This repository provides web services with advanced computer vision models that can be used with PhotoPrism and other applications.
The currently integrated models, each with its own endpoint, are kosmos-2, vit-gpt2-image-captioning, and blip-image-captioning large:
Komsos-2 is the most accurate model of the three. It was developed by Microsoft, and this application uses the transformers implementation of the original model, as described in its Huggingface. This model was released in June 2023, and offers object detection and spatial reasoning. Kosmos-2 has very accurate accurate image captions (a .04-.1 increase in clip score when compared to the other two models offered), and is the default model used.
This model was released by nlpconnect. This model combined VIT and GPT-2 to create a multi-modal image captioning model. I have found this to be the least performing of the three, but your mileage may vary.
This model was released by Salesforce in 2022. The primary purpose for this model was to increase both image understanding and text generation using novel techniques. It has achieved a 2.8% CIDEr result, and I've found this model to be more performant than VIT-GPT2, but Kosmos-2 to be slightly better (a .4 increase in CLIP score).
Flask is the framework that is used for the API. It allows for API creation with Python, which is key for this application as it utilizes ML.
PyTorch is key for working with the ML models to generate the outputs. It also enables GPU processing, speeding up the image processing with the models. PyTorch primarily creates and handles tensors, which are crucial for the function of the models.
Transformers is used for downloading and loading the models. In addition to this it is used in the image processing with the models.
Pillow is used to take the supplied URl and convert it into the format needed to input into the models.
Numpy could be used for further hardware acceleration. It isn't included in the application by default to save space and keep from installing unnecessary dependencies. Numpy can be configured to use the GPU for computations. PyTorch already enables GPU processing, so numpy may not make a signficant difference.
Before installing the Python dependencies, please make sure that you have Git and Python 3.12 (incl. pip) installed on your system, e.g. by running the following command on Ubuntu/Debian Linux:
sudo apt-get install -y git python3 python3-pip python3-venv python3-wheel
You can then install the required libraries in a virtual environment by either using the Makefiles we provide (i.e. run make
in the main project directory or a subdirectory) or by manually running the following commands in a service directory, for example:
git clone [email protected]:photoprism/photoprism-vision.git
cd photoprism-vision/describe
python3 -m venv ./venv
. ./venv/bin/activate
./venv/bin/pip install --disable-pip-version-check --upgrade pip
./venv/bin/pip install --disable-pip-version-check -r requirements.txt
Run the Python file app.py
in the describe
subdirectory to start the describe service after you have installed the dependencies (more services, e.g. for OCR and tag generation, may follow):
./venv/bin/python app.py
The service then listens on port 5000 by default and its API endpoints for generating captions support both GET
and POST
requests. It can be tested with the curl
command (curl.exe
on Windows) as shown in the example below:
curl -v -H "Content-Type: application/json" \
--data '{"url":"https://dl.photoprism.app/img/team/avatar.jpg"}' \
-X POST http://localhost:5000/api/v1/vision/describe
At a minimum, a valid image url
must be specified for this. In addition, a model
name and an arbitrary id
can be passed. The API will return the same id
in the response. If no id
is passed, a randomly generated UUID will be returned instead.
If your client submits POST
requests, the request body must be JSON-encoded, e.g.:
{
"id": "3487da77-246e-4b4c-9437-67507177bcd7",
"url": "https://dl.photoprism.app/img/team/avatar.jpg"
}
Alternatively, you can perform GET
requests with URL-encoded query parameters, which is easier to test without an HTTP client:
This is the default endpoint of the API. An image url should be passed in with the key "url", and optionally a "model" and/or "id" value can be passed in. The "model" key allows the user to specify which of the three models they would like to use. If no model is given, the application will default to using the kosmos-2 model.
This is the endpoint for the Kosmos-2 model. An image url should be passed in with the key "url", and optionally a "model" and/or "id" value can be passed in.
This is the endpoint for the VIT GPT-2 model. An image url should be passed in with the key "url", and optionally an "id" value can be passed in.
This is the endpoint for the BLIP model. An image url should be passed in with the key "url", and an "id" value can be passed in.
POST /api/v1/vision/describe
{
"id": "b0db2187-7a09-438c-8649-a9c6c0f7b8a1",
"model": "kosmos-2"
"url": "https://dl.photoprism.app/img/team/avatar.jpg",
}
{
"id": "b0db2187-7a09-438c-8649-a9c6c0f7b8a1",
"model": {
"name": "kosmos-2",
"version": "patch14-224"
},
"result": {
"caption": "An image of a man in a suit smiling."
}
}
MODEL_DIR = "models"
KOSMOS_MODEL_PATH = os.path.join(MODEL_DIR, "kosmos-2-patch14-224")
VIT_MODEL_PATH = os.path.join(MODEL_DIR, "vit-gpt2-image-captioning")
BLIP_MODEL_PATH = os.path.join(MODEL_DIR, "blip-image-captioning-large")
This code block creates the paths for the models. This will be useful when downloading/loading the models. It uses os.path to assemble the correct path depending on if the system is Windows-based or UNIX-based.
def download_model(model_name, save_path):
if not os.path.exists(save_path):
print(f"Downloading {model_name}...")
if model_name == "microsoft/kosmos-2-patch14-224":
AutoModelForVision2Seq.from_pretrained(model_name).save_pretrained(save_path)
AutoProcessor.from_pretrained(model_name).save_pretrained(save_path)
elif model_name == "nlpconnect/vit-gpt2-image-captioning":
VisionEncoderDecoderModel.from_pretrained(model_name).save_pretrained(save_path)
ViTImageProcessor.from_pretrained(model_name).save_pretrained(save_path)
AutoTokenizer.from_pretrained(model_name).save_pretrained(save_path)
elif model_name == "Salesforce/blip-image-captioning-large":
BlipForConditionalGeneration.from_pretrained(model_name).save_pretrained(save_path)
BlipProcessor.from_pretrained(model_name).save_pretrained(save_path)
print(f"{model_name} downloaded and saved to {save_path}")
else:
print(f"{model_name} already exists at {save_path}")
Here the code is checking if the models already exist or not. If they don't exist it is downloading them, if they do it is skipping the downloading.
os.makedirs(MODEL_DIR, exist_ok=True)
download_model("microsoft/kosmos-2-patch14-224", KOSMOS_MODEL_PATH)
download_model("nlpconnect/vit-gpt2-image-captioning", VIT_MODEL_PATH)
download_model("Salesforce/blip-image-captioning-large", BLIP_MODEL_PATH)
Here the code is downloading the models by calling the function in the previous block.
print("Loading models...")
kosmosModel = AutoModelForVision2Seq.from_pretrained(KOSMOS_MODEL_PATH)
kosmosProcessor = AutoProcessor.from_pretrained(KOSMOS_MODEL_PATH)
vitModel = VisionEncoderDecoderModel.from_pretrained(VIT_MODEL_PATH)
vitFeature_extractor = ViTImageProcessor.from_pretrained(VIT_MODEL_PATH)
vitTokenizer = AutoTokenizer.from_pretrained(VIT_MODEL_PATH)
blipProcessor = BlipProcessor.from_pretrained(BLIP_MODEL_PATH)
blipModel = BlipForConditionalGeneration.from_pretrained(BLIP_MODEL_PATH)
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
vitModel.to(device)
Here the models are being loaded after they have been saved.
def kosmosGenerateResponse(url):
try:
image = Image.open(requests.get(url, stream=True).raw)
except Exception as e:
return "fetchError", f"Unable to fetch image: {str(e)}"
prompt = "<grounding>An image of"
try:
inputs = kosmosProcessor(text=prompt, images=image, return_tensors="pt")
generated_ids = kosmosModel.generate(
pixel_values=inputs["pixel_values"],
input_ids=inputs["input_ids"],
attention_mask=inputs["attention_mask"],
image_embeds=None,
image_embeds_position_mask=inputs["image_embeds_position_mask"],
use_cache=True,
max_new_tokens=128,
)
generated_text = kosmosProcessor.batch_decode(generated_ids, skip_special_tokens=True)[0]
processed_text, entities = kosmosProcessor.post_process_generation(generated_text)
except Exception as e:
return "processingError", f"Error during processing: {str(e)}"
return "ok", processed_text
def vitGenerateResponse(url):
vitModel.to(device)
max_length = 16
num_beams = 4
gen_kwargs = {"max_length": max_length, "num_beams": num_beams}
def predict_step(url):
image = Image.open(requests.get(url, stream=True).raw)
images = []
if image.mode != "RGB":
image = image.convert(mode="RGB")
images.append(image)
pixel_values = vitFeature_extractor(images=images, return_tensors="pt").pixel_values
pixel_values = pixel_values.to(device)
output_ids = vitModel.generate(pixel_values, **gen_kwargs)
preds = vitTokenizer.batch_decode(output_ids, skip_special_tokens=True)
preds = [pred.strip() for pred in preds]
return preds
processed_text = predict_step(url) # returns prediction
return "ok", processed_text
def blipGenerateResponse(url):
img_url = url
raw_image = Image.open(requests.get(img_url, stream=True).raw).convert('RGB')
inputs = blipProcessor(raw_image, return_tensors="pt")
out = blipModel.generate(**inputs)
processed_text = blipProcessor.decode(out[0], skip_special_tokens=True)
return "ok", processed_text
These are the services to generate the captions. There is a function for each model.
@app.route('/api/v1/vision/describe', methods=['POST', 'GET'])
def generateResponse():
if request.method == 'POST':
if not request.is_json:
return jsonify({"error": "Request must be JSON"}), 400
data = request.get_json()
elif request.method == 'GET':
data = request.args
url = data.get('url')
model = data.get('model')
id = data.get('id')
if not url:
return jsonify({"error": "URL is required"}), 400
if model == "kosmos-2" or not model:
status, result = kosmosGenerateResponse(url)
if status == "fetchError":
return jsonify({"error": result}), 500
elif status == "processingError":
return jsonify({"error": result}), 500
elif status == "ok":
if id:
return jsonify({"id": id, "result": {"caption": result}, "model": {"name": "kosmos-2", "version": "patch14-224"}}), 200
return jsonify({"id": uuid.uuid4(), "result": {"caption": result}, "model": {"name": "kosmos-2", "version": "patch14-224"}}), 200
elif model == "vit-gpt2-image-captioning":
status, result = vitGenerateResponse(url)
if status == "ok":
if id:
return jsonify({"id": id, "result": {"caption": result}, "model": {"name": model, "version": "latest"}}), 200
return jsonify({"id": uuid.uuid4(), "result": {"caption": result}, "model": {"name": model, "version": "latest"}}), 200
return jsonify({"error": "Error during processing"})
elif model == "blip-image-captioning-large":
status, result = blipGenerateResponse(url)
if status =='ok':
if id:
return jsonify({"id": id, "result": {"caption": result}, "model": {"name": model, "version": "latest"}}), 200
return jsonify({"id": uuid.uuid4(), "result": {"caption": result}, "model": {"name": model, "version": "latest"}}), 200
return jsonify({"error": "Error during processing"})
This is the default endpoint. It checks to see if a model is specified, and if it is it calls the service associated with that model and returns the respose with the data. If a model isn't specified it uses kosmos-2.
@app.route('/api/v1/vision/describe/kosmos-2/patch14-224', methods=['POST', 'GET'])
def kosmosController():
if request.method == 'POST':
if not request.is_json:
return jsonify({"error": "Request must be JSON"}), 400
data = request.get_json()
elif request.method == 'GET':
data = request.args
url = data.get('url')
id = data.get('id')
if not url:
return jsonify({"error": "URL is required"}), 400
status, result = kosmosGenerateResponse(url)
if status == "fetchError":
return jsonify({"error": result}), 500
elif status == "processingError":
return jsonify({"error": result}), 500
elif status == "ok":
if id:
return jsonify({"id": id, "result": {"caption": result}, "model": {"name": "kosmos-2", "version": "patch14-224"}}), 200
return jsonify({"id": uuid.uuid4(), "result": {"caption": result}, "model": {"name": "kosmos-2", "version": "patch14-224"}}), 200
@app.route('/api/v1/vision/describe/vit-gpt2-image-captioning', methods=['POST', 'GET'])
def vitController():
if request.method == 'POST':
if not request.is_json:
return jsonify({"error": "Request must be JSON"}), 400
data = request.get_json()
elif request.method == 'GET':
data = request.args
url = data.get('url')
id = data.get('id')
if not url:
return jsonify({"error": "URL is required"}), 400
status, result = vitGenerateResponse(url)
if status == "ok":
if id:
return jsonify({"id": id, "result": {"caption": result}, "model": {"name": "vit-gpt2-image-captioning", "version": "latest"}}), 200
return jsonify({"id": uuid.uuid4(), "result": {"caption": result}, "model": {"name": "vit-gpt2-image-captioning", "version": "latest"}}), 200
return jsonify({"error": "Error during processing"})
@app.route('/api/v1/vision/describe/blip-image-captioning-large', methods=['POST', 'GET'])
def blipController():
if request.method == 'POST':
if not request.is_json:
return jsonify({"error": "Request must be JSON"}), 400
data = request.get_json()
elif request.method == 'GET':
data = request.args
url = data.get('url')
id = data.get('id')
if not url:
return jsonify({"error": "URL is required"}), 400
status, result = blipGenerateResponse(url)
if status == "ok":
if id:
return jsonify({"id": id, "result": {"caption": result}, "model": {"name": "blip-image-captioning-large", "version": "latest"}}), 200
return jsonify({"id": uuid.uuid4(), "result": {"caption": result}, "model": {"name": "blip-image-captioning-large", "version": "latest"}}), 200
return jsonify({"error", "Error during processing"})
These are the endpoints for each model. They do some error handling, run the service, and return the response.
We would like to thank everyone involved, especially Aatif Dawawala who got things rolling and contributed much of the initial code:
Follow our step-by-step guide to learn how to submit new features, bug fixes, and documentation enhancements.
The files in this repository are licensed under the Apache License, Version 2.0 (the “License”).
Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
Copyright © 2024 PhotoPrism UG. By using the software and services we provide, you agree to our Terms of Service, Privacy Policy, and Code of Conduct. PhotoPrism® is a registered trademark.