notebooks/stable-diffusion-v2/stable-diffusion-v2-text-to-image.ipynb

{
 "cells": [
  {
   "attachments": {},
   "cell_type": "markdown",
   "id": "e5501f221fd5d19f",
   "metadata": {},
   "source": [
    "# Text-to-Image Generation with Stable Diffusion v2 and OpenVINO™\n",
    "\n",
    "Stable Diffusion v2 is the next generation of Stable Diffusion model a Text-to-Image latent diffusion model created by the researchers and engineers from [Stability AI](https://stability.ai/) and [LAION](https://laion.ai/). \n",
    "\n",
    "General diffusion models are machine learning systems that are trained to denoise random gaussian noise step by step, to get to a sample of interest, such as an image.\n",
    "Diffusion models have shown to achieve state-of-the-art results for generating image data. But one downside of diffusion models is that the reverse denoising process is slow. In addition, these models consume a lot of memory because they operate in pixel space, which becomes unreasonably expensive when generating high-resolution images. Therefore, it is challenging to train these models and also use them for inference. OpenVINO brings capabilities to run model inference on Intel hardware and opens the door to the fantastic world of diffusion models for everyone!\n",
    "\n",
    "In previous notebooks, we already discussed how to run [Text-to-Image generation and Image-to-Image generation using Stable Diffusion v1](../stable-diffusion-text-to-image/stable-diffusion-text-to-image.ipynb) and [controlling its generation process using ControlNet](./controlnet-stable-diffusion/controlnet-stable-diffusion.ipynb). Now is turn of Stable Diffusion v2.\n",
    "\n",
    "## Stable Diffusion v2: What’s new?\n",
    "\n",
    "The new stable diffusion model offers a bunch of new features inspired by the other models that have emerged since the introduction of the first iteration. Some of the features that can be found in the new model are:\n",
    "\n",
    "* The model comes with a new robust encoder, OpenCLIP, created by LAION and aided by Stability AI; this version v2 significantly enhances the produced photos over the V1 versions. \n",
    "* The model can now generate images in a 768x768 resolution, offering more information to be shown in the generated images.\n",
    "* The model finetuned with [v-objective](https://arxiv.org/abs/2202.00512). The v-parameterization is particularly useful for numerical stability throughout the diffusion process to enable progressive distillation for models. For models that operate at higher resolution, it is also discovered that the v-parameterization avoids color shifting artifacts that are known to affect high resolution diffusion models, and in the video setting it avoids temporal color shifting that sometimes appears with epsilon-prediction used in Stable Diffusion v1. \n",
    "* The model also comes with a new diffusion model capable of running upscaling on the images generated. Upscaled images can be adjusted up to 4 times the original image. Provided as separated model, for more details please check [stable-diffusion-x4-upscaler](https://huggingface.co/stabilityai/stable-diffusion-x4-upscaler)\n",
    "* The model comes with a new refined depth architecture capable of preserving context from prior generation layers in an image-to-image setting. This structure preservation helps generate images that preserving forms and shadow of objects, but with different content.\n",
    "* The model comes with an updated inpainting module built upon the previous model. This text-guided inpainting makes switching out parts in the image easier than before.\n",
    "\n",
    "This notebook demonstrates how to convert and run Stable Diffusion v2 model using OpenVINO.\n",
    "\n",
    "Notebook contains the following steps:\n",
    "\n",
    "1. Create PyTorch models pipeline using Diffusers library.\n",
    "2. Convert PyTorch models to OpenVINO IR format, using model conversion API.\n",
    "3. Apply hybrid post-training quantization to UNet model with [NNCF](https://github.com/openvinotoolkit/nncf/).\n",
    "4. Run Stable Diffusion v2 Text-to-Image pipeline with OpenVINO.\n",
    "\n",
    "**Note:** This is the full version of the Stable Diffusion text-to-image implementation. If you would like to get started and run the notebook quickly, check out [stable-diffusion-v2-text-to-image-demo notebook](../stable-diffusion-v2/stable-diffusion-v2-text-to-image-demo.ipynb).\n",
    "\n",
    "\n",
    "<img referrerpolicy=\"no-referrer-when-downgrade\" src=\"https://static.scarf.sh/a.png?x-pxid=5b5a4db0-7875-4bfb-bdbd-01698b5b1a77&file=notebooks/stable-diffusion-v2/stable-diffusion-v2-text-to-image.ipynb\" />\n"
   ]
  },
  {
   "attachments": {},
   "cell_type": "markdown",
   "id": "e2bceddde9f8d526",
   "metadata": {},
   "source": [
    "#### Table of contents:\n",
    "\n",
    "- [Prerequisites](#Prerequisites)\n",
    "- [Stable Diffusion v2 for Text-to-Image Generation](#Stable-Diffusion-v2-for-Text-to-Image-Generation)\n",
    "    - [Stable Diffusion in Diffusers library](#Stable-Diffusion-in-Diffusers-library)\n",
    "    - [Convert models to OpenVINO Intermediate representation (IR) format](#Convert-models-to-OpenVINO-Intermediate-representation-(IR)-format)\n",
    "    - [Text Encoder](#Text-Encoder)\n",
    "    - [U-Net](#U-Net)\n",
    "    - [VAE](#VAE)\n",
    "    - [Prepare Inference Pipeline](#Prepare-Inference-Pipeline)\n",
    "    - [Configure Inference Pipeline](#Configure-Inference-Pipeline)\n",
    "- [Quantization](#Quantization)\n",
    "    - [Prepare calibration dataset](#Prepare-calibration-dataset)\n",
    "    - [Run Hybrid Model Quantization](#Run-Hybrid-Model-Quantization)\n",
    "    - [Compare inference time of the FP16 and INT8 pipelines](#Compare-inference-time-of-the-FP16-and-INT8-pipelines)\n",
    "- [Run Text-to-Image generation](#Run-Text-to-Image-generation)\n",
    "\n",
    "\n",
    "### Installation Instructions\n",
    "\n",
    "This is a self-contained example that relies solely on its own code.\n",
    "\n",
    "We recommend  running the notebook in a virtual environment. You only need a Jupyter server to start.\n",
    "For details, please refer to [Installation Guide](https://github.com/openvinotoolkit/openvino_notebooks/blob/latest/README.md#-installation-guide)."
   ]
  },
  {
   "attachments": {},
   "cell_type": "markdown",
   "id": "1a571d16e81bf3c4",
   "metadata": {},
   "source": [
    "## Prerequisites\n",
    "[back to top ⬆️](#Table-of-contents:)\n",
    "\n",
    "install required packages"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "6723fa8e346926b7",
   "metadata": {
    "ExecuteTime": {
     "end_time": "2024-02-13T13:29:10.765014Z",
     "start_time": "2024-02-13T13:29:08.472566Z"
    }
   },
   "outputs": [],
   "source": [
    "%pip install -q \"diffusers>=0.14.0\" \"openvino>=2023.1.0\" \"datasets>=2.14.6\" \"transformers>=4.25.1\" \"gradio>=4.19\" \"torch>=2.1\" Pillow opencv-python --extra-index-url https://download.pytorch.org/whl/cpu\n",
    "%pip install -q \"nncf>=2.9.0\""
   ]
  },
  {
   "attachments": {},
   "cell_type": "markdown",
   "id": "3c4a678aa7817277",
   "metadata": {},
   "source": [
    "## Stable Diffusion v2 for Text-to-Image Generation\n",
    "[back to top ⬆️](#Table-of-contents:)\n",
    "\n",
    "To start, let's look on Text-to-Image process for Stable Diffusion v2. We will use [Stable Diffusion v2-1](https://huggingface.co/stabilityai/stable-diffusion-2-1) model for these purposes. The main difference from Stable Diffusion v2 and Stable Diffusion v2.1 is usage of more data, more training, and less restrictive filtering of the dataset, that gives promising results for selecting wide range of input text prompts. More details about model can be found in [Stability AI blog post](https://stability.ai/blog/stablediffusion2-1-release7-dec-2022) and original model [repository](https://github.com/Stability-AI/stablediffusion).\n",
    "\n",
    "### Stable Diffusion in Diffusers library\n",
    "[back to top ⬆️](#Table-of-contents:)\n",
    "To work with Stable Diffusion v2, we will use Hugging Face [Diffusers](https://github.com/huggingface/diffusers) library. To experiment with Stable Diffusion models, Diffusers exposes the [`StableDiffusionPipeline`](https://huggingface.co/docs/diffusers/using-diffusers/conditional_image_generation) similar to the [other Diffusers pipelines](https://huggingface.co/docs/diffusers/api/pipelines/overview).  The code below demonstrates how to create `StableDiffusionPipeline` using `stable-diffusion-2-1`:"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 2,
   "id": "d277beedf7e92c43",
   "metadata": {
    "ExecuteTime": {
     "end_time": "2024-02-13T13:29:13.431319Z",
     "start_time": "2024-02-13T13:29:10.766275Z"
    }
   },
   "outputs": [
    {
     "data": {
      "application/vnd.jupyter.widget-view+json": {
       "model_id": "35c151ae50214855ac10300a2e2bb495",
       "version_major": 2,
       "version_minor": 0
      },
      "text/plain": [
       "Loading pipeline components...:   0%|          | 0/6 [00:00<?, ?it/s]"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    }
   ],
   "source": [
    "from diffusers import StableDiffusionPipeline\n",
    "\n",
    "pipe = StableDiffusionPipeline.from_pretrained(\"stabilityai/stable-diffusion-2-1-base\").to(\"cpu\")\n",
    "\n",
    "# for reducing memory consumption get all components from pipeline independently\n",
    "text_encoder = pipe.text_encoder\n",
    "text_encoder.eval()\n",
    "unet = pipe.unet\n",
    "unet.eval()\n",
    "vae = pipe.vae\n",
    "vae.eval()\n",
    "\n",
    "conf = pipe.scheduler.config\n",
    "\n",
    "del pipe"
   ]
  },
  {
   "attachments": {},
   "cell_type": "markdown",
   "id": "a636f5007c3781e",
   "metadata": {},
   "source": [
    "### Convert models to OpenVINO Intermediate representation (IR) format\n",
    "[back to top ⬆️](#Table-of-contents:)\n",
    "\n",
    "Starting from 2023.0 release, OpenVINO supports PyTorch models directly via Model Conversion API. `ov.convert_model` function accepts instance of PyTorch model and example inputs for tracing and returns object of `ov.Model` class, ready to use or save on disk using `ov.save_model` function. \n",
    "\n",
    "\n",
    "The pipeline consists of three important parts:\n",
    "\n",
    "* Text Encoder to create condition to generate an image from a text prompt.\n",
    "* U-Net for step-by-step denoising latent image representation.\n",
    "* Autoencoder (VAE) for decoding latent space to image.\n",
    "\n",
    "Let us convert each part:"
   ]
  },
  {
   "attachments": {},
   "cell_type": "markdown",
   "id": "d4f14633b43a9034",
   "metadata": {},
   "source": [
    "### Text Encoder\n",
    "[back to top ⬆️](#Table-of-contents:)\n",
    "\n",
    "The text-encoder is responsible for transforming the input prompt, for example, \"a photo of an astronaut riding a horse\" into an embedding space that can be understood by the U-Net. It is usually a simple transformer-based encoder that maps a sequence of input tokens to a sequence of latent text embeddings.\n",
    "\n",
    "The input of the text encoder is tensor `input_ids`, which contains indexes of tokens from text processed by the tokenizer and padded to the maximum length accepted by the model. Model outputs are two tensors: `last_hidden_state` - hidden state from the last MultiHeadAttention layer in the model and `pooler_out` - pooled output for whole model hidden states."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 3,
   "id": "3fd5f7ff0751a146",
   "metadata": {
    "ExecuteTime": {
     "end_time": "2024-02-13T13:29:13.434571Z",
     "start_time": "2024-02-13T13:29:13.432325Z"
    }
   },
   "outputs": [],
   "source": [
    "from pathlib import Path\n",
    "\n",
    "sd2_1_model_dir = Path(\"sd2.1\")\n",
    "sd2_1_model_dir.mkdir(exist_ok=True)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 4,
   "id": "19a36c5ca10a998a",
   "metadata": {
    "ExecuteTime": {
     "end_time": "2024-02-13T13:29:20.605486Z",
     "start_time": "2024-02-13T13:29:13.435752Z"
    }
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Text encoder will be loaded from sd2.1/text_encoder.xml\n"
     ]
    }
   ],
   "source": [
    "import gc\n",
    "import torch\n",
    "import openvino as ov\n",
    "\n",
    "TEXT_ENCODER_OV_PATH = sd2_1_model_dir / \"text_encoder.xml\"\n",
    "\n",
    "\n",
    "def cleanup_torchscript_cache():\n",
    "    \"\"\"\n",
    "    Helper for removing cached model representation\n",
    "    \"\"\"\n",
    "    torch._C._jit_clear_class_registry()\n",
    "    torch.jit._recursive.concrete_type_store = torch.jit._recursive.ConcreteTypeStore()\n",
    "    torch.jit._state._clear_class_state()\n",
    "\n",
    "\n",
    "def convert_encoder(text_encoder: torch.nn.Module, ir_path: Path):\n",
    "    \"\"\"\n",
    "    Convert Text Encoder model to IR.\n",
    "    Function accepts pipeline, prepares example inputs for conversion\n",
    "    Parameters:\n",
    "        text_encoder (torch.nn.Module): text encoder PyTorch model\n",
    "        ir_path (Path): File for storing model\n",
    "    Returns:\n",
    "        None\n",
    "    \"\"\"\n",
    "    if not ir_path.exists():\n",
    "        input_ids = torch.ones((1, 77), dtype=torch.long)\n",
    "        # switch model to inference mode\n",
    "        text_encoder.eval()\n",
    "\n",
    "        # disable gradients calculation for reducing memory consumption\n",
    "        with torch.no_grad():\n",
    "            # export model\n",
    "            ov_model = ov.convert_model(\n",
    "                text_encoder,  # model instance\n",
    "                example_input=input_ids,  # example inputs for model tracing\n",
    "                input=([1, 77],),  # input shape for conversion\n",
    "            )\n",
    "            ov.save_model(ov_model, ir_path)\n",
    "            del ov_model\n",
    "            cleanup_torchscript_cache()\n",
    "        print(\"Text Encoder successfully converted to IR\")\n",
    "\n",
    "\n",
    "if not TEXT_ENCODER_OV_PATH.exists():\n",
    "    convert_encoder(text_encoder, TEXT_ENCODER_OV_PATH)\n",
    "else:\n",
    "    print(f\"Text encoder will be loaded from {TEXT_ENCODER_OV_PATH}\")\n",
    "\n",
    "del text_encoder\n",
    "gc.collect();"
   ]
  },
  {
   "attachments": {},
   "cell_type": "markdown",
   "id": "ccaba8325efa1177",
   "metadata": {},
   "source": [
    "### U-Net\n",
    "[back to top ⬆️](#Table-of-contents:)\n",
    "\n",
    "U-Net model gradually denoises latent image representation guided by text encoder hidden state.\n",
    "\n",
    "U-Net model has three inputs:\n",
    "\n",
    "* `sample` - latent image sample from previous step. Generation process has not been started yet, so you will use random noise.\n",
    "* `timestep` - current scheduler step.\n",
    "* `encoder_hidden_state` - hidden state of text encoder.\n",
    "\n",
    "Model predicts the `sample` state for the next step.\n",
    "\n",
    "Generally, U-Net model conversion process remain the same like in Stable Diffusion v1, expect small changes in input sample size.  Our model was pretrained to generate images with resolution 768x768, initial latent sample size for this case is 96x96. Besides that, for different use cases like inpainting and depth to image generation model also can accept additional image information: depth map or mask as channel-wise concatenation with initial latent sample. For converting U-Net model for such use cases required to modify number of input channels."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 5,
   "id": "dbf8ba161bad414",
   "metadata": {
    "ExecuteTime": {
     "end_time": "2024-02-13T13:29:44.429898Z",
     "start_time": "2024-02-13T13:29:20.606525Z"
    }
   },
   "outputs": [],
   "source": [
    "import numpy as np\n",
    "\n",
    "UNET_OV_PATH = sd2_1_model_dir / \"unet.xml\"\n",
    "\n",
    "\n",
    "def convert_unet(\n",
    "    unet: torch.nn.Module,\n",
    "    ir_path: Path,\n",
    "    num_channels: int = 4,\n",
    "    width: int = 64,\n",
    "    height: int = 64,\n",
    "):\n",
    "    \"\"\"\n",
    "    Convert Unet model to IR format.\n",
    "    Function accepts pipeline, prepares example inputs for conversion\n",
    "    Parameters:\n",
    "        unet (torch.nn.Module): UNet PyTorch model\n",
    "        ir_path (Path): File for storing model\n",
    "        num_channels (int, optional, 4): number of input channels\n",
    "        width (int, optional, 64): input width\n",
    "        height (int, optional, 64): input height\n",
    "    Returns:\n",
    "        None\n",
    "    \"\"\"\n",
    "    dtype_mapping = {torch.float32: ov.Type.f32, torch.float64: ov.Type.f64}\n",
    "    if not ir_path.exists():\n",
    "        # prepare inputs\n",
    "        encoder_hidden_state = torch.ones((2, 77, 1024))\n",
    "        latents_shape = (2, num_channels, width, height)\n",
    "        latents = torch.randn(latents_shape)\n",
    "        t = torch.from_numpy(np.array(1, dtype=np.float32))\n",
    "        unet.eval()\n",
    "        dummy_inputs = (latents, t, encoder_hidden_state)\n",
    "        input_info = []\n",
    "        for input_tensor in dummy_inputs:\n",
    "            shape = ov.PartialShape(tuple(input_tensor.shape))\n",
    "            element_type = dtype_mapping[input_tensor.dtype]\n",
    "            input_info.append((shape, element_type))\n",
    "\n",
    "        with torch.no_grad():\n",
    "            ov_model = ov.convert_model(unet, example_input=dummy_inputs, input=input_info)\n",
    "        ov.save_model(ov_model, ir_path)\n",
    "        del ov_model\n",
    "        cleanup_torchscript_cache()\n",
    "        print(\"U-Net successfully converted to IR\")\n",
    "\n",
    "\n",
    "if not UNET_OV_PATH.exists():\n",
    "    convert_unet(unet, UNET_OV_PATH, width=96, height=96)\n",
    "    del unet\n",
    "    gc.collect()\n",
    "else:\n",
    "    del unet\n",
    "gc.collect();"
   ]
  },
  {
   "attachments": {},
   "cell_type": "markdown",
   "id": "e52e6090fd6a87b3",
   "metadata": {},
   "source": [
    "### VAE\n",
    "[back to top ⬆️](#Table-of-contents:)\n",
    "\n",
    "The VAE model has two parts, an encoder and a decoder. The encoder is used to convert the image into a low dimensional latent representation, which will serve as the input to the U-Net model. The decoder, conversely, transforms the latent representation back into an image.\n",
    "\n",
    "During latent diffusion training, the encoder is used to get the latent representations (latents) of the images for the forward diffusion process, which applies more and more noise at each step. During inference, the denoised latents generated by the reverse diffusion process are converted back into images using the VAE decoder. When you run inference for Text-to-Image, there is no initial image as a starting point. You can skip this step and directly generate initial random noise.\n",
    "\n",
    "When running Text-to-Image pipeline, we will see that we **only need the VAE decoder**, but preserve VAE encoder conversion, it will be useful in next chapter of our tutorial. \n",
    "\n",
    "Note: This process will take a few minutes and use significant amount of RAM (recommended at least 32GB)."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 6,
   "id": "442461d2f19e8c99",
   "metadata": {
    "ExecuteTime": {
     "end_time": "2024-02-13T13:30:16.030912Z",
     "start_time": "2024-02-13T13:29:44.431075Z"
    }
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "VAE encoder will be loaded from sd2.1/vae_encoder.xml\n",
      "VAE decoder will be loaded from sd2.1/vae_decoder.xml\n"
     ]
    }
   ],
   "source": [
    "VAE_ENCODER_OV_PATH = sd2_1_model_dir / \"vae_encoder.xml\"\n",
    "\n",
    "\n",
    "def convert_vae_encoder(vae: torch.nn.Module, ir_path: Path, width: int = 512, height: int = 512):\n",
    "    \"\"\"\n",
    "    Convert VAE model to IR format.\n",
    "    VAE model, creates wrapper class for export only necessary for inference part,\n",
    "    prepares example inputs for onversion\n",
    "    Parameters:\n",
    "        vae (torch.nn.Module): VAE PyTorch model\n",
    "        ir_path (Path): File for storing model\n",
    "        width (int, optional, 512): input width\n",
    "        height (int, optional, 512): input height\n",
    "    Returns:\n",
    "        None\n",
    "    \"\"\"\n",
    "\n",
    "    class VAEEncoderWrapper(torch.nn.Module):\n",
    "        def __init__(self, vae):\n",
    "            super().__init__()\n",
    "            self.vae = vae\n",
    "\n",
    "        def forward(self, image):\n",
    "            return self.vae.encode(x=image)[\"latent_dist\"].sample()\n",
    "\n",
    "    if not ir_path.exists():\n",
    "        vae_encoder = VAEEncoderWrapper(vae)\n",
    "        vae_encoder.eval()\n",
    "        image = torch.zeros((1, 3, width, height))\n",
    "        with torch.no_grad():\n",
    "            ov_model = ov.convert_model(vae_encoder, example_input=image, input=([1, 3, width, height],))\n",
    "        ov.save_model(ov_model, ir_path)\n",
    "        del ov_model\n",
    "        cleanup_torchscript_cache()\n",
    "        print(\"VAE encoder successfully converted to IR\")\n",
    "\n",
    "\n",
    "def convert_vae_decoder(vae: torch.nn.Module, ir_path: Path, width: int = 64, height: int = 64):\n",
    "    \"\"\"\n",
    "    Convert VAE decoder model to IR format.\n",
    "    Function accepts VAE model, creates wrapper class for export only necessary for inference part,\n",
    "    prepares example inputs for conversion\n",
    "    Parameters:\n",
    "        vae (torch.nn.Module): VAE model\n",
    "        ir_path (Path): File for storing model\n",
    "        width (int, optional, 64): input width\n",
    "        height (int, optional, 64): input height\n",
    "    Returns:\n",
    "        None\n",
    "    \"\"\"\n",
    "\n",
    "    class VAEDecoderWrapper(torch.nn.Module):\n",
    "        def __init__(self, vae):\n",
    "            super().__init__()\n",
    "            self.vae = vae\n",
    "\n",
    "        def forward(self, latents):\n",
    "            return self.vae.decode(latents)\n",
    "\n",
    "    if not ir_path.exists():\n",
    "        vae_decoder = VAEDecoderWrapper(vae)\n",
    "        latents = torch.zeros((1, 4, width, height))\n",
    "\n",
    "        vae_decoder.eval()\n",
    "        with torch.no_grad():\n",
    "            ov_model = ov.convert_model(vae_decoder, example_input=latents, input=([1, 4, width, height],))\n",
    "        ov.save_model(ov_model, ir_path)\n",
    "        del ov_model\n",
    "        cleanup_torchscript_cache()\n",
    "        print(\"VAE decoder successfully converted to IR\")\n",
    "\n",
    "\n",
    "if not VAE_ENCODER_OV_PATH.exists():\n",
    "    convert_vae_encoder(vae, VAE_ENCODER_OV_PATH, 768, 768)\n",
    "else:\n",
    "    print(f\"VAE encoder will be loaded from {VAE_ENCODER_OV_PATH}\")\n",
    "\n",
    "VAE_DECODER_OV_PATH = sd2_1_model_dir / \"vae_decoder.xml\"\n",
    "\n",
    "if not VAE_DECODER_OV_PATH.exists():\n",
    "    convert_vae_decoder(vae, VAE_DECODER_OV_PATH, 96, 96)\n",
    "else:\n",
    "    print(f\"VAE decoder will be loaded from {VAE_DECODER_OV_PATH}\")\n",
    "\n",
    "del vae\n",
    "gc.collect();"
   ]
  },
  {
   "attachments": {},
   "cell_type": "markdown",
   "id": "2009916bbc6c3ec8",
   "metadata": {},
   "source": [
    "### Prepare Inference Pipeline\n",
    "[back to top ⬆️](#Table-of-contents:)\n",
    "\n",
    "Putting it all together, let us now take a closer look at how the model works in inference by illustrating the logical flow.\n",
    "\n",
    "![text2img-stable-diffusion v2](https://github.com/openvinotoolkit/openvino_notebooks/assets/22090501/ec454103-0d28-48e3-a18e-b55da3fab381)\n",
    "\n",
    "The stable diffusion model takes both a latent seed and a text prompt as input. The latent seed is then used to generate random latent image representations of size $96 \\times 96$ where as the text prompt is transformed to text embeddings of size $77 \\times 1024$ via OpenCLIP's text encoder.\n",
    "\n",
    "Next, the U-Net iteratively *denoises* the random latent image representations while being conditioned on the text embeddings. The output of the U-Net, being the noise residual, is used to compute a denoised latent image representation via a scheduler algorithm. Many different scheduler algorithms can be used for this computation, each having its pros and cons. For Stable Diffusion, it is recommended to use one of:\n",
    "\n",
    "- [PNDM scheduler](https://github.com/huggingface/diffusers/blob/main/src/diffusers/schedulers/scheduling_pndm.py)\n",
    "- [DDIM scheduler](https://github.com/huggingface/diffusers/blob/main/src/diffusers/schedulers/scheduling_ddim.py)\n",
    "- [K-LMS scheduler](https://github.com/huggingface/diffusers/blob/main/src/diffusers/schedulers/scheduling_lms_discrete.py)\n",
    "\n",
    "Theory on how the scheduler algorithm function works is out of scope for this notebook, but in short, you should remember that they compute the predicted denoised image representation from the previous noise representation and the predicted noise residual.\n",
    "For more information, it is recommended to look into [Elucidating the Design Space of Diffusion-Based Generative Models](https://arxiv.org/abs/2206.00364).\n",
    "\n",
    "\n",
    "The chart above looks very similar to Stable Diffusion V1 from [notebook](../stable-diffusion-text-to-image/stable-diffusion-text-to-image.ipynb), but there is some small difference in details:\n",
    "\n",
    "* Changed input resolution for U-Net model.\n",
    "* Changed text encoder and as the result size of its hidden state embeddings.\n",
    "* Additionally, to improve image generation quality authors introduced negative prompting. Technically, positive prompt steers the diffusion toward the images associated with it, while negative prompt steers the diffusion away from it.In other words, negative prompt declares undesired concepts for generation image, e.g. if we want to have colorful and bright image, gray scale image will be result which we want to avoid, in this case gray scale can be treated as negative prompt. The positive and negative prompt are in equal footing. You can always use one with or without the other. More explanation of how it works can be found in this [article](https://stable-diffusion-art.com/how-negative-prompt-work/). "
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 7,
   "id": "ba60aaf6be593f39",
   "metadata": {
    "ExecuteTime": {
     "end_time": "2024-02-13T13:30:16.082794Z",
     "start_time": "2024-02-13T13:30:16.032084Z"
    }
   },
   "outputs": [],
   "source": [
    "import inspect\n",
    "from typing import List, Optional, Union, Dict\n",
    "\n",
    "import PIL\n",
    "import cv2\n",
    "import torch\n",
    "\n",
    "from transformers import CLIPTokenizer\n",
    "from diffusers import DiffusionPipeline\n",
    "from diffusers.schedulers import DDIMScheduler, LMSDiscreteScheduler, PNDMScheduler\n",
    "\n",
    "\n",
    "def scale_fit_to_window(dst_width: int, dst_height: int, image_width: int, image_height: int):\n",
    "    \"\"\"\n",
    "    Preprocessing helper function for calculating image size for resize with peserving original aspect ratio\n",
    "    and fitting image to specific window size\n",
    "\n",
    "    Parameters:\n",
    "      dst_width (int): destination window width\n",
    "      dst_height (int): destination window height\n",
    "      image_width (int): source image width\n",
    "      image_height (int): source image height\n",
    "    Returns:\n",
    "      result_width (int): calculated width for resize\n",
    "      result_height (int): calculated height for resize\n",
    "    \"\"\"\n",
    "    im_scale = min(dst_height / image_height, dst_width / image_width)\n",
    "    return int(im_scale * image_width), int(im_scale * image_height)\n",
    "\n",
    "\n",
    "def preprocess(image: PIL.Image.Image):\n",
    "    \"\"\"\n",
    "    Image preprocessing function. Takes image in PIL.Image format, resizes it to keep aspect ration and fits to model input window 512x512,\n",
    "    then converts it to np.ndarray and adds padding with zeros on right or bottom side of image (depends from aspect ratio), after that\n",
    "    converts data to float32 data type and change range of values from [0, 255] to [-1, 1], finally, converts data layout from planar NHWC to NCHW.\n",
    "    The function returns preprocessed input tensor and padding size, which can be used in postprocessing.\n",
    "\n",
    "    Parameters:\n",
    "      image (PIL.Image.Image): input image\n",
    "    Returns:\n",
    "       image (np.ndarray): preprocessed image tensor\n",
    "       meta (Dict): dictionary with preprocessing metadata info\n",
    "    \"\"\"\n",
    "    src_width, src_height = image.size\n",
    "    dst_width, dst_height = scale_fit_to_window(512, 512, src_width, src_height)\n",
    "    image = np.array(image.resize((dst_width, dst_height), resample=PIL.Image.Resampling.LANCZOS))[None, :]\n",
    "    pad_width = 512 - dst_width\n",
    "    pad_height = 512 - dst_height\n",
    "    pad = ((0, 0), (0, pad_height), (0, pad_width), (0, 0))\n",
    "    image = np.pad(image, pad, mode=\"constant\")\n",
    "    image = image.astype(np.float32) / 255.0\n",
    "    image = 2.0 * image - 1.0\n",
    "    image = image.transpose(0, 3, 1, 2)\n",
    "    return image, {\"padding\": pad, \"src_width\": src_width, \"src_height\": src_height}\n",
    "\n",
    "\n",
    "class OVStableDiffusionPipeline(DiffusionPipeline):\n",
    "    def __init__(\n",
    "        self,\n",
    "        vae_decoder: ov.Model,\n",
    "        text_encoder: ov.Model,\n",
    "        tokenizer: CLIPTokenizer,\n",
    "        unet: ov.Model,\n",
    "        scheduler: Union[DDIMScheduler, PNDMScheduler, LMSDiscreteScheduler],\n",
    "        vae_encoder: ov.Model = None,\n",
    "    ):\n",
    "        \"\"\"\n",
    "        Pipeline for text-to-image generation using Stable Diffusion.\n",
    "        Parameters:\n",
    "            vae_decoder (Model):\n",
    "                Variational Auto-Encoder (VAE) Model to decode images to and from latent representations.\n",
    "            text_encoder (Model):\n",
    "                Frozen text-encoder. Stable Diffusion uses the text portion of\n",
    "                [CLIP](https://huggingface.co/docs/transformers/model_doc/clip#transformers.CLIPTextModel), specifically\n",
    "                the clip-vit-large-patch14(https://huggingface.co/openai/clip-vit-large-patch14) variant.\n",
    "            tokenizer (CLIPTokenizer):\n",
    "                Tokenizer of class CLIPTokenizer(https://huggingface.co/docs/transformers/v4.21.0/en/model_doc/clip#transformers.CLIPTokenizer).\n",
    "            unet (Model): Conditional U-Net architecture to denoise the encoded image latents.\n",
    "            vae_encoder (Model):\n",
    "                Variational Auto-Encoder (VAE) Model to encode images to latent representation.\n",
    "            scheduler (SchedulerMixin):\n",
    "                A scheduler to be used in combination with unet to denoise the encoded image latents. Can be one of\n",
    "                DDIMScheduler, LMSDiscreteScheduler, or PNDMScheduler.\n",
    "        \"\"\"\n",
    "        super().__init__()\n",
    "        self.scheduler = scheduler\n",
    "        self.vae_decoder = vae_decoder\n",
    "        self.vae_encoder = vae_encoder\n",
    "        self.text_encoder = text_encoder\n",
    "        self.unet = unet\n",
    "        self.register_to_config(unet=unet)\n",
    "        self._text_encoder_output = text_encoder.output(0)\n",
    "        self._unet_output = unet.output(0)\n",
    "        self._vae_d_output = vae_decoder.output(0)\n",
    "        self._vae_e_output = vae_encoder.output(0) if vae_encoder is not None else None\n",
    "        self.height = self.unet.input(0).shape[2] * 8\n",
    "        self.width = self.unet.input(0).shape[3] * 8\n",
    "        self.tokenizer = tokenizer\n",
    "\n",
    "    def __call__(\n",
    "        self,\n",
    "        prompt: Union[str, List[str]],\n",
    "        image: PIL.Image.Image = None,\n",
    "        negative_prompt: Union[str, List[str]] = None,\n",
    "        num_inference_steps: Optional[int] = 50,\n",
    "        guidance_scale: Optional[float] = 7.5,\n",
    "        eta: Optional[float] = 0.0,\n",
    "        output_type: Optional[str] = \"pil\",\n",
    "        seed: Optional[int] = None,\n",
    "        strength: float = 1.0,\n",
    "    ):\n",
    "        \"\"\"\n",
    "        Function invoked when calling the pipeline for generation.\n",
    "        Parameters:\n",
    "            prompt (str or List[str]):\n",
    "                The prompt or prompts to guide the image generation.\n",
    "            image (PIL.Image.Image, *optional*, None):\n",
    "                 Intinal image for generation.\n",
    "            negative_prompt (str or List[str]):\n",
    "                The negative prompt or prompts to guide the image generation.\n",
    "            num_inference_steps (int, *optional*, defaults to 50):\n",
    "                The number of denoising steps. More denoising steps usually lead to a higher quality image at the\n",
    "                expense of slower inference.\n",
    "            guidance_scale (float, *optional*, defaults to 7.5):\n",
    "                Guidance scale as defined in Classifier-Free Diffusion Guidance(https://arxiv.org/abs/2207.12598).\n",
    "                guidance_scale is defined as `w` of equation 2.\n",
    "                Higher guidance scale encourages to generate images that are closely linked to the text prompt,\n",
    "                usually at the expense of lower image quality.\n",
    "            eta (float, *optional*, defaults to 0.0):\n",
    "                Corresponds to parameter eta (η) in the DDIM paper: https://arxiv.org/abs/2010.02502. Only applies to\n",
    "                [DDIMScheduler], will be ignored for others.\n",
    "            output_type (`str`, *optional*, defaults to \"pil\"):\n",
    "                The output format of the generate image. Choose between\n",
    "                [PIL](https://pillow.readthedocs.io/en/stable/): PIL.Image.Image or np.array.\n",
    "            seed (int, *optional*, None):\n",
    "                Seed for random generator state initialization.\n",
    "            strength (int, *optional*, 1.0):\n",
    "                strength between initial image and generated in Image-to-Image pipeline, do not used in Text-to-Image\n",
    "        Returns:\n",
    "            Dictionary with keys:\n",
    "                sample - the last generated image PIL.Image.Image or np.array\n",
    "        \"\"\"\n",
    "        if seed is not None:\n",
    "            np.random.seed(seed)\n",
    "        # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2)\n",
    "        # of the Imagen paper: https://arxiv.org/pdf/2205.11487.pdf . `guidance_scale = 1`\n",
    "        # corresponds to doing no classifier free guidance.\n",
    "        do_classifier_free_guidance = guidance_scale > 1.0\n",
    "        # get prompt text embeddings\n",
    "        text_embeddings = self._encode_prompt(\n",
    "            prompt,\n",
    "            do_classifier_free_guidance=do_classifier_free_guidance,\n",
    "            negative_prompt=negative_prompt,\n",
    "        )\n",
    "        # set timesteps\n",
    "        accepts_offset = \"offset\" in set(inspect.signature(self.scheduler.set_timesteps).parameters.keys())\n",
    "        extra_set_kwargs = {}\n",
    "        if accepts_offset:\n",
    "            extra_set_kwargs[\"offset\"] = 1\n",
    "\n",
    "        self.scheduler.set_timesteps(num_inference_steps, **extra_set_kwargs)\n",
    "        timesteps, num_inference_steps = self.get_timesteps(num_inference_steps, strength)\n",
    "        latent_timestep = timesteps[:1]\n",
    "\n",
    "        # get the initial random noise unless the user supplied it\n",
    "        latents, meta = self.prepare_latents(image, latent_timestep)\n",
    "\n",
    "        # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature\n",
    "        # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers.\n",
    "        # eta corresponds to η in DDIM paper: https://arxiv.org/abs/2010.02502\n",
    "        # and should be between [0, 1]\n",
    "        accepts_eta = \"eta\" in set(inspect.signature(self.scheduler.step).parameters.keys())\n",
    "        extra_step_kwargs = {}\n",
    "        if accepts_eta:\n",
    "            extra_step_kwargs[\"eta\"] = eta\n",
    "\n",
    "        for t in self.progress_bar(timesteps):\n",
    "            # expand the latents if we are doing classifier free guidance\n",
    "            latent_model_input = np.concatenate([latents] * 2) if do_classifier_free_guidance else latents\n",
    "            latent_model_input = self.scheduler.scale_model_input(latent_model_input, t)\n",
    "\n",
    "            # predict the noise residual\n",
    "            noise_pred = self.unet([latent_model_input, np.array(t, dtype=np.float32), text_embeddings])[self._unet_output]\n",
    "            # perform guidance\n",
    "            if do_classifier_free_guidance:\n",
    "                noise_pred_uncond, noise_pred_text = noise_pred[0], noise_pred[1]\n",
    "                noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond)\n",
    "\n",
    "            # compute the previous noisy sample x_t -> x_t-1\n",
    "            latents = self.scheduler.step(torch.from_numpy(noise_pred), t, torch.from_numpy(latents), **extra_step_kwargs)[\"prev_sample\"].numpy()\n",
    "        # scale and decode the image latents with vae\n",
    "        image = self.vae_decoder(latents * (1 / 0.18215))[self._vae_d_output]\n",
    "\n",
    "        image = self.postprocess_image(image, meta, output_type)\n",
    "        return {\"sample\": image}\n",
    "\n",
    "    def _encode_prompt(\n",
    "        self,\n",
    "        prompt: Union[str, List[str]],\n",
    "        num_images_per_prompt: int = 1,\n",
    "        do_classifier_free_guidance: bool = True,\n",
    "        negative_prompt: Union[str, List[str]] = None,\n",
    "    ):\n",
    "        \"\"\"\n",
    "        Encodes the prompt into text encoder hidden states.\n",
    "\n",
    "        Parameters:\n",
    "            prompt (str or list(str)): prompt to be encoded\n",
    "            num_images_per_prompt (int): number of images that should be generated per prompt\n",
    "            do_classifier_free_guidance (bool): whether to use classifier free guidance or not\n",
    "            negative_prompt (str or list(str)): negative prompt to be encoded\n",
    "        Returns:\n",
    "            text_embeddings (np.ndarray): text encoder hidden states\n",
    "        \"\"\"\n",
    "        batch_size = len(prompt) if isinstance(prompt, list) else 1\n",
    "\n",
    "        # tokenize input prompts\n",
    "        text_inputs = self.tokenizer(\n",
    "            prompt,\n",
    "            padding=\"max_length\",\n",
    "            max_length=self.tokenizer.model_max_length,\n",
    "            truncation=True,\n",
    "            return_tensors=\"np\",\n",
    "        )\n",
    "        text_input_ids = text_inputs.input_ids\n",
    "\n",
    "        text_embeddings = self.text_encoder(text_input_ids)[self._text_encoder_output]\n",
    "\n",
    "        # duplicate text embeddings for each generation per prompt\n",
    "        if num_images_per_prompt != 1:\n",
    "            bs_embed, seq_len, _ = text_embeddings.shape\n",
    "            text_embeddings = np.tile(text_embeddings, (1, num_images_per_prompt, 1))\n",
    "            text_embeddings = np.reshape(text_embeddings, (bs_embed * num_images_per_prompt, seq_len, -1))\n",
    "\n",
    "        # get unconditional embeddings for classifier free guidance\n",
    "        if do_classifier_free_guidance:\n",
    "            uncond_tokens: List[str]\n",
    "            max_length = text_input_ids.shape[-1]\n",
    "            if negative_prompt is None:\n",
    "                uncond_tokens = [\"\"] * batch_size\n",
    "            elif isinstance(negative_prompt, str):\n",
    "                uncond_tokens = [negative_prompt]\n",
    "            else:\n",
    "                uncond_tokens = negative_prompt\n",
    "            uncond_input = self.tokenizer(\n",
    "                uncond_tokens,\n",
    "                padding=\"max_length\",\n",
    "                max_length=max_length,\n",
    "                truncation=True,\n",
    "                return_tensors=\"np\",\n",
    "            )\n",
    "\n",
    "            uncond_embeddings = self.text_encoder(uncond_input.input_ids)[self._text_encoder_output]\n",
    "\n",
    "            # duplicate unconditional embeddings for each generation per prompt, using mps friendly method\n",
    "            seq_len = uncond_embeddings.shape[1]\n",
    "            uncond_embeddings = np.tile(uncond_embeddings, (1, num_images_per_prompt, 1))\n",
    "            uncond_embeddings = np.reshape(uncond_embeddings, (batch_size * num_images_per_prompt, seq_len, -1))\n",
    "\n",
    "            # For classifier free guidance, we need to do two forward passes.\n",
    "            # Here we concatenate the unconditional and text embeddings into a single batch\n",
    "            # to avoid doing two forward passes\n",
    "            text_embeddings = np.concatenate([uncond_embeddings, text_embeddings])\n",
    "\n",
    "        return text_embeddings\n",
    "\n",
    "    def prepare_latents(self, image: PIL.Image.Image = None, latent_timestep: torch.Tensor = None):\n",
    "        \"\"\"\n",
    "        Function for getting initial latents for starting generation\n",
    "\n",
    "        Parameters:\n",
    "            image (PIL.Image.Image, *optional*, None):\n",
    "                Input image for generation, if not provided randon noise will be used as starting point\n",
    "            latent_timestep (torch.Tensor, *optional*, None):\n",
    "                Predicted by scheduler initial step for image generation, required for latent image mixing with nosie\n",
    "        Returns:\n",
    "            latents (np.ndarray):\n",
    "                Image encoded in latent space\n",
    "        \"\"\"\n",
    "        latents_shape = (1, 4, self.height // 8, self.width // 8)\n",
    "        noise = np.random.randn(*latents_shape).astype(np.float32)\n",
    "        if image is None:\n",
    "            # if we use LMSDiscreteScheduler, let's make sure latents are mulitplied by sigmas\n",
    "            if isinstance(self.scheduler, LMSDiscreteScheduler):\n",
    "                noise = noise * self.scheduler.sigmas[0].numpy()\n",
    "            return noise, {}\n",
    "        input_image, meta = preprocess(image)\n",
    "        latents = self.vae_encoder(input_image)[self._vae_e_output]\n",
    "        latents = latents * 0.18215\n",
    "        latents = self.scheduler.add_noise(torch.from_numpy(latents), torch.from_numpy(noise), latent_timestep).numpy()\n",
    "        return latents, meta\n",
    "\n",
    "    def postprocess_image(self, image: np.ndarray, meta: Dict, output_type: str = \"pil\"):\n",
    "        \"\"\"\n",
    "        Postprocessing for decoded image. Takes generated image decoded by VAE decoder, unpad it to initila image size (if required),\n",
    "        normalize and convert to [0, 255] pixels range. Optionally, convertes it from np.ndarray to PIL.Image format\n",
    "\n",
    "        Parameters:\n",
    "            image (np.ndarray):\n",
    "                Generated image\n",
    "            meta (Dict):\n",
    "                Metadata obtained on latents preparing step, can be empty\n",
    "            output_type (str, *optional*, pil):\n",
    "                Output format for result, can be pil or numpy\n",
    "        Returns:\n",
    "            image (List of np.ndarray or PIL.Image.Image):\n",
    "                Postprocessed images\n",
    "        \"\"\"\n",
    "        if \"padding\" in meta:\n",
    "            pad = meta[\"padding\"]\n",
    "            (_, end_h), (_, end_w) = pad[1:3]\n",
    "            h, w = image.shape[2:]\n",
    "            unpad_h = h - end_h\n",
    "            unpad_w = w - end_w\n",
    "            image = image[:, :, :unpad_h, :unpad_w]\n",
    "        image = np.clip(image / 2 + 0.5, 0, 1)\n",
    "        image = np.transpose(image, (0, 2, 3, 1))\n",
    "        # 9. Convert to PIL\n",
    "        if output_type == \"pil\":\n",
    "            image = self.numpy_to_pil(image)\n",
    "            if \"src_height\" in meta:\n",
    "                orig_height, orig_width = meta[\"src_height\"], meta[\"src_width\"]\n",
    "                image = [img.resize((orig_width, orig_height), PIL.Image.Resampling.LANCZOS) for img in image]\n",
    "        else:\n",
    "            if \"src_height\" in meta:\n",
    "                orig_height, orig_width = meta[\"src_height\"], meta[\"src_width\"]\n",
    "                image = [cv2.resize(img, (orig_width, orig_width)) for img in image]\n",
    "        return image\n",
    "\n",
    "    def get_timesteps(self, num_inference_steps: int, strength: float):\n",
    "        \"\"\"\n",
    "        Helper function for getting scheduler timesteps for generation\n",
    "        In case of image-to-image generation, it updates number of steps according to strength\n",
    "\n",
    "        Parameters:\n",
    "           num_inference_steps (int):\n",
    "              number of inference steps for generation\n",
    "           strength (float):\n",
    "               value between 0.0 and 1.0, that controls the amount of noise that is added to the input image.\n",
    "               Values that approach 1.0 allow for lots of variations but will also produce images that are not semantically consistent with the input.\n",
    "        \"\"\"\n",
    "        # get the original timestep using init_timestep\n",
    "        init_timestep = min(int(num_inference_steps * strength), num_inference_steps)\n",
    "\n",
    "        t_start = max(num_inference_steps - init_timestep, 0)\n",
    "        timesteps = self.scheduler.timesteps[t_start:]\n",
    "\n",
    "        return timesteps, num_inference_steps - t_start"
   ]
  },
  {
   "attachments": {},
   "cell_type": "markdown",
   "id": "8ae7b18e9f8c995a",
   "metadata": {},
   "source": [
    "### Configure Inference Pipeline\n",
    "[back to top ⬆️](#Table-of-contents:)\n",
    "\n",
    "First, you should create instances of OpenVINO Model."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 8,
   "id": "635200a7b5d84bf3",
   "metadata": {
    "ExecuteTime": {
     "end_time": "2024-02-13T13:30:16.325848Z",
     "start_time": "2024-02-13T13:30:16.083685Z"
    }
   },
   "outputs": [
    {
     "data": {
      "application/vnd.jupyter.widget-view+json": {
       "model_id": "68783cc794cb49a1b5ca674b689805d1",
       "version_major": 2,
       "version_minor": 0
      },
      "text/plain": [
       "Dropdown(description='Device:', index=4, options=('CPU', 'GPU.0', 'GPU.1', 'GPU.2', 'AUTO'), value='AUTO')"
      ]
     },
     "execution_count": 8,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "import ipywidgets as widgets\n",
    "\n",
    "core = ov.Core()\n",
    "device = widgets.Dropdown(\n",
    "    options=core.available_devices + [\"AUTO\"],\n",
    "    value=\"AUTO\",\n",
    "    description=\"Device:\",\n",
    "    disabled=False,\n",
    ")\n",
    "\n",
    "device"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 9,
   "id": "ab277c23095fa5a7",
   "metadata": {
    "ExecuteTime": {
     "end_time": "2024-02-13T13:30:19.234103Z",
     "start_time": "2024-02-13T13:30:16.326662Z"
    }
   },
   "outputs": [],
   "source": [
    "ov_config = {\"INFERENCE_PRECISION_HINT\": \"f32\"} if device.value != \"CPU\" else {}\n",
    "\n",
    "text_enc = core.compile_model(TEXT_ENCODER_OV_PATH, device.value)\n",
    "unet_model = core.compile_model(UNET_OV_PATH, device.value)\n",
    "vae_decoder = core.compile_model(VAE_DECODER_OV_PATH, device.value, ov_config)\n",
    "vae_encoder = core.compile_model(VAE_ENCODER_OV_PATH, device.value, ov_config)"
   ]
  },
  {
   "attachments": {},
   "cell_type": "markdown",
   "id": "e0cdcdb5a1c10f1a",
   "metadata": {},
   "source": [
    "Model tokenizer and scheduler are also important parts of the pipeline. Let us define them and put all components together."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 10,
   "id": "25897d22da482d91",
   "metadata": {
    "ExecuteTime": {
     "end_time": "2024-02-13T13:30:19.432106Z",
     "start_time": "2024-02-13T13:30:19.236629Z"
    }
   },
   "outputs": [],
   "source": [
    "from transformers import CLIPTokenizer\n",
    "\n",
    "scheduler = DDIMScheduler.from_config(conf)  # DDIMScheduler is used because UNet quantization produces better results with it\n",
    "tokenizer = CLIPTokenizer.from_pretrained(\"openai/clip-vit-large-patch14\")\n",
    "\n",
    "ov_pipe = OVStableDiffusionPipeline(\n",
    "    tokenizer=tokenizer,\n",
    "    text_encoder=text_enc,\n",
    "    unet=unet_model,\n",
    "    vae_encoder=vae_encoder,\n",
    "    vae_decoder=vae_decoder,\n",
    "    scheduler=scheduler,\n",
    ")"
   ]
  },
  {
   "attachments": {},
   "cell_type": "markdown",
   "id": "5a5f91dfea77c91",
   "metadata": {},
   "source": [
    "## Quantization\n",
    "[back to top ⬆️](#Table-of-contents:)\n",
    "\n",
    "[NNCF](https://github.com/openvinotoolkit/nncf/) enables post-training quantization by adding quantization layers into model graph and then using a subset of the training dataset to initialize the parameters of these additional quantization layers. Quantized operations are executed in `INT8` instead of `FP32`/`FP16` making model inference faster.\n",
    "\n",
    "According to `Stable Diffusion v2` structure, the UNet model takes up significant portion of the overall pipeline execution time. Now we will show you how to optimize the UNet part using [NNCF](https://github.com/openvinotoolkit/nncf/) to reduce computation cost and speed up the pipeline. Quantizing the rest of the pipeline does not significantly improve inference performance but can lead to a substantial degradation of accuracy.\n",
    "\n",
    "For this model we apply quantization in hybrid mode which means that we quantize: (1) weights of MatMul and Embedding layers and (2) activations of other layers. The steps are the following:\n",
    "\n",
    "1. Create a calibration dataset for quantization.\n",
    "2. Collect operations with weights.\n",
    "3. Run `nncf.compress_model()` to compress only the model weights.\n",
    "4. Run `nncf.quantize()` on the compressed model with weighted operations ignored by providing `ignored_scope` parameter.\n",
    "5. Save the `INT8` model using `openvino.save_model()` function.\n",
    "\n",
    "Please select below whether you would like to run quantization to improve model inference speed.\n",
    "\n",
    "> **NOTE**: Quantization is time and memory consuming operation. Running quantization code below may take some time."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 11,
   "id": "d5850cecdf32cb56",
   "metadata": {
    "ExecuteTime": {
     "end_time": "2024-02-13T13:30:19.437075Z",
     "start_time": "2024-02-13T13:30:19.433071Z"
    }
   },
   "outputs": [
    {
     "data": {
      "application/vnd.jupyter.widget-view+json": {
       "model_id": "a254008cba184c21844c547636b96931",
       "version_major": 2,
       "version_minor": 0
      },
      "text/plain": [
       "Checkbox(value=True, description='Quantization')"
      ]
     },
     "execution_count": 11,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "to_quantize = widgets.Checkbox(\n",
    "    value=True,\n",
    "    description=\"Quantization\",\n",
    "    disabled=False,\n",
    ")\n",
    "\n",
    "to_quantize"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 12,
   "id": "7f9846bbc13089cd",
   "metadata": {
    "ExecuteTime": {
     "end_time": "2024-02-13T13:30:19.441469Z",
     "start_time": "2024-02-13T13:30:19.437809Z"
    }
   },
   "outputs": [],
   "source": [
    "# Fetch `skip_kernel_extension` module\n",
    "import requests\n",
    "\n",
    "r = requests.get(\n",
    "    url=\"https://raw.githubusercontent.com/openvinotoolkit/openvino_notebooks/latest/utils/skip_kernel_extension.py\",\n",
    ")\n",
    "open(\"skip_kernel_extension.py\", \"w\").write(r.text)\n",
    "\n",
    "int8_ov_pipe = None\n",
    "\n",
    "%load_ext skip_kernel_extension"
   ]
  },
  {
   "attachments": {},
   "cell_type": "markdown",
   "id": "75734990a15dcf4b",
   "metadata": {},
   "source": [
    "### Prepare calibration dataset\n",
    "[back to top ⬆️](#Table-of-contents:)\n",
    "\n",
    "We use a portion of [conceptual_captions](https://huggingface.co/datasets/google-research-datasets/conceptual_captions) dataset from Hugging Face as calibration data.\n",
    "To collect intermediate model inputs for calibration we should customize `CompiledModel`."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 13,
   "id": "e277c7ebebde9973",
   "metadata": {
    "ExecuteTime": {
     "end_time": "2024-02-13T13:30:20.034818Z",
     "start_time": "2024-02-13T13:30:19.442211Z"
    }
   },
   "outputs": [],
   "source": [
    "%%skip not $to_quantize.value\n",
    "\n",
    "import datasets\n",
    "import numpy as np\n",
    "from tqdm.notebook import tqdm\n",
    "from typing import Any, Dict, List\n",
    "\n",
    "\n",
    "def disable_progress_bar(pipeline, disable=True):\n",
    "    if not hasattr(pipeline, \"_progress_bar_config\"):\n",
    "        pipeline._progress_bar_config = {'disable': disable}\n",
    "    else:\n",
    "        pipeline._progress_bar_config['disable'] = disable\n",
    "\n",
    "\n",
    "class CompiledModelDecorator(ov.CompiledModel):\n",
    "    def __init__(self, compiled_model: ov.CompiledModel, data_cache: List[Any] = None, keep_prob: float = 0.5):\n",
    "        super().__init__(compiled_model)\n",
    "        self.data_cache = data_cache if data_cache is not None else []\n",
    "        self.keep_prob = keep_prob\n",
    "\n",
    "    def __call__(self, *args, **kwargs):\n",
    "        if np.random.rand() <= self.keep_prob:\n",
    "            self.data_cache.append(*args)\n",
    "        return super().__call__(*args, **kwargs)\n",
    "\n",
    "\n",
    "def collect_calibration_data(ov_pipe, calibration_dataset_size: int, num_inference_steps: int) -> List[Dict]:\n",
    "    original_unet = ov_pipe.unet\n",
    "    calibration_data = []\n",
    "    ov_pipe.unet = CompiledModelDecorator(original_unet, calibration_data, keep_prob=0.7)\n",
    "    disable_progress_bar(ov_pipe)\n",
    "\n",
    "    dataset = datasets.load_dataset(\"google-research-datasets/conceptual_captions\", split=\"train\", trust_remote_code=True).shuffle(seed=42)\n",
    "\n",
    "    # Run inference for data collection\n",
    "    pbar = tqdm(total=calibration_dataset_size)\n",
    "    for batch in dataset:\n",
    "        prompt = batch[\"caption\"]\n",
    "        if len(prompt) > ov_pipe.tokenizer.model_max_length:\n",
    "            continue\n",
    "        ov_pipe(prompt, num_inference_steps=num_inference_steps, seed=1)\n",
    "        pbar.update(len(calibration_data) - pbar.n)\n",
    "        if pbar.n >= calibration_dataset_size:\n",
    "            break\n",
    "\n",
    "    disable_progress_bar(ov_pipe, disable=False)\n",
    "    ov_pipe.unet = original_unet\n",
    "    return calibration_data"
   ]
  },
  {
   "attachments": {},
   "cell_type": "markdown",
   "id": "1a5d9c1dcbebb1a8",
   "metadata": {},
   "source": [
    "### Run Hybrid Model Quantization\n",
    "[back to top ⬆️](#Table-of-contents:)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 14,
   "id": "586920796dacd8db",
   "metadata": {
    "ExecuteTime": {
     "end_time": "2024-02-13T14:31:36.656460Z",
     "start_time": "2024-02-13T13:32:03.144519Z"
    },
    "test_replace": {
     "calibration_dataset_size = 300": "calibration_dataset_size = 10",
     "num_inference_steps=50)": "num_inference_steps=10)"
    }
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "INFO:nncf:NNCF initialized successfully. Supported frameworks detected: torch, onnx, openvino\n"
     ]
    }
   ],
   "source": [
    "%%skip not $to_quantize.value\n",
    "\n",
    "from collections import deque\n",
    "from transformers import set_seed\n",
    "import nncf\n",
    "\n",
    "def get_operation_const_op(operation, const_port_id: int):\n",
    "    node = operation.input_value(const_port_id).get_node()\n",
    "    queue = deque([node])\n",
    "    constant_node = None\n",
    "    allowed_propagation_types_list = [\"Convert\", \"FakeQuantize\", \"Reshape\"]\n",
    "\n",
    "    while len(queue) != 0:\n",
    "        curr_node = queue.popleft()\n",
    "        if curr_node.get_type_name() == \"Constant\":\n",
    "            constant_node = curr_node\n",
    "            break\n",
    "        if len(curr_node.inputs()) == 0:\n",
    "            break\n",
    "        if curr_node.get_type_name() in allowed_propagation_types_list:\n",
    "            queue.append(curr_node.input_value(0).get_node())\n",
    "\n",
    "    return constant_node\n",
    "\n",
    "\n",
    "def is_embedding(node) -> bool:\n",
    "    allowed_types_list = [\"f16\", \"f32\", \"f64\"]\n",
    "    const_port_id = 0\n",
    "    input_tensor = node.input_value(const_port_id)\n",
    "    if input_tensor.get_element_type().get_type_name() in allowed_types_list:\n",
    "        const_node = get_operation_const_op(node, const_port_id)\n",
    "        if const_node is not None:\n",
    "            return True\n",
    "\n",
    "    return False\n",
    "\n",
    "\n",
    "def collect_ops_with_weights(model):\n",
    "    ops_with_weights = []\n",
    "    for op in model.get_ops():\n",
    "        if op.get_type_name() == \"MatMul\":\n",
    "            constant_node_0 = get_operation_const_op(op, const_port_id=0)\n",
    "            constant_node_1 = get_operation_const_op(op, const_port_id=1)\n",
    "            if constant_node_0 or constant_node_1:\n",
    "                ops_with_weights.append(op.get_friendly_name())\n",
    "        if op.get_type_name() == \"Gather\" and is_embedding(op):\n",
    "            ops_with_weights.append(op.get_friendly_name())\n",
    "\n",
    "    return ops_with_weights\n",
    "\n",
    "UNET_INT8_OV_PATH = sd2_1_model_dir / 'unet_optimized.xml'\n",
    "if not UNET_INT8_OV_PATH.exists():\n",
    "    calibration_dataset_size = 300\n",
    "    set_seed(1)\n",
    "    unet_calibration_data = collect_calibration_data(ov_pipe,\n",
    "                                                     calibration_dataset_size=calibration_dataset_size,\n",
    "                                                     num_inference_steps=50)\n",
    "\n",
    "    unet = core.read_model(UNET_OV_PATH)\n",
    "    \n",
    "    # Collect operations which weights will be compressed\n",
    "    unet_ignored_scope = collect_ops_with_weights(unet)\n",
    "    \n",
    "    # Compress model weights\n",
    "    compressed_unet = nncf.compress_weights(unet, ignored_scope=nncf.IgnoredScope(types=['Convolution']))\n",
    "    \n",
    "    # Quantize both weights and activations of Convolution layers\n",
    "    quantized_unet = nncf.quantize(\n",
    "        model=compressed_unet,\n",
    "        calibration_dataset=nncf.Dataset(unet_calibration_data),\n",
    "        subset_size=calibration_dataset_size,\n",
    "        model_type=nncf.ModelType.TRANSFORMER,\n",
    "        ignored_scope=nncf.IgnoredScope(names=unet_ignored_scope),\n",
    "        advanced_parameters=nncf.AdvancedQuantizationParameters(smooth_quant_alpha=-1)\n",
    "    )\n",
    "    \n",
    "    ov.save_model(quantized_unet, UNET_INT8_OV_PATH)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 15,
   "id": "b5d398a5a7b87506",
   "metadata": {
    "ExecuteTime": {
     "end_time": "2024-02-13T14:32:27.192039Z",
     "start_time": "2024-02-13T14:32:23.567293Z"
    }
   },
   "outputs": [],
   "source": [
    "%%skip not $to_quantize.value\n",
    "\n",
    "int8_unet_model = core.compile_model(UNET_INT8_OV_PATH, device.value)\n",
    "int8_ov_pipe = OVStableDiffusionPipeline(\n",
    "    tokenizer=tokenizer,\n",
    "    text_encoder=text_enc,\n",
    "    unet=int8_unet_model,\n",
    "    vae_encoder=vae_encoder,\n",
    "    vae_decoder=vae_decoder,\n",
    "    scheduler=scheduler\n",
    ")"
   ]
  },
  {
   "attachments": {},
   "cell_type": "markdown",
   "id": "73151859000e9305",
   "metadata": {},
   "source": [
    "### Compare UNet file size"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 16,
   "id": "4cf380af026bfaac",
   "metadata": {
    "ExecuteTime": {
     "end_time": "2024-02-13T14:32:27.198596Z",
     "start_time": "2024-02-13T14:32:27.195616Z"
    }
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "FP16 model size: 1691232.51 KB\n",
      "INT8 model size: 846918.58 KB\n",
      "Model compression rate: 1.997\n"
     ]
    }
   ],
   "source": [
    "%%skip not $to_quantize.value\n",
    "\n",
    "fp16_ir_model_size = UNET_OV_PATH.with_suffix(\".bin\").stat().st_size / 1024\n",
    "quantized_model_size = UNET_INT8_OV_PATH.with_suffix(\".bin\").stat().st_size / 1024\n",
    "\n",
    "print(f\"FP16 model size: {fp16_ir_model_size:.2f} KB\")\n",
    "print(f\"INT8 model size: {quantized_model_size:.2f} KB\")\n",
    "print(f\"Model compression rate: {fp16_ir_model_size / quantized_model_size:.3f}\")"
   ]
  },
  {
   "attachments": {},
   "cell_type": "markdown",
   "id": "637babca1053f709",
   "metadata": {},
   "source": [
    "### Compare inference time of the FP16 and INT8 pipelines\n",
    "[back to top ⬆️](#Table-of-contents:)\n",
    "\n",
    "To measure the inference performance of the `FP16` and `INT8` pipelines, we use median inference time on calibration subset.\n",
    "\n",
    "> **NOTE**: For the most accurate performance estimation, it is recommended to run `benchmark_app` in a terminal/command prompt after closing other applications."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 17,
   "id": "b653073baccacaa1",
   "metadata": {
    "ExecuteTime": {
     "end_time": "2024-02-13T14:32:29.452397Z",
     "start_time": "2024-02-13T14:32:29.449204Z"
    }
   },
   "outputs": [],
   "source": [
    "%%skip not $to_quantize.value\n",
    "\n",
    "import time\n",
    "\n",
    "def calculate_inference_time(pipeline, validation_data):\n",
    "    inference_time = []\n",
    "    pipeline.set_progress_bar_config(disable=True)\n",
    "    for prompt in validation_data:\n",
    "        start = time.perf_counter()\n",
    "        _ = pipeline(prompt, num_inference_steps=10, seed=0)\n",
    "        end = time.perf_counter()\n",
    "        delta = end - start\n",
    "        inference_time.append(delta)\n",
    "    return np.median(inference_time)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 18,
   "id": "cd270c546efe7b33",
   "metadata": {
    "ExecuteTime": {
     "end_time": "2024-02-13T14:44:53.370640Z",
     "start_time": "2024-02-13T14:32:30.218862Z"
    }
   },
   "outputs": [
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "/home/nsavel/venvs/ov_notebooks_tmp/lib/python3.8/site-packages/datasets/load.py:1429: FutureWarning: The repository for conceptual_captions contains custom code which must be executed to correctly load the dataset. You can inspect the repository content at https://hf.co/datasets/conceptual_captions\n",
      "You can avoid this message in future by passing the argument `trust_remote_code=True`.\n",
      "Passing `trust_remote_code=True` will be mandatory to load this dataset from the next major release of `datasets`.\n",
      "  warnings.warn(\n"
     ]
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Performance speed-up: 1.232\n"
     ]
    }
   ],
   "source": [
    "%%skip not $to_quantize.value\n",
    "\n",
    "validation_size = 10\n",
    "validation_dataset = datasets.load_dataset(\"google-research-datasets/conceptual_captions\", split=\"train\", streaming=True, trust_remote_code=True).take(validation_size)\n",
    "validation_data = [batch[\"caption\"] for batch in validation_dataset]\n",
    "\n",
    "fp_latency = calculate_inference_time(ov_pipe, validation_data)\n",
    "int8_latency = calculate_inference_time(int8_ov_pipe, validation_data)\n",
    "print(f\"Performance speed-up: {fp_latency / int8_latency:.3f}\")"
   ]
  },
  {
   "attachments": {},
   "cell_type": "markdown",
   "id": "3b7f93ec1bd55a2f",
   "metadata": {},
   "source": [
    "## Run Text-to-Image generation\n",
    "[back to top ⬆️](#Table-of-contents:)\n",
    "\n",
    "Now, you can define a text prompts for image generation and run inference pipeline.\n",
    "Optionally, you can also change the random generator seed for latent state initialization and number of steps.\n",
    "\n",
    "> **Note**: Consider increasing `steps` to get more precise results. A suggested value is `50`, but it will take longer time to process."
   ]
  },
  {
   "attachments": {},
   "cell_type": "markdown",
   "id": "cd89fdab77ca1f8c",
   "metadata": {},
   "source": [
    "Please select below whether you would like to use the quantized model to launch the interactive demo."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 19,
   "id": "edd734469b49f2f8",
   "metadata": {
    "ExecuteTime": {
     "end_time": "2024-02-13T14:44:53.376186Z",
     "start_time": "2024-02-13T14:44:53.371802Z"
    }
   },
   "outputs": [
    {
     "data": {
      "application/vnd.jupyter.widget-view+json": {
       "model_id": "2bb8194dc4f24c1087fc33e4efe5e5e9",
       "version_major": 2,
       "version_minor": 0
      },
      "text/plain": [
       "Checkbox(value=True, description='Use quantized model')"
      ]
     },
     "execution_count": 19,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "quantized_model_present = int8_ov_pipe is not None\n",
    "\n",
    "use_quantized_model = widgets.Checkbox(\n",
    "    value=True if quantized_model_present else False,\n",
    "    description=\"Use quantized model\",\n",
    "    disabled=not quantized_model_present,\n",
    ")\n",
    "\n",
    "use_quantized_model"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "f1136e88de27fec2",
   "metadata": {
    "ExecuteTime": {
     "end_time": "2024-02-13T14:44:55.067276Z",
     "start_time": "2024-02-13T14:44:53.376929Z"
    }
   },
   "outputs": [],
   "source": [
    "if not Path(\"gradio_helper.py\").exists():\n",
    "    r = requests.get(url=\"https://raw.githubusercontent.com/openvinotoolkit/openvino_notebooks/latest/notebooks/stable-diffusion-v2/gradio_helper.py\")\n",
    "    open(\"gradio_helper.py\", \"w\").write(r.text)\n",
    "\n",
    "from gradio_helper import make_demo\n",
    "\n",
    "pipeline = int8_ov_pipe if use_quantized_model.value else ov_pipe\n",
    "\n",
    "demo = make_demo(pipeline)\n",
    "\n",
    "try:\n",
    "    demo.queue().launch()\n",
    "except Exception:\n",
    "    demo.queue().launch(share=True)"
   ]
  }
 ],
 "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": "3.8.10"
  },
  "openvino_notebooks": {
   "imageUrl": "https://github.com/openvinotoolkit/openvino_notebooks/blob/latest/notebooks/stable-diffusion-v2/stable-diffusion-v2-optimum-demo.png?raw=true",
   "tags": {
    "categories": [
     "Model Demos",
     "AI Trends"
    ],
    "libraries": [],
    "other": [
     "Stable Diffusion"
    ],
    "tasks": [
     "Text-to-Image"
    ]
   }
  },
  "widgets": {
   "application/vnd.jupyter.widget-state+json": {
    "state": {},
    "version_major": 2,
    "version_minor": 0
   }
  }
 },
 "nbformat": 4,
 "nbformat_minor": 5
}