feat: sam inference example

.gitignore

@@ -0,0 +1,2 @@
+*.npy
+mask.png

Makefile

@@ -0,0 +1,14 @@
+PYTHON=3.9
+BASENAME=$(shell basename $(CURDIR))
+CURRENT_DIR = $(shell pwd)
+
+env:
+	conda create -n $(BASENAME) -y python=$(PYTHON)
+
+setup:
+	pip install -r requirements.txt
+	pip install git+https://github.com/facebookresearch/segment-anything.git
+
+load-model:
+	mkdir -p models
+	curl -O https://dl.fbaipublicfiles.com/segment_anything/sam_vit_h_4b8939.pth && mv sam_vit_h_4b8939.pth models/sam_vit_h_4b8939.pth

README.md

@@ -10,4 +10,18 @@ pinned: false
 license: apache-2.0
 ---
 
-Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference
+## Before you started
+- requirements: Conda
+```
+make env
+conda activate sam-inference
+make setup
+make load-model
+
+```
+
+## Example inference script
+```
+
+python example.py --image samples/bears.jpg
+```

models/sam_vit_h_4b8939.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a7bf3b02f3ebf1267aba913ff637d9a2d5c33d3173bb679e46d9f338c26f262e
+size 2564550879

requirements.txt

@@ -0,0 +1,5 @@
+opencv-python 
+matplotlib
+gradio
+torch
+torchvision

scripts/example.py

@@ -0,0 +1,47 @@
+import argparse
+import cv2
+import numpy as np
+from segment_anything import SamPredictor, sam_model_registry
+
+# Argument parser
+parser = argparse.ArgumentParser()
+parser.add_argument("-i", "--image", required=True, help="Path to the image")
+args = parser.parse_args()
+
+# Set hyperparameters
+sam_checkpoint = "./models/sam_vit_h_4b8939.pth"
+model_type = "vit_h"
+device = "cpu"
+
+# Load model
+sam = sam_model_registry[model_type](checkpoint=sam_checkpoint)
+sam.to(device=device)
+predictor = SamPredictor(sam)
+
+# Preprocessing the image
+image = cv2.imread(args.image)
+predictor.set_image(image)
+
+# SAM Encoder for embedding
+embedding = predictor.get_image_embedding()
+np.save("models/embedding.npy", embedding)
+
+
+# SAM Decoder for segmentation
+input_point = np.array([[1300, 950]])
+input_label = np.array([1])
+mask, score, logit = predictor.predict(
+    point_coords=input_point,
+    point_labels=input_label,
+    multimask_output=False,
+)
+
+# Save output
+h, w = mask.shape[-2:]
+mask = mask.reshape(h, w, 1)
+
+## Mask has a 255 or 0 value
+mask = (mask * 255).astype(np.uint8)
+
+## Save mask image
+cv2.imwrite("mask.png", mask[:, :])