![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F2721247%2F0497057948a150f8b90e8cf3f622b14f%2Faptoswrong.bmp?generation=1567982632049414&alt=media\")
After competition ends I want to do some reproducing and experiments. But I find a weird situation:
\n\n
The bottom result is what i run the inference kernel. The top result is that I just copy the prediction from the inference kernel and do fast submission. why there is difference in public LB?
", "messages": [ { "id": "621758", "postDate": "09/08/2019 22:45:23", "content": "After competition ends I want to do some reproducing and experiments. But I find a weird situation:
\n\n
The bottom result is what i run the inference kernel. The top result is that I just copy the prediction from the inference kernel and do fast submission. why there is difference in public LB?
", "rawMarkdown": "After competition ends I want to do some reproducing and experiments. But I find a weird situation:\n\n\n\nThe bottom result is what i run the inference kernel. The top result is that I just copy the prediction from the inference kernel and do fast submission. why there is difference in public LB?", "votes": null }, { "id": "621762", "postDate": "09/08/2019 23:00:50", "content": "The submission.csv you can see only contains data from the public leaderboard set, so you only made predictions on that set.
", "rawMarkdown": "The submission.csv you can see only contains data from the public leaderboard set, so you only made predictions on that set.", "votes": null }, { "id": "621764", "postDate": "09/08/2019 23:03:42", "content": "But that should be no difference at public LB. Or you mean I should use test.csv instead of sample_submission.csv?
", "rawMarkdown": "But that should be no difference at public LB. Or you mean I should use test.csv instead of sample_submission.csv?", "votes": null }, { "id": "621782", "postDate": "09/09/2019 00:17:36", "content": "Ah I see, then it is not explained by that, I'm not sure then
", "rawMarkdown": "Ah I see, then it is not explained by that, I'm not sure then", "votes": null }, { "id": "621800", "postDate": "09/09/2019 00:53:54", "content": "findout the reason:\nthese 2 function to crop: \n```python\ndef resize_images(img, desired_size = 1024):
\n img = cv2.copyMakeBorder(img,10,10,10,10,cv2.BORDER_CONSTANT,value=[0,0,0])\n gray = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY)\n ret,gray = cv2.threshold(gray,10,255,cv2.THRESH_BINARY)
contours,hierarchy = cv2.findContours(gray,\n cv2.RETR_EXTERNAL,\n cv2.CHAIN_APPROX_SIMPLE)\ncontours = max(contours, key=cv2.contourArea)\nx,y,w,h = cv2.boundingRect(contours)\n\nif w>200 and h>200:\n new_img = img[y:y+h,x:x+w]\n height, width, _= new_img.shape\n\n if max([height, width]) > desired_size:\n ratio = float(desired_size/max([height, width]))\n new_img = cv2.resize(new_img, \n tuple([int(width*ratio), int(height*ratio)]), \n interpolation = cv2.INTER_CUBIC)\n\n return new_img\nelse:\n print(f'No bounding for {name}')\n return new_img\n\nand\npython\ndef crop(gray, img, percent_smaller):
thresh = 8\n\ntop = 0\nleft = 0\nbottom = gray.shape[0] - 1\nright = gray.shape[1] - 1\n\n# work in from the top and bottom along the middle collumn\nmiddleCol = gray[:, int(gray.shape[1]/2)] > thresh\nwhile middleCol[top] == 0:\n top += 1\nwhile middleCol[bottom] == 0:\n bottom -= 1\n\n# work in from the sides along the middle row\nmiddleRow = gray[int(gray.shape[0]/2)] > thresh\nwhile middleRow[left] == 0:\n left += 1\nwhile middleRow[right] == 0:\n right -= 1\n\nheight = bottom - top\nwidth = right - left\n\nbottom -= int(percent_smaller*height)\ntop += int(percent_smaller*height)\nright -= int(percent_smaller*width)\nleft += int(percent_smaller*width)\n\nif height < 100 or width < 100:\n print(\"Error: squareUp: bottom:\", bottom, \"top:\", top)\n print(\"Error: squareUp: right:\", right, \"left:\", left)\n return img\n\nreturn img[top:bottom, left:right]\n```\nhas trouble to deal with some private test picture.\nAny team used these 2 function or some similar functions might have a heart break trouble.
", "rawMarkdown": "findout the reason:\nthese 2 function to crop: \n```python\ndef resize_images(img, desired_size = 1024): \n img = cv2.copyMakeBorder(img,10,10,10,10,cv2.BORDER_CONSTANT,value=[0,0,0])\n gray = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY)\n ret,gray = cv2.threshold(gray,10,255,cv2.THRESH_BINARY)\n \n contours,hierarchy = cv2.findContours(gray,\n cv2.RETR_EXTERNAL,\n cv2.CHAIN_APPROX_SIMPLE)\n contours = max(contours, key=cv2.contourArea)\n x,y,w,h = cv2.boundingRect(contours)\n\n if w>200 and h>200:\n new_img = img[y:y+h,x:x+w]\n height, width, _= new_img.shape\n\n if max([height, width]) > desired_size:\n ratio = float(desired_size/max([height, width]))\n new_img = cv2.resize(new_img, \n tuple([int(width*ratio), int(height*ratio)]), \n interpolation = cv2.INTER_CUBIC)\n \n return new_img\n else:\n print(f'No bounding for {name}')\n return new_img\n```\nand\n```python\ndef crop(gray, img, percent_smaller):\n \n thresh = 8\n \n top = 0\n left = 0\n bottom = gray.shape[0] - 1\n right = gray.shape[1] - 1\n \n # work in from the top and bottom along the middle collumn\n middleCol = gray[:, int(gray.shape[1]/2)] > thresh\n while middleCol[top] == 0:\n top += 1\n while middleCol[bottom] == 0:\n bottom -= 1\n \n # work in from the sides along the middle row\n middleRow = gray[int(gray.shape[0]/2)] > thresh\n while middleRow[left] == 0:\n left += 1\n while middleRow[right] == 0:\n right -= 1\n \n height = bottom - top\n width = right - left\n \n bottom -= int(percent_smaller*height)\n top += int(percent_smaller*height)\n right -= int(percent_smaller*width)\n left += int(percent_smaller*width)\n \n if height < 100 or width < 100:\n print(\"Error: squareUp: bottom:\", bottom, \"top:\", top)\n print(\"Error: squareUp: right:\", right, \"left:\", left)\n return img\n \n return img[top:bottom, left:right]\n```\nhas trouble to deal with some private test picture.\nAny team used these 2 function or some similar functions might have a heart break trouble.", "votes": null } ], "comments": [ { "id": 621762, "author_name": "gzuidhof", "author_url": "", "post_date": "09/08/2019 23:00:50", "content": "The submission.csv you can see only contains data from the public leaderboard set, so you only made predictions on that set.
", "votes": null, "replies": [ { "id": 621764, "author_name": "httpwwwfszyc", "author_url": "", "post_date": "09/08/2019 23:03:42", "content": "But that should be no difference at public LB. Or you mean I should use test.csv instead of sample_submission.csv?
", "votes": null, "replies": [] }, { "id": 621782, "author_name": "gzuidhof", "author_url": "", "post_date": "09/09/2019 00:17:36", "content": "Ah I see, then it is not explained by that, I'm not sure then
", "votes": null, "replies": [] } ] }, { "id": 621800, "author_name": "httpwwwfszyc", "author_url": "", "post_date": "09/09/2019 00:53:54", "content": "findout the reason:\nthese 2 function to crop: \n```python\ndef resize_images(img, desired_size = 1024):
\n img = cv2.copyMakeBorder(img,10,10,10,10,cv2.BORDER_CONSTANT,value=[0,0,0])\n gray = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY)\n ret,gray = cv2.threshold(gray,10,255,cv2.THRESH_BINARY)
contours,hierarchy = cv2.findContours(gray,\n cv2.RETR_EXTERNAL,\n cv2.CHAIN_APPROX_SIMPLE)\ncontours = max(contours, key=cv2.contourArea)\nx,y,w,h = cv2.boundingRect(contours)\n\nif w>200 and h>200:\n new_img = img[y:y+h,x:x+w]\n height, width, _= new_img.shape\n\n if max([height, width]) > desired_size:\n ratio = float(desired_size/max([height, width]))\n new_img = cv2.resize(new_img, \n tuple([int(width*ratio), int(height*ratio)]), \n interpolation = cv2.INTER_CUBIC)\n\n return new_img\nelse:\n print(f'No bounding for {name}')\n return new_img\n\nand\npython\ndef crop(gray, img, percent_smaller):
thresh = 8\n\ntop = 0\nleft = 0\nbottom = gray.shape[0] - 1\nright = gray.shape[1] - 1\n\n# work in from the top and bottom along the middle collumn\nmiddleCol = gray[:, int(gray.shape[1]/2)] > thresh\nwhile middleCol[top] == 0:\n top += 1\nwhile middleCol[bottom] == 0:\n bottom -= 1\n\n# work in from the sides along the middle row\nmiddleRow = gray[int(gray.shape[0]/2)] > thresh\nwhile middleRow[left] == 0:\n left += 1\nwhile middleRow[right] == 0:\n right -= 1\n\nheight = bottom - top\nwidth = right - left\n\nbottom -= int(percent_smaller*height)\ntop += int(percent_smaller*height)\nright -= int(percent_smaller*width)\nleft += int(percent_smaller*width)\n\nif height < 100 or width < 100:\n print(\"Error: squareUp: bottom:\", bottom, \"top:\", top)\n print(\"Error: squareUp: right:\", right, \"left:\", left)\n return img\n\nreturn img[top:bottom, left:right]\n```\nhas trouble to deal with some private test picture.\nAny team used these 2 function or some similar functions might have a heart break trouble.
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "621758": "After competition ends I want to do some reproducing and experiments. But I find a weird situation:\n\n\n\nThe bottom result is what i run the inference kernel. The top result is that I just copy the prediction from the inference kernel and do fast submission. why there is difference in public LB?", "621762": "The submission.csv you can see only contains data from the public leaderboard set, so you only made predictions on that set.", "621764": "But that should be no difference at public LB. Or you mean I should use test.csv instead of sample_submission.csv?", "621782": "Ah I see, then it is not explained by that, I'm not sure then", "621800": "findout the reason:\nthese 2 function to crop: \n```python\ndef resize_images(img, desired_size = 1024): \n img = cv2.copyMakeBorder(img,10,10,10,10,cv2.BORDER_CONSTANT,value=[0,0,0])\n gray = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY)\n ret,gray = cv2.threshold(gray,10,255,cv2.THRESH_BINARY)\n \n contours,hierarchy = cv2.findContours(gray,\n cv2.RETR_EXTERNAL,\n cv2.CHAIN_APPROX_SIMPLE)\n contours = max(contours, key=cv2.contourArea)\n x,y,w,h = cv2.boundingRect(contours)\n\n if w>200 and h>200:\n new_img = img[y:y+h,x:x+w]\n height, width, _= new_img.shape\n\n if max([height, width]) > desired_size:\n ratio = float(desired_size/max([height, width]))\n new_img = cv2.resize(new_img, \n tuple([int(width*ratio), int(height*ratio)]), \n interpolation = cv2.INTER_CUBIC)\n \n return new_img\n else:\n print(f'No bounding for {name}')\n return new_img\n```\nand\n```python\ndef crop(gray, img, percent_smaller):\n \n thresh = 8\n \n top = 0\n left = 0\n bottom = gray.shape[0] - 1\n right = gray.shape[1] - 1\n \n # work in from the top and bottom along the middle collumn\n middleCol = gray[:, int(gray.shape[1]/2)] > thresh\n while middleCol[top] == 0:\n top += 1\n while middleCol[bottom] == 0:\n bottom -= 1\n \n # work in from the sides along the middle row\n middleRow = gray[int(gray.shape[0]/2)] > thresh\n while middleRow[left] == 0:\n left += 1\n while middleRow[right] == 0:\n right -= 1\n \n height = bottom - top\n width = right - left\n \n bottom -= int(percent_smaller*height)\n top += int(percent_smaller*height)\n right -= int(percent_smaller*width)\n left += int(percent_smaller*width)\n \n if height < 100 or width < 100:\n print(\"Error: squareUp: bottom:\", bottom, \"top:\", top)\n print(\"Error: squareUp: right:\", right, \"left:\", left)\n return img\n \n return img[top:bottom, left:right]\n```\nhas trouble to deal with some private test picture.\nAny team used these 2 function or some similar functions might have a heart break trouble." }, "source": "meta" }