136 lines
3.6 KiB
Plaintext
136 lines
3.6 KiB
Plaintext
{
|
|
"nbformat": 4,
|
|
"nbformat_minor": 0,
|
|
"metadata": {
|
|
"colab": {
|
|
"provenance": [],
|
|
"collapsed_sections": [
|
|
"0-LmeiloOQFI"
|
|
],
|
|
"gpuType": "T4"
|
|
},
|
|
"kernelspec": {
|
|
"name": "python3",
|
|
"display_name": "Python 3"
|
|
},
|
|
"language_info": {
|
|
"name": "python"
|
|
},
|
|
"accelerator": "GPU"
|
|
},
|
|
"cells": [
|
|
{
|
|
"cell_type": "markdown",
|
|
"source": [
|
|
"## Download COCO_2017 and sample 20k images (highest resolution ones)"
|
|
],
|
|
"metadata": {
|
|
"id": "sA8GqeiqlF99"
|
|
}
|
|
},
|
|
{
|
|
"cell_type": "code",
|
|
"execution_count": null,
|
|
"metadata": {
|
|
"id": "xU5fwAjdJ1m1"
|
|
},
|
|
"outputs": [],
|
|
"source": [
|
|
"%cd /content\n",
|
|
"!git clone --recursive https://github.com/verlab/accelerated_features.git\n",
|
|
"!pip install kornia==0.6.12 --no-deps\n",
|
|
"!pip install -r accelerated_features/requirements.txt --no-deps\n",
|
|
"\n",
|
|
"#Download precomputed coco-20k used in the paper.\n",
|
|
"#Please refer to https://cocodataset.org/#home for the original, complete dataset.\n",
|
|
"!gdown '1ijYsPq7dtLQSl-oEsUOGH1fAy21YLc7H'\n",
|
|
"!unzip -qq coco_20k.zip"
|
|
]
|
|
},
|
|
{
|
|
"cell_type": "markdown",
|
|
"source": [
|
|
"## If you want to build COCO_20k from scratch, try this below."
|
|
],
|
|
"metadata": {
|
|
"id": "0-LmeiloOQFI"
|
|
}
|
|
},
|
|
{
|
|
"cell_type": "code",
|
|
"source": [
|
|
"# !wget http://images.cocodataset.org/zips/train2017.zip #18 GB images\n",
|
|
"# !unzip train2017.zip\n",
|
|
"# !rm train2017.zip\n",
|
|
"\n",
|
|
"# import glob\n",
|
|
"# import random\n",
|
|
"# import cv2\n",
|
|
"# import os\n",
|
|
"# import tqdm\n",
|
|
"\n",
|
|
"# files = sorted(glob.glob('/content/train2017/*.jpg'))\n",
|
|
"\n",
|
|
"# !mkdir -p /content/coco_20k\n",
|
|
"\n",
|
|
"# random.seed(1111)\n",
|
|
"# random.shuffle(files)\n",
|
|
"\n",
|
|
"# max_imgs = 20_000\n",
|
|
"# cnt = 0\n",
|
|
"\n",
|
|
"# for f in files:\n",
|
|
"# with tqdm.tqdm(total=max_imgs) as pbar:\n",
|
|
"# im = cv2.imread(f)\n",
|
|
"# h, w = im.shape[:2]\n",
|
|
"# if w >= 600 and h >= 400:\n",
|
|
"# cv2.imwrite('/content/coco_20k/' + os.path.basename(f), im)\n",
|
|
"# cnt+=1\n",
|
|
"# pbar.update(1)\n",
|
|
"\n",
|
|
"# if cnt >= max_imgs:\n",
|
|
"# break\n"
|
|
],
|
|
"metadata": {
|
|
"id": "miB7u24KKJXh"
|
|
},
|
|
"execution_count": null,
|
|
"outputs": []
|
|
},
|
|
{
|
|
"cell_type": "markdown",
|
|
"source": [
|
|
"## Train XFeat on COCO_20k (Synthetic training)\n",
|
|
"For demonstration purposes we start to train XFeat from scratch with synthetic data only."
|
|
],
|
|
"metadata": {
|
|
"id": "VDSpQtdZlQk8"
|
|
}
|
|
},
|
|
{
|
|
"cell_type": "code",
|
|
"source": [
|
|
"# Load the TensorBoard notebook extension\n",
|
|
"%load_ext tensorboard\n",
|
|
"%tensorboard --logdir /content/ckpts"
|
|
],
|
|
"metadata": {
|
|
"id": "z3U4MZXqlPmO"
|
|
},
|
|
"execution_count": null,
|
|
"outputs": []
|
|
},
|
|
{
|
|
"cell_type": "code",
|
|
"source": [
|
|
"%cd accelerated_features\n",
|
|
"!python3 -m modules.training.train --training_type xfeat_synthetic --synthetic_root_path /content/coco_20k --ckpt_save_path /content/ckpts"
|
|
],
|
|
"metadata": {
|
|
"id": "nOP_BRppLrWm"
|
|
},
|
|
"execution_count": null,
|
|
"outputs": []
|
|
}
|
|
]
|
|
} |