Spaces:
Runtime error
Runtime error
initial commit with Florence-2
Browse files- .gitignore +2 -0
- app.py +74 -0
- requirements-local.txt +9 -0
- requirements.txt +8 -0
- utils/__init__.py +0 -0
- utils/florence.py +55 -0
.gitignore
ADDED
@@ -0,0 +1,2 @@
|
|
|
|
|
|
|
1 |
+
/venv
|
2 |
+
/.idea
|
app.py
ADDED
@@ -0,0 +1,74 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
from typing import Tuple
|
2 |
+
|
3 |
+
import gradio as gr
|
4 |
+
import supervision as sv
|
5 |
+
import torch
|
6 |
+
from PIL import Image
|
7 |
+
|
8 |
+
from utils.florence import load_model, run_inference, FLORENCE_DETAILED_CAPTION_TASK, \
|
9 |
+
FLORENCE_CAPTION_TO_PHRASE_GROUNDING_TASK
|
10 |
+
|
11 |
+
MARKDOWN = """
|
12 |
+
# Florence-2 + SAM2 🔥
|
13 |
+
"""
|
14 |
+
|
15 |
+
DEVICE = torch.device("cuda")
|
16 |
+
|
17 |
+
FLORENCE_MODEL, FLORENCE_PROCESSOR = load_model(device=DEVICE)
|
18 |
+
BOX_ANNOTATOR = sv.BoxAnnotator(color_lookup=sv.ColorLookup.INDEX)
|
19 |
+
LABEL_ANNOTATOR = sv.LabelAnnotator(color_lookup=sv.ColorLookup.INDEX)
|
20 |
+
|
21 |
+
|
22 |
+
def process(
|
23 |
+
image_input,
|
24 |
+
) -> Tuple[Image.Image, str]:
|
25 |
+
_, result = run_inference(
|
26 |
+
model=FLORENCE_MODEL,
|
27 |
+
processor=FLORENCE_PROCESSOR,
|
28 |
+
device=DEVICE,
|
29 |
+
image=image_input,
|
30 |
+
task=FLORENCE_DETAILED_CAPTION_TASK
|
31 |
+
)
|
32 |
+
caption = result[FLORENCE_DETAILED_CAPTION_TASK]
|
33 |
+
_, result = run_inference(
|
34 |
+
model=FLORENCE_MODEL,
|
35 |
+
processor=FLORENCE_PROCESSOR,
|
36 |
+
device=DEVICE,
|
37 |
+
image=image_input,
|
38 |
+
task=FLORENCE_CAPTION_TO_PHRASE_GROUNDING_TASK,
|
39 |
+
text=caption
|
40 |
+
)
|
41 |
+
detections = sv.Detections.from_lmm(
|
42 |
+
lmm=sv.LMM.FLORENCE_2,
|
43 |
+
result=result,
|
44 |
+
resolution_wh=image_input.size
|
45 |
+
)
|
46 |
+
|
47 |
+
output_image = image_input.copy()
|
48 |
+
output_image = BOX_ANNOTATOR.annotate(output_image, detections)
|
49 |
+
output_image = LABEL_ANNOTATOR.annotate(output_image, detections)
|
50 |
+
return output_image, caption
|
51 |
+
|
52 |
+
|
53 |
+
with gr.Blocks() as demo:
|
54 |
+
gr.Markdown(MARKDOWN)
|
55 |
+
with gr.Row():
|
56 |
+
with gr.Column():
|
57 |
+
image_input_component = gr.Image(
|
58 |
+
type='pil', label='Upload image')
|
59 |
+
submit_button_component = gr.Button(value='Submit', variant='primary')
|
60 |
+
|
61 |
+
with gr.Column():
|
62 |
+
image_output_component = gr.Image(type='pil', label='Image output')
|
63 |
+
text_output_component = gr.Textbox(label='Caption output')
|
64 |
+
|
65 |
+
submit_button_component.click(
|
66 |
+
fn=process,
|
67 |
+
inputs=[image_input_component],
|
68 |
+
outputs=[
|
69 |
+
image_output_component,
|
70 |
+
text_output_component
|
71 |
+
]
|
72 |
+
)
|
73 |
+
|
74 |
+
demo.launch(debug=False, show_error=True, max_threads=1)
|
requirements-local.txt
ADDED
@@ -0,0 +1,9 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
torch
|
2 |
+
einops
|
3 |
+
spaces
|
4 |
+
timm
|
5 |
+
transformers
|
6 |
+
samv2
|
7 |
+
gradio
|
8 |
+
supervision
|
9 |
+
opencv-python
|
requirements.txt
ADDED
@@ -0,0 +1,8 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
einops
|
2 |
+
spaces
|
3 |
+
timm
|
4 |
+
transformers
|
5 |
+
samv2
|
6 |
+
gradio
|
7 |
+
supervision
|
8 |
+
opencv-python
|
utils/__init__.py
ADDED
File without changes
|
utils/florence.py
ADDED
@@ -0,0 +1,55 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import os
|
2 |
+
from typing import Union, Any, Tuple, Dict
|
3 |
+
from unittest.mock import patch
|
4 |
+
|
5 |
+
import torch
|
6 |
+
from PIL import Image
|
7 |
+
from transformers import AutoModelForCausalLM, AutoProcessor
|
8 |
+
from transformers.dynamic_module_utils import get_imports
|
9 |
+
|
10 |
+
FLORENCE_CHECKPOINT = "microsoft/Florence-2-base"
|
11 |
+
FLORENCE_DETAILED_CAPTION_TASK = '<MORE_DETAILED_CAPTION>'
|
12 |
+
FLORENCE_CAPTION_TO_PHRASE_GROUNDING_TASK = '<CAPTION_TO_PHRASE_GROUNDING>'
|
13 |
+
|
14 |
+
|
15 |
+
def fixed_get_imports(filename: Union[str, os.PathLike]) -> list[str]:
|
16 |
+
"""Work around for https://huggingface.co/microsoft/phi-1_5/discussions/72."""
|
17 |
+
if not str(filename).endswith("/modeling_florence2.py"):
|
18 |
+
return get_imports(filename)
|
19 |
+
imports = get_imports(filename)
|
20 |
+
imports.remove("flash_attn")
|
21 |
+
return imports
|
22 |
+
|
23 |
+
|
24 |
+
def load_model(
|
25 |
+
device: torch.device, checkpoint: str = FLORENCE_CHECKPOINT
|
26 |
+
) -> Tuple[Any, Any]:
|
27 |
+
with patch("transformers.dynamic_module_utils.get_imports", fixed_get_imports):
|
28 |
+
model = AutoModelForCausalLM.from_pretrained(
|
29 |
+
checkpoint, trust_remote_code=True).to(device).eval()
|
30 |
+
processor = AutoProcessor.from_pretrained(
|
31 |
+
checkpoint, trust_remote_code=True)
|
32 |
+
return model, processor
|
33 |
+
|
34 |
+
|
35 |
+
def run_inference(
|
36 |
+
model: Any,
|
37 |
+
processor: Any,
|
38 |
+
device: torch.device,
|
39 |
+
image: Image,
|
40 |
+
task: str,
|
41 |
+
text: str = ""
|
42 |
+
) -> Tuple[str, Dict]:
|
43 |
+
prompt = task + text
|
44 |
+
inputs = processor(text=prompt, images=image, return_tensors="pt").to(device)
|
45 |
+
generated_ids = model.generate(
|
46 |
+
input_ids=inputs["input_ids"],
|
47 |
+
pixel_values=inputs["pixel_values"],
|
48 |
+
max_new_tokens=1024,
|
49 |
+
num_beams=3
|
50 |
+
)
|
51 |
+
generated_text = processor.batch_decode(
|
52 |
+
generated_ids, skip_special_tokens=False)[0]
|
53 |
+
response = processor.post_process_generation(
|
54 |
+
generated_text, task=task, image_size=image.size)
|
55 |
+
return generated_text, response
|