🚀 [Add] the example: using sliding windows in YOLO

examples/notebook_inference.ipynb

@@ -25,15 +25,14 @@
    "source": [
     "CONFIG_PATH = \"../yolo/config\"\n",
     "CONFIG_NAME = \"config\"\n",
+    "MODEL = \"v9-c\"\n",
     "\n",
     "DEVICE = 'cuda:0'\n",
     "CLASS_NUM = 80\n",
-    "WEIGHT_PATH = '../weights/v9-c.pt' \n",
     "IMAGE_PATH = '../demo/images/inference/image.png'\n",
     "\n",
     "custom_logger()\n",
-    "device = torch.device(DEVICE)\n",
-    "image = Image.open(IMAGE_PATH)"
+    "device = torch.device(DEVICE)"
    ]
   },
   {
@@ -43,8 +42,8 @@
    "outputs": [],
    "source": [
     "with initialize(config_path=CONFIG_PATH, version_base=None, job_name=\"notebook_job\"):\n",
-    "    cfg: Config = compose(config_name=CONFIG_NAME, overrides=[\"task=inference\", f\"task.data.source={IMAGE_PATH}\", \"model=v9-m\"])\n",
-    "    model = create_model(cfg.model, class_num=CLASS_NUM, weight_path=WEIGHT_PATH, device = device)\n",
+    "    cfg: Config = compose(config_name=CONFIG_NAME, overrides=[\"task=inference\", f\"task.data.source={IMAGE_PATH}\", f\"model={MODEL}\"])\n",
+    "    model = create_model(cfg.model, class_num=CLASS_NUM).to(device)\n",
     "    transform = AugmentationComposer([], cfg.image_size)\n",
     "    vec2box = Vec2Box(model, cfg.image_size, device)"
    ]
@@ -55,8 +54,10 @@
    "metadata": {},
    "outputs": [],
    "source": [
-    "image, bbox = transform(image, torch.zeros(0, 5))\n",
-    "image = image.to(device)[None]"
+    "pil_image = Image.open(IMAGE_PATH)\n",
+    "image, bbox, rev_tensor = transform(pil_image)\n",
+    "image = image.to(device)[None]\n",
+    "rev_tensor = rev_tensor.to(device)"
    ]
   },
   {
@@ -67,10 +68,11 @@
    "source": [
     "with torch.no_grad():\n",
     "    predict = model(image)\n",
-    "    predict = vec2box(predict[\"Main\"])\n",
+    "    pred_class, _, pred_bbox = vec2box(predict[\"Main\"])\n",
     "\n",
-    "predict_box = bbox_nms(predict[0], predict[2], cfg.task.nms)\n",
-    "draw_bboxes(image, predict_box, idx2label=cfg.class_list)"
+    "pred_bbox = (pred_bbox / rev_tensor[0] - rev_tensor[None, None, 1:]) \n",
+    "pred_bbox = bbox_nms(pred_class, pred_bbox, cfg.task.nms)\n",
+    "draw_bboxes(pil_image, pred_bbox, idx2label=cfg.class_list)"
    ]
   },
   {

examples/notebook_smallobject.ipynb

@@ -0,0 +1,161 @@
+{
+ "cells": [
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "%load_ext autoreload\n",
+    "%autoreload 2"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "import sys\n",
+    "from pathlib import Path\n",
+    "\n",
+    "import torch\n",
+    "from hydra import compose, initialize\n",
+    "from PIL import Image \n",
+    "from einops import rearrange\n",
+    "\n",
+    "# Ensure that the necessary repository is cloned and installed. You may need to run: \n",
+    "# git clone [email protected]:WongKinYiu/YOLO.git\n",
+    "# cd YOLO \n",
+    "# pip install .\n",
+    "project_root = Path().resolve().parent\n",
+    "sys.path.append(str(project_root))\n",
+    "from yolo.config.config import NMSConfig\n",
+    "from yolo import AugmentationComposer, bbox_nms, Config, create_model, custom_logger, draw_bboxes, Vec2Box"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "CONFIG_PATH = \"../yolo/config\"\n",
+    "CONFIG_NAME = \"config\"\n",
+    "MODEL = \"v9-c\"\n",
+    "\n",
+    "DEVICE = 'cuda:0'\n",
+    "CLASS_NUM = 80\n",
+    "IMAGE_PATH = '../image.png'\n",
+    "SLIDE = 4\n",
+    "\n",
+    "custom_logger()\n",
+    "device = torch.device(DEVICE)"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "with initialize(config_path=CONFIG_PATH, version_base=None, job_name=\"notebook_job\"):\n",
+    "    cfg: Config = compose(config_name=CONFIG_NAME, overrides=[\"task=inference\", f\"task.data.source={IMAGE_PATH}\", f\"model={MODEL}\"])\n",
+    "    model = create_model(cfg.model, class_num=CLASS_NUM).to(device)\n",
+    "    transform = AugmentationComposer([], cfg.image_size)\n",
+    "    vec2box = Vec2Box(model, cfg.image_size, device)"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "pil_image = Image.open(IMAGE_PATH)\n",
+    "image, bbox, rev_tensor = transform(pil_image)\n",
+    "image = image.to(device)[None]\n",
+    "rev_tensor = rev_tensor.to(device)"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "def slide_image(image, slide = 4, device = device):\n",
+    "    up_image = torch.nn.functional.interpolate(image, scale_factor=slide)\n",
+    "    image_list = [image]\n",
+    "    shift_list = []\n",
+    "    *_, w, h = up_image.shape\n",
+    "    for x_slide in range(slide):\n",
+    "        for y_slide in range(slide):\n",
+    "            left_w, right_w = w // slide * x_slide, w // slide * (x_slide + 1)\n",
+    "            left_h, right_h = h // slide * y_slide, h // slide * (y_slide + 1)\n",
+    "            slide_image = up_image[:, :, left_w: right_w, left_h: right_h]\n",
+    "            image_list.append(slide_image)\n",
+    "            shift_list.append(torch.Tensor([left_h, left_w, left_h, left_w]))\n",
+    "    total_image = torch.concat(image_list)\n",
+    "    total_shift = torch.stack(shift_list).to(device)\n",
+    "\n",
+    "    return total_image, total_shift"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "with torch.no_grad():\n",
+    "    total_image, total_shift = slide_image(image)\n",
+    "    predict = model(total_image)\n",
+    "    pred_class, _, pred_bbox = vec2box(predict[\"Main\"])\n",
+    "pred_bbox[1:] = (pred_bbox[1: ] + total_shift[:, None]) / SLIDE\n",
+    "pred_bbox = pred_bbox.view(1, -1, 4)\n",
+    "pred_class = pred_class.view(1, -1, 80)"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "pred_bbox = (pred_bbox / rev_tensor[0] - rev_tensor[None, None, 1:]) "
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "predict_box = bbox_nms(pred_class, pred_bbox, NMSConfig(0.5, 0.5))\n",
+    "draw_bboxes(pil_image, predict_box, idx2label=cfg.class_list)"
+   ]
+  }
+ ],
+ "metadata": {
+  "kernelspec": {
+   "display_name": "yolomit",
+   "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": "3.10.14"
+  }
+ },
+ "nbformat": 4,
+ "nbformat_minor": 2
+}

yolo/model/yolo.py

@@ -119,7 +119,7 @@ class YOLO(nn.Module):
             raise ValueError(f"Unsupported layer type: {layer_type}")
 
 
-def create_model(model_cfg: ModelConfig, weight_path: Union[bool, str], class_num: int = 80) -> YOLO:
+def create_model(model_cfg: ModelConfig, weight_path: Union[bool, str] = True, class_num: int = 80) -> YOLO:
     """Constructs and returns a model from a Dictionary configuration file.
 
     Args: