Spaces:
Runtime error
Runtime error
ttengwang
commited on
Commit
·
ccb14a3
1
Parent(s):
cd2f644
support "segment everything in a paragraph"
Browse files- app.py +89 -65
- caption_anything/captioner/base_captioner.py +38 -30
- caption_anything/captioner/blip.py +12 -5
- caption_anything/captioner/blip2.py +24 -11
- caption_anything/captioner/git.py +14 -7
- caption_anything/model.py +210 -71
- caption_anything/segmenter/__init__.py +11 -2
- caption_anything/segmenter/base_segmenter.py +8 -3
- caption_anything/utils/densecap_painter.py +64 -0
- caption_anything/utils/parser.py +6 -0
- caption_anything/utils/utils.py +31 -0
- requirements.txt +4 -1
app.py
CHANGED
|
@@ -1,6 +1,5 @@
|
|
| 1 |
import os
|
| 2 |
import json
|
| 3 |
-
import PIL
|
| 4 |
import gradio as gr
|
| 5 |
import numpy as np
|
| 6 |
from gradio import processing_utils
|
|
@@ -11,7 +10,7 @@ import functools
|
|
| 11 |
|
| 12 |
from caption_anything.model import CaptionAnything
|
| 13 |
from caption_anything.utils.image_editing_utils import create_bubble_frame
|
| 14 |
-
from caption_anything.utils.utils import mask_painter, seg_model_map, prepare_segmenter
|
| 15 |
from caption_anything.utils.parser import parse_augment
|
| 16 |
from caption_anything.captioner import build_captioner
|
| 17 |
from caption_anything.text_refiner import build_text_refiner
|
|
@@ -23,6 +22,7 @@ from segment_anything import sam_model_registry
|
|
| 23 |
args = parse_augment()
|
| 24 |
args.segmenter = "huge"
|
| 25 |
args.segmenter_checkpoint = "sam_vit_h_4b8939.pth"
|
|
|
|
| 26 |
if args.segmenter_checkpoint is None:
|
| 27 |
_, segmenter_checkpoint = prepare_segmenter(args.segmenter)
|
| 28 |
else:
|
|
@@ -53,9 +53,7 @@ class ImageSketcher(gr.Image):
|
|
| 53 |
mask = np.zeros((height, width, 4), dtype=np.uint8)
|
| 54 |
mask[..., -1] = 255
|
| 55 |
mask = self.postprocess(mask)
|
| 56 |
-
|
| 57 |
x['mask'] = mask
|
| 58 |
-
|
| 59 |
return super().preprocess(x)
|
| 60 |
|
| 61 |
|
|
@@ -74,16 +72,19 @@ def init_openai_api_key(api_key=""):
|
|
| 74 |
if api_key and len(api_key) > 30:
|
| 75 |
try:
|
| 76 |
text_refiner = build_text_refiner(args.text_refiner, args.device, args, api_key)
|
| 77 |
-
text_refiner.llm('hi')
|
| 78 |
visual_chatgpt = ConversationBot(shared_chatbot_tools, api_key)
|
| 79 |
except:
|
| 80 |
text_refiner = None
|
| 81 |
visual_chatgpt = None
|
| 82 |
openai_available = text_refiner is not None
|
| 83 |
-
|
| 84 |
-
|
| 85 |
-
|
| 86 |
-
|
|
|
|
|
|
|
|
|
|
| 87 |
|
| 88 |
def get_click_prompt(chat_input, click_state, click_mode):
|
| 89 |
inputs = json.loads(chat_input)
|
|
@@ -130,18 +131,15 @@ def chat_input_callback(*args):
|
|
| 130 |
state = state + [(chat_input, response)]
|
| 131 |
return state, state
|
| 132 |
|
|
|
|
|
|
|
| 133 |
def upload_callback(image_input, state, visual_chatgpt=None):
|
| 134 |
|
| 135 |
if isinstance(image_input, dict): # if upload from sketcher_input, input contains image and mask
|
| 136 |
image_input, mask = image_input['image'], image_input['mask']
|
| 137 |
|
| 138 |
click_state = [[], [], []]
|
| 139 |
-
|
| 140 |
-
width, height = image_input.size
|
| 141 |
-
ratio = min(1.0 * res / max(width, height), 1.0)
|
| 142 |
-
if ratio < 1.0:
|
| 143 |
-
image_input = image_input.resize((int(width * ratio), int(height * ratio)))
|
| 144 |
-
print('Scaling input image to {}'.format(image_input.size))
|
| 145 |
|
| 146 |
model = build_caption_anything_with_models(
|
| 147 |
args,
|
|
@@ -159,8 +157,8 @@ def upload_callback(image_input, state, visual_chatgpt=None):
|
|
| 159 |
new_image_path = get_new_image_name('chat_image', func_name='upload')
|
| 160 |
image_input.save(new_image_path)
|
| 161 |
visual_chatgpt.current_image = new_image_path
|
| 162 |
-
img_caption
|
| 163 |
-
Human_prompt = f'\nHuman:
|
| 164 |
AI_prompt = "Received."
|
| 165 |
visual_chatgpt.global_prompt = Human_prompt + 'AI: ' + AI_prompt
|
| 166 |
visual_chatgpt.agent.memory.buffer = visual_chatgpt.agent.memory.buffer + visual_chatgpt.global_prompt
|
|
@@ -201,11 +199,10 @@ def inference_click(image_input, point_prompt, click_mode, enable_wiki, language
|
|
| 201 |
model.setup(image_embedding, original_size, input_size, is_image_set=True)
|
| 202 |
|
| 203 |
enable_wiki = True if enable_wiki in ['True', 'TRUE', 'true', True, 'Yes', 'YES', 'yes'] else False
|
| 204 |
-
out = model.inference(image_input, prompt, controls, disable_gpt=True, enable_wiki=enable_wiki)
|
| 205 |
|
| 206 |
state = state + [("Image point: {}, Input label: {}".format(prompt["input_point"], prompt["input_label"]), None)]
|
| 207 |
state = state + [(None, "raw_caption: {}".format(out['generated_captions']['raw_caption']))]
|
| 208 |
-
wiki = out['generated_captions'].get('wiki', "")
|
| 209 |
update_click_state(click_state, out['generated_captions']['raw_caption'], click_mode)
|
| 210 |
text = out['generated_captions']['raw_caption']
|
| 211 |
input_mask = np.array(out['mask'].convert('P'))
|
|
@@ -221,21 +218,22 @@ def inference_click(image_input, point_prompt, click_mode, enable_wiki, language
|
|
| 221 |
point_prompt = f'You should primarly use tools on the selected regional image (description: {text}, path: {new_crop_save_path}), which is a part of the whole image (path: {visual_chatgpt.current_image}). If human mentioned some objects not in the selected region, you can use tools on the whole image.'
|
| 222 |
visual_chatgpt.point_prompt = point_prompt
|
| 223 |
|
| 224 |
-
yield state, state, click_state, image_input
|
| 225 |
if not args.disable_gpt and model.text_refiner:
|
| 226 |
refined_caption = model.text_refiner.inference(query=text, controls=controls, context=out['context_captions'],
|
| 227 |
enable_wiki=enable_wiki)
|
| 228 |
# new_cap = 'Original: ' + text + '. Refined: ' + refined_caption['caption']
|
| 229 |
new_cap = refined_caption['caption']
|
| 230 |
-
|
|
|
|
| 231 |
state = state + [(None, f"caption: {new_cap}")]
|
| 232 |
refined_image_input = create_bubble_frame(origin_image_input, new_cap, (click_index[0], click_index[1]),
|
| 233 |
input_mask,
|
| 234 |
input_points=input_points, input_labels=input_labels)
|
| 235 |
-
yield state, state, click_state, refined_image_input
|
| 236 |
|
| 237 |
|
| 238 |
-
def get_sketch_prompt(mask:
|
| 239 |
"""
|
| 240 |
Get the prompt for the sketcher.
|
| 241 |
TODO: This is a temporary solution. We should cluster the sketch and get the bounding box of each cluster.
|
|
@@ -282,12 +280,11 @@ def inference_traject(sketcher_image, enable_wiki, language, sentiment, factuali
|
|
| 282 |
model.setup(image_embedding, original_size, input_size, is_image_set=True)
|
| 283 |
|
| 284 |
enable_wiki = True if enable_wiki in ['True', 'TRUE', 'true', True, 'Yes', 'YES', 'yes'] else False
|
| 285 |
-
out = model.inference(image_input, prompt, controls, disable_gpt=True, enable_wiki=enable_wiki)
|
| 286 |
|
| 287 |
# Update components and states
|
| 288 |
state.append((f'Box: {boxes}', None))
|
| 289 |
state.append((None, f'raw_caption: {out["generated_captions"]["raw_caption"]}'))
|
| 290 |
-
wiki = out['generated_captions'].get('wiki', "")
|
| 291 |
text = out['generated_captions']['raw_caption']
|
| 292 |
input_mask = np.array(out['mask'].convert('P'))
|
| 293 |
image_input = mask_painter(np.array(image_input), input_mask)
|
|
@@ -297,18 +294,19 @@ def inference_traject(sketcher_image, enable_wiki, language, sentiment, factuali
|
|
| 297 |
fake_click_index = (int((boxes[0][0] + boxes[0][2]) / 2), int((boxes[0][1] + boxes[0][3]) / 2))
|
| 298 |
image_input = create_bubble_frame(image_input, text, fake_click_index, input_mask)
|
| 299 |
|
| 300 |
-
yield state, state, image_input
|
| 301 |
|
| 302 |
if not args.disable_gpt and model.text_refiner:
|
| 303 |
refined_caption = model.text_refiner.inference(query=text, controls=controls, context=out['context_captions'],
|
| 304 |
enable_wiki=enable_wiki)
|
| 305 |
|
| 306 |
new_cap = refined_caption['caption']
|
| 307 |
-
|
|
|
|
| 308 |
state = state + [(None, f"caption: {new_cap}")]
|
| 309 |
refined_image_input = create_bubble_frame(origin_image_input, new_cap, fake_click_index, input_mask)
|
| 310 |
|
| 311 |
-
yield state, state, refined_image_input
|
| 312 |
|
| 313 |
def clear_chat_memory(visual_chatgpt, keep_global=False):
|
| 314 |
if visual_chatgpt is not None:
|
|
@@ -319,7 +317,26 @@ def clear_chat_memory(visual_chatgpt, keep_global=False):
|
|
| 319 |
else:
|
| 320 |
visual_chatgpt.current_image = None
|
| 321 |
visual_chatgpt.global_prompt = ""
|
| 322 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 323 |
def get_style():
|
| 324 |
current_version = version.parse(gr.__version__)
|
| 325 |
if current_version <= version.parse('3.24.1'):
|
|
@@ -400,7 +417,7 @@ def create_ui():
|
|
| 400 |
with gr.Row():
|
| 401 |
submit_button_sketcher = gr.Button(value="Submit", interactive=True)
|
| 402 |
|
| 403 |
-
with gr.Column(visible=False) as
|
| 404 |
with gr.Row(scale=1.0):
|
| 405 |
language = gr.Dropdown(
|
| 406 |
['English', 'Chinese', 'French', "Spanish", "Arabic", "Portuguese", "Cantonese"],
|
|
@@ -431,26 +448,31 @@ def create_ui():
|
|
| 431 |
value="No",
|
| 432 |
label="Enable Wiki",
|
| 433 |
interactive=True)
|
| 434 |
-
with gr.Column(visible=True) as modules_not_need_gpt3:
|
| 435 |
-
|
| 436 |
-
|
| 437 |
-
|
| 438 |
-
|
| 439 |
with gr.Column(scale=0.5):
|
| 440 |
-
|
| 441 |
-
|
| 442 |
-
|
| 443 |
-
|
| 444 |
-
|
| 445 |
-
|
| 446 |
-
|
| 447 |
-
|
| 448 |
-
|
| 449 |
-
|
| 450 |
-
|
| 451 |
-
|
| 452 |
-
|
| 453 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 454 |
with gr.Column(visible=False) as modules_need_gpt3:
|
| 455 |
chat_input = gr.Textbox(show_label=False, placeholder="Enter text and press Enter").style(
|
| 456 |
container=False)
|
|
@@ -459,36 +481,38 @@ def create_ui():
|
|
| 459 |
submit_button_text = gr.Button(value="Submit", interactive=True, variant="primary")
|
| 460 |
|
| 461 |
openai_api_key.submit(init_openai_api_key, inputs=[openai_api_key],
|
| 462 |
-
outputs=[
|
| 463 |
-
modules_not_need_gpt2,
|
| 464 |
enable_chatGPT_button.click(init_openai_api_key, inputs=[openai_api_key],
|
| 465 |
-
outputs=[
|
| 466 |
modules_not_need_gpt,
|
| 467 |
-
modules_not_need_gpt2,
|
| 468 |
-
disable_chatGPT_button.click(
|
| 469 |
-
outputs=[
|
| 470 |
modules_not_need_gpt,
|
| 471 |
-
modules_not_need_gpt2,
|
| 472 |
-
|
| 473 |
enable_chatGPT_button.click(
|
| 474 |
lambda: (None, [], [], [[], [], []], "", "", ""),
|
| 475 |
[],
|
| 476 |
-
[image_input, chatbot, state, click_state,
|
| 477 |
queue=False,
|
| 478 |
show_progress=False
|
| 479 |
)
|
| 480 |
openai_api_key.submit(
|
| 481 |
lambda: (None, [], [], [[], [], []], "", "", ""),
|
| 482 |
[],
|
| 483 |
-
[image_input, chatbot, state, click_state,
|
| 484 |
queue=False,
|
| 485 |
show_progress=False
|
| 486 |
)
|
|
|
|
|
|
|
| 487 |
|
| 488 |
clear_button_click.click(
|
| 489 |
-
lambda x: ([[], [], []], x
|
| 490 |
[origin_image],
|
| 491 |
-
[click_state, image_input
|
| 492 |
queue=False,
|
| 493 |
show_progress=False
|
| 494 |
)
|
|
@@ -496,7 +520,7 @@ def create_ui():
|
|
| 496 |
clear_button_image.click(
|
| 497 |
lambda: (None, [], [], [[], [], []], "", "", ""),
|
| 498 |
[],
|
| 499 |
-
[image_input, chatbot, state, click_state,
|
| 500 |
queue=False,
|
| 501 |
show_progress=False
|
| 502 |
)
|
|
@@ -513,7 +537,7 @@ def create_ui():
|
|
| 513 |
image_input.clear(
|
| 514 |
lambda: (None, [], [], [[], [], []], "", "", ""),
|
| 515 |
[],
|
| 516 |
-
[image_input, chatbot, state, click_state,
|
| 517 |
queue=False,
|
| 518 |
show_progress=False
|
| 519 |
)
|
|
@@ -544,7 +568,7 @@ def create_ui():
|
|
| 544 |
origin_image, point_prompt, click_mode, enable_wiki, language, sentiment, factuality, length,
|
| 545 |
image_embedding, state, click_state, original_size, input_size, text_refiner, visual_chatgpt
|
| 546 |
],
|
| 547 |
-
outputs=[chatbot, state, click_state, image_input
|
| 548 |
show_progress=False, queue=True
|
| 549 |
)
|
| 550 |
|
|
@@ -554,7 +578,7 @@ def create_ui():
|
|
| 554 |
sketcher_input, enable_wiki, language, sentiment, factuality, length, image_embedding, state,
|
| 555 |
original_size, input_size, text_refiner
|
| 556 |
],
|
| 557 |
-
outputs=[chatbot, state, sketcher_input
|
| 558 |
show_progress=False, queue=True
|
| 559 |
)
|
| 560 |
|
|
|
|
| 1 |
import os
|
| 2 |
import json
|
|
|
|
| 3 |
import gradio as gr
|
| 4 |
import numpy as np
|
| 5 |
from gradio import processing_utils
|
|
|
|
| 10 |
|
| 11 |
from caption_anything.model import CaptionAnything
|
| 12 |
from caption_anything.utils.image_editing_utils import create_bubble_frame
|
| 13 |
+
from caption_anything.utils.utils import mask_painter, seg_model_map, prepare_segmenter, image_resize
|
| 14 |
from caption_anything.utils.parser import parse_augment
|
| 15 |
from caption_anything.captioner import build_captioner
|
| 16 |
from caption_anything.text_refiner import build_text_refiner
|
|
|
|
| 22 |
args = parse_augment()
|
| 23 |
args.segmenter = "huge"
|
| 24 |
args.segmenter_checkpoint = "sam_vit_h_4b8939.pth"
|
| 25 |
+
|
| 26 |
if args.segmenter_checkpoint is None:
|
| 27 |
_, segmenter_checkpoint = prepare_segmenter(args.segmenter)
|
| 28 |
else:
|
|
|
|
| 53 |
mask = np.zeros((height, width, 4), dtype=np.uint8)
|
| 54 |
mask[..., -1] = 255
|
| 55 |
mask = self.postprocess(mask)
|
|
|
|
| 56 |
x['mask'] = mask
|
|
|
|
| 57 |
return super().preprocess(x)
|
| 58 |
|
| 59 |
|
|
|
|
| 72 |
if api_key and len(api_key) > 30:
|
| 73 |
try:
|
| 74 |
text_refiner = build_text_refiner(args.text_refiner, args.device, args, api_key)
|
| 75 |
+
assert len(text_refiner.llm('hi')) > 0 # test
|
| 76 |
visual_chatgpt = ConversationBot(shared_chatbot_tools, api_key)
|
| 77 |
except:
|
| 78 |
text_refiner = None
|
| 79 |
visual_chatgpt = None
|
| 80 |
openai_available = text_refiner is not None
|
| 81 |
+
if openai_available:
|
| 82 |
+
return [gr.update(visible=True)]*6 + [gr.update(visible=False)]*2 + [text_refiner, visual_chatgpt, None]
|
| 83 |
+
else:
|
| 84 |
+
return [gr.update(visible=False)]*6 + [gr.update(visible=True)]*2 + [text_refiner, visual_chatgpt, 'Your OpenAI API Key is not available']
|
| 85 |
+
|
| 86 |
+
def init_wo_openai_api_key():
|
| 87 |
+
return [gr.update(visible=False)]*4 + [gr.update(visible=True)]*2 + [gr.update(visible=False)]*2 + [None, None, None]
|
| 88 |
|
| 89 |
def get_click_prompt(chat_input, click_state, click_mode):
|
| 90 |
inputs = json.loads(chat_input)
|
|
|
|
| 131 |
state = state + [(chat_input, response)]
|
| 132 |
return state, state
|
| 133 |
|
| 134 |
+
|
| 135 |
+
|
| 136 |
def upload_callback(image_input, state, visual_chatgpt=None):
|
| 137 |
|
| 138 |
if isinstance(image_input, dict): # if upload from sketcher_input, input contains image and mask
|
| 139 |
image_input, mask = image_input['image'], image_input['mask']
|
| 140 |
|
| 141 |
click_state = [[], [], []]
|
| 142 |
+
image_input = image_resize(image_input, res=1024)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 143 |
|
| 144 |
model = build_caption_anything_with_models(
|
| 145 |
args,
|
|
|
|
| 157 |
new_image_path = get_new_image_name('chat_image', func_name='upload')
|
| 158 |
image_input.save(new_image_path)
|
| 159 |
visual_chatgpt.current_image = new_image_path
|
| 160 |
+
img_caption = model.captioner.inference(image_input, filter=False, args={'text_prompt':''})['caption']
|
| 161 |
+
Human_prompt = f'\nHuman: The description of the image with path {new_image_path} is: {img_caption}. This information helps you to understand this image, but you should use tools to finish following tasks, rather than directly imagine from my description. If you understand, say \"Received\". \n'
|
| 162 |
AI_prompt = "Received."
|
| 163 |
visual_chatgpt.global_prompt = Human_prompt + 'AI: ' + AI_prompt
|
| 164 |
visual_chatgpt.agent.memory.buffer = visual_chatgpt.agent.memory.buffer + visual_chatgpt.global_prompt
|
|
|
|
| 199 |
model.setup(image_embedding, original_size, input_size, is_image_set=True)
|
| 200 |
|
| 201 |
enable_wiki = True if enable_wiki in ['True', 'TRUE', 'true', True, 'Yes', 'YES', 'yes'] else False
|
| 202 |
+
out = model.inference(image_input, prompt, controls, disable_gpt=True, enable_wiki=enable_wiki, verbose=True, args={'clip_filter': False})[0]
|
| 203 |
|
| 204 |
state = state + [("Image point: {}, Input label: {}".format(prompt["input_point"], prompt["input_label"]), None)]
|
| 205 |
state = state + [(None, "raw_caption: {}".format(out['generated_captions']['raw_caption']))]
|
|
|
|
| 206 |
update_click_state(click_state, out['generated_captions']['raw_caption'], click_mode)
|
| 207 |
text = out['generated_captions']['raw_caption']
|
| 208 |
input_mask = np.array(out['mask'].convert('P'))
|
|
|
|
| 218 |
point_prompt = f'You should primarly use tools on the selected regional image (description: {text}, path: {new_crop_save_path}), which is a part of the whole image (path: {visual_chatgpt.current_image}). If human mentioned some objects not in the selected region, you can use tools on the whole image.'
|
| 219 |
visual_chatgpt.point_prompt = point_prompt
|
| 220 |
|
| 221 |
+
yield state, state, click_state, image_input
|
| 222 |
if not args.disable_gpt and model.text_refiner:
|
| 223 |
refined_caption = model.text_refiner.inference(query=text, controls=controls, context=out['context_captions'],
|
| 224 |
enable_wiki=enable_wiki)
|
| 225 |
# new_cap = 'Original: ' + text + '. Refined: ' + refined_caption['caption']
|
| 226 |
new_cap = refined_caption['caption']
|
| 227 |
+
if refined_caption['wiki']:
|
| 228 |
+
state = state + [(None, "Wiki: {}".format(refined_caption['wiki']))]
|
| 229 |
state = state + [(None, f"caption: {new_cap}")]
|
| 230 |
refined_image_input = create_bubble_frame(origin_image_input, new_cap, (click_index[0], click_index[1]),
|
| 231 |
input_mask,
|
| 232 |
input_points=input_points, input_labels=input_labels)
|
| 233 |
+
yield state, state, click_state, refined_image_input
|
| 234 |
|
| 235 |
|
| 236 |
+
def get_sketch_prompt(mask: Image.Image):
|
| 237 |
"""
|
| 238 |
Get the prompt for the sketcher.
|
| 239 |
TODO: This is a temporary solution. We should cluster the sketch and get the bounding box of each cluster.
|
|
|
|
| 280 |
model.setup(image_embedding, original_size, input_size, is_image_set=True)
|
| 281 |
|
| 282 |
enable_wiki = True if enable_wiki in ['True', 'TRUE', 'true', True, 'Yes', 'YES', 'yes'] else False
|
| 283 |
+
out = model.inference(image_input, prompt, controls, disable_gpt=True, enable_wiki=enable_wiki)[0]
|
| 284 |
|
| 285 |
# Update components and states
|
| 286 |
state.append((f'Box: {boxes}', None))
|
| 287 |
state.append((None, f'raw_caption: {out["generated_captions"]["raw_caption"]}'))
|
|
|
|
| 288 |
text = out['generated_captions']['raw_caption']
|
| 289 |
input_mask = np.array(out['mask'].convert('P'))
|
| 290 |
image_input = mask_painter(np.array(image_input), input_mask)
|
|
|
|
| 294 |
fake_click_index = (int((boxes[0][0] + boxes[0][2]) / 2), int((boxes[0][1] + boxes[0][3]) / 2))
|
| 295 |
image_input = create_bubble_frame(image_input, text, fake_click_index, input_mask)
|
| 296 |
|
| 297 |
+
yield state, state, image_input
|
| 298 |
|
| 299 |
if not args.disable_gpt and model.text_refiner:
|
| 300 |
refined_caption = model.text_refiner.inference(query=text, controls=controls, context=out['context_captions'],
|
| 301 |
enable_wiki=enable_wiki)
|
| 302 |
|
| 303 |
new_cap = refined_caption['caption']
|
| 304 |
+
if refined_caption['wiki']:
|
| 305 |
+
state = state + [(None, "Wiki: {}".format(refined_caption['wiki']))]
|
| 306 |
state = state + [(None, f"caption: {new_cap}")]
|
| 307 |
refined_image_input = create_bubble_frame(origin_image_input, new_cap, fake_click_index, input_mask)
|
| 308 |
|
| 309 |
+
yield state, state, refined_image_input
|
| 310 |
|
| 311 |
def clear_chat_memory(visual_chatgpt, keep_global=False):
|
| 312 |
if visual_chatgpt is not None:
|
|
|
|
| 317 |
else:
|
| 318 |
visual_chatgpt.current_image = None
|
| 319 |
visual_chatgpt.global_prompt = ""
|
| 320 |
+
|
| 321 |
+
def cap_everything(image_input, visual_chatgpt, text_refiner):
|
| 322 |
+
|
| 323 |
+
model = build_caption_anything_with_models(
|
| 324 |
+
args,
|
| 325 |
+
api_key="",
|
| 326 |
+
captioner=shared_captioner,
|
| 327 |
+
sam_model=shared_sam_model,
|
| 328 |
+
text_refiner=text_refiner,
|
| 329 |
+
session_id=iface.app_id
|
| 330 |
+
)
|
| 331 |
+
paragraph = model.inference_cap_everything(image_input, verbose=True)
|
| 332 |
+
# state = state + [(None, f"Caption Everything: {paragraph}")]
|
| 333 |
+
Human_prompt = f'\nThe description of the image with path {visual_chatgpt.current_image} is:\n{paragraph}\nThis information helps you to understand this image, but you should use tools to finish following tasks, rather than directly imagine from my description. If you understand, say \"Received\". \n'
|
| 334 |
+
AI_prompt = "Received."
|
| 335 |
+
visual_chatgpt.global_prompt = Human_prompt + 'AI: ' + AI_prompt
|
| 336 |
+
visual_chatgpt.agent.memory.buffer = visual_chatgpt.agent.memory.buffer + visual_chatgpt.global_prompt
|
| 337 |
+
return paragraph
|
| 338 |
+
|
| 339 |
+
|
| 340 |
def get_style():
|
| 341 |
current_version = version.parse(gr.__version__)
|
| 342 |
if current_version <= version.parse('3.24.1'):
|
|
|
|
| 417 |
with gr.Row():
|
| 418 |
submit_button_sketcher = gr.Button(value="Submit", interactive=True)
|
| 419 |
|
| 420 |
+
with gr.Column(visible=False) as modules_need_gpt1:
|
| 421 |
with gr.Row(scale=1.0):
|
| 422 |
language = gr.Dropdown(
|
| 423 |
['English', 'Chinese', 'French', "Spanish", "Arabic", "Portuguese", "Cantonese"],
|
|
|
|
| 448 |
value="No",
|
| 449 |
label="Enable Wiki",
|
| 450 |
interactive=True)
|
| 451 |
+
# with gr.Column(visible=True) as modules_not_need_gpt3:
|
| 452 |
+
gr.Examples(
|
| 453 |
+
examples=examples,
|
| 454 |
+
inputs=[example_image],
|
| 455 |
+
)
|
| 456 |
with gr.Column(scale=0.5):
|
| 457 |
+
with gr.Column(visible=True) as module_key_input:
|
| 458 |
+
openai_api_key = gr.Textbox(
|
| 459 |
+
placeholder="Input openAI API key",
|
| 460 |
+
show_label=False,
|
| 461 |
+
label="OpenAI API Key",
|
| 462 |
+
lines=1,
|
| 463 |
+
type="password")
|
| 464 |
+
with gr.Row(scale=0.5):
|
| 465 |
+
enable_chatGPT_button = gr.Button(value="Run with ChatGPT", interactive=True, variant='primary')
|
| 466 |
+
disable_chatGPT_button = gr.Button(value="Run without ChatGPT (Faster)", interactive=True,
|
| 467 |
+
variant='primary')
|
| 468 |
+
with gr.Column(visible=False) as module_notification_box:
|
| 469 |
+
notification_box = gr.Textbox(lines=1, label="Notification", max_lines=5, show_label=False)
|
| 470 |
+
with gr.Column(visible=False) as modules_need_gpt2:
|
| 471 |
+
paragraph_output = gr.Textbox(lines=7, label="Describe Everything", max_lines=7)
|
| 472 |
+
with gr.Column(visible=False) as modules_need_gpt0:
|
| 473 |
+
cap_everything_button = gr.Button(value="Caption Everything in a Paragraph", interactive=True)
|
| 474 |
+
with gr.Column(visible=False) as modules_not_need_gpt2:
|
| 475 |
+
chatbot = gr.Chatbot(label="Chatbox", ).style(height=550, scale=0.5)
|
| 476 |
with gr.Column(visible=False) as modules_need_gpt3:
|
| 477 |
chat_input = gr.Textbox(show_label=False, placeholder="Enter text and press Enter").style(
|
| 478 |
container=False)
|
|
|
|
| 481 |
submit_button_text = gr.Button(value="Submit", interactive=True, variant="primary")
|
| 482 |
|
| 483 |
openai_api_key.submit(init_openai_api_key, inputs=[openai_api_key],
|
| 484 |
+
outputs=[modules_need_gpt0, modules_need_gpt1, modules_need_gpt2, modules_need_gpt3, modules_not_need_gpt,
|
| 485 |
+
modules_not_need_gpt2, module_key_input, module_notification_box, text_refiner, visual_chatgpt, notification_box])
|
| 486 |
enable_chatGPT_button.click(init_openai_api_key, inputs=[openai_api_key],
|
| 487 |
+
outputs=[modules_need_gpt0, modules_need_gpt1, modules_need_gpt2, modules_need_gpt3,
|
| 488 |
modules_not_need_gpt,
|
| 489 |
+
modules_not_need_gpt2, module_key_input, module_notification_box, text_refiner, visual_chatgpt, notification_box])
|
| 490 |
+
disable_chatGPT_button.click(init_wo_openai_api_key,
|
| 491 |
+
outputs=[modules_need_gpt0, modules_need_gpt1, modules_need_gpt2, modules_need_gpt3,
|
| 492 |
modules_not_need_gpt,
|
| 493 |
+
modules_not_need_gpt2, module_key_input, module_notification_box, text_refiner, visual_chatgpt, notification_box])
|
| 494 |
+
|
| 495 |
enable_chatGPT_button.click(
|
| 496 |
lambda: (None, [], [], [[], [], []], "", "", ""),
|
| 497 |
[],
|
| 498 |
+
[image_input, chatbot, state, click_state, paragraph_output, origin_image],
|
| 499 |
queue=False,
|
| 500 |
show_progress=False
|
| 501 |
)
|
| 502 |
openai_api_key.submit(
|
| 503 |
lambda: (None, [], [], [[], [], []], "", "", ""),
|
| 504 |
[],
|
| 505 |
+
[image_input, chatbot, state, click_state, paragraph_output, origin_image],
|
| 506 |
queue=False,
|
| 507 |
show_progress=False
|
| 508 |
)
|
| 509 |
+
|
| 510 |
+
cap_everything_button.click(cap_everything, [origin_image, visual_chatgpt, text_refiner], [paragraph_output])
|
| 511 |
|
| 512 |
clear_button_click.click(
|
| 513 |
+
lambda x: ([[], [], []], x),
|
| 514 |
[origin_image],
|
| 515 |
+
[click_state, image_input],
|
| 516 |
queue=False,
|
| 517 |
show_progress=False
|
| 518 |
)
|
|
|
|
| 520 |
clear_button_image.click(
|
| 521 |
lambda: (None, [], [], [[], [], []], "", "", ""),
|
| 522 |
[],
|
| 523 |
+
[image_input, chatbot, state, click_state, paragraph_output, origin_image],
|
| 524 |
queue=False,
|
| 525 |
show_progress=False
|
| 526 |
)
|
|
|
|
| 537 |
image_input.clear(
|
| 538 |
lambda: (None, [], [], [[], [], []], "", "", ""),
|
| 539 |
[],
|
| 540 |
+
[image_input, chatbot, state, click_state, paragraph_output, origin_image],
|
| 541 |
queue=False,
|
| 542 |
show_progress=False
|
| 543 |
)
|
|
|
|
| 568 |
origin_image, point_prompt, click_mode, enable_wiki, language, sentiment, factuality, length,
|
| 569 |
image_embedding, state, click_state, original_size, input_size, text_refiner, visual_chatgpt
|
| 570 |
],
|
| 571 |
+
outputs=[chatbot, state, click_state, image_input],
|
| 572 |
show_progress=False, queue=True
|
| 573 |
)
|
| 574 |
|
|
|
|
| 578 |
sketcher_input, enable_wiki, language, sentiment, factuality, length, image_embedding, state,
|
| 579 |
original_size, input_size, text_refiner
|
| 580 |
],
|
| 581 |
+
outputs=[chatbot, state, sketcher_input],
|
| 582 |
show_progress=False, queue=True
|
| 583 |
)
|
| 584 |
|
caption_anything/captioner/base_captioner.py
CHANGED
|
@@ -5,7 +5,7 @@ import json
|
|
| 5 |
import pdb
|
| 6 |
import cv2
|
| 7 |
import numpy as np
|
| 8 |
-
from typing import Union
|
| 9 |
import time
|
| 10 |
import clip
|
| 11 |
|
|
@@ -16,13 +16,10 @@ def boundary(inputs):
|
|
| 16 |
col = inputs.shape[1]
|
| 17 |
inputs = inputs.reshape(-1)
|
| 18 |
lens = len(inputs)
|
| 19 |
-
|
| 20 |
start = np.argmax(inputs)
|
| 21 |
end = lens - 1 - np.argmax(np.flip(inputs))
|
| 22 |
-
|
| 23 |
top = start // col
|
| 24 |
bottom = end // col
|
| 25 |
-
|
| 26 |
return top, bottom
|
| 27 |
|
| 28 |
|
|
@@ -84,27 +81,27 @@ class BaseCaptioner:
|
|
| 84 |
self.enable_filter = enable_filter
|
| 85 |
if enable_filter:
|
| 86 |
self.filter, self.preprocess = clip.load('ViT-B/32', device)
|
| 87 |
-
self.threshold = 0.2
|
| 88 |
|
| 89 |
@torch.no_grad()
|
| 90 |
-
def filter_caption(self, image: Union[np.ndarray, Image.Image, str], caption: str):
|
| 91 |
-
|
| 92 |
image = load_image(image, return_type='pil')
|
| 93 |
-
|
| 94 |
image = self.preprocess(image).unsqueeze(0).to(self.device) # (1, 3, 224, 224)
|
| 95 |
-
|
|
|
|
|
|
|
|
|
|
| 96 |
image_features = self.filter.encode_image(image) # (1, 512)
|
| 97 |
-
text_features = self.filter.encode_text(text)
|
| 98 |
image_features /= image_features.norm(dim=-1, keepdim=True)
|
| 99 |
text_features /= text_features.norm(dim=-1, keepdim=True)
|
| 100 |
-
|
| 101 |
-
if
|
| 102 |
-
|
| 103 |
-
|
| 104 |
else:
|
| 105 |
-
|
| 106 |
print(f'Clip score of the caption is {similarity}')
|
| 107 |
-
return
|
| 108 |
|
| 109 |
def inference(self, image: Union[np.ndarray, Image.Image, str], filter: bool = False):
|
| 110 |
raise NotImplementedError()
|
|
@@ -112,7 +109,7 @@ class BaseCaptioner:
|
|
| 112 |
def inference_with_reduced_tokens(self, image: Union[np.ndarray, Image.Image, str], seg_mask, filter: bool = False):
|
| 113 |
raise NotImplementedError()
|
| 114 |
|
| 115 |
-
def inference_box(self, image: Union[np.ndarray, Image.Image, str], box: Union[list, np.ndarray], filter=False):
|
| 116 |
image = load_image(image, return_type="pil")
|
| 117 |
|
| 118 |
if np.array(box).size == 4:
|
|
@@ -123,23 +120,31 @@ class BaseCaptioner:
|
|
| 123 |
elif np.array(box).size == 8: # four corners of an irregular rectangle
|
| 124 |
image_crop = cut_box(np.array(image), box)
|
| 125 |
|
| 126 |
-
crop_save_path =
|
| 127 |
-
|
| 128 |
-
|
| 129 |
-
|
| 130 |
-
|
| 131 |
-
|
| 132 |
-
|
| 133 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 134 |
if seg_mask is None:
|
| 135 |
seg_mask = np.ones(image.size).astype(bool)
|
| 136 |
-
|
| 137 |
image = load_image(image, return_type="pil")
|
| 138 |
seg_mask = load_image(seg_mask, return_type="pil")
|
| 139 |
|
| 140 |
seg_mask = seg_mask.resize(image.size)
|
| 141 |
seg_mask = np.array(seg_mask) > 0
|
| 142 |
-
|
| 143 |
if crop_mode == "wo_bg":
|
| 144 |
image = np.array(image) * seg_mask[:, :, np.newaxis] + (1 - seg_mask[:, :, np.newaxis]) * 255
|
| 145 |
image = np.uint8(image)
|
|
@@ -150,10 +155,13 @@ class BaseCaptioner:
|
|
| 150 |
min_area_box = seg_to_box(seg_mask)
|
| 151 |
else:
|
| 152 |
min_area_box = new_seg_to_box(seg_mask)
|
| 153 |
-
return self.inference_box(image, min_area_box, filter)
|
| 154 |
|
| 155 |
-
def generate_seg_cropped_image(self,
|
| 156 |
-
|
|
|
|
|
|
|
|
|
|
| 157 |
image = load_image(image, return_type="pil")
|
| 158 |
seg_mask = load_image(seg_mask, return_type="pil")
|
| 159 |
|
|
|
|
| 5 |
import pdb
|
| 6 |
import cv2
|
| 7 |
import numpy as np
|
| 8 |
+
from typing import Any, Union, List
|
| 9 |
import time
|
| 10 |
import clip
|
| 11 |
|
|
|
|
| 16 |
col = inputs.shape[1]
|
| 17 |
inputs = inputs.reshape(-1)
|
| 18 |
lens = len(inputs)
|
|
|
|
| 19 |
start = np.argmax(inputs)
|
| 20 |
end = lens - 1 - np.argmax(np.flip(inputs))
|
|
|
|
| 21 |
top = start // col
|
| 22 |
bottom = end // col
|
|
|
|
| 23 |
return top, bottom
|
| 24 |
|
| 25 |
|
|
|
|
| 81 |
self.enable_filter = enable_filter
|
| 82 |
if enable_filter:
|
| 83 |
self.filter, self.preprocess = clip.load('ViT-B/32', device)
|
|
|
|
| 84 |
|
| 85 |
@torch.no_grad()
|
| 86 |
+
def filter_caption(self, image: Union[np.ndarray, Image.Image, str], caption: str, reference_caption: List[str]=[]):
|
|
|
|
| 87 |
image = load_image(image, return_type='pil')
|
|
|
|
| 88 |
image = self.preprocess(image).unsqueeze(0).to(self.device) # (1, 3, 224, 224)
|
| 89 |
+
captions = [caption]
|
| 90 |
+
if len(reference_caption):
|
| 91 |
+
captions.extend(reference_caption)
|
| 92 |
+
text = clip.tokenize(captions).to(self.device) # (>1, 77)
|
| 93 |
image_features = self.filter.encode_image(image) # (1, 512)
|
| 94 |
+
text_features = self.filter.encode_text(text) # # (>1, 512)
|
| 95 |
image_features /= image_features.norm(dim=-1, keepdim=True)
|
| 96 |
text_features /= text_features.norm(dim=-1, keepdim=True)
|
| 97 |
+
|
| 98 |
+
if len(reference_caption):
|
| 99 |
+
similarity = torch.matmul(image_features, text_features.transpose(1, 0)) / 0.07
|
| 100 |
+
similarity = similarity.softmax(dim=1)[0, 0].item()
|
| 101 |
else:
|
| 102 |
+
similarity = torch.matmul(image_features, text_features.transpose(1, 0)).item()
|
| 103 |
print(f'Clip score of the caption is {similarity}')
|
| 104 |
+
return similarity
|
| 105 |
|
| 106 |
def inference(self, image: Union[np.ndarray, Image.Image, str], filter: bool = False):
|
| 107 |
raise NotImplementedError()
|
|
|
|
| 109 |
def inference_with_reduced_tokens(self, image: Union[np.ndarray, Image.Image, str], seg_mask, filter: bool = False):
|
| 110 |
raise NotImplementedError()
|
| 111 |
|
| 112 |
+
def inference_box(self, image: Union[np.ndarray, Image.Image, str], box: Union[list, np.ndarray], filter=False, verbose=False, caption_args={}):
|
| 113 |
image = load_image(image, return_type="pil")
|
| 114 |
|
| 115 |
if np.array(box).size == 4:
|
|
|
|
| 120 |
elif np.array(box).size == 8: # four corners of an irregular rectangle
|
| 121 |
image_crop = cut_box(np.array(image), box)
|
| 122 |
|
| 123 |
+
crop_save_path = None
|
| 124 |
+
if verbose:
|
| 125 |
+
crop_save_path = f'result/crop_{time.time()}.png'
|
| 126 |
+
Image.fromarray(image_crop).save(crop_save_path)
|
| 127 |
+
print(f'croped image saved in {crop_save_path}')
|
| 128 |
+
caption = self.inference(image_crop, filter, caption_args)
|
| 129 |
+
caption.update({'crop_save_path': crop_save_path})
|
| 130 |
+
return caption
|
| 131 |
+
|
| 132 |
+
def inference_seg(self,
|
| 133 |
+
image: Union[np.ndarray, str],
|
| 134 |
+
seg_mask: Union[np.ndarray, Image.Image, str] = None,
|
| 135 |
+
crop_mode="w_bg",
|
| 136 |
+
filter=False,
|
| 137 |
+
disable_regular_box=False,
|
| 138 |
+
verbose=False,
|
| 139 |
+
caption_args={}):
|
| 140 |
if seg_mask is None:
|
| 141 |
seg_mask = np.ones(image.size).astype(bool)
|
| 142 |
+
|
| 143 |
image = load_image(image, return_type="pil")
|
| 144 |
seg_mask = load_image(seg_mask, return_type="pil")
|
| 145 |
|
| 146 |
seg_mask = seg_mask.resize(image.size)
|
| 147 |
seg_mask = np.array(seg_mask) > 0
|
|
|
|
| 148 |
if crop_mode == "wo_bg":
|
| 149 |
image = np.array(image) * seg_mask[:, :, np.newaxis] + (1 - seg_mask[:, :, np.newaxis]) * 255
|
| 150 |
image = np.uint8(image)
|
|
|
|
| 155 |
min_area_box = seg_to_box(seg_mask)
|
| 156 |
else:
|
| 157 |
min_area_box = new_seg_to_box(seg_mask)
|
| 158 |
+
return self.inference_box(image, min_area_box, filter, verbose, caption_args)
|
| 159 |
|
| 160 |
+
def generate_seg_cropped_image(self,
|
| 161 |
+
image: Union[np.ndarray, str],
|
| 162 |
+
seg_mask: Union[np.ndarray, Image.Image, str],
|
| 163 |
+
crop_mode="w_bg",
|
| 164 |
+
disable_regular_box=False):
|
| 165 |
image = load_image(image, return_type="pil")
|
| 166 |
seg_mask = load_image(seg_mask, return_type="pil")
|
| 167 |
|
caption_anything/captioner/blip.py
CHANGED
|
@@ -20,19 +20,24 @@ class BLIPCaptioner(BaseCaptioner):
|
|
| 20 |
torch_dtype=self.torch_dtype).to(self.device)
|
| 21 |
|
| 22 |
@torch.no_grad()
|
| 23 |
-
def inference(self, image: Union[np.ndarray, Image.Image, str], filter=False):
|
| 24 |
image = load_image(image, return_type="pil")
|
| 25 |
inputs = self.processor(image, return_tensors="pt").to(self.device, self.torch_dtype)
|
| 26 |
out = self.model.generate(**inputs, max_new_tokens=50)
|
| 27 |
captions = self.processor.decode(out[0], skip_special_tokens=True).strip()
|
|
|
|
|
|
|
| 28 |
if self.enable_filter and filter:
|
| 29 |
-
|
|
|
|
|
|
|
| 30 |
print(f"\nProcessed ImageCaptioning by BLIPCaptioner, Output Text: {captions}")
|
| 31 |
-
return captions
|
| 32 |
|
| 33 |
@torch.no_grad()
|
| 34 |
def inference_with_reduced_tokens(self, image: Union[np.ndarray, Image.Image, str], seg_mask, crop_mode="w_bg",
|
| 35 |
filter=False, disable_regular_box=False):
|
|
|
|
| 36 |
crop_save_path = self.generate_seg_cropped_image(image=image, seg_mask=seg_mask, crop_mode=crop_mode,
|
| 37 |
disable_regular_box=disable_regular_box)
|
| 38 |
image = load_image(image, return_type="pil")
|
|
@@ -47,9 +52,11 @@ class BLIPCaptioner(BaseCaptioner):
|
|
| 47 |
out = self.model.generate(pixel_values=pixel_values, pixel_masks=pixel_masks, max_new_tokens=50)
|
| 48 |
captions = self.processor.decode(out[0], skip_special_tokens=True).strip()
|
| 49 |
if self.enable_filter and filter:
|
| 50 |
-
|
|
|
|
|
|
|
| 51 |
print(f"\nProcessed ImageCaptioning by BLIPCaptioner, Output Text: {captions}")
|
| 52 |
-
return
|
| 53 |
|
| 54 |
|
| 55 |
if __name__ == '__main__':
|
|
|
|
| 20 |
torch_dtype=self.torch_dtype).to(self.device)
|
| 21 |
|
| 22 |
@torch.no_grad()
|
| 23 |
+
def inference(self, image: Union[np.ndarray, Image.Image, str], filter=False, args={}):
|
| 24 |
image = load_image(image, return_type="pil")
|
| 25 |
inputs = self.processor(image, return_tensors="pt").to(self.device, self.torch_dtype)
|
| 26 |
out = self.model.generate(**inputs, max_new_tokens=50)
|
| 27 |
captions = self.processor.decode(out[0], skip_special_tokens=True).strip()
|
| 28 |
+
|
| 29 |
+
result = {}
|
| 30 |
if self.enable_filter and filter:
|
| 31 |
+
clip_score = self.filter_caption(image, captions)
|
| 32 |
+
result['clip_score'] = clip_score
|
| 33 |
+
result.update({'caption':captions})
|
| 34 |
print(f"\nProcessed ImageCaptioning by BLIPCaptioner, Output Text: {captions}")
|
| 35 |
+
return {'caption': captions}
|
| 36 |
|
| 37 |
@torch.no_grad()
|
| 38 |
def inference_with_reduced_tokens(self, image: Union[np.ndarray, Image.Image, str], seg_mask, crop_mode="w_bg",
|
| 39 |
filter=False, disable_regular_box=False):
|
| 40 |
+
result = {}
|
| 41 |
crop_save_path = self.generate_seg_cropped_image(image=image, seg_mask=seg_mask, crop_mode=crop_mode,
|
| 42 |
disable_regular_box=disable_regular_box)
|
| 43 |
image = load_image(image, return_type="pil")
|
|
|
|
| 52 |
out = self.model.generate(pixel_values=pixel_values, pixel_masks=pixel_masks, max_new_tokens=50)
|
| 53 |
captions = self.processor.decode(out[0], skip_special_tokens=True).strip()
|
| 54 |
if self.enable_filter and filter:
|
| 55 |
+
clip_score = self.filter_caption(image, captions)
|
| 56 |
+
result['clip_score'] = clip_score
|
| 57 |
+
result.update({'caption':captions, 'crop_save_path':crop_save_path})
|
| 58 |
print(f"\nProcessed ImageCaptioning by BLIPCaptioner, Output Text: {captions}")
|
| 59 |
+
return result
|
| 60 |
|
| 61 |
|
| 62 |
if __name__ == '__main__':
|
caption_anything/captioner/blip2.py
CHANGED
|
@@ -20,18 +20,31 @@ class BLIP2Captioner(BaseCaptioner):
|
|
| 20 |
self.model = Blip2ForConditionalGeneration.from_pretrained("Salesforce/blip2-opt-2.7b", device_map='sequential', load_in_8bit=True)
|
| 21 |
|
| 22 |
@torch.no_grad()
|
| 23 |
-
def inference(self,
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 24 |
image = load_image(image, return_type="pil")
|
| 25 |
-
|
| 26 |
if not self.dialogue:
|
| 27 |
-
|
| 28 |
-
|
| 29 |
-
|
| 30 |
-
|
| 31 |
if self.enable_filter and filter:
|
| 32 |
-
|
| 33 |
-
|
| 34 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 35 |
else:
|
| 36 |
context = []
|
| 37 |
template = "Question: {} Answer: {}."
|
|
@@ -44,8 +57,8 @@ class BLIP2Captioner(BaseCaptioner):
|
|
| 44 |
out = self.model.generate(**inputs, max_new_tokens=50)
|
| 45 |
captions = self.processor.decode(out[0], skip_special_tokens=True).strip()
|
| 46 |
context.append((input_texts, captions))
|
| 47 |
-
|
| 48 |
-
|
| 49 |
|
| 50 |
if __name__ == '__main__':
|
| 51 |
|
|
|
|
| 20 |
self.model = Blip2ForConditionalGeneration.from_pretrained("Salesforce/blip2-opt-2.7b", device_map='sequential', load_in_8bit=True)
|
| 21 |
|
| 22 |
@torch.no_grad()
|
| 23 |
+
def inference(self,
|
| 24 |
+
image: Union[np.ndarray, Image.Image, str],
|
| 25 |
+
filter=False,
|
| 26 |
+
args={}):
|
| 27 |
+
args['return_ppl'] = args.get('return_ppl', False)
|
| 28 |
+
args['text_prompt'] = args.get('text_prompt', 'Question: what does the image show? Answer:')
|
| 29 |
+
args['reference_caption'] = args.get('reference_caption', [])
|
| 30 |
+
|
| 31 |
image = load_image(image, return_type="pil")
|
| 32 |
+
result = {}
|
| 33 |
if not self.dialogue:
|
| 34 |
+
inputs = self.processor(image, text = args['text_prompt'], return_tensors="pt").to(self.device, self.torch_dtype)
|
| 35 |
+
out = self.model.generate(**inputs, return_dict_in_generate=True, output_scores=True, max_new_tokens=50)
|
| 36 |
+
captions = self.processor.batch_decode(out.sequences, skip_special_tokens=True)
|
| 37 |
+
caption = [caption.strip() for caption in captions][0]
|
| 38 |
if self.enable_filter and filter:
|
| 39 |
+
print('reference caption: {}, caption: {}'.format(args['reference_caption'], caption))
|
| 40 |
+
clip_score = self.filter_caption(image, caption, args['reference_caption'])
|
| 41 |
+
result['clip_score'] = clip_score
|
| 42 |
+
if args['return_ppl']:
|
| 43 |
+
ppl_score = torch.stack(out.scores, dim=1).softmax(dim=2).log().max(dim=2)[0].sum(dim=1)[0]
|
| 44 |
+
result['ppl_score'] = ppl_score.item()
|
| 45 |
+
print(f"\nProcessed ImageCaptioning by BLIP2Captioner, Output Text: {caption}")
|
| 46 |
+
result['caption'] = caption
|
| 47 |
+
return result
|
| 48 |
else:
|
| 49 |
context = []
|
| 50 |
template = "Question: {} Answer: {}."
|
|
|
|
| 57 |
out = self.model.generate(**inputs, max_new_tokens=50)
|
| 58 |
captions = self.processor.decode(out[0], skip_special_tokens=True).strip()
|
| 59 |
context.append((input_texts, captions))
|
| 60 |
+
result['caption'] = captions
|
| 61 |
+
return result
|
| 62 |
|
| 63 |
if __name__ == '__main__':
|
| 64 |
|
caption_anything/captioner/git.py
CHANGED
|
@@ -19,19 +19,24 @@ class GITCaptioner(BaseCaptioner):
|
|
| 19 |
self.model = GitForCausalLM.from_pretrained("microsoft/git-large", torch_dtype=self.torch_dtype).to(self.device)
|
| 20 |
|
| 21 |
@torch.no_grad()
|
| 22 |
-
def inference(self, image: Union[np.ndarray, Image.Image, str], filter=False):
|
| 23 |
image = load_image(image, return_type="pil")
|
| 24 |
pixel_values = self.processor(images=image, return_tensors="pt").pixel_values.to(self.device, self.torch_dtype)
|
| 25 |
generated_ids = self.model.generate(pixel_values=pixel_values, max_new_tokens=50)
|
| 26 |
-
|
|
|
|
|
|
|
| 27 |
if self.enable_filter and filter:
|
| 28 |
-
|
| 29 |
-
|
| 30 |
-
|
|
|
|
|
|
|
| 31 |
|
| 32 |
@torch.no_grad()
|
| 33 |
def inference_with_reduced_tokens(self, image: Union[np.ndarray, Image.Image, str], seg_mask, crop_mode="w_bg",
|
| 34 |
filter=False, disable_regular_box=False):
|
|
|
|
| 35 |
crop_save_path = self.generate_seg_cropped_image(image=image, seg_mask=seg_mask, crop_mode=crop_mode,
|
| 36 |
disable_regular_box=disable_regular_box)
|
| 37 |
image = load_image(image, return_type="pil")
|
|
@@ -46,9 +51,11 @@ class GITCaptioner(BaseCaptioner):
|
|
| 46 |
out = self.model.generate(pixel_values=pixel_values, pixel_masks=pixel_masks, max_new_tokens=50)
|
| 47 |
captions = self.processor.decode(out[0], skip_special_tokens=True).strip()
|
| 48 |
if self.enable_filter and filter:
|
| 49 |
-
|
|
|
|
| 50 |
print(f"\nProcessed ImageCaptioning by BLIPCaptioner, Output Text: {captions}")
|
| 51 |
-
|
|
|
|
| 52 |
|
| 53 |
|
| 54 |
if __name__ == '__main__':
|
|
|
|
| 19 |
self.model = GitForCausalLM.from_pretrained("microsoft/git-large", torch_dtype=self.torch_dtype).to(self.device)
|
| 20 |
|
| 21 |
@torch.no_grad()
|
| 22 |
+
def inference(self, image: Union[np.ndarray, Image.Image, str], filter=False, args={}):
|
| 23 |
image = load_image(image, return_type="pil")
|
| 24 |
pixel_values = self.processor(images=image, return_tensors="pt").pixel_values.to(self.device, self.torch_dtype)
|
| 25 |
generated_ids = self.model.generate(pixel_values=pixel_values, max_new_tokens=50)
|
| 26 |
+
captions = self.processor.batch_decode(generated_ids, skip_special_tokens=True)[0].strip()
|
| 27 |
+
|
| 28 |
+
result = {}
|
| 29 |
if self.enable_filter and filter:
|
| 30 |
+
clip_score = self.filter_caption(image, captions)
|
| 31 |
+
result['clip_score'] = clip_score
|
| 32 |
+
result.update({'caption':captions})
|
| 33 |
+
print(f"\nProcessed ImageCaptioning by GITCaptioner, Output Text: {captions}")
|
| 34 |
+
return {'caption': captions}
|
| 35 |
|
| 36 |
@torch.no_grad()
|
| 37 |
def inference_with_reduced_tokens(self, image: Union[np.ndarray, Image.Image, str], seg_mask, crop_mode="w_bg",
|
| 38 |
filter=False, disable_regular_box=False):
|
| 39 |
+
result = {}
|
| 40 |
crop_save_path = self.generate_seg_cropped_image(image=image, seg_mask=seg_mask, crop_mode=crop_mode,
|
| 41 |
disable_regular_box=disable_regular_box)
|
| 42 |
image = load_image(image, return_type="pil")
|
|
|
|
| 51 |
out = self.model.generate(pixel_values=pixel_values, pixel_masks=pixel_masks, max_new_tokens=50)
|
| 52 |
captions = self.processor.decode(out[0], skip_special_tokens=True).strip()
|
| 53 |
if self.enable_filter and filter:
|
| 54 |
+
clip_score = self.filter_caption(image, captions)
|
| 55 |
+
result['clip_score'] = clip_score
|
| 56 |
print(f"\nProcessed ImageCaptioning by BLIPCaptioner, Output Text: {captions}")
|
| 57 |
+
result.update({'caption':captions, 'crop_save_path':crop_save_path})
|
| 58 |
+
return result
|
| 59 |
|
| 60 |
|
| 61 |
if __name__ == '__main__':
|
caption_anything/model.py
CHANGED
|
@@ -5,24 +5,33 @@ import time
|
|
| 5 |
from PIL import Image
|
| 6 |
import cv2
|
| 7 |
import numpy as np
|
|
|
|
|
|
|
|
|
|
| 8 |
from caption_anything.captioner import build_captioner, BaseCaptioner
|
| 9 |
-
from caption_anything.segmenter import build_segmenter
|
| 10 |
from caption_anything.text_refiner import build_text_refiner
|
| 11 |
-
|
| 12 |
-
|
|
|
|
|
|
|
| 13 |
class CaptionAnything:
|
| 14 |
def __init__(self, args, api_key="", captioner=None, segmenter=None, text_refiner=None):
|
| 15 |
self.args = args
|
| 16 |
self.captioner = build_captioner(args.captioner, args.device, args) if captioner is None else captioner
|
| 17 |
self.segmenter = build_segmenter(args.segmenter, args.device, args) if segmenter is None else segmenter
|
| 18 |
-
|
|
|
|
|
|
|
|
|
|
| 19 |
self.text_refiner = None
|
| 20 |
if not args.disable_gpt:
|
| 21 |
if text_refiner is not None:
|
| 22 |
self.text_refiner = text_refiner
|
| 23 |
-
|
| 24 |
self.init_refiner(api_key)
|
| 25 |
-
|
|
|
|
| 26 |
@property
|
| 27 |
def image_embedding(self):
|
| 28 |
return self.segmenter.image_embedding
|
|
@@ -61,65 +70,195 @@ class CaptionAnything:
|
|
| 61 |
self.text_refiner = None
|
| 62 |
print('OpenAI GPT is not available')
|
| 63 |
|
| 64 |
-
def inference(self, image, prompt, controls, disable_gpt=False, enable_wiki=False):
|
| 65 |
-
# TODO: Add support to multiple seg masks.
|
| 66 |
-
|
| 67 |
# segment with prompt
|
| 68 |
print("CA prompt: ", prompt, "CA controls", controls)
|
| 69 |
-
|
| 70 |
-
|
| 71 |
-
|
| 72 |
-
|
| 73 |
-
|
| 74 |
-
|
| 75 |
-
|
| 76 |
-
|
| 77 |
-
|
| 78 |
-
|
| 79 |
-
|
| 80 |
-
|
| 81 |
-
|
| 82 |
-
seg_mask_img = seg_mask_img.convert('RGB')
|
| 83 |
-
seg_mask_img.save(mask_save_path)
|
| 84 |
-
print('seg_mask path: ', mask_save_path)
|
| 85 |
-
print("seg_mask.shape: ", seg_mask.shape)
|
| 86 |
-
|
| 87 |
-
# captioning with mask
|
| 88 |
-
if self.args.enable_reduce_tokens:
|
| 89 |
-
caption, crop_save_path = self.captioner. \
|
| 90 |
-
inference_with_reduced_tokens(image, seg_mask,
|
| 91 |
-
crop_mode=self.args.seg_crop_mode,
|
| 92 |
-
filter=self.args.clip_filter,
|
| 93 |
-
disable_regular_box=self.args.disable_regular_box)
|
| 94 |
-
else:
|
| 95 |
-
caption, crop_save_path = self.captioner. \
|
| 96 |
-
inference_seg(image, seg_mask, crop_mode=self.args.seg_crop_mode,
|
| 97 |
-
filter=self.args.clip_filter,
|
| 98 |
-
disable_regular_box=self.args.disable_regular_box)
|
| 99 |
-
|
| 100 |
-
# refining with TextRefiner
|
| 101 |
-
context_captions = []
|
| 102 |
-
if self.args.context_captions:
|
| 103 |
-
context_captions.append(self.captioner.inference(image))
|
| 104 |
-
if not disable_gpt and self.text_refiner is not None:
|
| 105 |
-
refined_caption = self.text_refiner.inference(query=caption, controls=controls, context=context_captions,
|
| 106 |
-
enable_wiki=enable_wiki)
|
| 107 |
else:
|
| 108 |
-
|
| 109 |
-
|
| 110 |
-
|
| 111 |
-
|
| 112 |
-
|
| 113 |
-
|
| 114 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
| 115 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 116 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 117 |
if __name__ == "__main__":
|
| 118 |
from caption_anything.utils.parser import parse_augment
|
| 119 |
-
|
| 120 |
args = parse_augment()
|
| 121 |
-
|
| 122 |
-
|
| 123 |
prompts = [
|
| 124 |
{
|
| 125 |
"prompt_type": ["click"],
|
|
@@ -127,12 +266,12 @@ if __name__ == "__main__":
|
|
| 127 |
"input_label": [1, 0],
|
| 128 |
"multimask_output": "True",
|
| 129 |
},
|
| 130 |
-
{
|
| 131 |
-
|
| 132 |
-
|
| 133 |
-
|
| 134 |
-
|
| 135 |
-
}
|
| 136 |
]
|
| 137 |
controls = {
|
| 138 |
"length": "30",
|
|
@@ -143,11 +282,11 @@ if __name__ == "__main__":
|
|
| 143 |
}
|
| 144 |
|
| 145 |
model = CaptionAnything(args, os.environ['OPENAI_API_KEY'])
|
| 146 |
-
|
| 147 |
-
|
| 148 |
-
|
| 149 |
-
|
| 150 |
-
|
| 151 |
-
print('
|
| 152 |
-
|
| 153 |
-
|
|
|
|
| 5 |
from PIL import Image
|
| 6 |
import cv2
|
| 7 |
import numpy as np
|
| 8 |
+
from PIL import Image
|
| 9 |
+
import easyocr
|
| 10 |
+
import copy
|
| 11 |
from caption_anything.captioner import build_captioner, BaseCaptioner
|
| 12 |
+
from caption_anything.segmenter import build_segmenter, build_segmenter_densecap
|
| 13 |
from caption_anything.text_refiner import build_text_refiner
|
| 14 |
+
from caption_anything.utils.utils import prepare_segmenter, seg_model_map, load_image, get_image_shape
|
| 15 |
+
from caption_anything.utils.utils import mask_painter_foreground_all, mask_painter, xywh_to_x1y1x2y2, image_resize
|
| 16 |
+
from caption_anything.utils.densecap_painter import draw_bbox
|
| 17 |
+
|
| 18 |
class CaptionAnything:
|
| 19 |
def __init__(self, args, api_key="", captioner=None, segmenter=None, text_refiner=None):
|
| 20 |
self.args = args
|
| 21 |
self.captioner = build_captioner(args.captioner, args.device, args) if captioner is None else captioner
|
| 22 |
self.segmenter = build_segmenter(args.segmenter, args.device, args) if segmenter is None else segmenter
|
| 23 |
+
self.segmenter_densecap = build_segmenter_densecap(args.segmenter, args.device, args, model=self.segmenter.model)
|
| 24 |
+
|
| 25 |
+
self.lang = ["ch_tra", "en"]
|
| 26 |
+
self.reader = easyocr.Reader(self.lang)
|
| 27 |
self.text_refiner = None
|
| 28 |
if not args.disable_gpt:
|
| 29 |
if text_refiner is not None:
|
| 30 |
self.text_refiner = text_refiner
|
| 31 |
+
elif api_key != "":
|
| 32 |
self.init_refiner(api_key)
|
| 33 |
+
self.require_caption_prompt = args.captioner == 'blip2'
|
| 34 |
+
|
| 35 |
@property
|
| 36 |
def image_embedding(self):
|
| 37 |
return self.segmenter.image_embedding
|
|
|
|
| 70 |
self.text_refiner = None
|
| 71 |
print('OpenAI GPT is not available')
|
| 72 |
|
| 73 |
+
def inference(self, image, prompt, controls, disable_gpt=False, enable_wiki=False, verbose=False, is_densecap=False, args={}):
|
|
|
|
|
|
|
| 74 |
# segment with prompt
|
| 75 |
print("CA prompt: ", prompt, "CA controls", controls)
|
| 76 |
+
is_seg_everything = 'everything' in prompt['prompt_type']
|
| 77 |
+
|
| 78 |
+
args['seg_crop_mode'] = args.get('seg_crop_mode', self.args.seg_crop_mode)
|
| 79 |
+
args['clip_filter'] = args.get('clip_filter', self.args.clip_filter)
|
| 80 |
+
args['disable_regular_box'] = args.get('disable_regular_box', self.args.disable_regular_box)
|
| 81 |
+
args['context_captions'] = args.get('context_captions', self.args.context_captions)
|
| 82 |
+
args['enable_reduce_tokens'] = args.get('enable_reduce_tokens', self.args.enable_reduce_tokens)
|
| 83 |
+
args['enable_morphologyex'] = args.get('enable_morphologyex', self.args.enable_morphologyex)
|
| 84 |
+
args['topN'] = args.get('topN', 10) if is_seg_everything else 1
|
| 85 |
+
args['min_mask_area'] = args.get('min_mask_area', 0)
|
| 86 |
+
|
| 87 |
+
if not is_densecap:
|
| 88 |
+
seg_results = self.segmenter.inference(image, prompt)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 89 |
else:
|
| 90 |
+
seg_results = self.segmenter_densecap.inference(image, prompt)
|
| 91 |
+
|
| 92 |
+
seg_masks, seg_bbox, seg_area = seg_results if is_seg_everything else (seg_results, None, None)
|
| 93 |
+
|
| 94 |
+
if args['topN'] > 1: # sort by area
|
| 95 |
+
samples = list(zip(*[seg_masks, seg_bbox, seg_area]))
|
| 96 |
+
# top_samples = sorted(samples, key=lambda x: x[2], reverse=True)
|
| 97 |
+
# seg_masks, seg_bbox, seg_area = list(zip(*top_samples))
|
| 98 |
+
samples = list(filter(lambda x: x[2] > args['min_mask_area'], samples))
|
| 99 |
+
samples = samples[:args['topN']]
|
| 100 |
+
seg_masks, seg_bbox, seg_area = list(zip(*samples))
|
| 101 |
|
| 102 |
+
out_list = []
|
| 103 |
+
for i, seg_mask in enumerate(seg_masks):
|
| 104 |
+
if args['enable_morphologyex']:
|
| 105 |
+
seg_mask = 255 * seg_mask.astype(np.uint8)
|
| 106 |
+
seg_mask = np.stack([seg_mask, seg_mask, seg_mask], axis=-1)
|
| 107 |
+
seg_mask = cv2.morphologyEx(seg_mask, cv2.MORPH_OPEN, kernel=np.ones((6, 6), np.uint8))
|
| 108 |
+
seg_mask = cv2.morphologyEx(seg_mask, cv2.MORPH_CLOSE, kernel=np.ones((6, 6), np.uint8))
|
| 109 |
+
seg_mask = seg_mask[:, :, 0] > 0
|
| 110 |
|
| 111 |
+
seg_mask_img = Image.fromarray(seg_mask.astype('int') * 255.)
|
| 112 |
+
mask_save_path = None
|
| 113 |
+
|
| 114 |
+
if verbose:
|
| 115 |
+
mask_save_path = f'result/mask_{time.time()}.png'
|
| 116 |
+
if not os.path.exists(os.path.dirname(mask_save_path)):
|
| 117 |
+
os.makedirs(os.path.dirname(mask_save_path))
|
| 118 |
+
|
| 119 |
+
if seg_mask_img.mode != 'RGB':
|
| 120 |
+
seg_mask_img = seg_mask_img.convert('RGB')
|
| 121 |
+
seg_mask_img.save(mask_save_path)
|
| 122 |
+
print('seg_mask path: ', mask_save_path)
|
| 123 |
+
print("seg_mask.shape: ", seg_mask.shape)
|
| 124 |
+
|
| 125 |
+
|
| 126 |
+
# captioning with mask
|
| 127 |
+
if args['enable_reduce_tokens']:
|
| 128 |
+
result = self.captioner.inference_with_reduced_tokens(image, seg_mask,
|
| 129 |
+
crop_mode=args['seg_crop_mode'],
|
| 130 |
+
filter=args['clip_filter'],
|
| 131 |
+
disable_regular_box=args['disable_regular_box'],
|
| 132 |
+
verbose=verbose,
|
| 133 |
+
caption_args=args)
|
| 134 |
+
else:
|
| 135 |
+
result = self.captioner.inference_seg(image, seg_mask,
|
| 136 |
+
crop_mode=args['seg_crop_mode'],
|
| 137 |
+
filter=args['clip_filter'],
|
| 138 |
+
disable_regular_box=args['disable_regular_box'],
|
| 139 |
+
verbose=verbose,
|
| 140 |
+
caption_args=args)
|
| 141 |
+
caption = result.get('caption', None)
|
| 142 |
+
crop_save_path = result.get('crop_save_path', None)
|
| 143 |
+
|
| 144 |
+
# refining with TextRefiner
|
| 145 |
+
context_captions = []
|
| 146 |
+
if args['context_captions']:
|
| 147 |
+
context_captions.append(self.captioner.inference(image)['caption'])
|
| 148 |
+
if not disable_gpt and self.text_refiner is not None:
|
| 149 |
+
refined_caption = self.text_refiner.inference(query=caption, controls=controls, context=context_captions,
|
| 150 |
+
enable_wiki=enable_wiki)
|
| 151 |
+
else:
|
| 152 |
+
refined_caption = {'raw_caption': caption}
|
| 153 |
+
out = {'generated_captions': refined_caption,
|
| 154 |
+
'crop_save_path': crop_save_path,
|
| 155 |
+
'mask_save_path': mask_save_path,
|
| 156 |
+
'mask': seg_mask_img,
|
| 157 |
+
'bbox': seg_bbox[i] if seg_bbox is not None else None,
|
| 158 |
+
'area': seg_area[i] if seg_area is not None else None,
|
| 159 |
+
'context_captions': context_captions,
|
| 160 |
+
'ppl_score': result.get('ppl_score', -100.),
|
| 161 |
+
'clip_score': result.get('clip_score', 0.)
|
| 162 |
+
}
|
| 163 |
+
out_list.append(out)
|
| 164 |
+
return out_list
|
| 165 |
+
|
| 166 |
+
def parse_dense_caption(self, image, topN=10, reference_caption=[], verbose=False):
|
| 167 |
+
width, height = get_image_shape(image)
|
| 168 |
+
prompt = {'prompt_type': ['everything']}
|
| 169 |
+
densecap_args = {
|
| 170 |
+
'return_ppl': True,
|
| 171 |
+
'clip_filter': True,
|
| 172 |
+
'reference_caption': reference_caption,
|
| 173 |
+
'text_prompt': "", # 'Question: what does the image show? Answer:'
|
| 174 |
+
'seg_crop_mode': 'w_bg',
|
| 175 |
+
# 'text_prompt': "",
|
| 176 |
+
# 'seg_crop_mode': 'wo_bg',
|
| 177 |
+
'disable_regular_box': False,
|
| 178 |
+
'topN': topN,
|
| 179 |
+
'min_ppl_score': -1.8,
|
| 180 |
+
'min_clip_score': 0.30,
|
| 181 |
+
'min_mask_area': 2500,
|
| 182 |
+
}
|
| 183 |
+
|
| 184 |
+
dense_captions = self.inference(image, prompt,
|
| 185 |
+
controls=None,
|
| 186 |
+
disable_gpt=True,
|
| 187 |
+
verbose=verbose,
|
| 188 |
+
is_densecap=True,
|
| 189 |
+
args=densecap_args)
|
| 190 |
+
print('Process Dense Captioning: \n', dense_captions)
|
| 191 |
+
dense_captions = list(filter(lambda x: x['ppl_score'] / (1+len(x['generated_captions']['raw_caption'].split())) >= densecap_args['min_ppl_score'], dense_captions))
|
| 192 |
+
dense_captions = list(filter(lambda x: x['clip_score'] >= densecap_args['min_clip_score'], dense_captions))
|
| 193 |
+
dense_cap_prompt = []
|
| 194 |
+
for cap in dense_captions:
|
| 195 |
+
x, y, w, h = cap['bbox']
|
| 196 |
+
cx, cy = x + w/2, (y + h/2)
|
| 197 |
+
dense_cap_prompt.append("({}: X:{:.0f}, Y:{:.0f}, Width:{:.0f}, Height:{:.0f})".format(cap['generated_captions']['raw_caption'], cx, cy, w, h))
|
| 198 |
+
|
| 199 |
+
if verbose:
|
| 200 |
+
all_masks = [np.array(item['mask'].convert('P')) for item in dense_captions]
|
| 201 |
+
new_image = mask_painter_foreground_all(np.array(image), all_masks, background_alpha=0.4)
|
| 202 |
+
save_path = 'result/dense_caption_mask.png'
|
| 203 |
+
Image.fromarray(new_image).save(save_path)
|
| 204 |
+
print(f'Dense captioning mask saved in {save_path}')
|
| 205 |
+
|
| 206 |
+
vis_path = 'result/dense_caption_vis_{}.png'.format(time.time())
|
| 207 |
+
dense_cap_painter_input = [{'bbox': xywh_to_x1y1x2y2(cap['bbox']),
|
| 208 |
+
'caption': cap['generated_captions']['raw_caption']} for cap in dense_captions]
|
| 209 |
+
draw_bbox(load_image(image, return_type='numpy'), vis_path, dense_cap_painter_input, show_caption=True)
|
| 210 |
+
print(f'Dense Captioning visualization saved in {vis_path}')
|
| 211 |
+
return ','.join(dense_cap_prompt)
|
| 212 |
+
|
| 213 |
+
def parse_ocr(self, image, thres=0.2):
|
| 214 |
+
width, height = get_image_shape(image)
|
| 215 |
+
image = load_image(image, return_type='numpy')
|
| 216 |
+
bounds = self.reader.readtext(image)
|
| 217 |
+
bounds = [bound for bound in bounds if bound[2] > thres]
|
| 218 |
+
print('Process OCR Text:\n', bounds)
|
| 219 |
+
|
| 220 |
+
ocr_prompt = []
|
| 221 |
+
for box, text, conf in bounds:
|
| 222 |
+
p0, p1, p2, p3 = box
|
| 223 |
+
ocr_prompt.append('(\"{}\": X:{:.0f}, Y:{:.0f})'.format(text, (p0[0]+p1[0]+p2[0]+p3[0])/4, (p0[1]+p1[1]+p2[1]+p3[1])/4))
|
| 224 |
+
ocr_prompt = '\n'.join(ocr_prompt)
|
| 225 |
+
|
| 226 |
+
# ocr_prompt = self.text_refiner.llm(f'The image have some scene texts with their locations: {ocr_prompt}. Please group these individual words into one or several phrase based on their relative positions (only give me your answer, do not show explanination)').strip()
|
| 227 |
+
|
| 228 |
+
# ocr_prefix1 = f'The image have some scene texts with their locations: {ocr_prompt}. Please group these individual words into one or several phrase based on their relative positions (only give me your answer, do not show explanination)'
|
| 229 |
+
# ocr_prefix2 = f'Please group these individual words into 1-3 phrases, given scene texts with their locations: {ocr_prompt}. You return is one or several strings and infer their locations. (only give me your answer like (“man working”, X: value, Y: value), do not show explanination)'
|
| 230 |
+
# ocr_prefix4 = f'summarize the individual scene text words detected by OCR tools into a fluent sentence based on their positions and distances. You should strictly describe all of the given scene text words. Do not miss any given word. Do not create non-exist words. Do not appear numeric positions. The individual words are given:\n{ocr_prompt}\n'
|
| 231 |
+
# ocr_prefix3 = f'combine the individual scene text words detected by OCR tools into one/several fluent phrases/sentences based on their positions and distances. You should strictly copy or correct all of the given scene text words. Do not miss any given word. Do not create non-exist words. The response is several strings seperate with their location (X, Y), each of which represents a phrase. The individual words are given:\n{ocr_prompt}\n'
|
| 232 |
+
# response = self.text_refiner.llm(ocr_prefix3).strip() if len(ocr_prompt) else ""
|
| 233 |
+
return ocr_prompt
|
| 234 |
+
|
| 235 |
+
def inference_cap_everything(self, image, verbose=False):
|
| 236 |
+
image = load_image(image, return_type='pil')
|
| 237 |
+
image = image_resize(image, res=1024)
|
| 238 |
+
width, height = get_image_shape(image)
|
| 239 |
+
other_args = {'text_prompt': ""} if self.require_caption_prompt else {}
|
| 240 |
+
img_caption = self.captioner.inference(image, filter=False, args=other_args)['caption']
|
| 241 |
+
dense_caption_prompt = self.parse_dense_caption(image, topN=10, verbose=verbose, reference_caption=[])
|
| 242 |
+
scene_text_prompt = self.parse_ocr(image, thres=0.2)
|
| 243 |
+
# scene_text_prompt = "N/A"
|
| 244 |
+
|
| 245 |
+
# the summarize_prompt is modified from https://github.com/JialianW/GRiT and https://github.com/showlab/Image2Paragraph
|
| 246 |
+
summarize_prompt = "Imagine you are a blind but intelligent image captioner. You should generate a descriptive, coherent and human-like paragraph based on the given information (a,b,c,d) instead of imagination:\na) Image Resolution: {image_size}\nb) Image Caption:{image_caption}\nc) Dense Caption: {dense_caption}\nd) Scene Text: {scene_text}\nThere are some rules for your response: Show objects with their attributes (e.g. position, color, size, shape, texture).\nPrimarily describe common objects with large size.\nProvide context of the image.\nShow relative position between objects.\nLess than 6 sentences.\nDo not appear number.\nDo not describe any individual letter.\nDo not show the image resolution.\nIngore the white background."
|
| 247 |
+
prompt = summarize_prompt.format(**{
|
| 248 |
+
"image_size": "width {} height {}".format(width, height),
|
| 249 |
+
"image_caption":img_caption,
|
| 250 |
+
"dense_caption": dense_caption_prompt,
|
| 251 |
+
"scene_text": scene_text_prompt})
|
| 252 |
+
print(f'caption everything prompt: {prompt}')
|
| 253 |
+
response = self.text_refiner.llm(prompt).strip()
|
| 254 |
+
# chinese_response = self.text_refiner.llm('Translate it into Chinese: {}'.format(response)).strip()
|
| 255 |
+
return response
|
| 256 |
+
|
| 257 |
if __name__ == "__main__":
|
| 258 |
from caption_anything.utils.parser import parse_augment
|
|
|
|
| 259 |
args = parse_augment()
|
| 260 |
+
image_path = 'image/ocr/Untitled.png'
|
| 261 |
+
image = Image.open(image_path)
|
| 262 |
prompts = [
|
| 263 |
{
|
| 264 |
"prompt_type": ["click"],
|
|
|
|
| 266 |
"input_label": [1, 0],
|
| 267 |
"multimask_output": "True",
|
| 268 |
},
|
| 269 |
+
# {
|
| 270 |
+
# "prompt_type": ["click"],
|
| 271 |
+
# "input_point": [[300, 800]],
|
| 272 |
+
# "input_label": [1],
|
| 273 |
+
# "multimask_output": "True",
|
| 274 |
+
# }
|
| 275 |
]
|
| 276 |
controls = {
|
| 277 |
"length": "30",
|
|
|
|
| 282 |
}
|
| 283 |
|
| 284 |
model = CaptionAnything(args, os.environ['OPENAI_API_KEY'])
|
| 285 |
+
img_dir = 'test_images/memes'
|
| 286 |
+
for image_file in os.listdir(img_dir):
|
| 287 |
+
image_path = os.path.join(img_dir, image_file)
|
| 288 |
+
print('image_path:', image_path)
|
| 289 |
+
paragraph = model.inference_cap_everything(image_path, verbose=True)
|
| 290 |
+
print('Caption Everything:\n', paragraph)
|
| 291 |
+
ocr = model.parse_ocr(image_path)
|
| 292 |
+
print('OCR', ocr)
|
caption_anything/segmenter/__init__.py
CHANGED
|
@@ -1,5 +1,14 @@
|
|
| 1 |
from .base_segmenter import BaseSegmenter
|
| 2 |
from caption_anything.utils.utils import seg_model_map
|
|
|
|
| 3 |
|
| 4 |
-
def build_segmenter(model_name, device, args
|
| 5 |
-
return BaseSegmenter(device, args.segmenter_checkpoint, model_name, reuse_feature=not args.disable_reuse_features, model=model)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
from .base_segmenter import BaseSegmenter
|
| 2 |
from caption_anything.utils.utils import seg_model_map
|
| 3 |
+
import copy
|
| 4 |
|
| 5 |
+
def build_segmenter(model_name, device, args, model=None):
|
| 6 |
+
return BaseSegmenter(device, args.segmenter_checkpoint, model_name, reuse_feature=not args.disable_reuse_features, model=model, args=args)
|
| 7 |
+
|
| 8 |
+
def build_segmenter_densecap(model_name, device, args, model=None):
|
| 9 |
+
args_for_densecap = copy.deepcopy(args)
|
| 10 |
+
args_for_densecap.pred_iou_thresh = 0.88
|
| 11 |
+
args_for_densecap.min_mask_region_area = 400
|
| 12 |
+
args_for_densecap.stability_score_thresh = 0.95
|
| 13 |
+
args_for_densecap.box_nms_thresh = 0.3
|
| 14 |
+
return BaseSegmenter(device, args.segmenter_checkpoint, model_name, reuse_feature=not args.disable_reuse_features, model=model, args=args)
|
caption_anything/segmenter/base_segmenter.py
CHANGED
|
@@ -11,7 +11,7 @@ import PIL
|
|
| 11 |
|
| 12 |
|
| 13 |
class BaseSegmenter:
|
| 14 |
-
def __init__(self, device, checkpoint, model_name='huge', reuse_feature=True, model=None):
|
| 15 |
print(f"Initializing BaseSegmenter to {device}")
|
| 16 |
self.device = device
|
| 17 |
self.torch_dtype = torch.float16 if 'cuda' in device else torch.float32
|
|
@@ -26,7 +26,10 @@ class BaseSegmenter:
|
|
| 26 |
self.model = model
|
| 27 |
self.reuse_feature = reuse_feature
|
| 28 |
self.predictor = SamPredictor(self.model)
|
| 29 |
-
|
|
|
|
|
|
|
|
|
|
| 30 |
self.image_embedding = None
|
| 31 |
self.image = None
|
| 32 |
|
|
@@ -69,7 +72,9 @@ class BaseSegmenter:
|
|
| 69 |
if 'everything' in control['prompt_type']:
|
| 70 |
masks = self.mask_generator.generate(image)
|
| 71 |
new_masks = np.concatenate([mask["segmentation"][np.newaxis, :] for mask in masks])
|
| 72 |
-
|
|
|
|
|
|
|
| 73 |
else:
|
| 74 |
if not self.reuse_feature or self.image_embedding is None:
|
| 75 |
self.set_image(image)
|
|
|
|
| 11 |
|
| 12 |
|
| 13 |
class BaseSegmenter:
|
| 14 |
+
def __init__(self, device, checkpoint, model_name='huge', reuse_feature=True, model=None, args=None):
|
| 15 |
print(f"Initializing BaseSegmenter to {device}")
|
| 16 |
self.device = device
|
| 17 |
self.torch_dtype = torch.float16 if 'cuda' in device else torch.float32
|
|
|
|
| 26 |
self.model = model
|
| 27 |
self.reuse_feature = reuse_feature
|
| 28 |
self.predictor = SamPredictor(self.model)
|
| 29 |
+
|
| 30 |
+
sam_generator_keys = ['pred_iou_thresh', 'min_mask_region_area', 'stability_score_thresh', 'box_nms_thresh']
|
| 31 |
+
generator_args = {k:v for k,v in vars(args).items() if k in sam_generator_keys}
|
| 32 |
+
self.mask_generator = SamAutomaticMaskGenerator(model=self.model, **generator_args)
|
| 33 |
self.image_embedding = None
|
| 34 |
self.image = None
|
| 35 |
|
|
|
|
| 72 |
if 'everything' in control['prompt_type']:
|
| 73 |
masks = self.mask_generator.generate(image)
|
| 74 |
new_masks = np.concatenate([mask["segmentation"][np.newaxis, :] for mask in masks])
|
| 75 |
+
bbox = np.array([mask["bbox"] for mask in masks])
|
| 76 |
+
area = np.array([mask["area"] for mask in masks])
|
| 77 |
+
return new_masks, bbox, area
|
| 78 |
else:
|
| 79 |
if not self.reuse_feature or self.image_embedding is None:
|
| 80 |
self.set_image(image)
|
caption_anything/utils/densecap_painter.py
ADDED
|
@@ -0,0 +1,64 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
import cv2
|
| 2 |
+
import json
|
| 3 |
+
import numpy as np
|
| 4 |
+
from typing import List
|
| 5 |
+
import random
|
| 6 |
+
from typing import Union
|
| 7 |
+
|
| 8 |
+
def draw_bbox(img: Union[np.ndarray, str], save_name: str, bbox: List[dict], show_caption: bool = False):
|
| 9 |
+
"""
|
| 10 |
+
bbox: [{'image_id': str, 'bbox': [x1, y1, x2, y2], 'caption': str}, ...]
|
| 11 |
+
"""
|
| 12 |
+
if isinstance(img, str):
|
| 13 |
+
img = cv2.imread(img)
|
| 14 |
+
|
| 15 |
+
RGB = [0, 50, 100, 150, 200, 250]
|
| 16 |
+
for box in bbox:
|
| 17 |
+
box['bbox'] = [int(_) for _ in box['bbox']]
|
| 18 |
+
x1, y1, x2, y2 = box['bbox']
|
| 19 |
+
caption = box['caption']
|
| 20 |
+
box_color = random.choices(RGB, k = 3)
|
| 21 |
+
(text_width, text_height), _ = cv2.getTextSize(caption, cv2.FONT_HERSHEY_SIMPLEX, fontScale = 0.5, thickness = 2)
|
| 22 |
+
cv2.rectangle(img, (x1, y1), (x2, y2), color = box_color, thickness = 2)
|
| 23 |
+
if show_caption:
|
| 24 |
+
cv2.putText(img, caption, (x1, y1 + text_height), cv2.FONT_HERSHEY_SIMPLEX, fontScale = 0.5, color = box_color, thickness = 2)
|
| 25 |
+
|
| 26 |
+
cv2.imwrite(save_name, img)
|
| 27 |
+
# cv2.imshow('visualise', img)
|
| 28 |
+
# cv2.waitKey(0)
|
| 29 |
+
|
| 30 |
+
def parse_bbox(anno, image_id: int = None):
|
| 31 |
+
|
| 32 |
+
with open(anno, 'r') as f:
|
| 33 |
+
predictions = json.load(f)
|
| 34 |
+
|
| 35 |
+
if image_id is None:
|
| 36 |
+
image_id = next(iter(predictions))
|
| 37 |
+
|
| 38 |
+
return predictions[image_id]
|
| 39 |
+
|
| 40 |
+
def gt_bbox(anno, img_name: int = None):
|
| 41 |
+
|
| 42 |
+
with open(anno, 'r') as f:
|
| 43 |
+
annotations = json.load(f)
|
| 44 |
+
annotations = annotations['annotations']
|
| 45 |
+
|
| 46 |
+
gt = []
|
| 47 |
+
img_name = int(img_name[:-4])
|
| 48 |
+
for annotation in annotations:
|
| 49 |
+
if annotation['image_id'] == 63:
|
| 50 |
+
x1, y1, w, h = annotation['bbox']
|
| 51 |
+
gt.append({'bbox': [x1, y1, x1 + w, y1 + h], 'caption': annotation['caption']})
|
| 52 |
+
return gt
|
| 53 |
+
|
| 54 |
+
if __name__ == '__main__':
|
| 55 |
+
|
| 56 |
+
img_name = '63.jpg'
|
| 57 |
+
show_caption = True
|
| 58 |
+
anno = 'vg_dense_captioning_blip2_top48_0.88_1000_0.96_debugTrue_predictions_shard_all.json'
|
| 59 |
+
|
| 60 |
+
img = cv2.imread(img_name)
|
| 61 |
+
examp_bbox = parse_bbox(anno)
|
| 62 |
+
ground_truth_bbox = gt_bbox('test.json', img_name)
|
| 63 |
+
draw_bbox(img, 'GT.jpg', ground_truth_bbox, show_caption)
|
| 64 |
+
draw_bbox(img, 'Pred.jpg', examp_bbox, show_caption)
|
caption_anything/utils/parser.py
CHANGED
|
@@ -22,6 +22,12 @@ def parse_augment():
|
|
| 22 |
parser.add_argument('--disable_reuse_features', action="store_true", default=False)
|
| 23 |
parser.add_argument('--enable_morphologyex', action="store_true", default=False)
|
| 24 |
parser.add_argument('--chat_tools_dict', type=str, default='VisualQuestionAnswering_cuda:0', help='Visual ChatGPT tools, only useful when running gradio applications')
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 25 |
args = parser.parse_args()
|
| 26 |
|
| 27 |
if args.debug:
|
|
|
|
| 22 |
parser.add_argument('--disable_reuse_features', action="store_true", default=False)
|
| 23 |
parser.add_argument('--enable_morphologyex', action="store_true", default=False)
|
| 24 |
parser.add_argument('--chat_tools_dict', type=str, default='VisualQuestionAnswering_cuda:0', help='Visual ChatGPT tools, only useful when running gradio applications')
|
| 25 |
+
|
| 26 |
+
parser.add_argument('--pred_iou_thresh', type=float, default=0.88, help="sam post-precessing")
|
| 27 |
+
parser.add_argument('--min_mask_region_area', type=int, default=0, help="sam post-precessing")
|
| 28 |
+
parser.add_argument('--stability_score_thresh', type=float, default=0.95, help='sam post-processing')
|
| 29 |
+
parser.add_argument('--box_nms_thresh', type=float, default=0.7, help='sam post-processing')
|
| 30 |
+
|
| 31 |
args = parser.parse_args()
|
| 32 |
|
| 33 |
if args.debug:
|
caption_anything/utils/utils.py
CHANGED
|
@@ -29,6 +29,9 @@ def load_image(image: Union[np.ndarray, Image.Image, str], return_type='numpy'):
|
|
| 29 |
elif isinstance(image, np.ndarray):
|
| 30 |
image = Image.fromarray(image)
|
| 31 |
|
|
|
|
|
|
|
|
|
|
| 32 |
if return_type == 'pil':
|
| 33 |
return image
|
| 34 |
elif return_type == 'numpy':
|
|
@@ -37,6 +40,34 @@ def load_image(image: Union[np.ndarray, Image.Image, str], return_type='numpy'):
|
|
| 37 |
raise NotImplementedError()
|
| 38 |
|
| 39 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 40 |
def is_platform_win():
|
| 41 |
return sys.platform == "win32"
|
| 42 |
|
|
|
|
| 29 |
elif isinstance(image, np.ndarray):
|
| 30 |
image = Image.fromarray(image)
|
| 31 |
|
| 32 |
+
if image.mode == "RGBA":
|
| 33 |
+
image = image.convert("RGB")
|
| 34 |
+
|
| 35 |
if return_type == 'pil':
|
| 36 |
return image
|
| 37 |
elif return_type == 'numpy':
|
|
|
|
| 40 |
raise NotImplementedError()
|
| 41 |
|
| 42 |
|
| 43 |
+
def image_resize(image: Image.Image, res=1024):
|
| 44 |
+
width, height = org_size = image.size
|
| 45 |
+
ratio = min(1.0 * res / max(width, height), 1.0)
|
| 46 |
+
if ratio < 1.0:
|
| 47 |
+
image = image.resize((int(width * ratio), int(height * ratio)))
|
| 48 |
+
print('Scaling image from {} to {}'.format(org_size, image.size))
|
| 49 |
+
return image
|
| 50 |
+
|
| 51 |
+
def xywh_to_x1y1x2y2(bbox):
|
| 52 |
+
x, y, w, h = bbox
|
| 53 |
+
return x,y,x+w,y+h
|
| 54 |
+
|
| 55 |
+
|
| 56 |
+
def x1y1x2y2_to_xywh(bbox):
|
| 57 |
+
x1, y1, x2, y2 = bbox
|
| 58 |
+
return x1,y1,x2-x1,y2-y1
|
| 59 |
+
|
| 60 |
+
|
| 61 |
+
def get_image_shape(image):
|
| 62 |
+
if isinstance(image, str):
|
| 63 |
+
return Image.open(image).size
|
| 64 |
+
elif isinstance(image, np.ndarray):
|
| 65 |
+
return image.shape
|
| 66 |
+
elif isinstance(image, Image.Image):
|
| 67 |
+
return image.size
|
| 68 |
+
else:
|
| 69 |
+
raise NotImplementedError
|
| 70 |
+
|
| 71 |
def is_platform_win():
|
| 72 |
return sys.platform == "win32"
|
| 73 |
|
requirements.txt
CHANGED
|
@@ -17,4 +17,7 @@ onnxruntime
|
|
| 17 |
onnx
|
| 18 |
https://gradio-builds.s3.amazonaws.com/3e68e5e882a6790ac5b457bd33f4edf9b695af90/gradio-3.24.1-py3-none-any.whl
|
| 19 |
accelerate
|
| 20 |
-
bitsandbytes
|
|
|
|
|
|
|
|
|
|
|
|
| 17 |
onnx
|
| 18 |
https://gradio-builds.s3.amazonaws.com/3e68e5e882a6790ac5b457bd33f4edf9b695af90/gradio-3.24.1-py3-none-any.whl
|
| 19 |
accelerate
|
| 20 |
+
bitsandbytes
|
| 21 |
+
packaging~=23.1
|
| 22 |
+
easyocr
|
| 23 |
+
tensorboardX
|