{ "cells": [ { "cell_type": "markdown", "metadata": { "id": "view-in-github", "colab_type": "text" }, "source": [ "\"Open" ] }, { "cell_type": "markdown", "metadata": { "id": "620o1BxdNbgq" }, "source": [ "# **Stable Diffusion 2.1**\n", "Gradio app for [Stable Diffusion 2](https://huggingface.co/stabilityai/stable-diffusion-2) by [Stability AI](https://stability.ai/) (v2-1_768-ema-pruned.ckpt).\n", "It uses [Hugging Face](https://huggingface.co/) Diffusers🧨 implementation.\n", "\n", "Currently supported pipelines are `text-to-image`, `image-to-image`, `inpainting`, `4x upscaling` and `depth-to-image`.\n", "\n", "
\n", "\n", "Colab by [anzorq](https://twitter.com/hahahahohohe). If you like it, please consider supporting me:\n", "\n", "[\"Buy](https://www.buymeacoffee.com/anzorq)\n", "
\n", "[![GitHub Repo stars](https://img.shields.io/github/stars/qunash/stable-diffusion-2-gui?style=social)](https://github.com/qunash/stable-diffusion-2-gui)\n", "\n", "![visitors](https://visitor-badge.glitch.me/badge?page_id=anzorq.sd-2-colab-header)" ] }, { "cell_type": "markdown", "metadata": { "id": "KQI4RX20DW_8" }, "source": [ "# Install dependencies (~1.5 mins)" ] }, { "cell_type": "code", "execution_count": null, "metadata": { "id": "78HoqRAB-cES", "cellView": "form" }, "outputs": [], "source": [ "!pip install --upgrade git+https://github.com/huggingface/diffusers.git\n", "# !pip install diffusers\n", "!pip install --upgrade git+https://github.com/huggingface/transformers/\n", "# !pip install transformers\n", "!pip install accelerate==0.12.0\n", "!pip install scipy\n", "!pip install ftfy\n", "!pip install gradio -q\n", "\n", "#@markdown ### ⬅️ Run this cell\n", "#@markdown ---\n", "#@markdown ### Install **xformers**?\n", "#@markdown This will take an additional ~3.5 mins.
But images will generate 25-40% faster.\n", "install_xformers = False #@param {type:\"boolean\"}\n", "\n", "if install_xformers:\n", " import os\n", " from subprocess import getoutput\n", "\n", " os.system(\"pip install --extra-index-url https://download.pytorch.org/whl/cu113 torch torchvision==0.13.1+cu113\")\n", " os.system(\"pip install triton==2.0.0.dev20220701\")\n", " gpu_info = getoutput('nvidia-smi')\n", " if(\"A10G\" in gpu_info):\n", " os.system(f\"pip install -q https://github.com/camenduru/stable-diffusion-webui-colab/releases/download/0.0.15/xformers-0.0.15.dev0+4c06c79.d20221205-cp38-cp38-linux_x86_64.whl\")\n", " elif(\"T4\" in gpu_info):\n", " os.system(f\"pip install -q https://github.com/camenduru/stable-diffusion-webui-colab/releases/download/0.0.15/xformers-0.0.15.dev0+1515f77.d20221130-cp38-cp38-linux_x86_64.whl\")\n", "\n", "\n", "# ### install xformers\n", "# from IPython.utils import capture\n", "# from subprocess import getoutput\n", "# from re import search\n", "\n", "# with capture.capture_output() as cap:\n", " \n", "# smi_out = getoutput('nvidia-smi')\n", "# supported = search('(T4|P100|V100|A100|K80)', smi_out)\n", "\n", "# if not supported:\n", "# while True:\n", "# print(\"\\x1b[1;31mThe current GPU is not supported, try starting a new session.\\x1b[0m\")\n", "# else:\n", "# supported = supported.group(0)\n", "\n", "# !pip install -q https://github.com/TheLastBen/fast-stable-diffusion/raw/main/precompiled/{supported}/xformers-0.0.13.dev0-py3-none-any.whl\n", "# !pip install -q https://github.com/ShivamShrirao/xformers-wheels/releases/download/4c06c79/xformers-0.0.15.dev0+4c06c79.d20221201-cp38-cp38-linux_x86_64.whl" ] }, { "cell_type": "markdown", "metadata": { "id": "OOPHNsFYDbc0" }, "source": [ "# Run the app" ] }, { "cell_type": "code", "execution_count": null, "metadata": { "cellView": "form", "id": "gId0-asCBVwL" }, "outputs": [], "source": [ "#@title ⬇️🖼️\n", "from diffusers import StableDiffusionPipeline, StableDiffusionImg2ImgPipeline, StableDiffusionUpscalePipeline, DiffusionPipeline, StableDiffusionDepth2ImgPipeline, DPMSolverMultistepScheduler\n", "import gradio as gr\n", "import torch\n", "from PIL import Image\n", "import random\n", "\n", "state = None\n", "current_steps = 25\n", "attn_slicing_enabled = True\n", "mem_eff_attn_enabled = install_xformers\n", "\n", "# model_id = 'stabilityai/stable-diffusion-2'\n", "model_id = 'stabilityai/stable-diffusion-2-1'\n", "\n", "scheduler = DPMSolverMultistepScheduler.from_pretrained(model_id, subfolder=\"scheduler\")\n", "\n", "pipe = StableDiffusionPipeline.from_pretrained(\n", " model_id,\n", " revision=\"fp16\" if torch.cuda.is_available() else \"fp32\",\n", " torch_dtype=torch.float16 if torch.cuda.is_available() else torch.float32,\n", " scheduler=scheduler\n", " ).to(\"cuda\")\n", "pipe.enable_attention_slicing()\n", "if mem_eff_attn_enabled:\n", " pipe.enable_xformers_memory_efficient_attention()\n", "\n", "pipe_i2i = None\n", "pipe_upscale = None\n", "pipe_inpaint = None\n", "pipe_depth2img = None\n", "\n", "\n", "modes = {\n", " 'txt2img': 'Text to Image',\n", " 'img2img': 'Image to Image',\n", " 'inpaint': 'Inpainting',\n", " 'upscale4x': 'Upscale 4x',\n", " 'depth2img': 'Depth to Image'\n", "}\n", "current_mode = modes['txt2img']\n", "\n", "def error_str(error, title=\"Error\"):\n", " return f\"\"\"#### {title}\n", " {error}\"\"\" if error else \"\"\n", "\n", "def update_state(new_state):\n", " global state\n", " state = new_state\n", "\n", "def update_state_info(old_state):\n", " if state and state != old_state:\n", " return gr.update(value=state)\n", "\n", "def set_mem_optimizations(pipe):\n", " if attn_slicing_enabled:\n", " pipe.enable_attention_slicing()\n", " else:\n", " pipe.disable_attention_slicing()\n", " \n", " if mem_eff_attn_enabled:\n", " pipe.enable_xformers_memory_efficient_attention()\n", " else:\n", " pipe.disable_xformers_memory_efficient_attention()\n", "\n", "def get_i2i_pipe(scheduler):\n", " \n", " update_state(\"Loading image to image model...\")\n", "\n", " pipe = StableDiffusionImg2ImgPipeline.from_pretrained(\n", " model_id,\n", " revision=\"fp16\" if torch.cuda.is_available() else \"fp32\",\n", " torch_dtype=torch.float16 if torch.cuda.is_available() else torch.float32,\n", " scheduler=scheduler\n", " )\n", " set_mem_optimizations(pipe)\n", " pipe.to(\"cuda\")\n", " return pipe\n", "\n", "def get_inpaint_pipe():\n", " \n", " update_state(\"Loading inpainting model...\")\n", "\n", " pipe = DiffusionPipeline.from_pretrained(\n", " \"stabilityai/stable-diffusion-2-inpainting\",\n", " revision=\"fp16\" if torch.cuda.is_available() else \"fp32\",\n", " torch_dtype=torch.float16 if torch.cuda.is_available() else torch.float32,\n", " # scheduler=scheduler # TODO currently setting scheduler here messes up the end result. A bug in Diffusers🧨\n", " ).to(\"cuda\")\n", " pipe.scheduler = DPMSolverMultistepScheduler.from_config(pipe.scheduler.config)\n", " pipe.enable_attention_slicing()\n", " pipe.enable_xformers_memory_efficient_attention()\n", " return pipe\n", "\n", "def get_upscale_pipe(scheduler):\n", " \n", " update_state(\"Loading upscale model...\")\n", "\n", " pipe = StableDiffusionUpscalePipeline.from_pretrained(\n", " \"stabilityai/stable-diffusion-x4-upscaler\",\n", " revision=\"fp16\" if torch.cuda.is_available() else \"fp32\",\n", " torch_dtype=torch.float16 if torch.cuda.is_available() else torch.float32,\n", " # scheduler=scheduler\n", " )\n", " # pipe.scheduler = DPMSolverMultistepScheduler.from_config(pipe.scheduler.config)\n", " set_mem_optimizations(pipe)\n", " pipe.to(\"cuda\")\n", " return pipe\n", " \n", "def get_depth2img_pipe():\n", " \n", " update_state(\"Loading depth to image model...\")\n", "\n", " pipe = StableDiffusionDepth2ImgPipeline.from_pretrained(\n", " \"stabilityai/stable-diffusion-2-depth\",\n", " revision=\"fp16\" if torch.cuda.is_available() else \"fp32\",\n", " torch_dtype=torch.float16 if torch.cuda.is_available() else torch.float32,\n", " # scheduler=scheduler\n", " )\n", " pipe.scheduler = DPMSolverMultistepScheduler.from_config(pipe.scheduler.config)\n", " set_mem_optimizations(pipe)\n", " pipe.to(\"cuda\")\n", " return pipe\n", "\n", "def switch_attention_slicing(attn_slicing):\n", " global attn_slicing_enabled\n", " attn_slicing_enabled = attn_slicing\n", "\n", "def switch_mem_eff_attn(mem_eff_attn):\n", " global mem_eff_attn_enabled\n", " mem_eff_attn_enabled = mem_eff_attn\n", "\n", "def pipe_callback(step: int, timestep: int, latents: torch.FloatTensor):\n", " update_state(f\"{step}/{current_steps} steps\")#\\nTime left, sec: {timestep/100:.0f}\")\n", "\n", "def inference(inf_mode, prompt, n_images, guidance, steps, width=768, height=768, seed=0, img=None, strength=0.5, neg_prompt=\"\"):\n", "\n", " update_state(\" \")\n", "\n", " global current_mode\n", " if inf_mode != current_mode:\n", " pipe.to(\"cuda\" if inf_mode == modes['txt2img'] else \"cpu\")\n", "\n", " if pipe_i2i is not None:\n", " pipe_i2i.to(\"cuda\" if inf_mode == modes['img2img'] else \"cpu\")\n", "\n", " if pipe_inpaint is not None:\n", " pipe_inpaint.to(\"cuda\" if inf_mode == modes['inpaint'] else \"cpu\")\n", "\n", " if pipe_upscale is not None:\n", " pipe_upscale.to(\"cuda\" if inf_mode == modes['upscale4x'] else \"cpu\")\n", " \n", " if pipe_depth2img is not None:\n", " pipe_depth2img.to(\"cuda\" if inf_mode == modes['depth2img'] else \"cpu\")\n", "\n", " current_mode = inf_mode\n", " \n", " if seed == 0:\n", " seed = random.randint(0, 2147483647)\n", "\n", " generator = torch.Generator('cuda').manual_seed(seed)\n", " prompt = prompt\n", "\n", " try:\n", " \n", " if inf_mode == modes['txt2img']:\n", " return txt_to_img(prompt, n_images, neg_prompt, guidance, steps, width, height, generator, seed), gr.update(visible=False, value=None)\n", " \n", " elif inf_mode == modes['img2img']:\n", " if img is None:\n", " return None, gr.update(visible=True, value=error_str(\"Image is required for Image to Image mode\"))\n", "\n", " return img_to_img(prompt, n_images, neg_prompt, img, strength, guidance, steps, width, height, generator, seed), gr.update(visible=False, value=None)\n", " \n", " elif inf_mode == modes['inpaint']:\n", " if img is None:\n", " return None, gr.update(visible=True, value=error_str(\"Image is required for Inpainting mode\"))\n", "\n", " return inpaint(prompt, n_images, neg_prompt, img, guidance, steps, width, height, generator, seed), gr.update(visible=False, value=None)\n", "\n", " elif inf_mode == modes['upscale4x']:\n", " if img is None:\n", " return None, gr.update(visible=True, value=error_str(\"Image is required for Upscale mode\"))\n", "\n", " return upscale(prompt, n_images, neg_prompt, img, guidance, steps, generator), gr.update(visible=False, value=None)\n", "\n", " elif inf_mode == modes['depth2img']:\n", " if img is None:\n", " return None, gr.update(visible=True, value=error_str(\"Image is required for Depth to Image mode\"))\n", "\n", " return depth2img(prompt, n_images, neg_prompt, img, guidance, steps, generator, seed), gr.update(visible=False, value=None)\n", "\n", " except Exception as e:\n", " return None, gr.update(visible=True, value=error_str(e))\n", "\n", "def txt_to_img(prompt, n_images, neg_prompt, guidance, steps, width, height, generator, seed):\n", "\n", " result = pipe(\n", " prompt,\n", " num_images_per_prompt = n_images,\n", " negative_prompt = neg_prompt,\n", " num_inference_steps = int(steps),\n", " guidance_scale = guidance,\n", " width = width,\n", " height = height,\n", " generator = generator,\n", " callback=pipe_callback).images\n", "\n", " update_state(f\"Done. Seed: {seed}\")\n", "\n", " return result\n", "\n", "def img_to_img(prompt, n_images, neg_prompt, img, strength, guidance, steps, width, height, generator, seed):\n", "\n", " global pipe_i2i\n", " if pipe_i2i is None:\n", " pipe_i2i = get_i2i_pipe(scheduler)\n", "\n", " img = img['image']\n", " ratio = min(height / img.height, width / img.width)\n", " img = img.resize((int(img.width * ratio), int(img.height * ratio)), Image.LANCZOS)\n", " result = pipe_i2i(\n", " prompt,\n", " num_images_per_prompt = n_images,\n", " negative_prompt = neg_prompt,\n", " image = img,\n", " num_inference_steps = int(steps),\n", " strength = strength,\n", " guidance_scale = guidance,\n", " # width = width,\n", " # height = height,\n", " generator = generator,\n", " callback=pipe_callback).images\n", "\n", " update_state(f\"Done. Seed: {seed}\")\n", " \n", " return result\n", "\n", "# TODO Currently supports only 512x512 images\n", "def inpaint(prompt, n_images, neg_prompt, img, guidance, steps, width, height, generator, seed):\n", "\n", " global pipe_inpaint\n", " if pipe_inpaint is None:\n", " pipe_inpaint = get_inpaint_pipe()\n", "\n", " inp_img = img['image']\n", " mask = img['mask']\n", " inp_img = square_padding(inp_img)\n", " mask = square_padding(mask)\n", "\n", " # # ratio = min(height / inp_img.height, width / inp_img.width)\n", " # ratio = min(512 / inp_img.height, 512 / inp_img.width)\n", " # inp_img = inp_img.resize((int(inp_img.width * ratio), int(inp_img.height * ratio)), Image.LANCZOS)\n", " # mask = mask.resize((int(mask.width * ratio), int(mask.height * ratio)), Image.LANCZOS)\n", "\n", " inp_img = inp_img.resize((512, 512))\n", " mask = mask.resize((512, 512))\n", "\n", " result = pipe_inpaint(\n", " prompt,\n", " image = inp_img,\n", " mask_image = mask,\n", " num_images_per_prompt = n_images,\n", " negative_prompt = neg_prompt,\n", " num_inference_steps = int(steps),\n", " guidance_scale = guidance,\n", " # width = width,\n", " # height = height,\n", " generator = generator,\n", " callback=pipe_callback).images\n", " \n", " update_state(f\"Done. Seed: {seed}\")\n", "\n", " return result\n", "\n", "def depth2img(prompt, n_images, neg_prompt, img, guidance, steps, generator, seed):\n", "\n", " global pipe_depth2img\n", " if pipe_depth2img is None:\n", " pipe_depth2img = get_depth2img_pipe()\n", "\n", " img = img['image']\n", " result = pipe_depth2img(\n", " prompt,\n", " num_images_per_prompt = n_images,\n", " negative_prompt = neg_prompt,\n", " image = img,\n", " num_inference_steps = int(steps),\n", " guidance_scale = guidance,\n", " # width = width,\n", " # height = height,\n", " generator = generator,\n", " callback=pipe_callback).images\n", "\n", " update_state(f\"Done. Seed: {seed}\")\n", " \n", " return result\n", "\n", "def square_padding(img):\n", " width, height = img.size\n", " if width == height:\n", " return img\n", " new_size = max(width, height)\n", " new_img = Image.new('RGB', (new_size, new_size), (0, 0, 0, 255))\n", " new_img.paste(img, ((new_size - width) // 2, (new_size - height) // 2))\n", " return new_img\n", "\n", "def upscale(prompt, n_images, neg_prompt, img, guidance, steps, generator):\n", "\n", " global pipe_upscale\n", " if pipe_upscale is None:\n", " pipe_upscale = get_upscale_pipe(scheduler)\n", "\n", " img = img['image']\n", " return upscale_tiling(prompt, neg_prompt, img, guidance, steps, generator)\n", "\n", " # result = pipe_upscale(\n", " # prompt,\n", " # image = img,\n", " # num_inference_steps = int(steps),\n", " # guidance_scale = guidance,\n", " # negative_prompt = neg_prompt,\n", " # num_images_per_prompt = n_images,\n", " # generator = generator).images[0]\n", "\n", " # return result\n", "\n", "def upscale_tiling(prompt, neg_prompt, img, guidance, steps, generator):\n", "\n", " width, height = img.size\n", "\n", " # calculate the padding needed to make the image dimensions a multiple of 128\n", " padding_x = 128 - (width % 128) if width % 128 != 0 else 0\n", " padding_y = 128 - (height % 128) if height % 128 != 0 else 0\n", "\n", " # create a white image of the right size to be used as padding\n", " padding_img = Image.new('RGB', (padding_x, padding_y), color=(255, 255, 255, 0))\n", "\n", " # paste the padding image onto the original image to add the padding\n", " img.paste(padding_img, (width, height))\n", "\n", " # update the image dimensions to include the padding\n", " width += padding_x\n", " height += padding_y\n", "\n", " if width > 128 or height > 128:\n", "\n", " num_tiles_x = int(width / 128)\n", " num_tiles_y = int(height / 128)\n", "\n", " upscaled_img = Image.new('RGB', (img.size[0] * 4, img.size[1] * 4))\n", " for x in range(num_tiles_x):\n", " for y in range(num_tiles_y):\n", " update_state(f\"Upscaling tile {x * num_tiles_y + y + 1}/{num_tiles_x * num_tiles_y}\")\n", " tile = img.crop((x * 128, y * 128, (x + 1) * 128, (y + 1) * 128))\n", "\n", " upscaled_tile = pipe_upscale(\n", " prompt=\"\",\n", " image=tile,\n", " num_inference_steps=steps,\n", " guidance_scale=guidance,\n", " # negative_prompt = neg_prompt,\n", " generator=generator,\n", " ).images[0]\n", "\n", " upscaled_img.paste(upscaled_tile, (x * upscaled_tile.size[0], y * upscaled_tile.size[1]))\n", "\n", " return [upscaled_img]\n", " else:\n", " return pipe_upscale(\n", " prompt=prompt,\n", " image=img,\n", " num_inference_steps=steps,\n", " guidance_scale=guidance,\n", " negative_prompt = neg_prompt,\n", " generator=generator,\n", " ).images\n", "\n", "\n", "\n", "def on_mode_change(mode):\n", " return gr.update(visible = mode in (modes['img2img'], modes['inpaint'], modes['upscale4x'], modes['depth2img'])), \\\n", " gr.update(visible = mode == modes['inpaint']), \\\n", " gr.update(visible = mode == modes['upscale4x']), \\\n", " gr.update(visible = mode == modes['img2img'])\n", "\n", "def on_steps_change(steps):\n", " global current_steps\n", " current_steps = steps\n", "\n", "css = \"\"\".main-div div{display:inline-flex;align-items:center;gap:.8rem;font-size:1.75rem}.main-div div h1{font-weight:900;margin-bottom:7px}.main-div p{margin-bottom:10px;font-size:94%}a{text-decoration:underline}.tabs{margin-top:0;margin-bottom:0}#gallery{min-height:20rem}\n", "\"\"\"\n", "with gr.Blocks(css=css) as demo:\n", " gr.HTML(\n", " f\"\"\"\n", "
\n", "
\n", "

Stable Diffusion 2.1

\n", "

\n", "

Model used: v2-1_768-ema-pruned.ckpt

\n", " Running on {\"GPU 🔥\" if torch.cuda.is_available() else \"CPU 🥶\"}\n", "
\n", " \"\"\"\n", " )\n", " with gr.Row():\n", " \n", " with gr.Column(scale=70):\n", " with gr.Group():\n", " with gr.Row():\n", " prompt = gr.Textbox(label=\"Prompt\", show_label=False, max_lines=2,placeholder=f\"Enter prompt\").style(container=False)\n", " generate = gr.Button(value=\"Generate\").style(rounded=(False, True, True, False))\n", "\n", " gallery = gr.Gallery(label=\"Generated images\", show_label=False).style(grid=[2], height=\"auto\")\n", " state_info = gr.Textbox(label=\"State\", show_label=False, max_lines=2).style(container=False)\n", " error_output = gr.Markdown(visible=False)\n", "\n", " with gr.Column(scale=30):\n", " inf_mode = gr.Radio(label=\"Inference Mode\", choices=list(modes.values()), value=modes['txt2img'])\n", " \n", " with gr.Group(visible=False) as i2i_options:\n", " image = gr.Image(label=\"Image\", height=128, type=\"pil\", tool='sketch')\n", " inpaint_info = gr.Markdown(\"Inpainting resizes and pads images to 512x512\", visible=False)\n", " upscale_info = gr.Markdown(\"\"\"Best for small images (128x128 or smaller).
\n", " Bigger images will be sliced into 128x128 tiles which will be upscaled individually.
\n", " This is done to avoid running out of GPU memory.\"\"\", visible=False)\n", " strength = gr.Slider(label=\"Transformation strength\", minimum=0, maximum=1, step=0.01, value=0.5)\n", "\n", " with gr.Group():\n", " neg_prompt = gr.Textbox(label=\"Negative prompt\", placeholder=\"What to exclude from the image\")\n", "\n", " n_images = gr.Slider(label=\"Number of images\", value=1, minimum=1, maximum=4, step=1)\n", " with gr.Row():\n", " guidance = gr.Slider(label=\"Guidance scale\", value=7.5, maximum=15)\n", " steps = gr.Slider(label=\"Steps\", value=current_steps, minimum=2, maximum=100, step=1)\n", "\n", " with gr.Row():\n", " width = gr.Slider(label=\"Width\", value=768, minimum=64, maximum=1024, step=8)\n", " height = gr.Slider(label=\"Height\", value=768, minimum=64, maximum=1024, step=8)\n", "\n", " seed = gr.Slider(0, 2147483647, label='Seed (0 = random)', value=0, step=1)\n", " with gr.Accordion(\"Memory optimization\"):\n", " attn_slicing = gr.Checkbox(label=\"Attention slicing (a bit slower, but uses less memory)\", value=attn_slicing_enabled)\n", " # mem_eff_attn = gr.Checkbox(label=\"Memory efficient attention (xformers)\", value=mem_eff_attn_enabled)\n", "\n", " inf_mode.change(on_mode_change, inputs=[inf_mode], outputs=[i2i_options, inpaint_info, upscale_info, strength], queue=False)\n", " steps.change(on_steps_change, inputs=[steps], outputs=[], queue=False)\n", " attn_slicing.change(lambda x: switch_attention_slicing(x), inputs=[attn_slicing], queue=False)\n", " # mem_eff_attn.change(lambda x: switch_mem_eff_attn(x), inputs=[mem_eff_attn], queue=False)\n", "\n", " inputs = [inf_mode, prompt, n_images, guidance, steps, width, height, seed, image, strength, neg_prompt]\n", " outputs = [gallery, error_output]\n", " prompt.submit(inference, inputs=inputs, outputs=outputs)\n", " generate.click(inference, inputs=inputs, outputs=outputs)\n", "\n", " demo.load(update_state_info, inputs=state_info, outputs=state_info, every=0.5, show_progress=False)\n", "\n", " gr.HTML(\"\"\"\n", "
\n", "
\n", "

Space by: \"Twitter


\n", "

Enjoying this app? Please consider supporting me

\n", " \"Buy

\n", " \"GitHub\n", "

\"visitors\"

\n", "
\n", " \"\"\")\n", "\n", "demo.queue()\n", "demo.launch(debug=True, share=True, height=768)\n" ] } ], "metadata": { "accelerator": "GPU", "colab": { "private_outputs": true, "provenance": [], "toc_visible": true, "include_colab_link": true }, "gpuClass": "standard", "kernelspec": { "display_name": "Python 3", "name": "python3" }, "language_info": { "name": "python" } }, "nbformat": 4, "nbformat_minor": 0 }