import gradio as gr from transformers import AutoModel, AutoTokenizer, AutoImageProcessor import torch import torchvision.transforms as T from PIL import Image from torchvision.transforms.functional import InterpolationMode # Define the path to your model path = 'h2oai/h2o-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 ).eval().cuda() tokenizer = AutoTokenizer.from_pretrained( path, trust_remote_code=True, use_fast=False ) tokenizer.pad_token = tokenizer.unk_token tokenizer.eos_token = "<|end|>" model.generation_config.pad_token_id = tokenizer.pad_token_id def inference(image, prompt): # Check if both image and prompt are provided if image is None or prompt.strip() == "": return "Please provide both an image and a prompt." # Process the image and get pixel_values pixel_values = load_image_msac(image) # Set generation config generation_config = dict( num_beams=1, max_new_tokens=2048, do_sample=False, ) # Generate the response response = model.chat( tokenizer, pixel_values, prompt, generation_config ) return response # Build the Gradio interface with gr.Blocks() as demo: gr.Markdown("H2O-Mississippi") with gr.Row(): image_input = gr.Image(type="pil", label="Upload an Image") prompt_input = gr.Textbox(label="Enter your prompt here") response_output = gr.Textbox(label="Model Response") with gr.Row(): submit_button = gr.Button("Submit") clear_button = gr.Button("Clear") # When the submit button is clicked, call the inference function submit_button.click( fn=inference, inputs=[image_input, prompt_input], outputs=response_output ) # Define the clear button action def clear_all(): return None, "", "" clear_button.click( fn=clear_all, inputs=None, outputs=[image_input, prompt_input, response_output] ) demo.launch()