1. Load libraries and dataframes with predictions","metadata":{}},{"cell_type":"code","source":"import warnings\nwarnings.filterwarnings('ignore')\n\nimport numpy as np\nimport pandas as pd\nimport os\nfrom tqdm import tqdm","metadata":{"_uuid":"8f2839f25d086af736a60e9eeb907d3b93b6e0e5","_cell_guid":"b1076dfc-b9ad-4769-8c92-a6c4dae69d19","trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"# Import outputs of each selected models\nyolo = pd.read_csv('../input/vinbigdatastack/yolov5.csv')\ndetectron = pd.read_csv('../input/vinbigdatastack/detectron2.csv')\nfasterrcnn = pd.read_csv('../input/vinbigdatastack/fasterrcnn.csv')\n\nimage_ids = yolo.image_id.values","metadata":{"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"#
2. Helper functions","metadata":{}},{"cell_type":"code","source":"def getitem(dataframe, img_id):\n \n \"\"\"\n Parameters\n ----------\n dataframe : pd.DataFrame\n img_id : str\n \n Returns\n -------\n Dictionary of radiographic observations\n \"\"\"\n\n pred = list(dataframe.loc[dataframe.image_id == img_id, \"PredictionString\"])[0].split(' ')\n nb_elm = len(pred)//6\n output = {}\n \n for elm in range(nb_elm):\n output[f'elm_{elm}'] = pred[elm*6 : (elm+1)*6]\n \n return output\n\n\ndef sortDictByProba(dict_):\n \n \"\"\"\n Parameters\n ----------\n dict_ : dict, Dictionary of radiographic observations\n \n Returns\n -------\n Dictionary of radiographic observations sorted by probabilities \n \"\"\"\n \n for key in dict_.keys():\n dict_[key] = list(map(lambda x: float(x), dict_[key]))\n \n # item[1][1] corresponds to the second element of the value (the confidence of the class identified)\n return {k: v for k, v in sorted(dict_.items(), key=lambda item: item[1][1], reverse = True)}\n\n\ndef getHighestProba(*list_of_dicts, n=3):\n \n \"\"\"\n Parameters\n ----------\n list_of_dicts : list[dict], List of dictionaries containing radiographic observations\n n : int, keep n highest elements of each list_of_dicts at most\n \n Returns\n -------\n Dict of merged top3 confidence interval in each dict of list_of_dicts\n \"\"\"\n \n output = {}\n for index, dict_ in enumerate(list_of_dicts):\n dict_length = len(dict_)\n for i in range(dict_length):\n if i < n:\n output[f\"elm_{i}_dict_{index}\"] =list(dict_.values())[i]\n \n return output\n\n\ndef getUnique(dict_):\n \n \"\"\"\n Parameters\n ----------\n dict_ : dict, Dictionary of radiographic observations\n \n Returns\n -------\n List of unique class_id, list of duplicates class_id\n \"\"\"\n \n dict_length = len(dict_)\n \n classes_non_unique = [list(dict_.values())[index][0] for index in range(dict_length)]\n classes_unique = list(set(classes_non_unique))\n \n uniques, counts = np.unique(classes_non_unique, return_counts=True)\n duplicates = uniques[counts > 1]\n singles = np.setdiff1d(classes_unique, duplicates)\n \n return singles, duplicates\n\n\ndef getKeysByValue(dictOfElements, valueToFind):\n \n \"\"\"\n Parameters\n ----------\n dictOfElements : dict, Dictionary of radiographic observations\n valueToFind : int, corresponds to class_id\n \n Returns\n -------\n List of keys of dictOfElements that contain valueToFind\n \"\"\"\n \n output = list()\n listOfItems = dictOfElements.items()\n \n for item in listOfItems:\n if item[1][0] == valueToFind:\n output.append(item[0])\n \n return output\n\n\ndef getListKeysByValue(dictOfElements, valuesToFind):\n \n \"\"\"\n Parameters\n ----------\n dictOfElements : dict, Dictionary of radiographic observations\n valuesToFind : list[int], list of class_id\n \n Returns\n -------\n List of lists of keys of dictOfElements for each value in valuesToFind\n \"\"\"\n \n output = []\n \n for value in valuesToFind:\n output.append(getKeysByValue(dictOfElements, value))\n \n return output\n\n\ndef averaging(from_dict, single_keys, dupl_keys):\n \n \"\"\"\n Parameters\n ----------\n from_dict : dict, dictionary to be filtered\n single_keys : list[str], list of keys that should be infered\n dupl_keys : list[str], list of class_id\n \n Returns\n -------\n A filtered dictionary with averaged probs and boxes\n \"\"\"\n \n output = {}\n \n # Infer single keys\n if len(np.ravel(single_keys)) != 0:\n for single in np.ravel(single_keys):\n output[single] = from_dict[single]\n\n # For each duplicates, get index of all occurences and average boxing\n if len(np.ravel(dupl_keys)) != 0:\n for index, list_of_duplicate_class in enumerate(dupl_keys):\n probs = [] \n boxing1 = []\n boxing2 = []\n boxing3 = []\n boxing4 = []\n \n for elm in list_of_duplicate_class:\n probs.append(from_dict[elm][1])\n boxing1.append(from_dict[elm][2])\n boxing2.append(from_dict[elm][3])\n boxing3.append(from_dict[elm][4])\n boxing4.append(from_dict[elm][5])\n \n output[f\"elm_{index}\"] = [from_dict[list_of_duplicate_class[0]][0],\n np.mean(probs),\n np.mean(boxing1),\n np.mean(boxing2),\n np.mean(boxing3),\n np.mean(boxing4)]\n \n return output\n\n\ndef toString(pred_list):\n \n \"\"\"\n Parameters\n ----------\n list_final : list[int], list of all radiographic observations\n \n Returns\n -------\n A string which fits with the expected output\n \"\"\"\n \n castedList = []\n for index, elm in enumerate(pred_list):\n if index%6 == 0:\n castedList.append(str(int(elm)))\n else:\n castedList.append(str(elm))\n \n output = \" \".join(castedList)\n \n return output","metadata":{"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"--------\n\n**
Back to summary**","metadata":{}},{"cell_type":"markdown","source":"#
3. Run ensembling with appropriate strategy","metadata":{"trusted":true}},{"cell_type":"markdown","source":"My strategy here consists in averaging observations that have at least one dupplicate among all models. Some filtering about boxing areas should be added. This will come in a future release","metadata":{}},{"cell_type":"code","source":"def main():\n \n output = pd.DataFrame(columns = [\"image_id\", \"PredictionString\"])\n \n for image_id in tqdm(image_ids):\n \n # For each model, get PredictionString of image_id as a dict\n fasterrcnn_pred = getitem(fasterrcnn, image_id)\n detectron_pred = getitem(detectron, image_id)\n yolo_pred = getitem(yolo, image_id) \n \n # Sort dicts by proba\n sorted_fasterrcnn = sortDictByProba(fasterrcnn_pred)\n sorted_detectron = sortDictByProba(detectron_pred)\n sorted_yolo = sortDictByProba(yolo_pred)\n\n # Filter dicts into one dict with at most top n probs\n highest_probs = getHighestProba(sorted_fasterrcnn, \n sorted_detectron, \n sorted_yolo,\n n = 3)\n \n # Get keys of unique and duplicates values in the filtered dict\n singles, duplicates = getUnique(highest_probs)\n single_keys = getListKeysByValue(highest_probs, singles)\n dupl_keys = getListKeysByValue(highest_probs, duplicates)\n \n # Apply averaging strategy\n stacked_dict = averaging(highest_probs, single_keys, dupl_keys)\n \n # Put string in right format\n prediction_int = np.ravel(list(stacked_dict.values()))\n prediction_string = toString(prediction_int)\n \n output = output.append({\"image_id\": image_id, \n \"PredictionString\": prediction_string},\n ignore_index=True)\n \n return output","metadata":{"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"Some other strategies will be tested in a future release:\n* OR method \n* AND method\n* Consensus method\n* Weighted Fusion","metadata":{}},{"cell_type":"markdown","source":"In the meantime, if you found this notebook usefull and you do have some suggestions on how this could be better implemented, do not hesitate to contribute, i'd really appreciate !","metadata":{}},{"cell_type":"markdown","source":"--------\n\n**
Back to summary**","metadata":{}},{"cell_type":"markdown","source":"#
4. Save results","metadata":{}},{"cell_type":"code","source":"final_sub = main()\nfinal_sub.to_csv(\"submission.csv\", index=False)","metadata":{"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"# References\n\n*
Article on object detection through ensemble of models\n* detectron2 : https://www.kaggle.com/c/vinbigdata-chest-xray-abnormalities-detection/code?competitionId=24800&sortBy=scoreDescending\n* fasterrcnn : https://www.kaggle.com/awsaf49/vinbigdata-cxr-ad-yolov5-14-class-infer\n* yolov5 : https://www.kaggle.com/basu369victor/chest-x-ray-abnormalities-detection-submission","metadata":{}},{"cell_type":"markdown","source":"
\n
Thank you for taking the time to read this notebook. I hope that I was able to answer your questions or your curiosity and that it was quite understandable. any constructive comments are welcome. They help me progress and motivate me to share better quality content. I am above all a passionate person who tries to advance my knowledge but also that of others. If you liked it, feel free to upvote and share my work.
\n
\n
Thank you and may passion guide you.
","metadata":{}}]}