TrOCR: Optimized for Mobile Deployment
Transformer based model for state-of-the-art optical character recognition (OCR) on both printed and handwritten text
End-to-end text recognition approach with pre-trained image transformer and text transformer models for both image understanding and wordpiece-level text generation.
This model is an implementation of TrOCR found here.
This repository provides scripts to run TrOCR on Qualcomm® devices. More details on model performance across various devices, can be found here.
Model Details
- Model Type: Image to text
- Model Stats:
- Model checkpoint: trocr-small-stage1
- Input resolution: 320x320
- Number of parameters (TrOCREncoder): 23.0M
- Model size (TrOCREncoder): 87.8 MB
- Number of parameters (TrOCRDecoder): 38.3M
- Model size (TrOCRDecoder): 146 MB
Model | Device | Chipset | Target Runtime | Inference Time (ms) | Peak Memory Range (MB) | Precision | Primary Compute Unit | Target Model |
---|---|---|---|---|---|---|---|---|
TrOCRDecoder | Samsung Galaxy S23 | Snapdragon® 8 Gen 2 | TFLITE | 2.061 ms | 0 - 30 MB | FP16 | NPU | TrOCR.tflite |
TrOCRDecoder | Samsung Galaxy S23 | Snapdragon® 8 Gen 2 | QNN | 2.1 ms | 1 - 26 MB | FP16 | NPU | TrOCR.so |
TrOCRDecoder | Samsung Galaxy S23 | Snapdragon® 8 Gen 2 | ONNX | 2.719 ms | 0 - 234 MB | FP16 | NPU | TrOCR.onnx |
TrOCRDecoder | Samsung Galaxy S24 | Snapdragon® 8 Gen 3 | TFLITE | 1.475 ms | 0 - 130 MB | FP16 | NPU | TrOCR.tflite |
TrOCRDecoder | Samsung Galaxy S24 | Snapdragon® 8 Gen 3 | QNN | 1.552 ms | 0 - 133 MB | FP16 | NPU | TrOCR.so |
TrOCRDecoder | Samsung Galaxy S24 | Snapdragon® 8 Gen 3 | ONNX | 2.063 ms | 0 - 135 MB | FP16 | NPU | TrOCR.onnx |
TrOCRDecoder | Snapdragon 8 Elite QRD | Snapdragon® 8 Elite | TFLITE | 1.367 ms | 0 - 60 MB | FP16 | NPU | TrOCR.tflite |
TrOCRDecoder | Snapdragon 8 Elite QRD | Snapdragon® 8 Elite | QNN | 1.241 ms | 1 - 63 MB | FP16 | NPU | Use Export Script |
TrOCRDecoder | Snapdragon 8 Elite QRD | Snapdragon® 8 Elite | ONNX | 1.674 ms | 7 - 155 MB | FP16 | NPU | TrOCR.onnx |
TrOCRDecoder | SA7255P ADP | SA7255P | TFLITE | 11.778 ms | 0 - 59 MB | FP16 | NPU | TrOCR.tflite |
TrOCRDecoder | SA7255P ADP | SA7255P | QNN | 11.685 ms | 7 - 16 MB | FP16 | NPU | Use Export Script |
TrOCRDecoder | SA8255 (Proxy) | SA8255P Proxy | TFLITE | 2.107 ms | 0 - 29 MB | FP16 | NPU | TrOCR.tflite |
TrOCRDecoder | SA8255 (Proxy) | SA8255P Proxy | QNN | 2.033 ms | 1 - 4 MB | FP16 | NPU | Use Export Script |
TrOCRDecoder | SA8295P ADP | SA8295P | TFLITE | 3.072 ms | 0 - 52 MB | FP16 | NPU | TrOCR.tflite |
TrOCRDecoder | SA8295P ADP | SA8295P | QNN | 3.001 ms | 7 - 25 MB | FP16 | NPU | Use Export Script |
TrOCRDecoder | SA8650 (Proxy) | SA8650P Proxy | TFLITE | 2.071 ms | 0 - 97 MB | FP16 | NPU | TrOCR.tflite |
TrOCRDecoder | SA8650 (Proxy) | SA8650P Proxy | QNN | 2.082 ms | 2 - 4 MB | FP16 | NPU | Use Export Script |
TrOCRDecoder | SA8775P ADP | SA8775P | TFLITE | 3.171 ms | 0 - 58 MB | FP16 | NPU | TrOCR.tflite |
TrOCRDecoder | SA8775P ADP | SA8775P | QNN | 3.005 ms | 7 - 17 MB | FP16 | NPU | Use Export Script |
TrOCRDecoder | QCS8275 (Proxy) | QCS8275 Proxy | TFLITE | 11.778 ms | 0 - 59 MB | FP16 | NPU | TrOCR.tflite |
TrOCRDecoder | QCS8275 (Proxy) | QCS8275 Proxy | QNN | 11.685 ms | 7 - 16 MB | FP16 | NPU | Use Export Script |
TrOCRDecoder | QCS8550 (Proxy) | QCS8550 Proxy | TFLITE | 2.084 ms | 0 - 26 MB | FP16 | NPU | TrOCR.tflite |
TrOCRDecoder | QCS8550 (Proxy) | QCS8550 Proxy | QNN | 2.204 ms | 3 - 6 MB | FP16 | NPU | Use Export Script |
TrOCRDecoder | QCS9075 (Proxy) | QCS9075 Proxy | TFLITE | 3.171 ms | 0 - 58 MB | FP16 | NPU | TrOCR.tflite |
TrOCRDecoder | QCS9075 (Proxy) | QCS9075 Proxy | QNN | 3.005 ms | 7 - 17 MB | FP16 | NPU | Use Export Script |
TrOCRDecoder | QCS8450 (Proxy) | QCS8450 Proxy | TFLITE | 2.711 ms | 0 - 121 MB | FP16 | NPU | TrOCR.tflite |
TrOCRDecoder | QCS8450 (Proxy) | QCS8450 Proxy | QNN | 2.838 ms | 4 - 126 MB | FP16 | NPU | Use Export Script |
TrOCRDecoder | Snapdragon X Elite CRD | Snapdragon® X Elite | QNN | 2.183 ms | 7 - 7 MB | FP16 | NPU | Use Export Script |
TrOCRDecoder | Snapdragon X Elite CRD | Snapdragon® X Elite | ONNX | 2.306 ms | 68 - 68 MB | FP16 | NPU | TrOCR.onnx |
TrOCREncoder | Samsung Galaxy S23 | Snapdragon® 8 Gen 2 | TFLITE | 38.67 ms | 7 - 31 MB | FP16 | NPU | TrOCR.tflite |
TrOCREncoder | Samsung Galaxy S23 | Snapdragon® 8 Gen 2 | QNN | 38.46 ms | 2 - 23 MB | FP16 | NPU | TrOCR.so |
TrOCREncoder | Samsung Galaxy S23 | Snapdragon® 8 Gen 2 | ONNX | 36.885 ms | 14 - 142 MB | FP16 | NPU | TrOCR.onnx |
TrOCREncoder | Samsung Galaxy S24 | Snapdragon® 8 Gen 3 | TFLITE | 30.36 ms | 5 - 169 MB | FP16 | NPU | TrOCR.tflite |
TrOCREncoder | Samsung Galaxy S24 | Snapdragon® 8 Gen 3 | QNN | 30.482 ms | 2 - 171 MB | FP16 | NPU | TrOCR.so |
TrOCREncoder | Samsung Galaxy S24 | Snapdragon® 8 Gen 3 | ONNX | 30.161 ms | 14 - 164 MB | FP16 | NPU | TrOCR.onnx |
TrOCREncoder | Snapdragon 8 Elite QRD | Snapdragon® 8 Elite | TFLITE | 24.168 ms | 5 - 169 MB | FP16 | NPU | TrOCR.tflite |
TrOCREncoder | Snapdragon 8 Elite QRD | Snapdragon® 8 Elite | QNN | 26.201 ms | 2 - 166 MB | FP16 | NPU | Use Export Script |
TrOCREncoder | Snapdragon 8 Elite QRD | Snapdragon® 8 Elite | ONNX | 28.217 ms | 12 - 171 MB | FP16 | NPU | TrOCR.onnx |
TrOCREncoder | SA7255P ADP | SA7255P | TFLITE | 256.471 ms | 7 - 170 MB | FP16 | NPU | TrOCR.tflite |
TrOCREncoder | SA7255P ADP | SA7255P | QNN | 247.564 ms | 2 - 11 MB | FP16 | NPU | Use Export Script |
TrOCREncoder | SA8255 (Proxy) | SA8255P Proxy | TFLITE | 39.462 ms | 7 - 35 MB | FP16 | NPU | TrOCR.tflite |
TrOCREncoder | SA8255 (Proxy) | SA8255P Proxy | QNN | 36.844 ms | 2 - 5 MB | FP16 | NPU | Use Export Script |
TrOCREncoder | SA8295P ADP | SA8295P | TFLITE | 53.223 ms | 7 - 169 MB | FP16 | NPU | TrOCR.tflite |
TrOCREncoder | SA8295P ADP | SA8295P | QNN | 50.12 ms | 2 - 19 MB | FP16 | NPU | Use Export Script |
TrOCREncoder | SA8650 (Proxy) | SA8650P Proxy | TFLITE | 39.025 ms | 7 - 27 MB | FP16 | NPU | TrOCR.tflite |
TrOCREncoder | SA8650 (Proxy) | SA8650P Proxy | QNN | 36.915 ms | 1 - 3 MB | FP16 | NPU | Use Export Script |
TrOCREncoder | SA8775P ADP | SA8775P | TFLITE | 47.642 ms | 7 - 169 MB | FP16 | NPU | TrOCR.tflite |
TrOCREncoder | SA8775P ADP | SA8775P | QNN | 42.601 ms | 2 - 11 MB | FP16 | NPU | Use Export Script |
TrOCREncoder | QCS8275 (Proxy) | QCS8275 Proxy | TFLITE | 256.471 ms | 7 - 170 MB | FP16 | NPU | TrOCR.tflite |
TrOCREncoder | QCS8275 (Proxy) | QCS8275 Proxy | QNN | 247.564 ms | 2 - 11 MB | FP16 | NPU | Use Export Script |
TrOCREncoder | QCS8550 (Proxy) | QCS8550 Proxy | TFLITE | 38.604 ms | 7 - 35 MB | FP16 | NPU | TrOCR.tflite |
TrOCREncoder | QCS8550 (Proxy) | QCS8550 Proxy | QNN | 36.951 ms | 2 - 5 MB | FP16 | NPU | Use Export Script |
TrOCREncoder | QCS9075 (Proxy) | QCS9075 Proxy | TFLITE | 47.642 ms | 7 - 169 MB | FP16 | NPU | TrOCR.tflite |
TrOCREncoder | QCS9075 (Proxy) | QCS9075 Proxy | QNN | 42.601 ms | 2 - 11 MB | FP16 | NPU | Use Export Script |
TrOCREncoder | QCS8450 (Proxy) | QCS8450 Proxy | TFLITE | 48.968 ms | 7 - 172 MB | FP16 | NPU | TrOCR.tflite |
TrOCREncoder | QCS8450 (Proxy) | QCS8450 Proxy | QNN | 50.164 ms | 2 - 169 MB | FP16 | NPU | Use Export Script |
TrOCREncoder | Snapdragon X Elite CRD | Snapdragon® X Elite | QNN | 34.307 ms | 2 - 2 MB | FP16 | NPU | Use Export Script |
TrOCREncoder | Snapdragon X Elite CRD | Snapdragon® X Elite | ONNX | 36.636 ms | 50 - 50 MB | FP16 | NPU | TrOCR.onnx |
Installation
Install the package via pip:
pip install "qai-hub-models[trocr]"
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.trocr.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.trocr.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.trocr.export
Profiling Results
------------------------------------------------------------
TrOCRDecoder
Device : Samsung Galaxy S23 (13)
Runtime : TFLITE
Estimated inference time (ms) : 2.1
Estimated peak memory usage (MB): [0, 30]
Total # Ops : 399
Compute Unit(s) : NPU (399 ops)
------------------------------------------------------------
TrOCREncoder
Device : Samsung Galaxy S23 (13)
Runtime : TFLITE
Estimated inference time (ms) : 38.7
Estimated peak memory usage (MB): [7, 31]
Total # Ops : 591
Compute Unit(s) : NPU (591 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.trocr import Model
# Load the model
torch_model = Model.from_pretrained()
# Device
device = hub.Device("Samsung Galaxy S24")
# Trace model
input_shape = torch_model.get_input_spec()
sample_inputs = torch_model.sample_inputs()
pt_model = torch.jit.trace(torch_model, [torch.tensor(data[0]) for _, data in sample_inputs.items()])
# Compile model on a specific device
compile_job = hub.submit_compile_job(
model=pt_model,
device=device,
input_specs=torch_model.get_input_spec(),
)
# Get target model to run on-device
target_model = 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.
profile_job = hub.submit_profile_job(
model=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.
input_data = torch_model.sample_inputs()
inference_job = hub.submit_inference_job(
model=target_model,
device=device,
inputs=input_data,
)
on_device_output = 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 TrOCR's performance across various devices here. Explore all available models on Qualcomm® AI Hub
License
- The license for the original implementation of TrOCR can be found here.
- The license for the compiled assets for on-device deployment can be found here
References
- TrOCR: Transformer-based Optical Character Recognition with Pre-trained Models
- Source Model Implementation
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.