vit

---

Vision Transformer (ViT) model

Introduction

The Vision Transformer (ViT) model was proposed in "An Image is Worth 16x16 Words, Transformers for Image Recognition at Scale". It’s the first paper that successfully trains a Transformer encoder on ImageNet, attaining very good results compared to familiar convolutional architectures. https://huggingface.co/docs/transformers/en/model_doc/vit

How to Run

To run the demo for question answering using the Bloom model, follow these instructions:

• For Imagenet-21K to test inference accuracy, use the following command to run the demo:

pytest --disable-warnings  models/demos/wormhole/vit/demo/demo_vit_ttnn_imagenet_inference.py

• For the inference overall rutime (end-2-end), use the following command to run the demo:

pytest --disable-warnings  models/demos/wormhole/vit/demo/demo_vit_ttnn_inference_perf_e2e_2cq_trace.py

• For running the inference device OPs analysis, use the following command to run the demo:

build_metal.sh --enable-profiler # need build to enable the profiler
./tt_metal/tools/profiler/profile_this.py -n vit -c "pytest --disable-warnings  models/demos/wormhole/vit/demo/demo_vit_ttnn_inference_device_OPs.py"

Results

• The Imagenet-21K inference accuracy is 79%
• Model runtime (host end-2-end) is 912 FPS

---