You can deploy this model and start earning money today!

munoncode/
edgetam_endpoint

endpoints_compatible

Views

No views yet

EdgeTAM Hugging Face Inference Endpoint

This repository contains a custom handler for deploying the EdgeTAM model as a Hugging Face Inference Endpoint.

📋 Overview

EdgeTAM is a segmentation model that generates masks based on image and point prompts. This handler allows you to deploy it as an API endpoint that accepts images and point coordinates, returning segmentation masks.

🚀 Deployment

Option 1: Deploy via Hugging Face Hub

Create a new model repository on Hugging Face
Upload these files:
- handler.py
- requirements.txt
Create an Inference Endpoint from the Hugging Face UI
Select your model and deploy!

Option 2: Test Locally First

bash

# Install dependencies
pip install -r requirements.txt

# Run local tests
python test_handler.py

📝 API Usage

Request Format

python

import requests
import base64
from PIL import Image
import io

# Load and encode your image
image = Image.open("your_image.jpg")
buffered = io.BytesIO()
image.save(buffered, format="JPEG")
img_base64 = base64.b64encode(buffered.getvalue()).decode()

# Prepare the request
data = {
    "inputs": img_base64,  # Base64 encoded image
    "points": [[[[x1, y1], [x2, y2]]]],  # Point coordinates
    "labels": [[[1, 1]]]  # 1 for positive (foreground), 0 for negative (background)
}

# Send request to your endpoint
response = requests.post(
    "https://YOUR-ENDPOINT-URL.endpoints.huggingface.cloud",
    json=data,
    headers={"Authorization": "Bearer YOUR_HF_TOKEN"}
)

result = response.json()

Input Parameters

inputs (required): Base64 encoded image string
points (required): 4D array of point coordinates
- Format: [[[[x1, y1], [x2, y2], ...]]]
- Dimensions: [image_dim, object_dim, points_per_object_dim, coordinates]
labels (required): 3D array of point labels
- Format: [[[1, 1, ...]]]
- Dimensions: [image_dim, object_dim, point_labels]
- Values: 1 = positive click (foreground), 0 = negative click (background)

Response Format

json

[
  {
    "masks": [[[...]]],
    "shape": [1, 3, 1200, 1800],
    "num_masks": 3,
    "description": "Generated 3 masks ranked by quality score"
  }
]

The model outputs 3 masks per object, ranked by quality score:

masks[0] = Best quality mask (use this one in most cases)
masks[1] = Second best
masks[2] = Third best

Output Dimensions

Shape: [num_objects, num_masks_per_object, height, width]
Each mask is a boolean array where True indicates the segmented region
Masks are returned in the original image dimensions

💡 Examples

Example 1: Single Object with Positive Points

python

data = {
    "inputs": "<base64_encoded_image>",
    "points": [[[[500, 375], [1000, 500]]]],  # Two points on the object
    "labels": [[[1, 1]]]  # Both positive (foreground)
}

Example 2: With Negative Points (Refinement)

python

data = {
    "inputs": "<base64_encoded_image>",
    "points": [[[[500, 375], [1000, 500], [200, 200]]]],
    "labels": [[[1, 1, 0]]]  # Two positive, one negative (background)
}

Example 3: Multiple Objects

python

data = {
    "inputs": "<base64_encoded_image>",
    "points": [[
        [[500, 375]],      # Points for object 1
        [[1000, 500]]      # Points for object 2
    ]],
    "labels": [[
        [1],               # Label for object 1
        [1]                # Label for object 2
    ]]
}

🔧 Advanced Usage

Processing the Masks

python

import numpy as np
from PIL import Image

# Get the best quality mask (first one)
mask = np.array(result[0]["masks"])[0, 0]  # [object_0, best_mask]

# Convert to PIL Image
mask_image = Image.fromarray((mask * 255).astype(np.uint8))

# Apply mask to original image
masked_image = np.array(original_image) * mask[:, :, np.newaxis]
masked_pil = Image.fromarray(masked_image.astype(np.uint8))

Error Handling

python

result = response.json()

if "error" in result[0]:
    print(f"Error: {result[0]['error']}")
    print(f"Type: {result[0]['type']}")
else:
    masks = result[0]["masks"]
    # Process masks...

📦 Dependencies

transformers>=4.40.0
torch>=2.0.0
pillow>=9.0.0
numpy>=1.21.0

🎯 Use Cases

Interactive Segmentation: Click on objects to segment them
Semi-Automatic Annotation: Provide rough points, get precise masks
Object Extraction: Isolate specific objects from images
Image Editing: Create masks for photo manipulation

📚 Model Information

Model: yonigozlan/EdgeTAM-hf
Base Architecture: SAM2 (Segment Anything Model 2)
License: Check the model card for licensing information

🐛 Troubleshooting

Common Issues

"Invalid base64 encoded image"
- Ensure your image is properly base64 encoded
- Check that you're sending the raw base64 string, not a data URL
"Invalid points format"
- Verify the 4D structure: [[[[x, y]]]]
- Ensure coordinates are within image bounds
"Invalid labels format"
- Verify the 3D structure: [[[1]]]
- Use only 0 or 1 as label values
Memory Issues
- Large images may cause OOM errors
- Consider resizing images before encoding
- Use smaller batch sizes

📄 License

This handler code is provided as-is. Please refer to the EdgeTAM model card for model-specific licensing.

munoncode/edgetam_endpoint

EdgeTAM Hugging Face Inference Endpoint

📋 Overview

🚀 Deployment

Option 1: Deploy via Hugging Face Hub

Option 2: Test Locally First

📝 API Usage

Request Format

Input Parameters

Response Format

Output Dimensions

💡 Examples

Example 1: Single Object with Positive Points

Example 2: With Negative Points (Refinement)

Example 3: Multiple Objects

🔧 Advanced Usage

Processing the Masks

Error Handling

📦 Dependencies

🎯 Use Cases

📚 Model Information

🐛 Troubleshooting

Common Issues

📄 License

munoncode/
edgetam_endpoint