{ "cells": [ { "cell_type": "code", "execution_count": null, "metadata": { "id": "VjYy0F2gZIPR" }, "outputs": [], "source": [ "!apt -y install -qq aria2\n", "!aria2c --console-log-level=error -c -x 16 -s 16 -k 1M https://huggingface.co/camenduru/joy-caption-alpha-one/raw/main/text_model/adapter_config.json -d /content/joy/text_model -o adapter_config.json\n", "!aria2c --console-log-level=error -c -x 16 -s 16 -k 1M https://huggingface.co/camenduru/joy-caption-alpha-one/resolve/main/text_model/adapter_model.safetensors -d /content/joy/text_model -o adapter_model.safetensors\n", "!aria2c --console-log-level=error -c -x 16 -s 16 -k 1M https://huggingface.co/camenduru/joy-caption-alpha-one/resolve/main/clip_model.pt -d /content/joy -o clip_model.pt\n", "!aria2c --console-log-level=error -c -x 16 -s 16 -k 1M https://huggingface.co/camenduru/joy-caption-alpha-one/raw/main/config.yaml -d /content/joy -o config.yaml\n", "!aria2c --console-log-level=error -c -x 16 -s 16 -k 1M https://huggingface.co/camenduru/joy-caption-alpha-one/resolve/main/image_adapter.pt -d /content/joy -o image_adapter.pt\n", "\n", "!pip install peft bitsandbytes" ] }, { "cell_type": "code", "execution_count": null, "metadata": { "id": "EBNKXBwIkJLk" }, "outputs": [], "source": [ "from huggingface_hub import InferenceClient\n", "from torch import nn\n", "from transformers import AutoModel, AutoProcessor, AutoTokenizer, PreTrainedTokenizer, PreTrainedTokenizerFast, AutoModelForCausalLM\n", "import torch\n", "import torch.amp.autocast_mode\n", "from PIL import Image\n", "import os\n", "import torchvision.transforms.functional as TVF\n", "\n", "CLIP_PATH = \"google/siglip-so400m-patch14-384\"\n", "MODEL_PATH = \"unsloth/Meta-Llama-3.1-8B\"\n", "CAPTION_TYPE_MAP = {\n", " (\"descriptive\", \"formal\", False, False): [\"Write a descriptive caption for this image in a formal tone.\"],\n", " (\"descriptive\", \"formal\", False, True): [\"Write a descriptive caption for this image in a formal tone within {word_count} words.\"],\n", " (\"descriptive\", \"formal\", True, False): [\"Write a {length} descriptive caption for this image in a formal tone.\"],\n", " (\"descriptive\", \"informal\", False, False): [\"Write a descriptive caption for this image in a casual tone.\"],\n", " (\"descriptive\", \"informal\", False, True): [\"Write a descriptive caption for this image in a casual tone within {word_count} words.\"],\n", " (\"descriptive\", \"informal\", True, False): [\"Write a {length} descriptive caption for this image in a casual tone.\"],\n", " (\"training_prompt\", \"formal\", False, False): [\"Write a stable diffusion prompt for this image.\"],\n", " (\"training_prompt\", \"formal\", False, True): [\"Write a stable diffusion prompt for this image within {word_count} words.\"],\n", " (\"training_prompt\", \"formal\", True, False): [\"Write a {length} stable diffusion prompt for this image.\"],\n", " (\"rng-tags\", \"formal\", False, False): [\"Write a list of Booru tags for this image.\"],\n", " (\"rng-tags\", \"formal\", False, True): [\"Write a list of Booru tags for this image within {word_count} words.\"],\n", " (\"rng-tags\", \"formal\", True, False): [\"Write a {length} list of Booru tags for this image.\"],\n", "}\n", "\n", "class ImageAdapter(nn.Module):\n", "\tdef __init__(self, input_features: int, output_features: int, ln1: bool, pos_emb: bool, num_image_tokens: int, deep_extract: bool):\n", "\t\tsuper().__init__()\n", "\t\tself.deep_extract = deep_extract\n", "\t\tif self.deep_extract:\n", "\t\t\tinput_features = input_features * 5\n", "\t\tself.linear1 = nn.Linear(input_features, output_features)\n", "\t\tself.activation = nn.GELU()\n", "\t\tself.linear2 = nn.Linear(output_features, output_features)\n", "\t\tself.ln1 = nn.Identity() if not ln1 else nn.LayerNorm(input_features)\n", "\t\tself.pos_emb = None if not pos_emb else nn.Parameter(torch.zeros(num_image_tokens, input_features))\n", "\t\tself.other_tokens = nn.Embedding(3, output_features)\n", "\t\tself.other_tokens.weight.data.normal_(mean=0.0, std=0.02) # Matches HF's implementation of llama3\n", "\tdef forward(self, vision_outputs: torch.Tensor):\n", "\t\tif self.deep_extract:\n", "\t\t\tx = torch.concat((\n", "\t\t\t\tvision_outputs[-2],\n", "\t\t\t\tvision_outputs[3],\n", "\t\t\t\tvision_outputs[7],\n", "\t\t\t\tvision_outputs[13],\n", "\t\t\t\tvision_outputs[20],\n", "\t\t\t), dim=-1)\n", "\t\t\tassert len(x.shape) == 3, f\"Expected 3, got {len(x.shape)}\" # batch, tokens, features\n", "\t\t\tassert x.shape[-1] == vision_outputs[-2].shape[-1] * 5, f\"Expected {vision_outputs[-2].shape[-1] * 5}, got {x.shape[-1]}\"\n", "\t\telse:\n", "\t\t\tx = vision_outputs[-2]\n", "\t\tx = self.ln1(x)\n", "\t\tif self.pos_emb is not None:\n", "\t\t\tassert x.shape[-2:] == self.pos_emb.shape, f\"Expected {self.pos_emb.shape}, got {x.shape[-2:]}\"\n", "\t\t\tx = x + self.pos_emb\n", "\t\tx = self.linear1(x)\n", "\t\tx = self.activation(x)\n", "\t\tx = self.linear2(x)\n", "\t\tother_tokens = self.other_tokens(torch.tensor([0, 1], device=self.other_tokens.weight.device).expand(x.shape[0], -1))\n", "\t\tassert other_tokens.shape == (x.shape[0], 2, x.shape[2]), f\"Expected {(x.shape[0], 2, x.shape[2])}, got {other_tokens.shape}\"\n", "\t\tx = torch.cat((other_tokens[:, 0:1], x, other_tokens[:, 1:2]), dim=1)\n", "\t\treturn x\n", "\tdef get_eot_embedding(self):\n", "\t\treturn self.other_tokens(torch.tensor([2], device=self.other_tokens.weight.device)).squeeze(0)\n", "\n", "clip_processor = AutoProcessor.from_pretrained(CLIP_PATH)\n", "clip_model = AutoModel.from_pretrained(CLIP_PATH)\n", "clip_model = clip_model.vision_model\n", "checkpoint = torch.load(\"/content/joy/clip_model.pt\", map_location='cpu')\n", "checkpoint = {k.replace(\"_orig_mod.module.\", \"\"): v for k, v in checkpoint.items()}\n", "clip_model.load_state_dict(checkpoint)\n", "# del checkpoint\n", "clip_model.eval()\n", "clip_model.requires_grad_(False)\n", "clip_model.to(\"cuda\")\n", "tokenizer = AutoTokenizer.from_pretrained(MODEL_PATH, use_fast=False)\n", "assert isinstance(tokenizer, PreTrainedTokenizer) or isinstance(tokenizer, PreTrainedTokenizerFast), f\"Tokenizer is of type {type(tokenizer)}\"\n", "text_model = AutoModelForCausalLM.from_pretrained(MODEL_PATH, load_in_8bit=True, device_map=\"auto\", torch_dtype=torch.bfloat16)\n", "text_model.load_adapter(\"/content/joy/text_model\")\n", "text_model.eval()\n", "image_adapter = ImageAdapter(clip_model.config.hidden_size, text_model.config.hidden_size, False, False, 38, False)\n", "image_adapter.load_state_dict(torch.load(\"/content/joy/image_adapter.pt\", map_location=\"cpu\"))\n", "image_adapter.eval()\n", "image_adapter.to(\"cuda\")\n", "\n", "@torch.no_grad()\n", "def stream_chat(input_image: Image.Image, caption_type: str, caption_tone: str, caption_length: str | int) -> str:\n", " torch.cuda.empty_cache()\n", " length = None if caption_length == \"any\" else caption_length\n", " if isinstance(length, str):\n", " try:\n", " length = int(length)\n", " except ValueError:\n", " pass\n", " if caption_type == \"rng-tags\" or caption_type == \"training_prompt\":\n", " caption_tone = \"formal\"\n", " prompt_key = (caption_type, caption_tone, isinstance(length, str), isinstance(length, int))\n", " if prompt_key not in CAPTION_TYPE_MAP:\n", " raise ValueError(f\"Invalid caption type: {prompt_key}\")\n", " prompt_str = CAPTION_TYPE_MAP[prompt_key][0].format(length=length, word_count=length)\n", " print(f\"Prompt: {prompt_str}\")\n", " image = input_image.resize((384, 384), Image.LANCZOS)\n", " pixel_values = TVF.pil_to_tensor(image).unsqueeze(0) / 255.0\n", " pixel_values = TVF.normalize(pixel_values, [0.5], [0.5])\n", " pixel_values = pixel_values.to('cuda')\n", " prompt = tokenizer.encode(prompt_str, return_tensors='pt', padding=False, truncation=False, add_special_tokens=False)\n", " with torch.amp.autocast_mode.autocast('cuda', enabled=True):\n", " vision_outputs = clip_model(pixel_values=pixel_values, output_hidden_states=True)\n", " image_features = vision_outputs.hidden_states\n", " embedded_images = image_adapter(image_features)\n", " embedded_images = embedded_images.to('cuda')\n", " prompt_embeds = text_model.model.embed_tokens(prompt.to('cuda'))\n", " assert prompt_embeds.shape == (1, prompt.shape[1], text_model.config.hidden_size), f\"Prompt shape is {prompt_embeds.shape}, expected {(1, prompt.shape[1], text_model.config.hidden_size)}\"\n", " embedded_bos = text_model.model.embed_tokens(torch.tensor([[tokenizer.bos_token_id]], device=text_model.device, dtype=torch.int64))\n", " eot_embed = image_adapter.get_eot_embedding().unsqueeze(0).to(dtype=text_model.dtype)\n", " inputs_embeds = torch.cat([\n", " embedded_bos.expand(embedded_images.shape[0], -1, -1),\n", " embedded_images.to(dtype=embedded_bos.dtype),\n", " prompt_embeds.expand(embedded_images.shape[0], -1, -1),\n", " eot_embed.expand(embedded_images.shape[0], -1, -1),\n", " ], dim=1)\n", " input_ids = torch.cat([\n", " torch.tensor([[tokenizer.bos_token_id]], dtype=torch.long),\n", " torch.zeros((1, embedded_images.shape[1]), dtype=torch.long),\n", " prompt,\n", " torch.tensor([[tokenizer.convert_tokens_to_ids(\"<|eot_id|>\")]], dtype=torch.long),\n", " ], dim=1).to('cuda')\n", " attention_mask = torch.ones_like(input_ids)\n", " generate_ids = text_model.generate(input_ids, inputs_embeds=inputs_embeds, attention_mask=attention_mask, max_new_tokens=300, do_sample=True, suppress_tokens=None) # Uses the default which is temp=0.6, top_p=0.9\n", " generate_ids = generate_ids[:, input_ids.shape[1]:]\n", " if generate_ids[0][-1] == tokenizer.eos_token_id or generate_ids[0][-1] == tokenizer.convert_tokens_to_ids(\"<|eot_id|>\"):\n", " generate_ids = generate_ids[:, :-1]\n", " caption = tokenizer.batch_decode(generate_ids, skip_special_tokens=False, clean_up_tokenization_spaces=False)[0]\n", " return caption.strip()" ] }, { "cell_type": "code", "execution_count": null, "metadata": { "id": "ldMOiaY7kJLp" }, "outputs": [], "source": [ "input_image = Image.open(\"/content/jmwknt.jpg\")\n", "# caption type (descriptive/training_prompt/rng-tags)\n", "# caption tone (formal/informal)\n", "# caption length (any/very short/short/medium-length/long/very long or a specific number)\n", "caption = stream_chat(input_image, \"descriptive\", \"formal\", \"any\")\n", "caption" ] } ], "metadata": { "accelerator": "GPU", "colab": { "gpuType": "T4", "provenance": [] }, "kernelspec": { "display_name": "Python 3", "name": "python3" }, "language_info": { "name": "python" } }, "nbformat": 4, "nbformat_minor": 0 }