Spaces:
Sleeping
Sleeping
Update main.py
Browse files
main.py
CHANGED
@@ -1,5 +1,4 @@
|
|
1 |
from fastapi import FastAPI, File, UploadFile, HTTPException
|
2 |
-
from fastapi.responses import JSONResponse
|
3 |
from pydantic import BaseModel
|
4 |
import base64
|
5 |
import io
|
@@ -17,31 +16,20 @@ from utils import (
|
|
17 |
get_som_labeled_img,
|
18 |
)
|
19 |
|
20 |
-
#
|
21 |
from ultralytics import YOLO
|
22 |
-
from transformers import AutoProcessor, AutoModelForCausalLM
|
23 |
-
|
24 |
-
# ---------------------------------------------------------------------------
|
25 |
-
# Load the YOLO model
|
26 |
-
# ---------------------------------------------------------------------------
|
27 |
-
try:
|
28 |
-
yolo_model = torch.load("weights/icon_detect/best.pt", map_location="cuda", weights_only=False)["model"]
|
29 |
-
yolo_model = yolo_model.to("cuda")
|
30 |
-
except Exception as e:
|
31 |
-
print("Error loading YOLO model on CUDA:", e)
|
32 |
-
yolo_model = torch.load("weights/icon_detect/best.pt", map_location="cpu", weights_only=False)["model"]
|
33 |
|
|
|
|
|
34 |
print(f"YOLO model type: {type(yolo_model)}")
|
35 |
|
36 |
-
# ---------------------------------------------------------------------------
|
37 |
# Load the captioning model (Florence-2)
|
38 |
-
|
|
|
39 |
device = "cuda" if torch.cuda.is_available() else "cpu"
|
40 |
dtype = torch.float16 if device == "cuda" else torch.float32
|
41 |
|
42 |
-
# Load the processor for the Florence-2 model
|
43 |
processor = AutoProcessor.from_pretrained("microsoft/Florence-2-base", trust_remote_code=True)
|
44 |
-
|
45 |
try:
|
46 |
model = AutoModelForCausalLM.from_pretrained(
|
47 |
"weights/icon_caption_florence",
|
@@ -50,14 +38,12 @@ try:
|
|
50 |
).to(device)
|
51 |
except Exception as e:
|
52 |
print(f"Error loading caption model: {str(e)}")
|
53 |
-
# Fallback to CPU with float32
|
54 |
model = AutoModelForCausalLM.from_pretrained(
|
55 |
"weights/icon_caption_florence",
|
56 |
torch_dtype=torch.float32,
|
57 |
trust_remote_code=True
|
58 |
).to("cpu")
|
59 |
|
60 |
-
# Force configuration for DaViT vision tower if missing
|
61 |
if not hasattr(model.config, 'vision_config'):
|
62 |
model.config.vision_config = {}
|
63 |
if 'model_type' not in model.config.vision_config:
|
@@ -66,9 +52,6 @@ if 'model_type' not in model.config.vision_config:
|
|
66 |
caption_model_processor = {"processor": processor, "model": model}
|
67 |
print("Finish loading caption model!")
|
68 |
|
69 |
-
# ---------------------------------------------------------------------------
|
70 |
-
# Create FastAPI application and response model
|
71 |
-
# ---------------------------------------------------------------------------
|
72 |
app = FastAPI()
|
73 |
|
74 |
class ProcessResponse(BaseModel):
|
@@ -76,18 +59,13 @@ class ProcessResponse(BaseModel):
|
|
76 |
parsed_content_list: str
|
77 |
label_coordinates: str
|
78 |
|
79 |
-
# ---------------------------------------------------------------------------
|
80 |
-
# Main processing function
|
81 |
-
# ---------------------------------------------------------------------------
|
82 |
def process(image_input: Image.Image, box_threshold: float, iou_threshold: float) -> ProcessResponse:
|
83 |
-
# Save the input image temporarily
|
84 |
image_save_path = "imgs/saved_image_demo.png"
|
85 |
os.makedirs(os.path.dirname(image_save_path), exist_ok=True)
|
86 |
image_input.save(image_save_path)
|
87 |
|
88 |
-
# Open the saved image for processing
|
89 |
image = Image.open(image_save_path)
|
90 |
-
box_overlay_ratio = image.size[0] / 3200
|
91 |
draw_bbox_config = {
|
92 |
"text_scale": 0.8 * box_overlay_ratio,
|
93 |
"text_thickness": max(int(2 * box_overlay_ratio), 1),
|
@@ -95,7 +73,6 @@ def process(image_input: Image.Image, box_threshold: float, iou_threshold: float
|
|
95 |
"thickness": max(int(3 * box_overlay_ratio), 1),
|
96 |
}
|
97 |
|
98 |
-
# Run OCR to get text and OCR bounding boxes
|
99 |
ocr_bbox_rslt, is_goal_filtered = check_ocr_box(
|
100 |
image_save_path,
|
101 |
display_img=False,
|
@@ -106,7 +83,6 @@ def process(image_input: Image.Image, box_threshold: float, iou_threshold: float
|
|
106 |
)
|
107 |
text, ocr_bbox = ocr_bbox_rslt
|
108 |
|
109 |
-
# Run YOLO and semantic processing to get the labeled image and bounding boxes
|
110 |
dino_labled_img, label_coordinates, parsed_content_list = get_som_labeled_img(
|
111 |
image_save_path,
|
112 |
yolo_model,
|
@@ -118,13 +94,10 @@ def process(image_input: Image.Image, box_threshold: float, iou_threshold: float
|
|
118 |
ocr_text=text,
|
119 |
iou_threshold=iou_threshold,
|
120 |
)
|
121 |
-
|
122 |
-
# Decode the base64-encoded image output from get_som_labeled_img
|
123 |
image = Image.open(io.BytesIO(base64.b64decode(dino_labled_img)))
|
124 |
print("Finish processing")
|
125 |
parsed_content_list_str = "\n".join(parsed_content_list)
|
126 |
|
127 |
-
# Encode final image to base64 string for response
|
128 |
buffered = io.BytesIO()
|
129 |
image.save(buffered, format="PNG")
|
130 |
img_str = base64.b64encode(buffered.getvalue()).decode("utf-8")
|
@@ -135,9 +108,6 @@ def process(image_input: Image.Image, box_threshold: float, iou_threshold: float
|
|
135 |
label_coordinates=str(label_coordinates),
|
136 |
)
|
137 |
|
138 |
-
# ---------------------------------------------------------------------------
|
139 |
-
# FastAPI endpoint for image processing
|
140 |
-
# ---------------------------------------------------------------------------
|
141 |
@app.post("/process_image", response_model=ProcessResponse)
|
142 |
async def process_image(
|
143 |
image_file: UploadFile = File(...),
|
@@ -148,13 +118,10 @@ async def process_image(
|
|
148 |
contents = await image_file.read()
|
149 |
image_input = Image.open(io.BytesIO(contents)).convert("RGB")
|
150 |
|
151 |
-
# Debug logging for file information
|
152 |
print(f"Processing image: {image_file.filename}")
|
153 |
print(f"Image size: {image_input.size}")
|
154 |
|
155 |
response = process(image_input, box_threshold, iou_threshold)
|
156 |
-
|
157 |
-
# Validate response
|
158 |
if not response.image:
|
159 |
raise ValueError("Empty image in response")
|
160 |
|
@@ -162,5 +129,5 @@ async def process_image(
|
|
162 |
|
163 |
except Exception as e:
|
164 |
import traceback
|
165 |
-
traceback.print_exc()
|
166 |
raise HTTPException(status_code=500, detail=str(e))
|
|
|
1 |
from fastapi import FastAPI, File, UploadFile, HTTPException
|
|
|
2 |
from pydantic import BaseModel
|
3 |
import base64
|
4 |
import io
|
|
|
16 |
get_som_labeled_img,
|
17 |
)
|
18 |
|
19 |
+
# Load the YOLO model using the ultralytics class instead of torch.load
|
20 |
from ultralytics import YOLO
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
21 |
|
22 |
+
# Use the YOLO constructor to load the model properly
|
23 |
+
yolo_model = YOLO("weights/icon_detect/best.pt")
|
24 |
print(f"YOLO model type: {type(yolo_model)}")
|
25 |
|
|
|
26 |
# Load the captioning model (Florence-2)
|
27 |
+
from transformers import AutoProcessor, AutoModelForCausalLM
|
28 |
+
|
29 |
device = "cuda" if torch.cuda.is_available() else "cpu"
|
30 |
dtype = torch.float16 if device == "cuda" else torch.float32
|
31 |
|
|
|
32 |
processor = AutoProcessor.from_pretrained("microsoft/Florence-2-base", trust_remote_code=True)
|
|
|
33 |
try:
|
34 |
model = AutoModelForCausalLM.from_pretrained(
|
35 |
"weights/icon_caption_florence",
|
|
|
38 |
).to(device)
|
39 |
except Exception as e:
|
40 |
print(f"Error loading caption model: {str(e)}")
|
|
|
41 |
model = AutoModelForCausalLM.from_pretrained(
|
42 |
"weights/icon_caption_florence",
|
43 |
torch_dtype=torch.float32,
|
44 |
trust_remote_code=True
|
45 |
).to("cpu")
|
46 |
|
|
|
47 |
if not hasattr(model.config, 'vision_config'):
|
48 |
model.config.vision_config = {}
|
49 |
if 'model_type' not in model.config.vision_config:
|
|
|
52 |
caption_model_processor = {"processor": processor, "model": model}
|
53 |
print("Finish loading caption model!")
|
54 |
|
|
|
|
|
|
|
55 |
app = FastAPI()
|
56 |
|
57 |
class ProcessResponse(BaseModel):
|
|
|
59 |
parsed_content_list: str
|
60 |
label_coordinates: str
|
61 |
|
|
|
|
|
|
|
62 |
def process(image_input: Image.Image, box_threshold: float, iou_threshold: float) -> ProcessResponse:
|
|
|
63 |
image_save_path = "imgs/saved_image_demo.png"
|
64 |
os.makedirs(os.path.dirname(image_save_path), exist_ok=True)
|
65 |
image_input.save(image_save_path)
|
66 |
|
|
|
67 |
image = Image.open(image_save_path)
|
68 |
+
box_overlay_ratio = image.size[0] / 3200
|
69 |
draw_bbox_config = {
|
70 |
"text_scale": 0.8 * box_overlay_ratio,
|
71 |
"text_thickness": max(int(2 * box_overlay_ratio), 1),
|
|
|
73 |
"thickness": max(int(3 * box_overlay_ratio), 1),
|
74 |
}
|
75 |
|
|
|
76 |
ocr_bbox_rslt, is_goal_filtered = check_ocr_box(
|
77 |
image_save_path,
|
78 |
display_img=False,
|
|
|
83 |
)
|
84 |
text, ocr_bbox = ocr_bbox_rslt
|
85 |
|
|
|
86 |
dino_labled_img, label_coordinates, parsed_content_list = get_som_labeled_img(
|
87 |
image_save_path,
|
88 |
yolo_model,
|
|
|
94 |
ocr_text=text,
|
95 |
iou_threshold=iou_threshold,
|
96 |
)
|
|
|
|
|
97 |
image = Image.open(io.BytesIO(base64.b64decode(dino_labled_img)))
|
98 |
print("Finish processing")
|
99 |
parsed_content_list_str = "\n".join(parsed_content_list)
|
100 |
|
|
|
101 |
buffered = io.BytesIO()
|
102 |
image.save(buffered, format="PNG")
|
103 |
img_str = base64.b64encode(buffered.getvalue()).decode("utf-8")
|
|
|
108 |
label_coordinates=str(label_coordinates),
|
109 |
)
|
110 |
|
|
|
|
|
|
|
111 |
@app.post("/process_image", response_model=ProcessResponse)
|
112 |
async def process_image(
|
113 |
image_file: UploadFile = File(...),
|
|
|
118 |
contents = await image_file.read()
|
119 |
image_input = Image.open(io.BytesIO(contents)).convert("RGB")
|
120 |
|
|
|
121 |
print(f"Processing image: {image_file.filename}")
|
122 |
print(f"Image size: {image_input.size}")
|
123 |
|
124 |
response = process(image_input, box_threshold, iou_threshold)
|
|
|
|
|
125 |
if not response.image:
|
126 |
raise ValueError("Empty image in response")
|
127 |
|
|
|
129 |
|
130 |
except Exception as e:
|
131 |
import traceback
|
132 |
+
traceback.print_exc()
|
133 |
raise HTTPException(status_code=500, detail=str(e))
|