File size: 5,308 Bytes
5fc15f5 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 | {
"cells": [
{
"cell_type": "markdown",
"metadata": {},
"source": [
"# LatentRecurrentFlow (LRF) v3 — Train & Generate\n",
"\n",
"**One notebook. Run top to bottom. Produces real images.**\n",
"\n",
"Architecture:\n",
"- **TAESD** (pre-trained, 2.4M params, frozen) as the VAE\n",
"- **1.47M-param Recursive Denoising Core** (4 shared blocks × 2 recursions = 8 effective layers)\n",
"- **Rectified flow** matching with SNR weighting and CFG dropout\n",
"- **EMA** for stable sampling\n",
"\n",
"Trains on CIFAR-10 in ~60 min on CPU, ~10 min on GPU."
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {},
"outputs": [],
"source": [
"!pip install -q torch torchvision einops diffusers safetensors huggingface_hub matplotlib pillow"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {},
"outputs": [],
"source": [
"# Download the self-contained training script\n",
"import os\n",
"if not os.path.exists('lrf_v3.py'):\n",
" !wget -q https://huggingface.co/krystv/LatentRecurrentFlow/resolve/main/lrf_v3.py\n",
"from lrf_v3 import *\n",
"print(f'Device: {DEVICE}')"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"## 1. Architecture"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {},
"outputs": [],
"source": [
"notebook_cell_2_architecture()"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"## 2. Train on CIFAR-10\n",
"\n",
"This will:\n",
"1. Load TAESD (pre-trained VAE)\n",
"2. Pre-compute CIFAR-10 latents (~4 min CPU)\n",
"3. Train the denoiser (30 epochs, ~60 min CPU / ~10 min GPU)\n",
"4. Generate class-conditional samples"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {},
"outputs": [],
"source": [
"model, vae, losses = train(epochs=30, bs=64, lr=3e-4, out='./lrf_out')"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"## 3. Visualize Results"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {},
"outputs": [],
"source": [
"import matplotlib.pyplot as plt\n",
"from PIL import Image\n",
"import numpy as np\n",
"\n",
"# Loss curve\n",
"plt.figure(figsize=(10, 3))\n",
"plt.plot(losses, 'b-')\n",
"plt.xlabel('Epoch'); plt.ylabel('Loss'); plt.title('Training Loss')\n",
"plt.grid(True, alpha=0.3); plt.show()\n",
"\n",
"# VAE reconstruction\n",
"plt.figure(figsize=(8, 4))\n",
"plt.imshow(np.array(Image.open('./lrf_out/vae_check.png')))\n",
"plt.title('VAE Check (top=original, bottom=TAESD reconstruction)')\n",
"plt.axis('off'); plt.show()\n",
"\n",
"# Final generation\n",
"plt.figure(figsize=(10, 25))\n",
"plt.imshow(np.array(Image.open('./lrf_out/final.png')))\n",
"classes = ['airplane','auto','bird','cat','deer','dog','frog','horse','ship','truck']\n",
"plt.title('Class-conditional generation\\n' + ', '.join(classes))\n",
"plt.axis('off'); plt.show()"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"## 4. Generate Custom Samples"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {},
"outputs": [],
"source": [
"import torchvision\n",
"\n",
"sched = FlowScheduler()\n",
"classes = ['airplane','auto','bird','cat','deer','dog','frog','horse','ship','truck']\n",
"\n",
"# Generate 8 images of each class\n",
"for cls_id, cls_name in enumerate(classes):\n",
" imgs = gen(model, vae, sched, DEVICE, n=8, steps=50, cfg=3.0, cls_id=cls_id)\n",
" grid = torchvision.utils.make_grid((imgs+1)/2, nrow=8, padding=2)\n",
" plt.figure(figsize=(16, 2))\n",
" plt.imshow(grid.permute(1,2,0).numpy())\n",
" plt.title(f'{cls_name} (class {cls_id})')\n",
" plt.axis('off'); plt.show()"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"## 5. Load Pre-Trained Model (skip training)"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {},
"outputs": [],
"source": [
"# Load pre-trained model from HuggingFace Hub\n",
"from huggingface_hub import hf_hub_download\n",
"import torch\n",
"\n",
"model_path = hf_hub_download('krystv/LatentRecurrentFlow', 'v3/model.pt')\n",
"ckpt = torch.load(model_path, map_location='cpu', weights_only=False)\n",
"\n",
"model = LRF(ckpt['cfg'])\n",
"model.load_state_dict(ckpt['state'])\n",
"model.eval()\n",
"\n",
"vae = get_taesd(DEVICE)\n",
"sched = FlowScheduler()\n",
"\n",
"# Generate\n",
"imgs = gen(model, vae, sched, DEVICE, n=16, steps=50, cfg=3.0)\n",
"save_grid(imgs, 'quick_gen.png', 4)\n",
"\n",
"from PIL import Image\n",
"import matplotlib.pyplot as plt\n",
"plt.imshow(np.array(Image.open('quick_gen.png')))\n",
"plt.axis('off'); plt.show()"
]
}
],
"metadata": {
"kernelspec": {"display_name": "Python 3", "language": "python", "name": "python3"},
"language_info": {"name": "python", "version": "3.10.0"}
},
"nbformat": 4,
"nbformat_minor": 4
}
|