{"metadata":{"kernelspec":{"language":"python","display_name":"Python 3","name":"python3"},"language_info":{"pygments_lexer":"ipython3","nbconvert_exporter":"python","version":"3.6.4","file_extension":".py","codemirror_mode":{"name":"ipython","version":3},"name":"python","mimetype":"text/x-python"}},"nbformat_minor":4,"nbformat":4,"cells":[{"cell_type":"markdown","source":"# Train YOLOX on COTS dataset (PART 1 - TRAINING)\n\nThis notebook shows how to train custom object detection model (COTS dataset) on Kaggle. It could be good starting point for build own custom model based on YOLOX detector. Full github repository you can find here - [YOLOX](https://github.com/Megvii-BaseDetection/YOLOX)\n\n\n\n**Steps covered in this notebook:**\n* Install YOLOX \n* Prepare COTS dataset for YOLOX object detection training\n* Download Pre-Trained Weights for YOLOX\n* Prepare configuration files\n* YOLOX training\n* Run YOLOX inference on test images\n* Export YOLOX weights for Tensorflow inference (soon)\n\nNow I created notebook for learning and prototyping in YOLOX. Next step is too create better model (play with YOLOX experimentation parameters).","metadata":{"execution":{"iopub.status.busy":"2021-11-29T13:34:31.033138Z","iopub.execute_input":"2021-11-29T13:34:31.033449Z","iopub.status.idle":"2021-11-29T13:34:33.455468Z","shell.execute_reply.started":"2021-11-29T13:34:31.033368Z","shell.execute_reply":"2021-11-29T13:34:33.454141Z"}}},{"cell_type":"markdown","source":"\nI found that there is no reference custom model training YOLOX notebook on Kaggle (or I am bad in searching ... ). Since we have such an opportunity this is my contribution to this competition. Feel free to use it and enjoy!\n I really appreciate if you upvote this notebook. Thank you! \n
\n\n\n\nThis work consists of two parts: \n
\n \n
\n For YOLOX_s I use input size 960x960 but you can change it for your experiments. \n
","metadata":{}},{"cell_type":"code","source":"config_file_template = '''\n\n#!/usr/bin/env python3\n# -*- coding:utf-8 -*-\n# Copyright (c) Megvii, Inc. and its affiliates.\n\nimport os\n\nfrom yolox.exp import Exp as MyExp\n\n\nclass Exp(MyExp):\n def __init__(self):\n super(Exp, self).__init__()\n self.depth = 0.33\n self.width = 0.50\n self.exp_name = os.path.split(os.path.realpath(__file__))[1].split(\".\")[0]\n \n # Define yourself dataset path\n self.data_dir = \"/kaggle/working/dataset/images\"\n self.train_ann = \"train.json\"\n self.val_ann = \"valid.json\"\n\n self.num_classes = 1\n\n self.max_epoch = $max_epoch\n self.data_num_workers = 2\n self.eval_interval = 1\n \n self.mosaic_prob = 1.0\n self.mixup_prob = 1.0\n self.hsv_prob = 1.0\n self.flip_prob = 0.5\n self.no_aug_epochs = 2\n self.albu=True\n \n self.input_size = (960, 960)\n self.mosaic_scale = (0.5, 1.5)\n self.random_size = (10, 20)\n self.test_size = (960, 960)\n def get_dataset(self, cache: bool = False, cache_type: str = \"ram\"):\n \"\"\"\n Get dataset according to cache and cache_type parameters.\n Args:\n cache (bool): Whether to cache imgs to ram or disk.\n cache_type (str, optional): Defaults to \"ram\".\n \"ram\" : Caching imgs to ram for fast training.\n \"disk\": Caching imgs to disk for fast training.\n \"\"\"\n from yolox.data import COCODataset, TrainTransform\n print('Cocodataset called with albu as ',self.albu)\n\n return COCODataset(\n data_dir=self.data_dir,\n json_file=self.train_ann,\n img_size=self.input_size,\n preproc=TrainTransform(\n max_labels=50,\n flip_prob=self.flip_prob,\n hsv_prob=self.hsv_prob,\n albu=self.albu\n ),\n cache=cache,\n cache_type=cache_type,\n )\n def get_data_loader(self, batch_size, is_distributed, no_aug=False, \n cache_img: str = None):\n \"\"\"\n Get dataloader according to cache_img parameter.\n Args:\n no_aug (bool, optional): Whether to turn off mosaic data enhancement. Defaults to False.\n cache_img (str, optional): cache_img is equivalent to cache_type. Defaults to None.\n \"ram\" : Caching imgs to ram for fast training.\n \"disk\": Caching imgs to disk for fast training.\n None: Do not use cache, in this case cache_data is also None.\n \"\"\"\n from yolox.data import (\n TrainTransform,\n YoloBatchSampler,\n DataLoader,\n InfiniteSampler,\n MosaicDetection,\n worker_init_reset_seed,\n )\n from yolox.utils import wait_for_the_master\n\n # if cache is True, we will create self.dataset before launch\n # else we will create self.dataset after launch\n if self.dataset is None:\n print('build dataset')\n with wait_for_the_master():\n assert cache_img is None, \\\n \"cache_img must be None if you didn't create self.dataset before launch\"\n self.dataset = self.get_dataset(cache=False, cache_type=cache_img)\n print('new data loader')\n self.dataset = MosaicDetection(\n dataset=self.dataset,\n mosaic=not no_aug,\n img_size=self.input_size,\n preproc=TrainTransform(\n max_labels=120,\n flip_prob=self.flip_prob,\n hsv_prob=self.hsv_prob,\n albu=self.albu),\n degrees=self.degrees,\n translate=self.translate,\n mosaic_scale=self.mosaic_scale,\n mixup_scale=self.mixup_scale,\n shear=self.shear,\n enable_mixup=self.enable_mixup,\n mosaic_prob=self.mosaic_prob,\n mixup_prob=self.mixup_prob,\n )\n\n if is_distributed:\n batch_size = batch_size // dist.get_world_size()\n\n sampler = InfiniteSampler(len(self.dataset), seed=self.seed if self.seed else 0)\n\n batch_sampler = YoloBatchSampler(\n sampler=sampler,\n batch_size=batch_size,\n drop_last=False,\n mosaic=not no_aug,\n )\n\n dataloader_kwargs = {\"num_workers\": self.data_num_workers, \"pin_memory\": True}\n dataloader_kwargs[\"batch_sampler\"] = batch_sampler\n\n # Make sure each process has different random seed, especially for 'fork' method.\n # Check https://github.com/pytorch/pytorch/issues/63311 for more details.\n dataloader_kwargs[\"worker_init_fn\"] = worker_init_reset_seed\n\n train_loader = DataLoader(self.dataset, **dataloader_kwargs)\n\n return train_loader\n\n'''","metadata":{"execution":{"iopub.status.busy":"2023-07-05T16:41:50.801415Z","iopub.execute_input":"2023-07-05T16:41:50.802005Z","iopub.status.idle":"2023-07-05T16:41:50.811964Z","shell.execute_reply.started":"2023-07-05T16:41:50.801966Z","shell.execute_reply":"2023-07-05T16:41:50.811259Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"Data augment.y","metadata":{}},{"cell_type":"code","source":"%%writefile /kaggle/working/YOLOX/yolox/data/data_augment.py\nimport albumentations as A\nclass Albumentations:\n # YOLOv5 Albumentations class (optional, only used if package is installed)\n def __init__(self, size=640):\n self.transform = None\n print('initi Albu')\n #prefix = colorstr('albumentations: ')\n try:\n import albumentations as A\n #check_version(A.__version__, '1.0.3', hard=True) # version requirement\n\n T = [\n \n A.OneOf([\n #A.HueSaturationValue(hue_shift_limit=(0.1,0.2), sat_shift_limit= (0.1,0.2), \n # val_shift_limit=(0.1,0.2), p=0.8), #0.9\n A.ColorJitter(brightness=0.2, contrast=0.2, saturation=0.2, hue=0.2, p=1),\n A.RandomBrightnessContrast(brightness_limit=(0.1,0.2), \n contrast_limit=(0.1,0.3), p=0.8), #0.9\n A.RandomGamma(gamma_limit=(30, 150),p=0.2) ],p=0.6),\n A.Cutout(max_h_size=int(25), max_w_size=int(50), num_holes=2, p=0.5,fill_value=256) ] # transforms\n #self.transform = A.Compose(T, \n # bbox_params=A.BboxParams(format='coco', \n # label_fields=['class_labels']))\n self.transform = A.Compose(T)\n #LOGGER.info(prefix + ', '.join(f'{x}'.replace('always_apply=False, ', '') for x in T if x.p))\n except ImportError: # package not installed, skip\n pass\n except Exception as e:\n print('Error message in Albu ',e)\n #LOGGER.info(f'{prefix}{e}')\n\n def __call__(self, im, boxes=None,label=None, p=1.0):\n if self.transform and random.random() < p:\n #print('im shape',im.shape, boxes)\n #new = self.transform(image=im, bboxes=boxes, \n # class_labels=label) # transformed\n new = self.transform(image=im)\n #im, labels = new['image'], np.array([[c, *b] for c, b in zip(new['class_labels'], new['bboxes'])])\n im=new['image']\n return im#, labels\n \n #!/usr/bin/env python3\n# -*- coding:utf-8 -*-\n# Copyright (c) Megvii, Inc. and its affiliates.\n\"\"\"\nData augmentation functionality. Passed as callable transformations to\nDataset classes.\n\nThe data augmentation procedures were interpreted from @weiliu89's SSD paper\nhttp://arxiv.org/abs/1512.02325\n\"\"\"\n\nimport math\nimport random\n\nimport cv2\nimport numpy as np\n\nfrom yolox.utils import xyxy2cxcywh\n\n\ndef augment_hsv(img, hgain=5, sgain=30, vgain=30):\n hsv_augs = np.random.uniform(-1, 1, 3) * [hgain, sgain, vgain] # random gains\n hsv_augs *= np.random.randint(0, 2, 3) # random selection of h, s, v\n hsv_augs = hsv_augs.astype(np.int16)\n img_hsv = cv2.cvtColor(img, cv2.COLOR_BGR2HSV).astype(np.int16)\n\n img_hsv[..., 0] = (img_hsv[..., 0] + hsv_augs[0]) % 180\n img_hsv[..., 1] = np.clip(img_hsv[..., 1] + hsv_augs[1], 0, 255)\n img_hsv[..., 2] = np.clip(img_hsv[..., 2] + hsv_augs[2], 0, 255)\n\n cv2.cvtColor(img_hsv.astype(img.dtype), cv2.COLOR_HSV2BGR, dst=img) # no return needed\n\n\ndef get_aug_params(value, center=0):\n if isinstance(value, float):\n return random.uniform(center - value, center + value)\n elif len(value) == 2:\n return random.uniform(value[0], value[1])\n else:\n raise ValueError(\n \"Affine params should be either a sequence containing two values\\\n or single float values. Got {}\".format(value)\n )\n\n\ndef get_affine_matrix(\n target_size,\n degrees=10,\n translate=0.1,\n scales=0.1,\n shear=10,\n):\n twidth, theight = target_size\n\n # Rotation and Scale\n angle = get_aug_params(degrees)\n scale = get_aug_params(scales, center=1.0)\n\n if scale <= 0.0:\n raise ValueError(\"Argument scale should be positive\")\n\n R = cv2.getRotationMatrix2D(angle=angle, center=(0, 0), scale=scale)\n\n M = np.ones([2, 3])\n # Shear\n shear_x = math.tan(get_aug_params(shear) * math.pi / 180)\n shear_y = math.tan(get_aug_params(shear) * math.pi / 180)\n\n M[0] = R[0] + shear_y * R[1]\n M[1] = R[1] + shear_x * R[0]\n\n # Translation\n translation_x = get_aug_params(translate) * twidth # x translation (pixels)\n translation_y = get_aug_params(translate) * theight # y translation (pixels)\n\n M[0, 2] = translation_x\n M[1, 2] = translation_y\n\n return M, scale\n\n\ndef apply_affine_to_bboxes(targets, target_size, M, scale):\n num_gts = len(targets)\n\n # warp corner points\n twidth, theight = target_size\n corner_points = np.ones((4 * num_gts, 3))\n corner_points[:, :2] = targets[:, [0, 1, 2, 3, 0, 3, 2, 1]].reshape(\n 4 * num_gts, 2\n ) # x1y1, x2y2, x1y2, x2y1\n corner_points = corner_points @ M.T # apply affine transform\n corner_points = corner_points.reshape(num_gts, 8)\n\n # create new boxes\n corner_xs = corner_points[:, 0::2]\n corner_ys = corner_points[:, 1::2]\n new_bboxes = (\n np.concatenate(\n (corner_xs.min(1), corner_ys.min(1), corner_xs.max(1), corner_ys.max(1))\n )\n .reshape(4, num_gts)\n .T\n )\n\n # clip boxes\n new_bboxes[:, 0::2] = new_bboxes[:, 0::2].clip(0, twidth)\n new_bboxes[:, 1::2] = new_bboxes[:, 1::2].clip(0, theight)\n\n targets[:, :4] = new_bboxes\n\n return targets\n\n\ndef random_affine(\n img,\n targets=(),\n target_size=(640, 640),\n degrees=10,\n translate=0.1,\n scales=0.1,\n shear=10,\n):\n M, scale = get_affine_matrix(target_size, degrees, translate, scales, shear)\n\n img = cv2.warpAffine(img, M, dsize=target_size, borderValue=(114, 114, 114))\n\n # Transform label coordinates\n if len(targets) > 0:\n targets = apply_affine_to_bboxes(targets, target_size, M, scale)\n\n return img, targets\n\n\ndef _mirror(image, boxes, prob=0.5):\n _, width, _ = image.shape\n if random.random() < prob:\n image = image[:, ::-1]\n boxes[:, 0::2] = width - boxes[:, 2::-2]\n return image, boxes\n\n\ndef preproc(img, input_size, swap=(2, 0, 1)):\n if len(img.shape) == 3:\n padded_img = np.ones((input_size[0], input_size[1], 3), dtype=np.uint8) * 114\n else:\n padded_img = np.ones(input_size, dtype=np.uint8) * 114\n\n r = min(input_size[0] / img.shape[0], input_size[1] / img.shape[1])\n resized_img = cv2.resize(\n img,\n (int(img.shape[1] * r), int(img.shape[0] * r)),\n interpolation=cv2.INTER_LINEAR,\n ).astype(np.uint8)\n padded_img[: int(img.shape[0] * r), : int(img.shape[1] * r)] = resized_img\n\n padded_img = padded_img.transpose(swap)\n padded_img = np.ascontiguousarray(padded_img, dtype=np.float32)\n return padded_img, r\n\n\nclass TrainTransform:\n def __init__(self, max_labels=50, flip_prob=0.5, hsv_prob=1.0,albu=False):\n self.max_labels = max_labels\n self.flip_prob = flip_prob\n self.hsv_prob = hsv_prob\n self.albu=albu\n \n self.albumentations = Albumentations() if albu else None\n print('self_albumentations',self.albumentations)\n\n def __call__(self, image, targets, input_dim):\n boxes = targets[:, :4].copy()\n labels = targets[:, 4].copy()\n if len(boxes) == 0:\n targets = np.zeros((self.max_labels, 5), dtype=np.float32)\n image, r_o = preproc(image, input_dim)\n return image, targets\n \n image_o = image.copy()\n targets_o = targets.copy()\n height_o, width_o, _ = image_o.shape\n boxes_o = targets_o[:, :4]\n labels_o = targets_o[:, 4]\n # bbox_o: [xyxy] to [c_x,c_y,w,h]\n boxes_o = xyxy2cxcywh(boxes_o)\n\n if random.random() < self.hsv_prob:\n augment_hsv(image)\n #print('self.albu',self.albu)\n if self.albu:\n #print('albu called',image.shape)\n #print('boxes',boxes)\n #boxes=[int(b) for b in box for box in boxes] \n box_int=[]\n '''\n for b in boxes:\n b[0]=int(b[0])\n b[1]=int(b[1])\n b[2]=min(int(b[2]),image.shape[1])\n b[3]=min(int(b[3]),image.shape[0])\n box_int.append(b)\n ''' \n #image, boxes = self.albumentations(image, boxes,labels)\n image = self.albumentations(image )\n image_t, boxes = _mirror(image, boxes, self.flip_prob)\n height, width, _ = image_t.shape\n image_t, r_ = preproc(image_t, input_dim)\n # boxes [xyxy] 2 [cx,cy,w,h]\n boxes = xyxy2cxcywh(boxes)\n boxes *= r_\n\n mask_b = np.minimum(boxes[:, 2], boxes[:, 3]) > 1\n boxes_t = boxes[mask_b]\n labels_t = labels[mask_b]\n\n if len(boxes_t) == 0:\n image_t, r_o = preproc(image_o, input_dim)\n boxes_o *= r_o\n boxes_t = boxes_o\n labels_t = labels_o\n\n labels_t = np.expand_dims(labels_t, 1)\n\n targets_t = np.hstack((labels_t, boxes_t))\n padded_labels = np.zeros((self.max_labels, 5))\n padded_labels[range(len(targets_t))[: self.max_labels]] = targets_t[\n : self.max_labels\n ]\n padded_labels = np.ascontiguousarray(padded_labels, dtype=np.float32)\n return image_t, padded_labels\n\n\nclass ValTransform:\n \"\"\"\n Defines the transformations that should be applied to test PIL image\n for input into the network\n\n dimension -> tensorize -> color adj\n\n Arguments:\n resize (int): input dimension to SSD\n rgb_means ((int,int,int)): average RGB of the dataset\n (104,117,123)\n swap ((int,int,int)): final order of channels\n\n Returns:\n transform (transform) : callable transform to be applied to test/val\n data\n \"\"\"\n\n def __init__(self, swap=(2, 0, 1), legacy=False):\n self.swap = swap\n self.legacy = legacy\n\n # assume input is cv2 img for now\n def __call__(self, img, res, input_size):\n img, _ = preproc(img, input_size, self.swap)\n if self.legacy:\n img = img[::-1, :, :].copy()\n img /= 255.0\n img -= np.array([0.485, 0.456, 0.406]).reshape(3, 1, 1)\n img /= np.array([0.229, 0.224, 0.225]).reshape(3, 1, 1)\n return img, np.zeros((1, 5))","metadata":{"execution":{"iopub.status.busy":"2023-07-05T16:51:06.343525Z","iopub.execute_input":"2023-07-05T16:51:06.343847Z","iopub.status.idle":"2023-07-05T16:51:06.358537Z","shell.execute_reply.started":"2023-07-05T16:51:06.343806Z","shell.execute_reply":"2023-07-05T16:51:06.357766Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"## 3B. YOLOX-NANO CONFIG FILE\n\n For YOLOX_nano I use input size 460x460 but you can change it for your experiments. \n\n I trained model for 20 EPOCHS only .... This is for DEMO purposes only. \n
","metadata":{}},{"cell_type":"code","source":"PIPELINE_CONFIG_PATH='cots_config.py'\n\npipeline = Template(config_file_template).substitute(max_epoch = 20)\n\nwith open(PIPELINE_CONFIG_PATH, 'w') as f:\n f.write(pipeline)","metadata":{"execution":{"iopub.status.busy":"2023-07-05T16:51:15.504121Z","iopub.execute_input":"2023-07-05T16:51:15.504789Z","iopub.status.idle":"2023-07-05T16:51:15.510394Z","shell.execute_reply.started":"2023-07-05T16:51:15.504719Z","shell.execute_reply":"2023-07-05T16:51:15.509210Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"# ./yolox/data/datasets/voc_classes.py\n\nvoc_cls = '''\nVOC_CLASSES = (\n \"starfish\",\n)\n'''\nwith open('./yolox/data/datasets/voc_classes.py', 'w') as f:\n f.write(voc_cls)\n\n# ./yolox/data/datasets/coco_classes.py\n\ncoco_cls = '''\nCOCO_CLASSES = (\n \"starfish\",\n)\n'''\nwith open('./yolox/data/datasets/coco_classes.py', 'w') as f:\n f.write(coco_cls)\n\n# check if everything is ok \n!more ./yolox/data/datasets/coco_classes.py","metadata":{"execution":{"iopub.status.busy":"2023-07-05T16:41:50.858063Z","iopub.execute_input":"2023-07-05T16:41:50.858281Z","iopub.status.idle":"2023-07-05T16:41:51.814895Z","shell.execute_reply.started":"2023-07-05T16:41:50.858255Z","shell.execute_reply":"2023-07-05T16:41:51.813962Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"# 4. DOWNLOAD PRETRAINED WEIGHTS","metadata":{}},{"cell_type":"markdown","source":"List of pretrained models:\n* YOLOX-s\n* YOLOX-m\n* YOLOX-nano for inference speed (!)\n* etc.","metadata":{}},{"cell_type":"code","source":"sh = 'wget https://github.com/Megvii-BaseDetection/storage/releases/download/0.0.1/yolox_s.pth'\nMODEL_FILE = 'yolox_s.pth'\n\nif NANO:\n sh = '''\n wget https://github.com/Megvii-BaseDetection/storage/releases/download/0.0.1/yolox_nano.pth\n '''\n MODEL_FILE = 'yolox_nano.pth'\n\nwith open('script.sh', 'w') as file:\n file.write(sh)\n\n!bash script.sh","metadata":{"execution":{"iopub.status.busy":"2023-07-05T16:41:51.818643Z","iopub.execute_input":"2023-07-05T16:41:51.818921Z","iopub.status.idle":"2023-07-05T16:41:57.926161Z","shell.execute_reply.started":"2023-07-05T16:41:51.818888Z","shell.execute_reply":"2023-07-05T16:41:57.925142Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"# 5. TRAIN MODEL","metadata":{}},{"cell_type":"code","source":"!cp ./tools/train.py ./","metadata":{"execution":{"iopub.status.busy":"2023-07-05T16:41:57.928981Z","iopub.execute_input":"2023-07-05T16:41:57.929929Z","iopub.status.idle":"2023-07-05T16:41:58.933510Z","shell.execute_reply.started":"2023-07-05T16:41:57.929883Z","shell.execute_reply":"2023-07-05T16:41:58.932474Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"!pip install --upgrade \"protobuf<=3.20.1\"","metadata":{},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"!python train.py \\\n -f cots_config.py \\\n -d 1 \\\n -b 32 \\\n --fp16 \\\n -o \\\n -c {MODEL_FILE} # Remember to chenge this line if you take different model eg. yolo_nano.pth, yolox_s.pth or yolox_m.pth","metadata":{"execution":{"iopub.status.busy":"2023-07-05T16:51:18.747300Z","iopub.execute_input":"2023-07-05T16:51:18.747591Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"!cat #/opt/conda/lib/python3.7/site-packages/albumentations/augmentations/bbox_utils.py","metadata":{"execution":{"iopub.status.busy":"2023-07-05T16:42:02.287321Z","iopub.execute_input":"2023-07-05T16:42:02.287621Z","iopub.status.idle":"2023-07-05T16:42:03.244451Z","shell.execute_reply.started":"2023-07-05T16:42:02.287578Z","shell.execute_reply":"2023-07-05T16:42:03.243307Z"},"collapsed":true,"jupyter":{"outputs_hidden":true},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"","metadata":{"execution":{"iopub.status.busy":"2023-07-05T16:44:57.178988Z","iopub.execute_input":"2023-07-05T16:44:57.179317Z","iopub.status.idle":"2023-07-05T16:45:07.428277Z","shell.execute_reply.started":"2023-07-05T16:44:57.179283Z","shell.execute_reply":"2023-07-05T16:45:07.427320Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"# 6. RUN INFERENCE\n\n## 6A. INFERENCE USING YOLOX TOOL","metadata":{}},{"cell_type":"code","source":"# I have to fix demo.py file because it:\n# - raises error in Kaggle (cvWaitKey does not work) \n# - saves result files in time named directory eg. /2021_11_29_22_51_08/ which is difficult then to automatically show results\n\n%cp ../../input/yolox-kaggle-fix-for-demo-inference/demo.py tools/demo.py","metadata":{"execution":{"iopub.status.busy":"2023-07-05T16:42:13.822354Z","iopub.execute_input":"2023-07-05T16:42:13.822675Z","iopub.status.idle":"2023-07-05T16:42:14.781338Z","shell.execute_reply.started":"2023-07-05T16:42:13.822631Z","shell.execute_reply":"2023-07-05T16:42:14.780274Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"TEST_IMAGE_PATH = \"/kaggle/working/dataset/images/val2017/0-4614.jpg\"\nMODEL_PATH = \"./YOLOX_outputs/cots_config/best_ckpt.pth\"\n\n!python tools/demo.py image \\\n -f cots_config.py \\\n -c {MODEL_PATH} \\\n --path {TEST_IMAGE_PATH} \\\n --conf 0.1 \\\n --nms 0.45 \\\n --tsize 960 \\\n --save_result \\\n --device gpu","metadata":{"execution":{"iopub.status.busy":"2023-07-05T16:42:14.787393Z","iopub.execute_input":"2023-07-05T16:42:14.787651Z","iopub.status.idle":"2023-07-05T16:42:23.157231Z","shell.execute_reply.started":"2023-07-05T16:42:14.787617Z","shell.execute_reply":"2023-07-05T16:42:23.156281Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"OUTPUT_IMAGE_PATH = \"./YOLOX_outputs/cots_config/vis_res/0-4614.jpg\" \nImage.open(OUTPUT_IMAGE_PATH)","metadata":{"execution":{"iopub.status.busy":"2023-07-05T16:42:23.160026Z","iopub.execute_input":"2023-07-05T16:42:23.161950Z","iopub.status.idle":"2023-07-05T16:42:23.321764Z","shell.execute_reply.started":"2023-07-05T16:42:23.161900Z","shell.execute_reply":"2023-07-05T16:42:23.320583Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"## 6B. INFERENCE USING CUSTOM SCRIPT (IT WOULD BE USED FOR COTS INFERENCE PART)\n\n### 6B.1 SETUP MODEL","metadata":{}},{"cell_type":"code","source":"from yolox.utils import postprocess\nfrom yolox.data.data_augment import ValTransform\n\nCOCO_CLASSES = (\n \"starfish\",\n)\n\n# get YOLOX experiment\ncurrent_exp = importlib.import_module('cots_config')\nexp = current_exp.Exp()\n\n# set inference parameters\ntest_size = (960, 960)\nnum_classes = 1\nconfthre = 0.1\nnmsthre = 0.45\n\n\n# get YOLOX model\nmodel = exp.get_model()\nmodel.cuda()\nmodel.eval()\n\n# get custom trained checkpoint\nckpt_file = \"./YOLOX_outputs/cots_config/best_ckpt.pth\"\nckpt = torch.load(ckpt_file, map_location=\"cpu\")\nmodel.load_state_dict(ckpt[\"model\"])","metadata":{"execution":{"iopub.status.busy":"2023-07-05T16:42:23.323414Z","iopub.status.idle":"2023-07-05T16:42:23.324155Z","shell.execute_reply.started":"2023-07-05T16:42:23.323878Z","shell.execute_reply":"2023-07-05T16:42:23.323906Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"### 6B.2 INFERENCE BBOXES","metadata":{}},{"cell_type":"code","source":"def yolox_inference(img, model, test_size): \n bboxes = []\n bbclasses = []\n scores = []\n \n preproc = ValTransform(legacy = False)\n\n tensor_img, _ = preproc(img, None, test_size)\n tensor_img = torch.from_numpy(tensor_img).unsqueeze(0)\n tensor_img = tensor_img.float()\n tensor_img = tensor_img.cuda()\n\n with torch.no_grad():\n outputs = model(tensor_img)\n outputs = postprocess(\n outputs, num_classes, confthre,\n nmsthre, class_agnostic=True\n )\n\n if outputs[0] is None:\n return [], [], []\n \n outputs = outputs[0].cpu()\n bboxes = outputs[:, 0:4]\n\n bboxes /= min(test_size[0] / img.shape[0], test_size[1] / img.shape[1])\n bbclasses = outputs[:, 6]\n scores = outputs[:, 4] * outputs[:, 5]\n \n return bboxes, bbclasses, scores","metadata":{"execution":{"iopub.status.busy":"2023-07-05T16:42:23.325612Z","iopub.status.idle":"2023-07-05T16:42:23.326348Z","shell.execute_reply.started":"2023-07-05T16:42:23.326065Z","shell.execute_reply":"2023-07-05T16:42:23.326092Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"### 6B.3 DRAW RESULT","metadata":{}},{"cell_type":"code","source":"def draw_yolox_predictions(img, bboxes, scores, bbclasses, confthre, classes_dict):\n for i in range(len(bboxes)):\n box = bboxes[i]\n cls_id = int(bbclasses[i])\n score = scores[i]\n if score < confthre:\n continue\n x0 = int(box[0])\n y0 = int(box[1])\n x1 = int(box[2])\n y1 = int(box[3])\n\n cv2.rectangle(img, (x0, y0), (x1, y1), (0, 255, 0), 2)\n cv2.putText(img, '{}:{:.1f}%'.format(classes_dict[cls_id], score * 100), (x0, y0 - 3), cv2.FONT_HERSHEY_PLAIN, 0.8, (0,255,0), thickness = 1)\n return img","metadata":{"execution":{"iopub.status.busy":"2023-07-05T16:42:23.327710Z","iopub.status.idle":"2023-07-05T16:42:23.328408Z","shell.execute_reply.started":"2023-07-05T16:42:23.328148Z","shell.execute_reply":"2023-07-05T16:42:23.328175Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"### 6B.4 ALL PUZZLES TOGETHER","metadata":{}},{"cell_type":"code","source":"TEST_IMAGE_PATH = \"/kaggle/working/dataset/images/val2017/0-4614.jpg\"\nimg = cv2.imread(TEST_IMAGE_PATH)\n\n# Get predictions\nbboxes, bbclasses, scores = yolox_inference(img, model, test_size)\n\n# Draw predictions\nout_image = draw_yolox_predictions(img, bboxes, scores, bbclasses, confthre, COCO_CLASSES)\n\n# Since we load image using OpenCV we have to convert it \nout_image = cv2.cvtColor(out_image, cv2.COLOR_BGR2RGB)\ndisplay(Image.fromarray(out_image))","metadata":{"execution":{"iopub.status.busy":"2023-07-05T16:42:23.329776Z","iopub.status.idle":"2023-07-05T16:42:23.330459Z","shell.execute_reply.started":"2023-07-05T16:42:23.330203Z","shell.execute_reply":"2023-07-05T16:42:23.330230Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"\n Find this notebook helpful? :) Please give me a vote ;) Thank you\n
","metadata":{}},{"cell_type":"markdown","source":"# 7. SUBMIT TO COTS COMPETITION AND EVALUATE","metadata":{}},{"cell_type":"code","source":"import greatbarrierreef\n\nenv = greatbarrierreef.make_env() # initialize the environment\niter_test = env.iter_test() ","metadata":{"execution":{"iopub.status.busy":"2023-07-05T16:42:23.331814Z","iopub.status.idle":"2023-07-05T16:42:23.332489Z","shell.execute_reply.started":"2023-07-05T16:42:23.332231Z","shell.execute_reply":"2023-07-05T16:42:23.332257Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"submission_dict = {\n 'id': [],\n 'prediction_string': [],\n}\n\nfor (image_np, sample_prediction_df) in iter_test:\n \n bboxes, bbclasses, scores = yolox_inference(image_np, model, test_size)\n \n predictions = []\n for i in range(len(bboxes)):\n box = bboxes[i]\n cls_id = int(bbclasses[i])\n score = scores[i]\n if score < confthre:\n continue\n x_min = int(box[0])\n y_min = int(box[1])\n x_max = int(box[2])\n y_max = int(box[3])\n \n bbox_width = x_max - x_min\n bbox_height = y_max - y_min\n \n predictions.append('{:.2f} {} {} {} {}'.format(score, x_min, y_min, bbox_width, bbox_height))\n \n prediction_str = ' '.join(predictions)\n sample_prediction_df['annotations'] = prediction_str\n env.predict(sample_prediction_df)\n\n print('Prediction:', prediction_str)","metadata":{"execution":{"iopub.status.busy":"2023-07-05T16:42:23.333858Z","iopub.status.idle":"2023-07-05T16:42:23.334544Z","shell.execute_reply.started":"2023-07-05T16:42:23.334281Z","shell.execute_reply":"2023-07-05T16:42:23.334308Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"sub_df = pd.read_csv('submission.csv')\nsub_df.head()","metadata":{"execution":{"iopub.status.busy":"2023-07-05T16:42:23.335880Z","iopub.status.idle":"2023-07-05T16:42:23.336557Z","shell.execute_reply.started":"2023-07-05T16:42:23.336300Z","shell.execute_reply":"2023-07-05T16:42:23.336326Z"},"trusted":true},"execution_count":null,"outputs":[]}]} 
![](\"https://github.com/Megvii-BaseDetection/YOLOX/raw/main/assets/git_fig.png\"/)