I tried to overcome the problem of small objects in the training dataset and the problem of large GPU needs, so I divided each image to 36 images with 214*120 pixels and correct labels of these images.
\nthen trained my YOLOV5l on this new data with image size 640 for 100 epochs.
the validation results were great, I got 0.92 F2 score.
\nthen I used sliding window algorithm that divide testing images into small images and run the model inference through these small images individually and then collect these predictions together.
\nbut I got very bad results nearly 0.185
\nI wondering if I made I mistake somewhere??
", "messages": [ { "id": "1660113", "postDate": "01/22/2022 12:22:45", "content": "I tried to overcome the problem of small objects in the training dataset and the problem of large GPU needs, so I divided each image to 36 images with 214*120 pixels and correct labels of these images.
\nthen trained my YOLOV5l on this new data with image size 640 for 100 epochs.
the validation results were great, I got 0.92 F2 score.
\nthen I used sliding window algorithm that divide testing images into small images and run the model inference through these small images individually and then collect these predictions together.
\nbut I got very bad results nearly 0.185
\nI wondering if I made I mistake somewhere??
", "rawMarkdown": "I tried to overcome the problem of small objects in the training dataset and the problem of large GPU needs, so I divided each image to 36 images with 214*120 pixels and correct labels of these images.\nthen trained my YOLOV5l on this new data with image size 640 for 100 epochs.\n\nthe validation results were great, I got 0.92 F2 score.\n\nthen I used sliding window algorithm that divide testing images into small images and run the model inference through these small images individually and then collect these predictions together.\n\nbut I got very bad results nearly 0.185\n\nI wondering if I made I mistake somewhere??", "votes": null }, { "id": "1660502", "postDate": "01/22/2022 17:52:35", "content": "It depends, I think 214x120 is too small and I have tried other levels with sahi. I recommend maybe doing around 768. I got some acceptable results using a modified detector class:
\nclass Yolov5DetectionModelHiRez(Yolov5DetectionModel):\n def perform_inference(self, image: np.ndarray, image_size: int = None):\n \"\"\"\n Prediction is performed using self.model and the prediction result is set to self._original_predictions.\n Args:\n image: np.ndarray\n A numpy array that contains the image to be predicted. 3 channel image should be in RGB order.\n image_size: int\n Inference input size.\n \"\"\"\n try:\n import yolov5\n except ImportError:\n raise ImportError('Please run \"pip install -U yolov5\" ' \"to install YOLOv5 first for YOLOv5 inference.\")\n\n # Confirm model is loaded\n assert self.model is not None, \"Model is not loaded, load it by calling .load_model()\"\n\n prediction_result = self.model(image, size=1280, augment=False) # Inference at a bit higher rez than original for each slize\n\n self._original_predictions = prediction_result\nBut results are still not great, it may be worth looking at doing a sahi with high overlap instead of augmentation. Even my results from splitting images smaller than 768 did not do better than high-resolution training. Hopefully, my tests help someone make something better.
", "rawMarkdown": "It depends, I think 214x120 is too small and I have tried other levels with [sahi](https://github.com/obss/sahi). I recommend maybe doing around 768. I got some acceptable results using a modified detector class:\n\n```\nclass Yolov5DetectionModelHiRez(Yolov5DetectionModel):\n def perform_inference(self, image: np.ndarray, image_size: int = None):\n \"\"\"\n Prediction is performed using self.model and the prediction result is set to self._original_predictions.\n Args:\n image: np.ndarray\n A numpy array that contains the image to be predicted. 3 channel image should be in RGB order.\n image_size: int\n Inference input size.\n \"\"\"\n try:\n import yolov5\n except ImportError:\n raise ImportError('Please run \"pip install -U yolov5\" ' \"to install YOLOv5 first for YOLOv5 inference.\")\n\n # Confirm model is loaded\n assert self.model is not None, \"Model is not loaded, load it by calling .load_model()\"\n \n prediction_result = self.model(image, size=1280, augment=False) # Inference at a bit higher rez than original for each slize\n\n self._original_predictions = prediction_result\n```\n\nBut results are still not great, it may be worth looking at doing a sahi with high overlap instead of augmentation. Even my results from splitting images smaller than 768 did not do better than high-resolution training. Hopefully, my tests help someone make something better.", "votes": null }, { "id": "1660507", "postDate": "01/22/2022 17:54:02", "content": "I will publish a notebook for it soon if I get higher results than the baseline yolov5, it depends if figure it out before the cutoff for posting notebooks since this competition will be ending in a few weeks.
", "rawMarkdown": "I will publish a notebook for it soon if I get higher results than the baseline yolov5, it depends if figure it out before the cutoff for posting notebooks since this competition will be ending in a few weeks.", "votes": null }, { "id": "1660675", "postDate": "01/22/2022 21:05:26", "content": "@outwrest but why I got 0.92 F2 score while getting these bad results on LB
", "rawMarkdown": "outwrest but why I got 0.92 F2 score while getting these bad results on LB", "votes": null }, { "id": "1660687", "postDate": "01/22/2022 21:23:41", "content": "I am not sure, it could be a lot of things that can result in you getting a very good F2 score on validation data. Make sure your data is split correctly and that you have a lot of empty images in validation to get an accurate result.
", "rawMarkdown": "I am not sure, it could be a lot of things that can result in you getting a very good F2 score on validation data. Make sure your data is split correctly and that you have a lot of empty images in validation to get an accurate result.", "votes": null }, { "id": "1663538", "postDate": "01/25/2022 07:38:02", "content": "0.92F2? Maybe your validation set is not split correctly.
", "rawMarkdown": "0.92F2? Maybe your validation set is not split correctly.", "votes": null } ], "comments": [ { "id": 1660502, "author_name": "outwrest", "author_url": "", "post_date": "01/22/2022 17:52:35", "content": "It depends, I think 214x120 is too small and I have tried other levels with sahi. I recommend maybe doing around 768. I got some acceptable results using a modified detector class:
\nclass Yolov5DetectionModelHiRez(Yolov5DetectionModel):\n def perform_inference(self, image: np.ndarray, image_size: int = None):\n \"\"\"\n Prediction is performed using self.model and the prediction result is set to self._original_predictions.\n Args:\n image: np.ndarray\n A numpy array that contains the image to be predicted. 3 channel image should be in RGB order.\n image_size: int\n Inference input size.\n \"\"\"\n try:\n import yolov5\n except ImportError:\n raise ImportError('Please run \"pip install -U yolov5\" ' \"to install YOLOv5 first for YOLOv5 inference.\")\n\n # Confirm model is loaded\n assert self.model is not None, \"Model is not loaded, load it by calling .load_model()\"\n\n prediction_result = self.model(image, size=1280, augment=False) # Inference at a bit higher rez than original for each slize\n\n self._original_predictions = prediction_result\nBut results are still not great, it may be worth looking at doing a sahi with high overlap instead of augmentation. Even my results from splitting images smaller than 768 did not do better than high-resolution training. Hopefully, my tests help someone make something better.
", "votes": null, "replies": [ { "id": 1660507, "author_name": "outwrest", "author_url": "", "post_date": "01/22/2022 17:54:02", "content": "I will publish a notebook for it soon if I get higher results than the baseline yolov5, it depends if figure it out before the cutoff for posting notebooks since this competition will be ending in a few weeks.
", "votes": null, "replies": [] } ] }, { "id": 1660675, "author_name": "mohameddonia2222", "author_url": "", "post_date": "01/22/2022 21:05:26", "content": "@outwrest but why I got 0.92 F2 score while getting these bad results on LB
", "votes": null, "replies": [ { "id": 1660687, "author_name": "outwrest", "author_url": "", "post_date": "01/22/2022 21:23:41", "content": "I am not sure, it could be a lot of things that can result in you getting a very good F2 score on validation data. Make sure your data is split correctly and that you have a lot of empty images in validation to get an accurate result.
", "votes": null, "replies": [] } ] }, { "id": 1663538, "author_name": "klawensliu", "author_url": "", "post_date": "01/25/2022 07:38:02", "content": "0.92F2? Maybe your validation set is not split correctly.
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "1660113": "I tried to overcome the problem of small objects in the training dataset and the problem of large GPU needs, so I divided each image to 36 images with 214*120 pixels and correct labels of these images.\nthen trained my YOLOV5l on this new data with image size 640 for 100 epochs.\n\nthe validation results were great, I got 0.92 F2 score.\n\nthen I used sliding window algorithm that divide testing images into small images and run the model inference through these small images individually and then collect these predictions together.\n\nbut I got very bad results nearly 0.185\n\nI wondering if I made I mistake somewhere??", "1660502": "It depends, I think 214x120 is too small and I have tried other levels with [sahi](https://github.com/obss/sahi). I recommend maybe doing around 768. I got some acceptable results using a modified detector class:\n\n```\nclass Yolov5DetectionModelHiRez(Yolov5DetectionModel):\n def perform_inference(self, image: np.ndarray, image_size: int = None):\n \"\"\"\n Prediction is performed using self.model and the prediction result is set to self._original_predictions.\n Args:\n image: np.ndarray\n A numpy array that contains the image to be predicted. 3 channel image should be in RGB order.\n image_size: int\n Inference input size.\n \"\"\"\n try:\n import yolov5\n except ImportError:\n raise ImportError('Please run \"pip install -U yolov5\" ' \"to install YOLOv5 first for YOLOv5 inference.\")\n\n # Confirm model is loaded\n assert self.model is not None, \"Model is not loaded, load it by calling .load_model()\"\n \n prediction_result = self.model(image, size=1280, augment=False) # Inference at a bit higher rez than original for each slize\n\n self._original_predictions = prediction_result\n```\n\nBut results are still not great, it may be worth looking at doing a sahi with high overlap instead of augmentation. Even my results from splitting images smaller than 768 did not do better than high-resolution training. Hopefully, my tests help someone make something better.", "1660507": "I will publish a notebook for it soon if I get higher results than the baseline yolov5, it depends if figure it out before the cutoff for posting notebooks since this competition will be ending in a few weeks.", "1660675": "outwrest but why I got 0.92 F2 score while getting these bad results on LB", "1660687": "I am not sure, it could be a lot of things that can result in you getting a very good F2 score on validation data. Make sure your data is split correctly and that you have a lot of empty images in validation to get an accurate result.", "1663538": "0.92F2? Maybe your validation set is not split correctly." }, "source": "meta" }