feat:update demo code of CLIP
This commit is contained in:
dian.yuan
parent
4bf4aafc73
commit
5478a8618b
12 changed files with 50385 additions and 694 deletions
|
|
@ -1,95 +1,159 @@
|
|||
## Demo Run
|
||||
|
||||
### CPP
|
||||
|
||||
#### 1. Compile
|
||||
|
||||
**Prerequisites:**
|
||||
- Android NDK (r25e recommended)
|
||||
- `ANDROID_NDK_PATH` environment variable set
|
||||
|
||||
**Build:**
|
||||
```bash
|
||||
# Build for arm64-v8a
|
||||
cd examples/clip/cpp
|
||||
./build-android.sh -a arm64-v8a
|
||||
```
|
||||
|
||||
The executable will be generated at `build/android_arm64-v8a/clip_demo` (Note: executable name may vary, verify in build folder).
|
||||
|
||||
#### 2. Run
|
||||
|
||||
```bash
|
||||
# Push executable to device
|
||||
adb push build/android_arm64-v8a/clip_demo /data/local/tmp/
|
||||
adb push model/vision_model_int8_A311D2.adla /data/local/tmp/
|
||||
adb push clip_datasets/ /data/local/tmp/
|
||||
adb push test_hat_0.jpg /data/local/tmp/
|
||||
|
||||
# Run on device
|
||||
adb shell
|
||||
cd /data/local/tmp
|
||||
chmod +x clip_demo
|
||||
export LD_LIBRARY_PATH=/vendor/lib64 or (/vendor/lib)
|
||||
|
||||
# Usage: ./clip_demo <model_path> [base_dir] [json_filename]
|
||||
./clip_demo vision_model_int8_A311D2.adla ./clip_datasets/ clip_text_res.json
|
||||
```
|
||||
|
||||
**Note:**
|
||||
- Replace `vision_model_int8_A311D2.adla` with your actual model file path.
|
||||
- The `base_dir` and `json_filename` parameters are optional. You can also use environment variables `CLIP_BASE_DIR` and `CLIP_JSON_FILENAME`.
|
||||
- The program will prompt you to enter image paths interactively. Enter "exit" to quit.
|
||||
|
||||
### Python
|
||||
|
||||
**Prerequisites:**
|
||||
- Python 3.10
|
||||
- Required packages: `numpy`, `Pillow`, `amlnnlite`
|
||||
|
||||
**Install dependencies:**
|
||||
```bash
|
||||
pip install numpy Pillow amlnnlite-1.0.0-cp310-cp310-linux_aarch64.whl
|
||||
```
|
||||
|
||||
**Run on device:**
|
||||
```bash
|
||||
# Basic usage (process current directory)
|
||||
python clip.py --model-path ./vision_model_int8_A311D2.adla
|
||||
|
||||
# Specify image directory or file
|
||||
python clip.py --model-path ./vision_model_int8_A311D2.adla --image-dir ./
|
||||
|
||||
# Specify base directory and JSON filename
|
||||
python clip.py --model-path ./vision_model_int8_A311D2.adla --base-dir ./clip_datasets/ --json-filename clip_text_res.json
|
||||
```
|
||||
|
||||
The script will automatically process all image files (`.jpg`, `.jpeg`, `.png`, `.bmp`) in the specified directory or process a single image file, and display the best matching dataset for each image.
|
||||
|
||||
5. Results
|
||||
|
||||
The program will print the best matching dataset path for each processed image. The program searches through all dataset folders in the base directory and finds the text feature with the highest similarity to the input image.
|
||||
|
||||
**Example output:**
|
||||
```
|
||||
# python demo result
|
||||
Model initialized successfully.
|
||||
|
||||
Found 2 image file(s) to process
|
||||
Searching in base directory: ./clip_datasets/
|
||||
|
||||
Processing image: test_jacket_0.jpg
|
||||
Best matching dataset: ./clip_datasets/shirt10_jacket7
|
||||
Searching in base directory: ./clip_datasets/
|
||||
|
||||
Processing image: test_hat_0.jpg
|
||||
Best matching dataset: ./clip_datasets/hat1_jd
|
||||
|
||||
Total results: 2
|
||||
Index[0]: ./clip_datasets/shirt10_jacket7
|
||||
Index[1]: ./clip_datasets/hat1_jd
|
||||
|
||||
Done.
|
||||
```
|
||||
|
||||
The program returns the dataset folder path that contains the text feature with the highest similarity to the input image. Each result represents the best matching dataset for the corresponding input image.
|
||||
# CLIP
|
||||
|
||||
## 1. Overview
|
||||
|
||||
This demo demonstrates how to run CLIP (Contrastive Language-Image Pre-Training) image-text matching using AMLNNLite. The CLIP model consists of two parts: a vision encoder and a text encoder, which work together to compute similarity between images and text descriptions.
|
||||
|
||||
## 2. Model Download
|
||||
|
||||
TO DO
|
||||
|
||||
## 3. Model Conversion
|
||||
|
||||
TO DO
|
||||
|
||||
## 4. Demo Run
|
||||
|
||||
### CPP
|
||||
|
||||
#### 1. Compile
|
||||
|
||||
**Prerequisites:**
|
||||
- Android NDK (r25e recommended)
|
||||
- `ANDROID_NDK_PATH` environment variable set
|
||||
|
||||
**Build:**
|
||||
```bash
|
||||
# Build for arm64-v8a
|
||||
cd examples/clip/cpp
|
||||
./build-android.sh -a arm64-v8a
|
||||
```
|
||||
|
||||
The executable will be generated at `build/android_arm64-v8a/clip_demo`.
|
||||
|
||||
#### 2. Run
|
||||
|
||||
```bash
|
||||
# Push executable and resources to device
|
||||
adb push build/android_arm64-v8a/clip_demo /data/local/tmp/
|
||||
adb push model/vision_model_int8_S905X5.adla /data/local/tmp/
|
||||
adb push model/text_model_int8_S905X5.adla /data/local/tmp/
|
||||
adb push tokenizer_path/ /data/local/tmp/
|
||||
|
||||
# Run on device
|
||||
adb shell
|
||||
cd /data/local/tmp
|
||||
chmod +x clip_demo
|
||||
export LD_LIBRARY_PATH=/vendor/lib64 or (/vendor/lib)
|
||||
|
||||
# Usage: ./clip_demo <vision_model> <text_model> <tokenizer_path> [--profiling]
|
||||
./clip_demo vision_model_int8_S905X5.adla text_model_int8_S905X5.adla ./tokenizer_path/
|
||||
```
|
||||
|
||||
The program will prompt for image paths and text descriptions interactively. Enter the path to an image file, then enter comma-separated text descriptions (or `skip` to use defaults). Type `exit` to quit.
|
||||
|
||||
**Argument Descriptions:**
|
||||
|
||||
| Argument | Description |
|
||||
| -------------- | ------------------------------------------------------------ |
|
||||
| vision_model | Path to vision encoder .adla model (required) |
|
||||
| text_model | Path to text encoder .adla model (required) |
|
||||
| tokenizer_path | Path to directory containing `vocab.json` and `merges.txt` (required) |
|
||||
| --profiling | Enable performance profiling output (optional) |
|
||||
|
||||
**Note:** The `tokenizer_path` should contain `vocab.json` and `merges.txt` files from the CLIP tokenizer (e.g., from `openai/clip-vit-base-patch32`).
|
||||
|
||||
### Python
|
||||
|
||||
**Prerequisites:**
|
||||
- Python 3.10
|
||||
- Required packages: `numpy`, `Pillow`, `transformers`, `amlnnlite`
|
||||
|
||||
**Install dependencies:**
|
||||
```bash
|
||||
pip install numpy Pillow transformers amlnnlite-1.0.0-cp310-cp310-linux_aarch64.whl
|
||||
```
|
||||
|
||||
**Run on device:**
|
||||
```bash
|
||||
python clip.py \
|
||||
--vision-model ./vision_model_int8_S905X5.adla \
|
||||
--text-model ./text_model_int8_S905X5.adla \
|
||||
--tokenizer-dir ./tokenizer_path \
|
||||
--image-path ./000000004505.jpg \
|
||||
--texts "a red handbag" "a blue jacket" "a red bus"
|
||||
```
|
||||
|
||||
**Interactive Mode (Recommended):**
|
||||
|
||||
If you don't provide `--image-path`, the program will run in interactive mode:
|
||||
|
||||
```bash
|
||||
python clip.py \
|
||||
--vision-model ./vision_model_int8_S905X5.adla \
|
||||
--text-model ./text_model_int8_S905X5.adla \
|
||||
--tokenizer-dir ./tokenizer_path
|
||||
```
|
||||
|
||||
The program will prompt for image paths and text descriptions. Enter an image path to process, then enter comma-separated texts to compare. Type `exit` to quit.
|
||||
|
||||
**Argument Descriptions:**
|
||||
|
||||
| Argument | Description |
|
||||
| ---------------- | ------------------------------------------------------------ |
|
||||
| --vision-model | Path to vision encoder .adla model (required) |
|
||||
| --text-model | Path to text encoder .adla model (required) |
|
||||
| --tokenizer-dir | Path to CLIPTokenizer directory (required) |
|
||||
| --image-path | Path to input image (.jpg, .png) - optional, will prompt if not provided |
|
||||
| --texts | List of text descriptions to compare (space-separated) |
|
||||
| --max-len | Maximum token sequence length, default is 64 |
|
||||
| --logit-scale | Logit scale factor, default is 100.0 |
|
||||
|
||||
**Note:** The `--tokenizer-dir` should point to the directory containing the CLIPTokenizer files. You can use a Hugging Face model ID (e.g., `openai/clip-vit-base-patch32`) or a local directory.
|
||||
|
||||
## 5. Results
|
||||
|
||||
**Performance Feedback**
|
||||
|
||||
By using the `--profiling` flag (C++) or setting the loglevel to INFO, the program provides real-time performance metrics upon completion. The console log will display essential hardware and execution details, including:
|
||||
- Hardware Information: System and ADLA library versions.
|
||||
- Model Overview: Basic input/output configurations.
|
||||
- NPU Metrics: Total inference time (latency) and total DRAM bandwidth consumption.
|
||||
|
||||
**Interactive Mode Example:**
|
||||
|
||||
```bash
|
||||
$ ./clip_demo vision_model_int8_S905X5.adla text_model_int8_S905X5.adla ./tokenizer_path
|
||||
|
||||
[Info] Models initialized successfully.
|
||||
|
||||
============================================================
|
||||
[Info] Image Path (or 'exit' to quit):
|
||||
000000004505.jpg
|
||||
[Info] Enter text descriptions (comma-separated, or 'skip' for defaults):
|
||||
a red handbag, a blue jacket, a red bus
|
||||
|
||||
[Info] Processing image: 000000004505.jpg
|
||||
[Info] Image embedding size: 512
|
||||
[Info] Processing 3 text(s)...
|
||||
[Info] Text embeddings size: 3 x 512
|
||||
|
||||
============================================================
|
||||
CLIP Image-Text Matching Results
|
||||
============================================================
|
||||
Image: 000000004505.jpg
|
||||
logit_scale: 100.000000
|
||||
------------------------------------------------------------
|
||||
[1] prob=0.999975 sim=0.327895 text='a red bus'
|
||||
[2] prob=0.000016 sim=0.217690 text='a red handbag'
|
||||
[3] prob=0.000008 sim=0.211029 text='a blue jacket'
|
||||
============================================================
|
||||
|
||||
============================================================
|
||||
[Info] Image Path (or 'exit' to quit):
|
||||
exit
|
||||
[Info] Exiting...
|
||||
Free vision model memory.
|
||||
Free text model memory.
|
||||
[Info] Done.
|
||||
```
|
||||
|
|
|
|||
Loading…
Add table
Add a link
Reference in a new issue