string-sg/visual-chatgpt
0
1import sys2import os3sys.path.append(os.path.dirname(os.path.realpath(__file__)))4os.makedirs('image', exist_ok=True)5sys.path.append(os.path.dirname(os.path.dirname(os.path.realpath(__file__))))6import gradio as gr7from transformers import AutoModelForCausalLM, AutoTokenizer, CLIPSegProcessor, CLIPSegForImageSegmentation8import torch9from diffusers import StableDiffusionPipeline10from diffusers import StableDiffusionInstructPix2PixPipeline, EulerAncestralDiscreteScheduler11from langchain.agents.initialize import initialize_agent12from langchain.agents.tools import Tool13from langchain.chains.conversation.memory import ConversationBufferMemory14from langchain.llms.openai import OpenAI15import re16import uuid17from diffusers import StableDiffusionInpaintPipeline18from diffusers import StableDiffusionControlNetPipeline, ControlNetModel19from diffusers import UniPCMultistepScheduler20from PIL import Image21import numpy as np22from omegaconf import OmegaConf23from transformers import pipeline, BlipProcessor, BlipForConditionalGeneration, BlipForQuestionAnswering24import cv225import einops26from pytorch_lightning import seed_everything27import random28 29VISUAL_CHATGPT_PREFIX = """Visual ChatGPT is designed to be able to assist with a wide range of text and visual related tasks, from answering simple questions to providing in-depth explanations and discussions on a wide range of topics. Visual ChatGPT is able to generate human-like text based on the input it receives, allowing it to engage in natural-sounding conversations and provide responses that are coherent and relevant to the topic at hand.30Visual ChatGPT is able to process and understand large amounts of text and images. As a language model, Visual ChatGPT can not directly read images, but it has a list of tools to finish different visual tasks. Each image will have a file name formed as "image/xxx.png", and Visual ChatGPT can invoke different tools to indirectly understand pictures. When talking about images, Visual ChatGPT is very strict to the file name and will never fabricate nonexistent files. When using tools to generate new image files, Visual ChatGPT is also known that the image may not be the same as the user's demand, and will use other visual question answering tools or description tools to observe the real image. Visual ChatGPT is able to use tools in a sequence, and is loyal to the tool observation outputs rather than faking the image content and image file name. It will remember to provide the file name from the last tool observation, if a new image is generated.31Human may provide new figures to Visual ChatGPT with a description. The description helps Visual ChatGPT to understand this image, but Visual ChatGPT should use tools to finish following tasks, rather than directly imagine from the description.32Overall, Visual ChatGPT is a powerful visual dialogue assistant tool that can help with a wide range of tasks and provide valuable insights and information on a wide range of topics. 33TOOLS:34------35Visual ChatGPT has access to the following tools:"""36 37VISUAL_CHATGPT_FORMAT_INSTRUCTIONS = """To use a tool, please use the following format:38```39Thought: Do I need to use a tool? Yes40Action: the action to take, should be one of [{tool_names}]41Action Input: the input to the action42Observation: the result of the action43```44When you have a response to say to the Human, or if you do not need to use a tool, you MUST use the format:45```46Thought: Do I need to use a tool? No47{ai_prefix}: [your response here]48```49"""50 51VISUAL_CHATGPT_SUFFIX = """You are very strict to the filename correctness and will never fake a file name if it does not exist.52You will remember to provide the image file name loyally if it's provided in the last tool observation.53Begin!54Previous conversation history:55{chat_history}56New input: {input}57Since Visual ChatGPT is a text language model, Visual ChatGPT must use tools to observe images rather than imagination.58The thoughts and observations are only visible for Visual ChatGPT, Visual ChatGPT should remember to repeat important information in the final response for Human. 59Thought: Do I need to use a tool? {agent_scratchpad}"""60 61def cut_dialogue_history(history_memory, keep_last_n_words=500):62 tokens = history_memory.split()63 n_tokens = len(tokens)64 print(f"hitory_memory:{history_memory}, n_tokens: {n_tokens}")65 if n_tokens < keep_last_n_words:66 return history_memory67 else:68 paragraphs = history_memory.split('\n')69 last_n_tokens = n_tokens70 while last_n_tokens >= keep_last_n_words:71 last_n_tokens = last_n_tokens - len(paragraphs[0].split(' '))72 paragraphs = paragraphs[1:]73 return '\n' + '\n'.join(paragraphs)74 75def get_new_image_name(org_img_name, func_name="update"):76 head_tail = os.path.split(org_img_name)77 head = head_tail[0]78 tail = head_tail[1]79 name_split = tail.split('.')[0].split('_')80 this_new_uuid = str(uuid.uuid4())[0:4]81 if len(name_split) == 1:82 most_org_file_name = name_split[0]83 recent_prev_file_name = name_split[0]84 new_file_name = '{}_{}_{}_{}.png'.format(this_new_uuid, func_name, recent_prev_file_name, most_org_file_name)85 else:86 assert len(name_split) == 487 most_org_file_name = name_split[3]88 recent_prev_file_name = name_split[0]89 new_file_name = '{}_{}_{}_{}.png'.format(this_new_uuid, func_name, recent_prev_file_name, most_org_file_name)90 return os.path.join(head, new_file_name)91 92def create_model(config_path, device):93 config = OmegaConf.load(config_path)94 OmegaConf.update(config, "model.params.cond_stage_config.params.device", device)95 model = instantiate_from_config(config.model).cpu()96 print(f'Loaded model config from [{config_path}]')97 return model98 99class MaskFormer:100 def __init__(self, device):101 self.device = device102 self.processor = CLIPSegProcessor.from_pretrained("CIDAS/clipseg-rd64-refined")103 self.model = CLIPSegForImageSegmentation.from_pretrained("CIDAS/clipseg-rd64-refined").to(device)104 105 def inference(self, image_path, text):106 threshold = 0.5107 min_area = 0.02108 padding = 20109 original_image = Image.open(image_path)110 image = original_image.resize((512, 512))111 inputs = self.processor(text=text, images=image, padding="max_length", return_tensors="pt",).to(self.device)112 with torch.no_grad():113 outputs = self.model(**inputs)114 mask = torch.sigmoid(outputs[0]).squeeze().cpu().numpy() > threshold115 area_ratio = len(np.argwhere(mask)) / (mask.shape[0] * mask.shape[1])116 if area_ratio < min_area:117 return None118 true_indices = np.argwhere(mask)119 mask_array = np.zeros_like(mask, dtype=bool)120 for idx in true_indices:121 padded_slice = tuple(slice(max(0, i - padding), i + padding + 1) for i in idx)122 mask_array[padded_slice] = True123 visual_mask = (mask_array * 255).astype(np.uint8)124 image_mask = Image.fromarray(visual_mask)125 return image_mask.resize(image.size)126 127class ImageEditing:128 def __init__(self, device):129 print("Initializing StableDiffusionInpaint to %s" % device)130 self.device = device131 self.mask_former = MaskFormer(device=self.device)132 self.inpainting = StableDiffusionInpaintPipeline.from_pretrained("runwayml/stable-diffusion-inpainting",).to(device)133 134 def remove_part_of_image(self, input):135 image_path, to_be_removed_txt = input.split(",")136 print(f'remove_part_of_image: to_be_removed {to_be_removed_txt}')137 return self.replace_part_of_image(f"{image_path},{to_be_removed_txt},background")138 139 def replace_part_of_image(self, input):140 image_path, to_be_replaced_txt, replace_with_txt = input.split(",")141 print(f'replace_part_of_image: replace_with_txt {replace_with_txt}')142 original_image = Image.open(image_path)143 mask_image = self.mask_former.inference(image_path, to_be_replaced_txt)144 updated_image = self.inpainting(prompt=replace_with_txt, image=original_image, mask_image=mask_image).images[0]145 updated_image_path = get_new_image_name(image_path, func_name="replace-something")146 updated_image.save(updated_image_path)147 return updated_image_path148 149class Pix2Pix:150 def __init__(self, device):151 print("Initializing Pix2Pix to %s" % device)152 self.device = device153 self.pipe = StableDiffusionInstructPix2PixPipeline.from_pretrained("timbrooks/instruct-pix2pix", torch_dtype=torch.float16, safety_checker=None).to(device)154 self.pipe.scheduler = EulerAncestralDiscreteScheduler.from_config(self.pipe.scheduler.config)155 156 def inference(self, inputs):157 """Change style of image."""158 print("===>Starting Pix2Pix Inference")159 image_path, instruct_text = inputs.split(",")[0], ','.join(inputs.split(',')[1:])160 original_image = Image.open(image_path)161 image = self.pipe(instruct_text,image=original_image,num_inference_steps=40,image_guidance_scale=1.2,).images[0]162 updated_image_path = get_new_image_name(image_path, func_name="pix2pix")163 image.save(updated_image_path)164 return updated_image_path165 166class T2I:167 def __init__(self, device):168 print("Initializing T2I to %s" % device)169 self.device = device170 self.pipe = StableDiffusionPipeline.from_pretrained("runwayml/stable-diffusion-v1-5", torch_dtype=torch.float16)171 self.text_refine_tokenizer = AutoTokenizer.from_pretrained("Gustavosta/MagicPrompt-Stable-Diffusion")172 self.text_refine_model = AutoModelForCausalLM.from_pretrained("Gustavosta/MagicPrompt-Stable-Diffusion")173 self.text_refine_gpt2_pipe = pipeline("text-generation", model=self.text_refine_model, tokenizer=self.text_refine_tokenizer, device=self.device)174 self.pipe.to(device)175 176 def inference(self, text):177 image_filename = os.path.join('image', str(uuid.uuid4())[0:8] + ".png")178 refined_text = self.text_refine_gpt2_pipe(text)[0]["generated_text"]179 print(f'{text} refined to {refined_text}')180 image = self.pipe(refined_text).images[0]181 image.save(image_filename)182 print(f"Processed T2I.run, text: {text}, image_filename: {image_filename}")183 return image_filename184 185class ImageCaptioning:186 def __init__(self, device):187 print("Initializing ImageCaptioning to %s" % device)188 self.device = device189 self.processor = BlipProcessor.from_pretrained("Salesforce/blip-image-captioning-base")190 self.model = BlipForConditionalGeneration.from_pretrained("Salesforce/blip-image-captioning-base").to(self.device)191 192 def inference(self, image_path):193 inputs = self.processor(Image.open(image_path), return_tensors="pt").to(self.device)194 out = self.model.generate(**inputs)195 captions = self.processor.decode(out[0], skip_special_tokens=True)196 return captions197 198class image2canny:199 def __init__(self):200 print("Direct detect canny.")201 self.low_thresh = 100202 self.high_thresh = 200203 204 def inference(self, inputs):205 print("===>Starting image2canny Inference")206 image = Image.open(inputs)207 image = np.array(image)208 209 image = cv2.Canny(image, low_threshold, high_threshold)210 image = image[:, :, None]211 image = np.concatenate([image, image, image], axis=2)212 canny_image = Image.fromarray(image)213 updated_image_path = get_new_image_name(inputs, func_name="edge")214 canny_image.save(updated_image_path)215 return updated_image_path216 217class canny2image:218 def __init__(self, device):219 print("Initialize the canny2image model.")220 low_threshold = 100221 high_threshold = 200222 223 # Models224 controlnet = ControlNetModel.from_pretrained("lllyasviel/sd-controlnet-canny", torch_dtype=torch.float16)225 self.pipe = StableDiffusionControlNetPipeline.from_pretrained(226 "runwayml/stable-diffusion-v1-5", controlnet=controlnet, safety_checker=None, torch_dtype=torch.float16227 )228 self.pipe.scheduler = UniPCMultistepScheduler.from_config(self.pipe.scheduler.config)229 230 # This command loads the individual model components on GPU on-demand. So, we don't231 # need to explicitly call pipe.to("cuda").232 self.pipe.enable_model_cpu_offload()233 234 self.pipe.enable_xformers_memory_efficient_attention()235 236 # Generator seed,237 self.generator = torch.manual_seed(0)238 239 240 def get_canny_filter(self,image):241 if not isinstance(image, np.ndarray):242 image = np.array(image) 243 image = cv2.Canny(image, low_threshold, high_threshold)244 image = image[:, :, None]245 image = np.concatenate([image, image, image], axis=2)246 canny_image = Image.fromarray(image)247 return canny_image248 249 def inference(self, inputs):250 print("===>Starting canny2image Inference")251 image_path, instruct_text = inputs.split(",")[0], ','.join(inputs.split(',')[1:])252 image = Image.open(image_path)253 image = np.array(image)254 prompt = instruct_text255 canny_image = self.get_canny_filter(image)256 output = self.pipe(prompt,canny_image,generator=self.generator,num_images_per_prompt=1,num_inference_steps=20)257 258 updated_image_path = get_new_image_name(image_path, func_name="canny2image")259 real_image = Image.fromarray(output.images[0]) # get default the index0 image260 real_image.save(updated_image_path)261 return updated_image_path262 263class BLIPVQA:264 def __init__(self, device):265 print("Initializing BLIP VQA to %s" % device)266 self.device = device267 self.processor = BlipProcessor.from_pretrained("Salesforce/blip-vqa-base")268 self.model = BlipForQuestionAnswering.from_pretrained("Salesforce/blip-vqa-base").to(self.device)269 270 def get_answer_from_question_and_image(self, inputs):271 image_path, question = inputs.split(",")272 raw_image = Image.open(image_path).convert('RGB')273 print(F'BLIPVQA :question :{question}')274 inputs = self.processor(raw_image, question, return_tensors="pt").to(self.device)275 out = self.model.generate(**inputs)276 answer = self.processor.decode(out[0], skip_special_tokens=True)277 return answer278 279class ConversationBot:280 def __init__(self):281 print("Initializing VisualChatGPT")282 #self.edit = ImageEditing(device="cuda:0")283 self.i2t = ImageCaptioning(device="cuda:0")284 self.t2i = T2I(device="cuda:0")285 self.image2canny = image2canny()286 #self.canny2image = canny2image(device="cuda:0")287 self.BLIPVQA = BLIPVQA(device="cuda:0")288 #self.pix2pix = Pix2Pix(device="cuda:0")289 self.memory = ConversationBufferMemory(memory_key="chat_history", output_key='output')290 self.tools = [291 Tool(name="Get Photo Description", func=self.i2t.inference,292 description="useful when you want to know what is inside the photo. receives image_path as input. "293 "The input to this tool should be a string, representing the image_path. "),294 Tool(name="Generate Image From User Input Text", func=self.t2i.inference,295 description="useful when you want to generate an image from a user input text and save it to a file. like: generate an image of an object or something, or generate an image that includes some objects. "296 "The input to this tool should be a string, representing the text used to generate image. "),297 #Tool(name="Remove Something From The Photo", func=self.edit.remove_part_of_image,298 # description="useful when you want to remove and object or something from the photo from its description or location. "299 # "The input to this tool should be a comma seperated string of two, representing the image_path and the object need to be removed. "),300 #Tool(name="Replace Something From The Photo", func=self.edit.replace_part_of_image,301 #description="useful when you want to replace an object from the object description or location with another object from its description. "302 #"The input to this tool should be a comma seperated string of three, representing the image_path, the object to be replaced, the object to be replaced with "),303 304 #Tool(name="Instruct Image Using Text", func=self.pix2pix.inference,305 # description="useful when you want to the style of the image to be like the text. like: make it look like a painting. or make it like a robot. "306 # "The input to this tool should be a comma seperated string of two, representing the image_path and the text. "),307 Tool(name="Answer Question About The Image", func=self.BLIPVQA.get_answer_from_question_and_image,308 description="useful when you need an answer for a question based on an image. like: what is the background color of the last image, how many cats in this figure, what is in this figure. "309 "The input to this tool should be a comma seperated string of two, representing the image_path and the question"),310 Tool(name="Edge Detection On Image", func=self.image2canny.inference,311 description="useful when you want to detect the edge of the image. like: detect the edges of this image, or canny detection on image, or peform edge detection on this image, or detect the canny image of this image. "312 "The input to this tool should be a string, representing the image_path"),313 #Tool(name="Generate Image Condition On Canny Image", func=self.canny2image.inference,314 # description="useful when you want to generate a new real image from both the user desciption and a canny image. like: generate a real image of a object or something from this canny image, or generate a new real image of a object or something from this edge image. "315 # "The input to this tool should be a comma seperated string of two, representing the image_path and the user description. "),316 #Tool(name="Line Detection On Image", func=self.image2line.inference,317 #description="useful when you want to detect the straight line of the image. like: detect the straight lines of this image, or straight line detection on image, or peform straight line detection on this image, or detect the straight line image of this image. "318 # "The input to this tool should be a string, representing the image_path"),319 #Tool(name="Generate Image Condition On Line Image", func=self.line2image.inference,320 #description="useful when you want to generate a new real image from both the user desciption and a straight line image. like: generate a real image of a object or something from this straight line image, or generate a new real image of a object or something from this straight lines. "321 # "The input to this tool should be a comma seperated string of two, representing the image_path and the user description. "),322 #Tool(name="Hed Detection On Image", func=self.image2hed.inference,323 #description="useful when you want to detect the soft hed boundary of the image. like: detect the soft hed boundary of this image, or hed boundary detection on image, or peform hed boundary detection on this image, or detect soft hed boundary image of this image. "324 # "The input to this tool should be a string, representing the image_path"),325 #Tool(name="Generate Image Condition On Soft Hed Boundary Image", func=self.hed2image.inference,326 #description="useful when you want to generate a new real image from both the user desciption and a soft hed boundary image. like: generate a real image of a object or something from this soft hed boundary image, or generate a new real image of a object or something from this hed boundary. "327 # "The input to this tool should be a comma seperated string of two, representing the image_path and the user description"),328 #Tool(name="Segmentation On Image", func=self.image2seg.inference,329 #description="useful when you want to detect segmentations of the image. like: segment this image, or generate segmentations on this image, or peform segmentation on this image. "330 #"The input to this tool should be a string, representing the image_path"),331 #Tool(name="Generate Image Condition On Segmentations", func=self.seg2image.inference,332 #description="useful when you want to generate a new real image from both the user desciption and segmentations. like: generate a real image of a object or something from this segmentation image, or generate a new real image of a object or something from these segmentations. "333 #"The input to this tool should be a comma seperated string of two, representing the image_path and the user description"),334 #Tool(name="Predict Depth On Image", func=self.image2depth.inference,335 #description="useful when you want to detect depth of the image. like: generate the depth from this image, or detect the depth map on this image, or predict the depth for this image. "336 #"The input to this tool should be a string, representing the image_path"),337 #Tool(name="Generate Image Condition On Depth", func=self.depth2image.inference,338 #description="useful when you want to generate a new real image from both the user desciption and depth image. like: generate a real image of a object or something from this depth image, or generate a new real image of a object or something from the depth map. "339 #"The input to this tool should be a comma seperated string of two, representing the image_path and the user description"),340 #Tool(name="Predict Normal Map On Image", func=self.image2normal.inference,341 #description="useful when you want to detect norm map of the image. like: generate normal map from this image, or predict normal map of this image. "342 #"The input to this tool should be a string, representing the image_path"),343 #Tool(name="Generate Image Condition On Normal Map", func=self.normal2image.inference,344 #description="useful when you want to generate a new real image from both the user desciption and normal map. like: generate a real image of a object or something from this normal map, or generate a new real image of a object or something from the normal map. "345 #"The input to this tool should be a comma seperated string of two, representing the image_path and the user description"),346 #Tool(name="Sketch Detection On Image", func=self.image2scribble.inference,347 #description="useful when you want to generate a scribble of the image. like: generate a scribble of this image, or generate a sketch from this image, detect the sketch from this image. "348 #"The input to this tool should be a string, representing the image_path"),349 #Tool(name="Generate Image Condition On Sketch Image", func=self.scribble2image.inference,350 #description="useful when you want to generate a new real image from both the user desciption and a scribble image or a sketch image. "351 #"The input to this tool should be a comma seperated string of two, representing the image_path and the user description"),352 #Tool(name="Pose Detection On Image", func=self.image2pose.inference,353 #description="useful when you want to detect the human pose of the image. like: generate human poses of this image, or generate a pose image from this image. "354 #"The input to this tool should be a string, representing the image_path"),355 #Tool(name="Generate Image Condition On Pose Image", func=self.pose2image.inference,356 #description="useful when you want to generate a new real image from both the user desciption and a human pose image. like: generate a real image of a human from this human pose image, or generate a new real image of a human from this pose. "357 #"The input to this tool should be a comma seperated string of two, representing the image_path and the user description")]358 ]359 360 def init_langchain(self,api_key):361 self.llm = OpenAI(temperature = 0, openai_api_key = api_key)362 self.agent = initialize_agent(363 self.tools,364 self.llm,365 agent="conversational-react-description",366 verbose=True,367 memory=self.memory,368 return_intermediate_steps=True,369 agent_kwargs={'prefix': VISUAL_CHATGPT_PREFIX, 'format_instructions': VISUAL_CHATGPT_FORMAT_INSTRUCTIONS, 'suffix': VISUAL_CHATGPT_SUFFIX}, )370 return gr.update(visible = True)371 372 def run_text(self, text, state):373 print("===============Running run_text =============")374 print("Inputs:", text, state)375 print("======>Previous memory:\n %s" % self.agent.memory)376 self.agent.memory.buffer = cut_dialogue_history(self.agent.memory.buffer, keep_last_n_words=500)377 res = self.agent({"input": text})378 print("======>Current memory:\n %s" % self.agent.memory)379 response = re.sub('(image/\S*png)', lambda m: f'})*{m.group(0)}*', res['output'])380 state = state + [(text, response)]381 print("Outputs:", state)382 return state, state383 384 def run_image(self, image, state, txt):385 print("===============Running run_image =============")386 print("Inputs:", image, state)387 print("======>Previous memory:\n %s" % self.agent.memory)388 image_filename = os.path.join('image', str(uuid.uuid4())[0:8] + ".png")389 print("======>Auto Resize Image...")390 img = Image.open(image.name)391 width, height = img.size392 ratio = min(512 / width, 512 / height)393 width_new, height_new = (round(width * ratio), round(height * ratio))394 img = img.resize((width_new, height_new))395 img = img.convert('RGB')396 img.save(image_filename, "PNG")397 print(f"Resize image form {width}x{height} to {width_new}x{height_new}")398 description = self.i2t.inference(image_filename)399 Human_prompt = "\nHuman: provide a figure named {}. The description is: {}. This information helps you to understand this image, but you should use tools to finish following tasks, " \400 "rather than directly imagine from my description. If you understand, say \"Received\". \n".format(image_filename, description)401 AI_prompt = "Received. "402 self.agent.memory.buffer = self.agent.memory.buffer + Human_prompt + 'AI: ' + AI_prompt403 print("======>Current memory:\n %s" % self.agent.memory)404 state = state + [(f"*{image_filename}*", AI_prompt)]405 print("Outputs:", state)406 return state, state, txt + ' ' + image_filename + ' '407 408 409bot = ConversationBot()410with gr.Blocks(css="#chatbot .overflow-y-auto{height:500px}") as demo:411 gr.Markdown("# Visual ChatGPT <p> Currently supports text, image captioning, image generation, and visual question answering</p>")412 openai_api_key_input = gr.Textbox(type = "password", label = "Enter your OpenAI API key here") 413 chatbot = gr.Chatbot(elem_id="chatbot", label="Visual ChatGPT")414 state = gr.State([])415 416 with gr.Row(visible = False) as input_row:417 with gr.Column(scale=0.7):418 txt = gr.Textbox(show_label=False, placeholder="Enter text and press enter, or upload an image").style(container=False)419 with gr.Column(scale=0.15, min_width=0):420 clear = gr.Button("Clear️")421 with gr.Column(scale=0.15, min_width=0):422 btn = gr.UploadButton("Upload", file_types=["image"])423 424 openai_api_key_input.submit(bot.init_langchain,openai_api_key_input,[input_row])425 txt.submit(bot.run_text, [txt, state], [chatbot, state])426 txt.submit(lambda: "", None, txt)427 btn.upload(bot.run_image, [btn, state, txt], [chatbot, state, txt])428 clear.click(bot.memory.clear)429 clear.click(lambda: [], None, chatbot)430 clear.click(lambda: [], None, state)431demo.launch()