Upload 2 files

app.py

@@ -5,6 +5,7 @@ import requests
 from caption_anything import CaptionAnything
 import torch
 import json
+from diffusers import StableDiffusionInpaintPipeline
 import sys
 import argparse
 from caption_anything import parse_augment
@@ -15,10 +16,7 @@ import copy
 from tools import mask_painter
 from PIL import Image
 import os
-from captioner import build_captioner
-from segment_anything import sam_model_registry
-from text_refiner import build_text_refiner
-from segmenter import build_segmenter
+import cv2
 
 def download_checkpoint(url, folder, filename):
     os.makedirs(folder, exist_ok=True)
@@ -39,8 +37,8 @@ filename = "sam_vit_h_4b8939.pth"
 download_checkpoint(checkpoint_url, folder, filename)
 
 
-title = """<h1 align="center">Caption-Anything</h1>"""
-description = """Gradio demo for Caption Anything, image to dense captioning generation with various language styles. To use it, simply upload your image, or click one of the examples to load them. Code: https://github.com/ttengwang/Caption-Anything
+title = """<h1 align="center">Edit Anything</h1>"""
+description = """Gradio demo for Segment Anything, image to dense Segment generation with various language styles. To use it, simply upload your image, or click one of the examples to load them. 
 """
 
 examples = [
@@ -55,108 +53,62 @@ examples = [
 
 args = parse_augment()
 # args.device = 'cuda:5'
-# args.disable_gpt = True
-# args.enable_reduce_tokens = False
+# args.disable_gpt = False
+# args.enable_reduce_tokens = True
 # args.port=20322
-# args.captioner = 'blip'
-# args.regular_box = True
-shared_captioner = build_captioner(args.captioner, args.device, args)
-shared_sam_model = sam_model_registry['vit_h'](checkpoint=args.segmenter_checkpoint).to(args.device)
+model = CaptionAnything(args)
 
+def init_openai_api_key(api_key):
+    # os.environ['OPENAI_API_KEY'] = api_key
+    model.init_refiner(api_key)
+    openai_available = model.text_refiner is not None
+    return gr.update(visible = openai_available), gr.update(visible = openai_available), gr.update(visible = openai_available), gr.update(visible = True), gr.update(visible = True)
 
-def build_caption_anything_with_models(args, api_key="", captioner=None, sam_model=None, text_refiner=None, session_id=None):
-    segmenter = build_segmenter(args.segmenter, args.device, args, model=sam_model)
-    captioner = captioner
-    if session_id is not None:
-        print('Init caption anything for session {}'.format(session_id))
-    return CaptionAnything(args, api_key, captioner=captioner, segmenter=segmenter, text_refiner=text_refiner)
-
-
-def init_openai_api_key(api_key=""):
-    text_refiner = None
-    if api_key and len(api_key) > 30:
-        try:
-            text_refiner = build_text_refiner(args.text_refiner, args.device, args, api_key)
-            text_refiner.llm('hi') # test
-        except:
-            text_refiner = None
-    openai_available = text_refiner is not None
-    return gr.update(visible = openai_available), gr.update(visible = openai_available), gr.update(visible = openai_available), gr.update(visible = True), gr.update(visible = True), gr.update(visible = True), text_refiner
-
-
-def get_prompt(chat_input, click_state, click_mode):
+def get_prompt(chat_input, click_state):    
+    points = click_state[0]
+    labels = click_state[1]
     inputs = json.loads(chat_input)
-    if click_mode == 'Continuous':
-        points = click_state[0]
-        labels = click_state[1]
-        for input in inputs:
-            points.append(input[:2])
-            labels.append(input[2])
-    elif click_mode == 'Single':
-        points = []
-        labels = []
-        for input in inputs:
-            points.append(input[:2])
-            labels.append(input[2])
-        click_state[0] = points
-        click_state[1] = labels
-    else:
-        raise NotImplementedError
+    for input in inputs:
+        points.append(input[:2])
+        labels.append(input[2])
     
     prompt = {
         "prompt_type":["click"],
-        "input_point":click_state[0],
-        "input_label":click_state[1],
+        "input_point":points,
+        "input_label":labels,
         "multimask_output":"True",
     }
     return prompt
 
-def update_click_state(click_state, caption, click_mode):
-    if click_mode == 'Continuous':
-        click_state[2].append(caption)
-    elif click_mode == 'Single':
-        click_state[2] = [caption]
-    else:
-        raise NotImplementedError     
-
-
-def chat_with_points(chat_input, click_state, state, text_refiner):
-    if text_refiner is None:
-        response = "Text refiner is not initilzed, please input openai api key."
-        state = state + [(chat_input, response)]
-        return state, state
+def chat_with_points(chat_input, click_state, state, mask_save_path,image_input):
     
     points, labels, captions = click_state
-    # point_chat_prompt = "I want you act as a chat bot in terms of image. I will give you some points (w, h) in the image and tell you what happed on the point in natural language. Note that (0, 0) refers to the top-left corner of the image, w refers to the width and h refers the height. You should chat with me based on the fact in the image instead of imagination. Now I tell you the points with their visual description:\n{points_with_caps}\nNow begin chatting! Human: {chat_input}\nAI: "
-    # # "The image is of width {width} and height {height}." 
-    point_chat_prompt = "a) Revised prompt: I am an AI trained to chat with you about an image based on specific points (w, h) you provide, along with their visual descriptions. Please note that (0, 0) refers to the top-left corner of the image, w refers to the width, and h refers to the height. Here are the points and their descriptions you've given me: {points_with_caps}. Now, let's chat! Human: {chat_input} AI:"
-    prev_visual_context = ""
-    pos_points = [f"{points[i][0]}, {points[i][1]}" for i in range(len(points)) if labels[i] == 1]
-    if len(captions):
-        prev_visual_context = ', '.join(pos_points) + captions[-1] + '\n'
-    else:
-        prev_visual_context = 'no point exists.'
-    chat_prompt = point_chat_prompt.format(**{"points_with_caps": prev_visual_context, "chat_input": chat_input})
-    response = text_refiner.llm(chat_prompt)
-    state = state + [(chat_input, response)]
-    return state, state
-
-def inference_seg_cap(image_input, point_prompt, click_mode, language, sentiment, factuality, 
-                length, image_embedding, state, click_state, original_size, input_size, text_refiner, evt:gr.SelectData):
     
-    model = build_caption_anything_with_models(
-        args,    
-        api_key="",
-        captioner=shared_captioner,
-        sam_model=shared_sam_model,
-        text_refiner=text_refiner,
-        session_id=iface.app_id
+    
+    # inpainting
+    pipe = StableDiffusionInpaintPipeline.from_pretrained(
+        "stabilityai/stable-diffusion-2-inpainting",
+        torch_dtype=torch.float32,
     )
+
+
+    pipe = pipe.to("cuda")
+    mask = cv2.imread(mask_save_path)
     
-    model.segmenter.image_embedding = image_embedding
-    model.segmenter.predictor.original_size = original_size
-    model.segmenter.predictor.input_size = input_size
-    model.segmenter.predictor.is_image_set = True
+    image_input = np.array(image_input)
+    h,w = image_input.shape[:2]
+    
+    image = cv2.resize(image_input,(512,512))
+    mask = cv2.resize(mask,(512,512)).astype(np.uint8)[:,:,0]
+    print(image.shape,mask.shape)
+    print("chat_input:",chat_input)
+    image = pipe(prompt=chat_input, image=image, mask_image=mask).images[0]
+    image = image.resize((w,h))
+    
+#     image = Image.fromarray(image, mode='RGB')
+    return state, state, image
+
+def inference_seg_cap(image_input, point_prompt, language, sentiment, factuality, length, state, click_state, evt:gr.SelectData):
 
     if point_prompt == 'Positive':
         coordinate = "[[{}, {}, 1]]".format(str(evt.index[0]), str(evt.index[1]))
@@ -170,32 +122,33 @@ def inference_seg_cap(image_input, point_prompt, click_mode, language, sentiment
 
     # click_coordinate = "[[{}, {}, 1]]".format(str(evt.index[0]), str(evt.index[1])) 
     # chat_input = click_coordinate
-    prompt = get_prompt(coordinate, click_state, click_mode)
+    prompt = get_prompt(coordinate, click_state)
     print('prompt: ', prompt, 'controls: ', controls)
 
-    out = model.inference(image_input, prompt, controls, disable_gpt=True)
+    out = model.inference(image_input, prompt, controls)
     state = state + [(None, "Image point: {}, Input label: {}".format(prompt["input_point"], prompt["input_label"]))]
     # for k, v in out['generated_captions'].items():
     #     state = state + [(f'{k}: {v}', None)]
-    state = state + [("caption: {}".format(out['generated_captions']['raw_caption']), None)]
-    wiki = out['generated_captions'].get('wiki', "")
-
-    update_click_state(click_state, out['generated_captions']['raw_caption'], click_mode)
-    text = out['generated_captions']['raw_caption']
+#     state = state + [("caption: {}".format(out['generated_captions']['raw_caption']), None)]
+#     wiki = out['generated_captions'].get('wiki', "")
+#     click_state[2].append(out['generated_captions']['raw_caption'])
+    
+#     text = out['generated_captions']['raw_caption']
     # draw = ImageDraw.Draw(image_input)
     # draw.text((evt.index[0], evt.index[1]), text, textcolor=(0,0,255), text_size=120)
     input_mask = np.array(out['mask'].convert('P'))
     image_input = mask_painter(np.array(image_input), input_mask)
     origin_image_input = image_input
+    text = "edit"
     image_input = create_bubble_frame(image_input, text, (evt.index[0], evt.index[1]))
 
-    yield state, state, click_state, chat_input, image_input, wiki
-    if not args.disable_gpt and model.text_refiner:
-        refined_caption = model.text_refiner.inference(query=text, controls=controls, context=out['context_captions'])
-        # new_cap = 'Original: ' + text + '. Refined: ' + refined_caption['caption']
-        new_cap = refined_caption['caption']
-        refined_image_input = create_bubble_frame(origin_image_input, new_cap, (evt.index[0], evt.index[1]))
-        yield state, state, click_state, chat_input, refined_image_input, wiki
+    yield state, state, click_state, image_input, out["mask_save_path"]
+#     if not args.disable_gpt and model.text_refiner:
+#         refined_caption = model.text_refiner.inference(query=text, controls=controls, context=out['context_captions'])
+#         # new_cap = 'Original: ' + text + '. Refined: ' + refined_caption['caption']
+#         new_cap = refined_caption['caption']
+#         refined_image_input = create_bubble_frame(origin_image_input, new_cap, (evt.index[0], evt.index[1]))
+#         yield state, state, click_state, chat_input, refined_image_input, wiki
 
 
 def upload_callback(image_input, state):
@@ -207,19 +160,10 @@ def upload_callback(image_input, state):
     if ratio < 1.0:
         image_input = image_input.resize((int(width * ratio), int(height * ratio)))
         print('Scaling input image to {}'.format(image_input.size))
-        
-    model = build_caption_anything_with_models(
-        args,    
-        api_key="",
-        captioner=shared_captioner,
-        sam_model=shared_sam_model,
-        session_id=iface.app_id
-    )
+    model.segmenter.image = None
+    model.segmenter.image_embedding = None
     model.segmenter.set_image(image_input)
-    image_embedding = model.segmenter.image_embedding
-    original_size = model.segmenter.predictor.original_size
-    input_size = model.segmenter.predictor.input_size
-    return state, state, image_input, click_state, image_input, image_embedding, original_size, input_size
+    return state, image_input, click_state, image_input
 
 with gr.Blocks(
     css='''
@@ -230,38 +174,28 @@ with gr.Blocks(
     state = gr.State([])
     click_state = gr.State([[],[],[]])
     origin_image = gr.State(None)
-    image_embedding = gr.State(None)
-    text_refiner = gr.State(None)
-    original_size = gr.State(None)
-    input_size = gr.State(None)
+    mask_save_path = gr.State(None)
 
     gr.Markdown(title)
     gr.Markdown(description)
 
     with gr.Row():
         with gr.Column(scale=1.0):
-            with gr.Column(visible=False) as modules_not_need_gpt:
+            with gr.Column(visible=True) as modules_not_need_gpt:
                 image_input = gr.Image(type="pil", interactive=True, elem_id="image_upload")
                 example_image = gr.Image(type="pil", interactive=False, visible=False)
                 with gr.Row(scale=1.0):
-                    with gr.Row(scale=0.4):
-                        point_prompt = gr.Radio(
-                            choices=["Positive",  "Negative"],
-                            value="Positive",
-                            label="Point Prompt",
-                            interactive=True)
-                        click_mode = gr.Radio(
-                            choices=["Continuous",  "Single"],
-                            value="Continuous",
-                            label="Clicking Mode",
-                            interactive=True)
-                    with gr.Row(scale=0.4):
-                        clear_button_clike = gr.Button(value="Clear Clicks", interactive=True)
-                        clear_button_image = gr.Button(value="Clear Image", interactive=True)
-            with gr.Column(visible=False) as modules_need_gpt:
+                    point_prompt = gr.Radio(
+                        choices=["Positive",  "Negative"],
+                        value="Positive",
+                        label="Point Prompt",
+                        interactive=True)
+                    clear_button_clike = gr.Button(value="Clear Clicks", interactive=True)
+                    clear_button_image = gr.Button(value="Clear Image", interactive=True)
+            with gr.Column(visible=True) as modules_need_gpt:
                 with gr.Row(scale=1.0):
                     language = gr.Dropdown(['English', 'Chinese', 'French', "Spanish", "Arabic", "Portuguese", "Cantonese"], value="English", label="Language", interactive=True)
-
+                
                     sentiment = gr.Radio(
                         choices=["Positive", "Natural", "Negative"],
                         value="Natural",
@@ -282,47 +216,40 @@ with gr.Blocks(
                         step=1,
                         interactive=True,
                         label="Length",
-                    )   
-            with gr.Column(visible=True) as modules_not_need_gpt3:
-                gr.Examples(
-                    examples=examples,
-                    inputs=[example_image],
-                )
+                    )
+        
         with gr.Column(scale=0.5):
-            openai_api_key = gr.Textbox(
-                placeholder="Input openAI API key",
-                show_label=False,
-                label = "OpenAI API Key",
-                lines=1,
-                type="password")
-            with gr.Row(scale=0.5):
-                enable_chatGPT_button = gr.Button(value="Run with ChatGPT", interactive=True, variant='primary')
-                disable_chatGPT_button = gr.Button(value="Run without ChatGPT (Faster)", interactive=True, variant='primary')
-            with gr.Column(visible=False) as modules_need_gpt2:
-                wiki_output = gr.Textbox(lines=5, label="Wiki", max_lines=5)
-            with gr.Column(visible=False) as modules_not_need_gpt2:
-                chatbot = gr.Chatbot(label="Chat about Selected Object",).style(height=550,scale=0.5)
-                with gr.Column(visible=False) as modules_need_gpt3:
-                    chat_input = gr.Textbox(lines=1, label="Chat Input")
+#             openai_api_key = gr.Textbox(
+#                 placeholder="Input openAI API key and press Enter (Input blank will disable GPT)",
+#                 show_label=False,
+#                 label = "OpenAI API Key",
+#                 lines=1,
+#                 type="password"
+#                 )
+#             with gr.Column(visible=True) as modules_need_gpt2:
+#                 wiki_output = gr.Textbox(lines=6, label="Wiki")
+            with gr.Column(visible=True) as modules_not_need_gpt2:
+                chatbot = gr.Chatbot(label="History",).style(height=450,scale=0.5)
+                with gr.Column(visible=True) as modules_need_gpt3:
+                    chat_input = gr.Textbox(lines=1, label="Edit Prompt")
                     with gr.Row():
                         clear_button_text = gr.Button(value="Clear Text", interactive=True)
                         submit_button_text = gr.Button(value="Submit", interactive=True, variant="primary")
-    
-    openai_api_key.submit(init_openai_api_key, inputs=[openai_api_key], outputs=[modules_need_gpt,modules_need_gpt2, modules_need_gpt3, modules_not_need_gpt, modules_not_need_gpt2, modules_not_need_gpt3, text_refiner])
-    enable_chatGPT_button.click(init_openai_api_key, inputs=[openai_api_key], outputs=[modules_need_gpt,modules_need_gpt2, modules_need_gpt3, modules_not_need_gpt, modules_not_need_gpt2, modules_not_need_gpt3, text_refiner])
-    disable_chatGPT_button.click(init_openai_api_key, outputs=[modules_need_gpt,modules_need_gpt2, modules_need_gpt3, modules_not_need_gpt, modules_not_need_gpt2, modules_not_need_gpt3, text_refiner])
+                    
+#     openai_api_key.submit(init_openai_api_key, inputs=[openai_api_key], outputs=[modules_need_gpt,modules_need_gpt2, modules_need_gpt3, modules_not_need_gpt, modules_not_need_gpt2])
     
     clear_button_clike.click(
         lambda x: ([[], [], []], x, ""),
         [origin_image],
-        [click_state, image_input, wiki_output],
+        [click_state, image_input],
         queue=False,
         show_progress=False
     )
+    
     clear_button_image.click(
         lambda: (None, [], [], [[], [], []], "", ""),
         [],
-        [image_input, chatbot, state, click_state, wiki_output, origin_image],
+        [image_input, chatbot, state, click_state, origin_image],
         queue=False,
         show_progress=False
     )
@@ -333,37 +260,50 @@ with gr.Blocks(
         queue=False,
         show_progress=False
     )
+    
+    
     image_input.clear(
         lambda: (None, [], [], [[], [], []], "", ""),
         [],
-        [image_input, chatbot, state, click_state, wiki_output, origin_image],
+        [image_input, chatbot, state, click_state, origin_image],
         queue=False,
         show_progress=False
     )
 
-    image_input.upload(upload_callback,[image_input, state], [chatbot, state, origin_image, click_state, image_input, image_embedding, original_size, input_size])
-    chat_input.submit(chat_with_points, [chat_input, click_state, state, text_refiner], [chatbot, state])
-    example_image.change(upload_callback,[example_image, state], [state, state, origin_image, click_state, image_input, image_embedding, original_size, input_size])
+    def example_callback(x):
+        model.image_embedding = None
+        return x
+        
+    gr.Examples(
+        examples=examples,
+        inputs=[example_image],
+    )
+    
+    submit_button_text.click(
+        chat_with_points,
+        [chat_input, click_state, state, mask_save_path,image_input],
+        [chatbot, state, image_input]
+    )
+    
+    
+    image_input.upload(upload_callback,[image_input, state], [state, origin_image, click_state, image_input])
+    chat_input.submit(chat_with_points, [chat_input, click_state, state, mask_save_path,image_input], [chatbot, state, image_input])
+    example_image.change(upload_callback,[example_image, state], [state, origin_image, click_state, image_input])
 
     # select coordinate
     image_input.select(inference_seg_cap, 
         inputs=[
         origin_image,
         point_prompt,
-        click_mode,
         language,
         sentiment,
         factuality,
         length,
-        image_embedding,
         state,
-        click_state,
-        original_size, 
-        input_size,
-        text_refiner
+        click_state
         ],
-        outputs=[chatbot, state, click_state, chat_input, image_input, wiki_output],
+        outputs=[chatbot, state, click_state, image_input, mask_save_path],
         show_progress=False, queue=True)
     
-iface.queue(concurrency_count=5, api_open=False, max_size=10)
-iface.launch(server_name="0.0.0.0", enable_queue=True)
+iface.queue(concurrency_count=1, api_open=False, max_size=10)
+iface.launch(server_name="0.0.0.0", enable_queue=True, server_port=args.port, share=True)  

caption_anything.py

@@ -1,45 +1,26 @@
-from captioner import build_captioner, BaseCaptioner
+
 from segmenter import build_segmenter
-from text_refiner import build_text_refiner
 import os
 import argparse
 import pdb
 import time
 from PIL import Image
-import cv2
-import numpy as np
+
+
 
 class CaptionAnything():
-    def __init__(self, args, api_key="", captioner=None, segmenter=None, text_refiner=None):
+    def __init__(self, args, api_key=""):
         self.args = args
-        self.captioner = build_captioner(args.captioner, args.device, args) if captioner is None else captioner
-        self.segmenter = build_segmenter(args.segmenter, args.device, args) if segmenter is None else segmenter
-        
+#         self.captioner = build_captioner(args.captioner, args.device, args)
+        self.segmenter = build_segmenter(args.segmenter, args.device, args)
         self.text_refiner = None
-        if not args.disable_gpt:
-            if text_refiner is not None:
-                self.text_refiner = text_refiner
-            else:
-                self.init_refiner(api_key)
-
-    def init_refiner(self, api_key):
-        try:
-            self.text_refiner = build_text_refiner(self.args.text_refiner, self.args.device, self.args, api_key)
-            self.text_refiner.llm('hi') # test
-        except:
-            self.text_refiner = None
-            print('OpenAI GPT is not available')
+        
     
     def inference(self, image, prompt, controls, disable_gpt=False):
         #  segment with prompt
         print("CA prompt: ", prompt, "CA controls",controls)
+        print(image)
         seg_mask = self.segmenter.inference(image, prompt)[0, ...]
-        if self.args.enable_morphologyex:
-            seg_mask = 255 * seg_mask.astype(np.uint8)
-            seg_mask = np.stack([seg_mask, seg_mask, seg_mask], axis = -1)
-            seg_mask = cv2.morphologyEx(seg_mask, cv2.MORPH_OPEN, kernel = np.ones((6, 6), np.uint8))
-            seg_mask = cv2.morphologyEx(seg_mask, cv2.MORPH_CLOSE, kernel = np.ones((6, 6), np.uint8))
-            seg_mask = seg_mask[:,:,0] > 0
         mask_save_path = f'result/mask_{time.time()}.png'
         if not os.path.exists(os.path.dirname(mask_save_path)):
             os.makedirs(os.path.dirname(mask_save_path))
@@ -49,24 +30,26 @@ class CaptionAnything():
         seg_mask_img.save(mask_save_path)
         print('seg_mask path: ', mask_save_path)
         print("seg_mask.shape: ", seg_mask.shape)
+        
+#         mask_image = mask_image(image,np.array(seg_mask_img))
+#         cv2.imwrite(f'result/mask_vis.png',mask_image)
         #  captioning with mask
-        if self.args.enable_reduce_tokens:
-            caption, crop_save_path = self.captioner.inference_with_reduced_tokens(image, seg_mask, crop_mode=self.args.seg_crop_mode, filter=self.args.clip_filter, disable_regular_box = self.args.disable_regular_box)
-        else:
-            caption, crop_save_path = self.captioner.inference_seg(image, seg_mask, crop_mode=self.args.seg_crop_mode, filter=self.args.clip_filter, disable_regular_box = self.args.disable_regular_box)
-        #  refining with TextRefiner
-        context_captions = []
-        if self.args.context_captions:
-            context_captions.append(self.captioner.inference(image))
-        if not disable_gpt and self.text_refiner is not None:
-            refined_caption = self.text_refiner.inference(query=caption, controls=controls, context=context_captions)
-        else:
-            refined_caption = {'raw_caption': caption}                
-        out = {'generated_captions': refined_caption,
-            'crop_save_path': crop_save_path,
+#         if self.args.enable_reduce_tokens:
+#             caption, crop_save_path = self.captioner.inference_with_reduced_tokens(image, seg_mask, crop_mode=self.args.seg_crop_mode, filter=self.args.clip_filter, disable_regular_box = self.args.disable_regular_box)
+#         else:
+#             caption, crop_save_path = self.captioner.inference_seg(image, seg_mask, crop_mode=self.args.seg_crop_mode, filter=self.args.clip_filter, disable_regular_box = self.args.disable_regular_box)
+            
+#         #  refining with TextRefiner
+#         context_captions = []
+#         if self.args.context_captions:
+#             context_captions.append(self.captioner.inference(image))
+#         if not disable_gpt and self.text_refiner is not None:
+#             refined_caption = self.text_refiner.inference(query=caption, controls=controls, context=context_captions)
+#         else:
+#             refined_caption = {'raw_caption': caption}                
+        out = {
             'mask_save_path': mask_save_path,
-            'mask': seg_mask_img,
-            'context_captions': context_captions}
+            'mask': seg_mask_img}
         return out
     
 def parse_augment():
@@ -86,7 +69,6 @@ def parse_augment():
     parser.add_argument('--disable_gpt', action="store_true")
     parser.add_argument('--enable_reduce_tokens', action="store_true", default=False)
     parser.add_argument('--disable_reuse_features', action="store_true", default=False)
-    parser.add_argument('--enable_morphologyex', action="store_true", default=False)
     args = parser.parse_args()
 
     if args.debug:
@@ -129,4 +111,4 @@ if __name__ == "__main__":
         print('Language controls:\n', controls)
         out = model.inference(image_path, prompt, controls)
     
-    
+