patch14-336-textual-float16.onnx	vit-large-patch14-336	textual	float16	✅	248
models/clip-vit-large-patch14-336-textual-qint8.onnx	vit-large-patch14-336	textual	qint8	✅	124
models/clip-vit-large-patch14-336-textual-quint8.onnx	vit-large-patch14-336	textual	quint8	✅	124