{ "cells": [ { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [], "source": [ "from __future__ import annotations\n", "\n", "import logging\n", "import math\n", "import os\n", "from typing import Dict, List, Optional\n", "\n", "import torch\n", "from PIL import Image\n", "from torch.utils.data import DataLoader\n", "from tqdm.autonotebook import tqdm\n", "from transformers import AutoModelForVision2Seq, AutoProcessor" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [], "source": [ "class GmeQwen2VL:\n", " def __init__(\n", " self,\n", " model_name: str = \"Alibaba-NLP/gme-Qwen2-VL-2B-Instruct\",\n", " model_path: Optional[str] = None,\n", " device: str = \"cuda\" if torch.cuda.is_available() else \"cpu\",\n", " min_image_tokens=256,\n", " max_image_tokens=1280,\n", " max_length=1800,\n", " **kwargs,\n", " ) -> None:\n", " model_name = model_path or model_name\n", " self.base = AutoModelForVision2Seq.from_pretrained(\n", " model_name, torch_dtype=torch.float16, **kwargs\n", " )\n", " self.base.eval()\n", " self.normalize = True\n", " self.device = device\n", " min_pixels = min_image_tokens * 28 * 28\n", " max_pixels = max_image_tokens * 28 * 28\n", " self.max_length = max_length\n", " self.processor = AutoProcessor.from_pretrained(\n", " model_name, min_pixels=min_pixels, max_pixels=max_pixels, **kwargs\n", " )\n", " self.processor.tokenizer.padding_side = 'right'\n", " self.defualt_instruction = 'You are a helpful assistant.'\n", " self.sep = ' '\n", "\n", " def forward(\n", " self,\n", " input_ids: Optional[torch.LongTensor] = None,\n", " attention_mask: Optional[torch.Tensor] = None,\n", " position_ids: Optional[torch.LongTensor] = None,\n", " past_key_values: Optional[List[torch.FloatTensor]] = None,\n", " inputs_embeds: Optional[torch.FloatTensor] = None,\n", " pixel_values: Optional[torch.Tensor] = None,\n", " # pixel_values_videos: Optional[torch.FloatTensor] = None,\n", " image_grid_thw: Optional[torch.LongTensor] = None,\n", " # video_grid_thw: Optional[torch.LongTensor] = None,\n", " pooling_mask: Optional[torch.LongTensor] = None,\n", " **kwargs\n", " ) -> torch.Tensor:\n", " if inputs_embeds is None:\n", " inputs_embeds = self.base.model.embed_tokens(input_ids)\n", " if pixel_values is not None:\n", " pixel_values = pixel_values.type(self.base.visual.get_dtype())\n", " image_embeds = self.base.visual(pixel_values, grid_thw=image_grid_thw).to(inputs_embeds.device)\n", " image_mask = input_ids == self.base.config.image_token_id\n", " inputs_embeds[image_mask] = image_embeds\n", " # if pixel_values_videos is not None:\n", " # pixel_values_videos = pixel_values_videos.type(self.base.visual.get_dtype())\n", " # video_embeds = self.base.visual(pixel_values_videos, grid_thw=video_grid_thw).to(inputs_embeds.device)\n", " # video_mask = input_ids == self.base.config.video_token_id\n", " # inputs_embeds[video_mask] = video_embeds\n", " if attention_mask is not None:\n", " attention_mask = attention_mask.to(inputs_embeds.device)\n", "\n", " outputs = self.base.model(\n", " input_ids=None,\n", " position_ids=position_ids,\n", " attention_mask=attention_mask,\n", " past_key_values=past_key_values,\n", " inputs_embeds=inputs_embeds,\n", " )\n", "\n", " pooling_mask = attention_mask if pooling_mask is None else pooling_mask\n", " left_padding = (pooling_mask[:, -1].sum() == pooling_mask.shape[0]) # TODO\n", " if left_padding:\n", " embeddings = outputs.last_hidden_state[:, -1]\n", " else:\n", " sequence_lengths = pooling_mask.sum(dim=1) - 1\n", " batch_size = outputs.last_hidden_state.shape[0]\n", " embeddings = outputs.last_hidden_state[torch.arange(\n", " batch_size, device=outputs.last_hidden_state.device\n", " ), sequence_lengths]\n", " if self.normalize:\n", " embeddings = torch.nn.functional.normalize(embeddings, p=2, dim=1)\n", " return embeddings.contiguous()\n", "\n", " def embed(self, texts: list[str], images: list[Image.Image], is_query=True, instruction=None, **kwargs):\n", " self.base.to(self.device)\n", " # Inputs must be batched\n", " input_texts, input_images = list(), list()\n", " for t, i in zip(texts, images):\n", " if not is_query or instruction is None:\n", " instruction = self.defualt_instruction\n", " input_str = ''\n", " if i is None:\n", " input_images = None # All examples in the same batch are consistent\n", " else:\n", " input_str += '<|vision_start|><|image_pad|><|vision_end|>'\n", " i = fetch_image(i)\n", " input_images.append(i)\n", " if t is not None:\n", " input_str += t\n", " msg = f'<|im_start|>system\\n{instruction}<|im_end|>\\n<|im_start|>user\\n{input_str}<|im_end|>\\n<|im_start|>assistant\\n<|endoftext|>'\n", " input_texts.append(msg)\n", "\n", " inputs = self.processor(\n", " text=input_texts,\n", " images=input_images,\n", " padding=True,\n", " truncation=True,\n", " max_length=self.max_length,\n", " return_tensors='pt'\n", " )\n", " inputs = {k: v.to(self.device) for k, v in inputs.items()} # TODO\n", " with torch.no_grad():\n", " embeddings = self.forward(**inputs)\n", " return embeddings\n", "\n", " def encode(self, sentences: list[str], *, prompt_name=None, **kwargs):\n", " return self.get_fused_embeddings(texts=sentences, prompt_name=prompt_name, **kwargs)\n", "\n", " def encode_queries(self, queries: List[str], **kwargs):\n", " embeddings = self.encode(queries, **kwargs)\n", " return embeddings\n", "\n", " def encode_corpus(self, corpus: List[Dict[str, str]], **kwargs):\n", " if type(corpus) is dict:\n", " sentences = [\n", " (corpus[\"title\"][i] + self.sep + corpus[\"text\"][i]).strip()\n", " if \"title\" in corpus\n", " else corpus[\"text\"][i].strip()\n", " for i in range(len(corpus[\"text\"]))\n", " ]\n", " else:\n", " sentences = [\n", " (doc[\"title\"] + self.sep + doc[\"text\"]).strip() if \"title\" in doc else doc[\"text\"].strip()\n", " for doc in corpus\n", " ]\n", " embeddings = self.encode(sentences, is_query=False, **kwargs)\n", " return embeddings\n", "\n", " def get_image_embeddings(self, images: list[Image.Image] | DataLoader, **kwargs):\n", " return self.get_fused_embeddings(images=images, **kwargs)\n", "\n", " def get_text_embeddings(self, texts: list[str], **kwargs):\n", " return self.get_fused_embeddings(texts=texts, **kwargs)\n", "\n", " def get_fused_embeddings(self, texts: list[str] = None, images: list[Image.Image] | DataLoader = None, **kwargs):\n", " if isinstance(images, DataLoader):\n", " image_loader = images\n", " batch_size = image_loader.batch_size\n", " image_loader.dataset.transform = None\n", " else:\n", " batch_size = kwargs.pop('batch_size', 32)\n", " if images is None:\n", " image_loader = None\n", " else:\n", " image_loader = DataLoader(\n", " images,\n", " batch_size=batch_size,\n", " shuffle=False,\n", " collate_fn=custom_collate_fn,\n", " num_workers=min(math.floor(os.cpu_count() / 2), 8),\n", " )\n", "\n", " if texts is None:\n", " assert image_loader is not None\n", " n_batch = len(image_loader)\n", " else:\n", " n_batch = len(texts) // batch_size + int(len(texts) % batch_size > 0)\n", " image_loader = image_loader or [None] * n_batch\n", "\n", " all_embeddings = list()\n", " none_batch = [None] * batch_size\n", " show_progress_bar = kwargs.pop('show_progress_bar', True)\n", " pbar = tqdm(total=n_batch, disable=not show_progress_bar, mininterval=1, miniters=10, desc='encode')\n", " for n, img_batch in zip(range(0, n_batch * batch_size, batch_size), image_loader):\n", " text_batch = none_batch if texts is None else texts[n: n+batch_size]\n", " img_batch = none_batch if img_batch is None else img_batch\n", " embeddings = self.embed(texts=text_batch, images=img_batch, **kwargs)\n", " pbar.update(1)\n", " all_embeddings.append(embeddings.cpu())\n", " pbar.close()\n", " all_embeddings = torch.cat(all_embeddings, dim=0)\n", " return all_embeddings\n", "\n", "\n", "def custom_collate_fn(batch):\n", " return batch\n" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [], "source": [ "import base64\n", "from io import BytesIO\n", "import requests\n", "\n", "IMAGE_FACTOR = 28\n", "MIN_PIXELS = 4 * 28 * 28\n", "MAX_PIXELS = 16384 * 28 * 28\n", "MAX_RATIO = 200\n", "\n", "\n", "def round_by_factor(number: int, factor: int) -> int:\n", " \"\"\"Returns the closest integer to 'number' that is divisible by 'factor'.\"\"\"\n", " return round(number / factor) * factor\n", "\n", "\n", "def ceil_by_factor(number: int, factor: int) -> int:\n", " \"\"\"Returns the smallest integer greater than or equal to 'number' that is divisible by 'factor'.\"\"\"\n", " return math.ceil(number / factor) * factor\n", "\n", "\n", "def floor_by_factor(number: int, factor: int) -> int:\n", " \"\"\"Returns the largest integer less than or equal to 'number' that is divisible by 'factor'.\"\"\"\n", " return math.floor(number / factor) * factor\n", "\n", "\n", "def smart_resize(\n", " height: int, width: int, factor: int = IMAGE_FACTOR, min_pixels: int = MIN_PIXELS, max_pixels: int = MAX_PIXELS\n", ") -> tuple[int, int]:\n", " \"\"\"\n", " Rescales the image so that the following conditions are met:\n", " 1. Both dimensions (height and width) are divisible by 'factor'.\n", " 2. The total number of pixels is within the range ['min_pixels', 'max_pixels'].\n", " 3. The aspect ratio of the image is maintained as closely as possible.\n", " \"\"\"\n", " h_bar = max(factor, round_by_factor(height, factor))\n", " w_bar = max(factor, round_by_factor(width, factor))\n", " if h_bar * w_bar > max_pixels:\n", " beta = math.sqrt((height * width) / max_pixels)\n", " h_bar = floor_by_factor(height / beta, factor)\n", " w_bar = floor_by_factor(width / beta, factor)\n", " elif h_bar * w_bar < min_pixels:\n", " beta = math.sqrt(min_pixels / (height * width))\n", " h_bar = ceil_by_factor(height * beta, factor)\n", " w_bar = ceil_by_factor(width * beta, factor)\n", "\n", " if max(h_bar, w_bar) / min(h_bar, w_bar) > MAX_RATIO:\n", " logging.warning(\n", " f\"Absolute aspect ratio must be smaller than {MAX_RATIO}, got {max(h_bar, w_bar) / min(h_bar, w_bar)}\"\n", " )\n", " if h_bar > w_bar:\n", " h_bar = w_bar * MAX_RATIO\n", " else:\n", " w_bar = h_bar * MAX_RATIO\n", " return h_bar, w_bar\n", "\n", "\n", "def fetch_image(image: str | Image.Image, size_factor: int = IMAGE_FACTOR) -> Image.Image:\n", " image_obj = None\n", " if isinstance(image, Image.Image):\n", " image_obj = image\n", " elif image.startswith(\"http://\") or image.startswith(\"https://\"):\n", " image_obj = Image.open(requests.get(image, stream=True).raw)\n", " elif image.startswith(\"file://\"):\n", " image_obj = Image.open(image[7:])\n", " elif image.startswith(\"data:image\"):\n", " if \"base64,\" in image:\n", " _, base64_data = image.split(\"base64,\", 1)\n", " data = base64.b64decode(base64_data)\n", " image_obj = Image.open(BytesIO(data))\n", " else:\n", " image_obj = Image.open(image)\n", " if image_obj is None:\n", " raise ValueError(f\"Unrecognized image input, support local path, http url, base64 and PIL.Image, got {image}\")\n", " image = image_obj.convert(\"RGB\")\n", " ## resize\n", " # if \"resized_height\" in ele and \"resized_width\" in ele:\n", " # resized_height, resized_width = smart_resize(\n", " # ele[\"resized_height\"],\n", " # ele[\"resized_width\"],\n", " # factor=size_factor,\n", " # )\n", " # else:\n", " width, height = image.size\n", " # min_pixels = ele.get(\"min_pixels\", MIN_PIXELS)\n", " # max_pixels = ele.get(\"max_pixels\", MAX_PIXELS)\n", " resized_height, resized_width = smart_resize(\n", " height,\n", " width,\n", " factor=size_factor,\n", " min_pixels=MIN_PIXELS,\n", " max_pixels=MAX_PIXELS,\n", " )\n", " image = image.resize((resized_width, resized_height))\n", "\n", " return image" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [], "source": [ "gme = GmeQwen2VL(\"ckpts/gme-Qwen2-VL-2B-Instruct\")" ] }, { "cell_type": "code", "execution_count": null, "metadata": {}, "outputs": [], "source": [ "texts = [\n", " \"What kind of car is this?\",\n", " \"The Tesla Cybertruck is a battery electric pickup truck built by Tesla, Inc. since 2023.\"\n", " ]\n", "\n", "images = [\n", " './examples/image1.png',\n", " './examples/image2.png',\n", "]\n", "\n", "# Single-modal embedding\n", "e_text = gme.get_text_embeddings(texts=texts)\n", "e_image = gme.get_image_embeddings(images=images)\n", "print((e_text * e_image).sum(-1))\n", "## tensor([0.2281, 0.6001], dtype=torch.float16)\n", "\n", "# How to set embedding instruction\n", "e_query = gme.get_text_embeddings(texts=texts, instruction='Find an image that matches the given text.')\n", "# If is_query=False, we always use the default instruction.\n", "e_corpus = gme.get_image_embeddings(images=images, is_query=False)\n", "print((e_query * e_corpus).sum(-1))\n", "## tensor([0.2433, 0.7051], dtype=torch.float16)\n", "\n", "# Fused-modal embedding\n", "e_fused = gme.get_fused_embeddings(texts=texts, images=images)\n", "print((e_fused[0] * e_fused[1]).sum())" ] } ], "metadata": { "kernelspec": { "display_name": "Python 3 (ipykernel)", "language": "python", "name": "python3" }, "language_info": { "codemirror_mode": { "name": "ipython", "version": 3 }, "file_extension": ".py", "mimetype": "text/x-python", "name": "python", "nbconvert_exporter": "python", "pygments_lexer": "ipython3", "version": "3.10.12" } }, "nbformat": 4, "nbformat_minor": 2 }