import gradio as gr from transformers import AutoModel, AutoTokenizer, AutoImageProcessor import torch import torchvision.transforms as T from PIL import Image import logging logging.basicConfig(level=logging.INFO) from torchvision.transforms.functional import InterpolationMode import os from huggingface_hub import login hf_token = os.environ.get('hf_token', None) # Define the path to your model path = "h2oai/h2ovl-mississippi-2b" # image preprocesing IMAGENET_MEAN = (0.485, 0.456, 0.406) IMAGENET_STD = (0.229, 0.224, 0.225) def build_transform(input_size): MEAN, STD = IMAGENET_MEAN, IMAGENET_STD transform = T.Compose([ T.Lambda(lambda img: img.convert('RGB') if img.mode != 'RGB' else img), T.Resize((input_size, input_size), interpolation=InterpolationMode.BICUBIC), T.ToTensor(), T.Normalize(mean=MEAN, std=STD) ]) return transform def find_closest_aspect_ratio(aspect_ratio, target_ratios, width, height, image_size): best_ratio_diff = float('inf') best_ratio = (1, 1) area = width * height for ratio in target_ratios: target_aspect_ratio = ratio[0] / ratio[1] ratio_diff = abs(aspect_ratio - target_aspect_ratio) if ratio_diff < best_ratio_diff: best_ratio_diff = ratio_diff best_ratio = ratio elif ratio_diff == best_ratio_diff: if area > 0.5 * image_size * image_size * ratio[0] * ratio[1]: best_ratio = ratio return best_ratio def dynamic_preprocess(image, min_num=1, max_num=6, image_size=448, use_thumbnail=False): orig_width, orig_height = image.size aspect_ratio = orig_width / orig_height # calculate the existing image aspect ratio target_ratios = set( (i, j) for n in range(min_num, max_num + 1) for i in range(1, n + 1) for j in range(1, n + 1) if i * j <= max_num and i * j >= min_num) target_ratios = sorted(target_ratios, key=lambda x: x[0] * x[1]) # find the closest aspect ratio to the target target_aspect_ratio = find_closest_aspect_ratio( aspect_ratio, target_ratios, orig_width, orig_height, image_size) # calculate the target width and height target_width = image_size * target_aspect_ratio[0] target_height = image_size * target_aspect_ratio[1] blocks = target_aspect_ratio[0] * target_aspect_ratio[1] # resize the image resized_img = image.resize((target_width, target_height)) processed_images = [] for i in range(blocks): box = ( (i % (target_width // image_size)) * image_size, (i // (target_width // image_size)) * image_size, ((i % (target_width // image_size)) + 1) * image_size, ((i // (target_width // image_size)) + 1) * image_size ) # split the image split_img = resized_img.crop(box) processed_images.append(split_img) assert len(processed_images) == blocks if use_thumbnail and len(processed_images) != 1: thumbnail_img = image.resize((image_size, image_size)) processed_images.append(thumbnail_img) return processed_images, target_aspect_ratio def dynamic_preprocess2(image, min_num=1, max_num=6, image_size=448, use_thumbnail=False, prior_aspect_ratio=None): orig_width, orig_height = image.size aspect_ratio = orig_width / orig_height # calculate the existing image aspect ratio target_ratios = set( (i, j) for n in range(min_num, max_num + 1) for i in range(1, n + 1) for j in range(1, n + 1) if i * j <= max_num and i * j >= min_num) target_ratios = sorted(target_ratios, key=lambda x: x[0] * x[1]) new_target_ratios = [] if prior_aspect_ratio is not None: for i in target_ratios: if prior_aspect_ratio[0]%i[0] != 0 and prior_aspect_ratio[1]%i[1] != 0: new_target_ratios.append(i) else: continue # find the closest aspect ratio to the target target_aspect_ratio = find_closest_aspect_ratio( aspect_ratio, new_target_ratios, orig_width, orig_height, image_size) # calculate the target width and height target_width = image_size * target_aspect_ratio[0] target_height = image_size * target_aspect_ratio[1] blocks = target_aspect_ratio[0] * target_aspect_ratio[1] # resize the image resized_img = image.resize((target_width, target_height)) processed_images = [] for i in range(blocks): box = ( (i % (target_width // image_size)) * image_size, (i // (target_width // image_size)) * image_size, ((i % (target_width // image_size)) + 1) * image_size, ((i // (target_width // image_size)) + 1) * image_size ) # split the image split_img = resized_img.crop(box) processed_images.append(split_img) assert len(processed_images) == blocks if use_thumbnail and len(processed_images) != 1: thumbnail_img = image.resize((image_size, image_size)) processed_images.append(thumbnail_img) return processed_images def load_image1(image_file, input_size=448, min_num=1, max_num=12): if isinstance(image_file, str): image = Image.open(image_file).convert('RGB') else: image = image_file transform = build_transform(input_size=input_size) images, target_aspect_ratio = dynamic_preprocess(image, image_size=input_size, use_thumbnail=True, min_num=min_num, max_num=max_num) pixel_values = [transform(image) for image in images] pixel_values = torch.stack(pixel_values) return pixel_values, target_aspect_ratio def load_image2(image_file, input_size=448, min_num=1, max_num=12, target_aspect_ratio=None): if isinstance(image_file, str): image = Image.open(image_file).convert('RGB') else: image = image_file transform = build_transform(input_size=input_size) images = dynamic_preprocess2(image, image_size=input_size, use_thumbnail=True, min_num=min_num, max_num=max_num, prior_aspect_ratio=target_aspect_ratio) pixel_values = [transform(image) for image in images] pixel_values = torch.stack(pixel_values) return pixel_values def load_image_msac(file_name): pixel_values, target_aspect_ratio = load_image1(file_name, min_num=1, max_num=6) pixel_values = pixel_values.to(torch.bfloat16).cuda() pixel_values2 = load_image2(file_name, min_num=3, max_num=6, target_aspect_ratio=target_aspect_ratio) pixel_values2 = pixel_values2.to(torch.bfloat16).cuda() pixel_values = torch.cat([pixel_values2[:-1], pixel_values[:-1], pixel_values2[-1:]], 0) return pixel_values # Load the model and tokenizer model = AutoModel.from_pretrained( path, torch_dtype=torch.bfloat16, low_cpu_mem_usage=True, trust_remote_code=True, use_auth_token=hf_token ).eval().cuda() tokenizer = AutoTokenizer.from_pretrained( path, trust_remote_code=True, use_fast=False, use_auth_token=hf_token ) tokenizer.pad_token = tokenizer.unk_token tokenizer.eos_token = "<|end|>" model.generation_config.pad_token_id = tokenizer.pad_token_id def inference(image, user_message, temperature, top_p, max_new_tokens, chatbot,state, image_state): # if image is provided, store it in image_state: if chatbot is None: chatbot = [] if image is not None: image_state = load_image_msac(image) else: # If image_state is None, then no image has been provided yet if image_state is None: chatbot.append(("System", "Please provide an image to start the conversation.")) return chatbot, state, image_state, "" # Initialize history (state) if it's None if state is None: state = None # model.chat function handles None as empty history # Append user message to chatbot chatbot.append((user_message, None)) # Set generation config do_sample = (float(temperature) != 0.0) generation_config = dict( num_beams=1, max_new_tokens=int(max_new_tokens), do_sample=do_sample, temperature= float(temperature), top_p= float(top_p), ) # Call model.chat with history response_text, new_state = model.chat( tokenizer, image_state, user_message, generation_config=generation_config, history=state, return_history=True ) # update the satet with new_state state = new_state # Update chatbot with the model's response chatbot[-1] = (user_message, response_text) return chatbot, state, image_state, "" def regenerate_response(chatbot, temperature, top_p, max_new_tokens, state, image_state): # Check if there is a previous user message if chatbot is None or len(chatbot) == 0: chatbot = [] chatbot.append(("System", "Nothing to regenerate. Please start a conversation first.")) return chatbot, state, image_state # Check if there is a previous user message if state is None or image_state is None or len(state) == 0: chatbot.append(("System", "Nothing to regenerate. Please start a conversation first.")) return chatbot, state, image_state # Get the last user message last_user_message, last_response = chatbot[-1] state = state[:-1] # Remove last assistant's response from history if len(state) == 0: state = None # Set generation config do_sample = (float(temperature) != 0.0) generation_config = dict( num_beams=1, max_new_tokens=int(max_new_tokens), do_sample=do_sample, temperature= float(temperature), top_p= float(top_p), ) # Regenerate the response response_text, new_state = model.chat( tokenizer, image_state, last_user_message, generation_config=generation_config, history=state, # Exclude last assistant's response return_history=True ) # Update the state with new_state state = new_state # Update chatbot with the regenerated response chatbot.append((last_user_message, response_text)) return chatbot, state, image_state def clear_all(): return [], None, None, None # Clear chatbot, state, image_state, image_input # Build the Gradio interface with gr.Blocks() as demo: gr.Markdown("# **H2OVL-Mississippi**") state= gr.State() image_state = gr.State() with gr.Row(equal_height=True): # First column with image input with gr.Column(scale=1): image_input = gr.Image(type="pil", label="Upload an Image") # Second column with chatbot and user input with gr.Column(scale=2): chatbot = gr.Chatbot(label="Conversation") user_input = gr.Textbox(label="What is your question", placeholder="Type your message here") with gr.Accordion('Parameters', open=False): with gr.Row(): temperature_input = gr.Slider( minimum=0.0, maximum=1.0, step=0.1, value=0.0, interactive=True, label="Temperature") top_p_input = gr.Slider( minimum=0.0, maximum=1.0, step=0.1, value=0.9, interactive=True, label="Top P") max_new_tokens_input = gr.Slider( minimum=0, maximum=4096, step=64, value=1024, interactive=True, label="Max New Tokens (default: 1024)" ) with gr.Row(): submit_button = gr.Button("Submit") regenerate_button = gr.Button("Regenerate") clear_button = gr.Button("Clear") # When the submit button is clicked, call the inference function submit_button.click( fn=inference, inputs=[image_input, user_input, temperature_input, top_p_input, max_new_tokens_input, chatbot, state, image_state], outputs=[chatbot, state, image_state, user_input] ) # When the regenerate button is clicked, re-run the last inference regenerate_button.click( fn=regenerate_response, inputs=[chatbot, temperature_input, top_p_input,max_new_tokens_input, state, image_state], outputs=[chatbot, state, image_state] ) clear_button.click( fn=clear_all, inputs=None, outputs=[chatbot, state, image_state, image_input] ) gr.Examples( examples=[ ["assets/driver_license.png", "Extract the text from the image and fill the following json {'license_number':'',\n'full_name':'',\n'date_of_birth':'',\n'address':'',\n'issue_date':'',\n'expiration_date':'',\n}"], # ["assets/receipt.jpg", "Read the text on the image"], ["assets/invoice.png", "Please extract the following fields, and return the result in JSON format: supplier_name, supplier_address, customer_name, customer_address, invoice_number, invoice_total_amount, invoice_tax_amount"], ["assets/CBA-1H23-Results-Presentation_wheel.png", "What is the efficiency of H2O.AI in document processing?"], ], inputs = [image_input, user_input], # outputs = [chatbot, state, image_state, user_input], # fn=inference, label = "examples", ) demo.launch()