Update README.md
Browse files
README.md
CHANGED
|
@@ -1,3 +1,151 @@
|
|
| 1 |
-
---
|
| 2 |
-
license: mit
|
| 3 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
license: mit
|
| 3 |
+
language:
|
| 4 |
+
- en
|
| 5 |
+
base_model:
|
| 6 |
+
- google/siglip-so400m-patch14-384
|
| 7 |
+
pipeline_tag: zero-shot-image-classification
|
| 8 |
+
tags:
|
| 9 |
+
- siglip
|
| 10 |
+
- Int8
|
| 11 |
+
---
|
| 12 |
+
|
| 13 |
+
# SigLIP (shape-optimized model)
|
| 14 |
+
|
| 15 |
+
SigLIP model pre-trained on WebLi at resolution 384x384. It was introduced in the paper Sigmoid Loss for Language Image Pre-Training by Zhai et al. and first released in this repository.
|
| 16 |
+
|
| 17 |
+
The Original repo is https://huggingface.co/google/siglip-so400m-patch14-384.
|
| 18 |
+
|
| 19 |
+
This model of SigLIP has been converted to run on the Axera NPU using **w8a16** quantization.
|
| 20 |
+
|
| 21 |
+
This model has been optimized with the following LoRA:
|
| 22 |
+
|
| 23 |
+
Compatible with Pulsar2 version: 3.4
|
| 24 |
+
|
| 25 |
+
## Convert tools links:
|
| 26 |
+
|
| 27 |
+
For those who are interested in model conversion, you can try to export axmodel through
|
| 28 |
+
|
| 29 |
+
|
| 30 |
+
- [The repo of AXera Platform](https://github.com/AXERA-TECH/SigLIP.axera), which you can get the detial of guide
|
| 31 |
+
|
| 32 |
+
- [Pulsar2 Link, How to Convert ONNX to axmodel](https://pulsar2-docs.readthedocs.io/en/latest/pulsar2/introduction.html)
|
| 33 |
+
|
| 34 |
+
|
| 35 |
+
## Support Platform
|
| 36 |
+
|
| 37 |
+
- AX650
|
| 38 |
+
- [M4N-Dock(η±θ―ζ΄ΎPro)](https://wiki.sipeed.com/hardware/zh/maixIV/m4ndock/m4ndock.html)
|
| 39 |
+
- [M.2 Accelerator card](https://axcl-docs.readthedocs.io/zh-cn/latest/doc_guide_hardware.html)
|
| 40 |
+
|
| 41 |
+
|
| 42 |
+
| Models | Raspberry Pi5 Only CPU | Intel i7-13700 | Raspberry Pi5 + M.2 Card |
|
| 43 |
+
| --------------------- | ---------------------- | -------------- | ------------------------ |
|
| 44 |
+
| Image Encoder | 8.3 s | 1.2 s | 0.19 s |
|
| 45 |
+
| Text Encoder | 1.3 s | 0.3 s | 0.05 s |
|
| 46 |
+
|
| 47 |
+
## How to use
|
| 48 |
+
|
| 49 |
+
Download all files from this repository to the device
|
| 50 |
+
|
| 51 |
+
```
|
| 52 |
+
(axcl) axera@raspberrypi:~/samples/siglip $ tree -L 2
|
| 53 |
+
.
|
| 54 |
+
βββ 000000039769.jpg
|
| 55 |
+
βββ ax650
|
| 56 |
+
βΒ Β βββ siglip_text_u16.axmodel
|
| 57 |
+
βΒ Β βββ siglip_vision_u16_fcu8.axmodel
|
| 58 |
+
βββ config.json
|
| 59 |
+
βββ onnx
|
| 60 |
+
βΒ Β βββ siglip-so400m-patch14-384_text.onnx
|
| 61 |
+
βΒ Β βββ siglip-so400m-patch14-384_vision.onnx
|
| 62 |
+
βββ python
|
| 63 |
+
βΒ Β βββ inference_axmodel.py
|
| 64 |
+
βΒ Β βββ inference_onnx.py
|
| 65 |
+
βΒ Β βββ requirements.txt
|
| 66 |
+
βββ tokenizer
|
| 67 |
+
βββ config.json
|
| 68 |
+
βββ preprocessor_config.json
|
| 69 |
+
βββ special_tokens_map.json
|
| 70 |
+
βββ spiece.model
|
| 71 |
+
βββ tokenizer_config.json
|
| 72 |
+
βββ tokenizer.json
|
| 73 |
+
|
| 74 |
+
5 directories, 15 files
|
| 75 |
+
```
|
| 76 |
+
|
| 77 |
+
### python env requirement
|
| 78 |
+
|
| 79 |
+
#### pyaxengine
|
| 80 |
+
|
| 81 |
+
https://github.com/AXERA-TECH/pyaxengine
|
| 82 |
+
|
| 83 |
+
```
|
| 84 |
+
wget https://github.com/AXERA-TECH/pyaxengine/releases/download/0.1.3rc0/axengine-0.1.3-py3-none-any.whl
|
| 85 |
+
pip install axengine-0.1.3-py3-none-any.whl
|
| 86 |
+
```
|
| 87 |
+
|
| 88 |
+
#### others
|
| 89 |
+
|
| 90 |
+
```
|
| 91 |
+
pip install -r python/requirements.txt
|
| 92 |
+
```
|
| 93 |
+
|
| 94 |
+
## Inputs
|
| 95 |
+
|
| 96 |
+
**Test**
|
| 97 |
+
```
|
| 98 |
+
"a photo of 2 cats", "a photo of 2 dogs"
|
| 99 |
+
```
|
| 100 |
+
|
| 101 |
+
**Image**
|
| 102 |
+

|
| 103 |
+
|
| 104 |
+
## Inference with AX650 Host, such as M4N-Dock(η±θ―ζ΄ΎPro)
|
| 105 |
+
|
| 106 |
+
```
|
| 107 |
+
root@ax650:/mnt/qtang/inner/SigLIP.axera# python3 python/inference_axmodel.py
|
| 108 |
+
[INFO] Available providers: ['AxEngineExecutionProvider']
|
| 109 |
+
[INFO] Using provider: AxEngineExecutionProvider
|
| 110 |
+
[INFO] Chip type: ChipType.MC50
|
| 111 |
+
[INFO] VNPU type: VNPUType.DISABLED
|
| 112 |
+
[INFO] Engine version: 2.7.2a
|
| 113 |
+
[INFO] Model type: 2 (triple core)
|
| 114 |
+
[INFO] Compiler version: 3.4-dirty 739e2b35-dirty
|
| 115 |
+
Model loading time: 3.86 seconds
|
| 116 |
+
[INFO] Using provider: AxEngineExecutionProvider
|
| 117 |
+
[INFO] Model type: 2 (triple core)
|
| 118 |
+
[INFO] Compiler version: 3.4-dirty 739e2b35-dirty
|
| 119 |
+
Model loading time: 3.22 seconds
|
| 120 |
+
Total model loading time: 7.08 seconds
|
| 121 |
+
Model inference time: 0.19 seconds
|
| 122 |
+
Model inference time: 0.05 seconds
|
| 123 |
+
Total inference time: 0.24 seconds
|
| 124 |
+
49.4% that image 0 is 'a photo of 2 cats'
|
| 125 |
+
root@ax650:/mnt/qtang/inner/SigLIP.axera#
|
| 126 |
+
```
|
| 127 |
+
|
| 128 |
+
## Inference with M.2 Accelerator card
|
| 129 |
+
|
| 130 |
+
[What is M.2 Accelerator card?](https://axcl-docs.readthedocs.io/zh-cn/latest/doc_guide_hardware.html), Show this DEMO based on Raspberry PI 5.
|
| 131 |
+
|
| 132 |
+
```
|
| 133 |
+
(axcl) axera@raspberrypi:~/samples/siglip $ python python/inference_axmodel.py
|
| 134 |
+
[INFO] Available providers: ['AXCLRTExecutionProvider']
|
| 135 |
+
[INFO] Using provider: AXCLRTExecutionProvider
|
| 136 |
+
[INFO] SOC Name: AX650N
|
| 137 |
+
[INFO] VNPU type: VNPUType.DISABLED
|
| 138 |
+
[INFO] Compiler version: 3.4-dirty 739e2b35-dirty
|
| 139 |
+
Model loading time: 12.31 seconds
|
| 140 |
+
[INFO] Using provider: AXCLRTExecutionProvider
|
| 141 |
+
[INFO] SOC Name: AX650N
|
| 142 |
+
[INFO] VNPU type: VNPUType.DISABLED
|
| 143 |
+
[INFO] Compiler version: 3.4-dirty 739e2b35-dirty
|
| 144 |
+
Model loading time: 12.37 seconds
|
| 145 |
+
Total model loading time: 24.68 seconds
|
| 146 |
+
Model inference time: 0.19 seconds
|
| 147 |
+
Model inference time: 0.05 seconds
|
| 148 |
+
Total inference time: 0.24 seconds
|
| 149 |
+
52.5% that image 0 is 'a photo of 2 cats'
|
| 150 |
+
(axcl) axera@raspberrypi:~/samples/siglip $
|
| 151 |
+
```
|