OpenAI-Clip: Optimized for Mobile Deployment

Multi-modal foundational model for vision and language tasks like image/text similarity and for zero-shot image classification

Contrastive Language-Image Pre-Training (CLIP) uses a ViT like transformer to get visual features and a causal language model to get the text features. Both the text and visual features can then be used for a variety of zero-shot learning tasks.

This model is an implementation of OpenAI-Clip found here.

This repository provides scripts to run OpenAI-Clip on Qualcomm® devices. More details on model performance across various devices, can be found here.

Model Details

Model Type: Image classification
Model Stats:
- Model checkpoint: ViT-B/16
- Image input resolution: 224x224
- Text context length: 77
- Number of parameters (CLIPTextEncoder): 76.0M
- Model size (CLIPTextEncoder): 290 MB
- Number of parameters (CLIPImageEncoder): 115M
- Model size (CLIPImageEncoder): 437 MB

Model	Device	Chipset	Target Runtime	Inference Time (ms)	Peak Memory Range (MB)	Precision	Primary Compute Unit	Target Model
CLIPImageEncoder	Samsung Galaxy S23	Snapdragon® 8 Gen 2	TFLITE	34.591 ms	0 - 57 MB	FP16	NPU	OpenAI-Clip.tflite
CLIPImageEncoder	Samsung Galaxy S23	Snapdragon® 8 Gen 2	QNN	26.472 ms	0 - 55 MB	FP16	NPU	OpenAI-Clip.so
CLIPImageEncoder	Samsung Galaxy S24	Snapdragon® 8 Gen 3	TFLITE	27.035 ms	0 - 264 MB	FP16	NPU	OpenAI-Clip.tflite
CLIPImageEncoder	Samsung Galaxy S24	Snapdragon® 8 Gen 3	QNN	20.808 ms	1 - 170 MB	FP16	NPU	OpenAI-Clip.so
CLIPImageEncoder	Snapdragon 8 Elite QRD	Snapdragon® 8 Elite	TFLITE	24.249 ms	0 - 266 MB	FP16	NPU	OpenAI-Clip.tflite
CLIPImageEncoder	Snapdragon 8 Elite QRD	Snapdragon® 8 Elite	QNN	18.669 ms	0 - 171 MB	FP16	NPU	Use Export Script
CLIPImageEncoder	QCS8550 (Proxy)	QCS8550 Proxy	TFLITE	33.984 ms	0 - 55 MB	FP16	NPU	OpenAI-Clip.tflite
CLIPImageEncoder	QCS8550 (Proxy)	QCS8550 Proxy	QNN	19.984 ms	1 - 2 MB	FP16	NPU	Use Export Script
CLIPImageEncoder	SA7255P ADP	SA7255P	TFLITE	327.04 ms	0 - 264 MB	FP16	NPU	OpenAI-Clip.tflite
CLIPImageEncoder	SA7255P ADP	SA7255P	QNN	265.55 ms	1 - 11 MB	FP16	NPU	Use Export Script
CLIPImageEncoder	SA8255 (Proxy)	SA8255P Proxy	TFLITE	34.335 ms	0 - 54 MB	FP16	NPU	OpenAI-Clip.tflite
CLIPImageEncoder	SA8255 (Proxy)	SA8255P Proxy	QNN	20.528 ms	1 - 2 MB	FP16	NPU	Use Export Script
CLIPImageEncoder	SA8295P ADP	SA8295P	TFLITE	40.114 ms	0 - 200 MB	FP16	NPU	OpenAI-Clip.tflite
CLIPImageEncoder	SA8295P ADP	SA8295P	QNN	30.939 ms	1 - 7 MB	FP16	NPU	Use Export Script
CLIPImageEncoder	SA8650 (Proxy)	SA8650P Proxy	TFLITE	34.062 ms	0 - 58 MB	FP16	NPU	OpenAI-Clip.tflite
CLIPImageEncoder	SA8650 (Proxy)	SA8650P Proxy	QNN	20.836 ms	1 - 2 MB	FP16	NPU	Use Export Script
CLIPImageEncoder	SA8775P ADP	SA8775P	TFLITE	42.508 ms	0 - 264 MB	FP16	NPU	OpenAI-Clip.tflite
CLIPImageEncoder	SA8775P ADP	SA8775P	QNN	29.748 ms	1 - 11 MB	FP16	NPU	Use Export Script
CLIPImageEncoder	QCS8450 (Proxy)	QCS8450 Proxy	TFLITE	34.902 ms	0 - 201 MB	FP16	NPU	OpenAI-Clip.tflite
CLIPImageEncoder	QCS8450 (Proxy)	QCS8450 Proxy	QNN	28.971 ms	0 - 169 MB	FP16	NPU	Use Export Script
CLIPImageEncoder	Snapdragon X Elite CRD	Snapdragon® X Elite	QNN	22.167 ms	1 - 1 MB	FP16	NPU	Use Export Script
CLIPTextEncoder	Samsung Galaxy S23	Snapdragon® 8 Gen 2	TFLITE	5.809 ms	0 - 24 MB	FP16	NPU	OpenAI-Clip.tflite
CLIPTextEncoder	Samsung Galaxy S23	Snapdragon® 8 Gen 2	QNN	4.636 ms	0 - 18 MB	FP16	NPU	OpenAI-Clip.so
CLIPTextEncoder	Samsung Galaxy S24	Snapdragon® 8 Gen 3	TFLITE	3.991 ms	0 - 83 MB	FP16	NPU	OpenAI-Clip.tflite
CLIPTextEncoder	Samsung Galaxy S24	Snapdragon® 8 Gen 3	QNN	3.281 ms	0 - 68 MB	FP16	NPU	OpenAI-Clip.so
CLIPTextEncoder	Snapdragon 8 Elite QRD	Snapdragon® 8 Elite	TFLITE	3.351 ms	0 - 83 MB	FP16	NPU	OpenAI-Clip.tflite
CLIPTextEncoder	Snapdragon 8 Elite QRD	Snapdragon® 8 Elite	QNN	3.197 ms	0 - 68 MB	FP16	NPU	Use Export Script
CLIPTextEncoder	QCS8550 (Proxy)	QCS8550 Proxy	TFLITE	5.613 ms	0 - 23 MB	FP16	NPU	OpenAI-Clip.tflite
CLIPTextEncoder	QCS8550 (Proxy)	QCS8550 Proxy	QNN	4.743 ms	0 - 1 MB	FP16	NPU	Use Export Script
CLIPTextEncoder	SA7255P ADP	SA7255P	TFLITE	61.341 ms	0 - 82 MB	FP16	NPU	OpenAI-Clip.tflite
CLIPTextEncoder	SA7255P ADP	SA7255P	QNN	51.576 ms	0 - 11 MB	FP16	NPU	Use Export Script
CLIPTextEncoder	SA8255 (Proxy)	SA8255P Proxy	TFLITE	5.729 ms	0 - 23 MB	FP16	NPU	OpenAI-Clip.tflite
CLIPTextEncoder	SA8255 (Proxy)	SA8255P Proxy	QNN	4.772 ms	0 - 1 MB	FP16	NPU	Use Export Script
CLIPTextEncoder	SA8295P ADP	SA8295P	TFLITE	7.632 ms	0 - 68 MB	FP16	NPU	OpenAI-Clip.tflite
CLIPTextEncoder	SA8295P ADP	SA8295P	QNN	6.53 ms	0 - 6 MB	FP16	NPU	Use Export Script
CLIPTextEncoder	SA8650 (Proxy)	SA8650P Proxy	TFLITE	5.678 ms	0 - 19 MB	FP16	NPU	OpenAI-Clip.tflite
CLIPTextEncoder	SA8650 (Proxy)	SA8650P Proxy	QNN	4.872 ms	0 - 1 MB	FP16	NPU	Use Export Script
CLIPTextEncoder	SA8775P ADP	SA8775P	TFLITE	8.137 ms	0 - 81 MB	FP16	NPU	OpenAI-Clip.tflite
CLIPTextEncoder	SA8775P ADP	SA8775P	QNN	6.947 ms	0 - 6 MB	FP16	NPU	Use Export Script
CLIPTextEncoder	QCS8450 (Proxy)	QCS8450 Proxy	TFLITE	6.349 ms	0 - 74 MB	FP16	NPU	OpenAI-Clip.tflite
CLIPTextEncoder	QCS8450 (Proxy)	QCS8450 Proxy	QNN	5.399 ms	0 - 71 MB	FP16	NPU	Use Export Script
CLIPTextEncoder	Snapdragon X Elite CRD	Snapdragon® X Elite	QNN	5.08 ms	0 - 0 MB	FP16	NPU	Use Export Script

Installation

This model can be installed as a Python package via pip.

pip install "qai-hub-models[openai_clip]"

Configure Qualcomm® AI Hub to run this model on a cloud-hosted device

Sign-in to Qualcomm® AI Hub with your Qualcomm® ID. Once signed in navigate to Account -> Settings -> API Token.

With this API token, you can configure your client to run models on the cloud hosted devices.

qai-hub configure --api_token API_TOKEN

Navigate to docs for more information.

Demo off target

The package contains a simple end-to-end demo that downloads pre-trained weights and runs this model on a sample input.

python -m qai_hub_models.models.openai_clip.demo

The above demo runs a reference implementation of pre-processing, model inference, and post processing.

NOTE: If you want running in a Jupyter Notebook or Google Colab like environment, please add the following to your cell (instead of the above).

%run -m qai_hub_models.models.openai_clip.demo

Run model on a cloud-hosted device

In addition to the demo, you can also run the model on a cloud-hosted Qualcomm® device. This script does the following:

Performance check on-device on a cloud-hosted device
Downloads compiled assets that can be deployed on-device for Android.
Accuracy check between PyTorch and on-device outputs.

python -m qai_hub_models.models.openai_clip.export

Profiling Results
------------------------------------------------------------
CLIPImageEncoder
Device                          : Samsung Galaxy S23 (13)
Runtime                         : TFLITE                 
Estimated inference time (ms)   : 34.6                   
Estimated peak memory usage (MB): [0, 57]                
Total # Ops                     : 659                    
Compute Unit(s)                 : NPU (659 ops)          

------------------------------------------------------------
CLIPTextEncoder
Device                          : Samsung Galaxy S23 (13)  
Runtime                         : TFLITE                   
Estimated inference time (ms)   : 5.8                      
Estimated peak memory usage (MB): [0, 24]                  
Total # Ops                     : 660                      
Compute Unit(s)                 : NPU (658 ops) CPU (2 ops)

How does this work?

This export script leverages Qualcomm® AI Hub to optimize, validate, and deploy this model on-device. Lets go through each step below in detail:

Step 1: Compile model for on-device deployment

To compile a PyTorch model for on-device deployment, we first trace the model in memory using the jit.trace and then call the submit_compile_job API.

import torch

import qai_hub as hub
from qai_hub_models.models.openai_clip import Model

# Load the model
model = Model.from_pretrained()
image_encoder_model = model.image_encoder
text_encoder_model = model.text_encoder

# Device
device = hub.Device("Samsung Galaxy S23")

# Trace model
image_encoder_input_shape = image_encoder_model.get_input_spec()
image_encoder_sample_inputs = image_encoder_model.sample_inputs()

traced_image_encoder_model = torch.jit.trace(image_encoder_model, [torch.tensor(data[0]) for _, data in image_encoder_sample_inputs.items()])

# Compile model on a specific device
image_encoder_compile_job = hub.submit_compile_job(
    model=traced_image_encoder_model ,
    device=device,
    input_specs=image_encoder_model.get_input_spec(),
)

# Get target model to run on-device
image_encoder_target_model = image_encoder_compile_job.get_target_model()
# Trace model
text_encoder_input_shape = text_encoder_model.get_input_spec()
text_encoder_sample_inputs = text_encoder_model.sample_inputs()

traced_text_encoder_model = torch.jit.trace(text_encoder_model, [torch.tensor(data[0]) for _, data in text_encoder_sample_inputs.items()])

# Compile model on a specific device
text_encoder_compile_job = hub.submit_compile_job(
    model=traced_text_encoder_model ,
    device=device,
    input_specs=text_encoder_model.get_input_spec(),
)

# Get target model to run on-device
text_encoder_target_model = text_encoder_compile_job.get_target_model()

Step 2: Performance profiling on cloud-hosted device

After compiling models from step 1. Models can be profiled model on-device using the target_model. Note that this scripts runs the model on a device automatically provisioned in the cloud. Once the job is submitted, you can navigate to a provided job URL to view a variety of on-device performance metrics.

image_encoder_profile_job = hub.submit_profile_job(
    model=image_encoder_target_model,
    device=device,
)
text_encoder_profile_job = hub.submit_profile_job(
    model=text_encoder_target_model,
    device=device,
)

Step 3: Verify on-device accuracy

To verify the accuracy of the model on-device, you can run on-device inference on sample input data on the same cloud hosted device.

image_encoder_input_data = image_encoder_model.sample_inputs()
image_encoder_inference_job = hub.submit_inference_job(
    model=image_encoder_target_model,
    device=device,
    inputs=image_encoder_input_data,
)
image_encoder_inference_job.download_output_data()
text_encoder_input_data = text_encoder_model.sample_inputs()
text_encoder_inference_job = hub.submit_inference_job(
    model=text_encoder_target_model,
    device=device,
    inputs=text_encoder_input_data,
)
text_encoder_inference_job.download_output_data()

With the output of the model, you can compute like PSNR, relative errors or spot check the output with expected output.

Note: This on-device profiling and inference requires access to Qualcomm® AI Hub. Sign up for access.

Deploying compiled model to Android

The models can be deployed using multiple runtimes:

TensorFlow Lite (.tflite export): This tutorial provides a guide to deploy the .tflite model in an Android application.
QNN (.so export ): This sample app provides instructions on how to use the .so shared library in an Android application.

View on Qualcomm® AI Hub

Get more details on OpenAI-Clip's performance across various devices here. Explore all available models on Qualcomm® AI Hub

License

The license for the original implementation of OpenAI-Clip can be found here.
The license for the compiled assets for on-device deployment can be found here

References

Community

Join our AI Hub Slack community to collaborate, post questions and learn more about on-device AI.
For questions or feedback please reach out to us.