![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F1696976%2F11388298dc71e325102614fd932723da%2F.png?generation=1579422596035225&alt=media\")
Hi, everyone! Been struggling to make a successful submission. There are the timeouts, kaggle error, etc. So I decide to give the discussion a try.
\n\nHere is a kernel I forked and edited. \nhttps://www.kaggle.com/roguekk007/another-stab\nThe resnet34 achieves .970 CV locally, but when running it I got submission CSV not found. I am running at 224x224 resolution which I found helps.
\n\nCommiting the kernel gives me a submission.csv. Why is there this error?\nAny help is appreciated, thanks👍
", "messages": [ { "id": "722241", "postDate": "01/18/2020 09:56:27", "content": "Hi, everyone! Been struggling to make a successful submission. There are the timeouts, kaggle error, etc. So I decide to give the discussion a try.
\n\nHere is a kernel I forked and edited. \nhttps://www.kaggle.com/roguekk007/another-stab\nThe resnet34 achieves .970 CV locally, but when running it I got submission CSV not found. I am running at 224x224 resolution which I found helps.
\n\nCommiting the kernel gives me a submission.csv. Why is there this error?\nAny help is appreciated, thanks👍
", "rawMarkdown": "Hi, everyone! Been struggling to make a successful submission. There are the timeouts, kaggle error, etc. So I decide to give the discussion a try.\n\nHere is a kernel I forked and edited. \nhttps://www.kaggle.com/roguekk007/another-stab\nThe resnet34 achieves .970 CV locally, but when running it I got **submission CSV not found**. I am running at 224x224 resolution which I found helps. \n\nCommiting the kernel gives me a submission.csv. Why is there this error?\nAny help is appreciated, thanks👍", "votes": null }, { "id": "722247", "postDate": "01/18/2020 10:04:54", "content": "This code is not optimized yet. You probably ran out of memory or time. Try a test commit with the train parquet files to see where is the issue.
\n\nYou can find my optimization tips here. (Implementation notebook included)
", "rawMarkdown": "This code is not optimized yet. You probably ran out of memory or time. Try a test commit with the train parquet files to see where is the issue. \n\nYou can find my optimization tips [here](https://www.kaggle.com/c/bengaliai-cv19/discussion/126054). (Implementation notebook included)", "votes": null }, { "id": "722259", "postDate": "01/18/2020 10:23:31", "content": "Hi Peter, thanks for your answer! I was trying to keep the code minimal. It is probably not a timeout error I have received an explicit \"Notebook timeout error\" before. The \"Did not find CSV\" error is really confusing. I have read your optimization tips and found them very useful :)
", "rawMarkdown": "Hi Peter, thanks for your answer! I was trying to keep the code minimal. It is probably not a timeout error I have received an explicit \"Notebook timeout error\" before. The \"Did not find CSV\" error is really confusing. I have read your optimization tips and found them very useful :)", "votes": null }, { "id": "722354", "postDate": "01/18/2020 13:01:34", "content": "There, I also managed to get a submission scoring error...I think I just obtained a collection of all make-able kaggle submission errors in my log now :)
", "rawMarkdown": "There, I also managed to get a submission scoring error...I think I just obtained a collection of all make-able kaggle submission errors in my log now :)", "votes": null }, { "id": "722872", "postDate": "01/19/2020 08:30:28", "content": "This competition has two challenges. First of them is to make a submission :)
\n\n
There is no way to find out what exact exception was raised during submission. And I cannot reproduce it, since kernel's commit is successful. It would be really helpful to show the stack trace in this case, since there is no way to debug this issue.
\n\n", "rawMarkdown": "There is no way to find out what exact exception was raised during submission. And I cannot reproduce it, since kernel's commit is successful. It would be really helpful to show the stack trace in this case, since there is no way to debug this issue.\n\n@addisonhoward @sohier", "votes": null }, { "id": "723428", "postDate": "01/20/2020 02:52:31", "content": "Exactly the same here! My commit kernel produces a valid submission but I got a \"submission CSV not found\" error. Debugging the submission process is really frustrating.\nLooking forward to solving this problem.\n@addisonhoward @sohier
\n\nHere is the tragic submission log for me\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F1943421%2F8f9a6e3e0776692335d9660feb4f42cd%2Fsubmission_error.jpg?generation=1579488879291317&alt=media\")
In order to avoid memory issues, i first saved the model as a h5 file. Now in my kernel i am loading this model making a prediction and submitting, even that results in Error, and no log is provided. Can someone help pls.
", "rawMarkdown": "In order to avoid memory issues, i first saved the model as a h5 file. Now in my kernel i am loading this model making a prediction and submitting, even that results in Error, and no log is provided. Can someone help pls.", "votes": null }, { "id": "724318", "postDate": "01/21/2020 03:21:22", "content": "Hi! I solved my submission error by doing 9 controlled experiments (yes, I failed 8 times) and 29 previously failed submissions👋 . Seems like the first challenge in the competition is indeed making a successful submission :) Here is what worked for me, hope it helps.\n I found that the crop_resize function (which I edited a little) and transformation led to failure. I added a backup transform without crop_resize with try:...except:... and solved the problem. Single model which achieves .972 CV leads to my current position.
\n\nOriginal Code (fails)\n```\nclass GraphemeDataset(Dataset):\n def init(self, fname, transform):\n self.df = pd.read_parquet(fname)\n self.data = 255 - self.df.iloc[:, 1:].values.reshape(-1, HEIGHT, WIDTH).astype(np.uint8)\n self.transform = transform
\n\ndef __len__(self):\n return len(self.data)\n\ndef __getitem__(self, idx):\n name = self.df.iloc[idx,0]\n img = self.data[idx].astype(np.uint8)\n img = self.transform(image=img)['image']\n return img, name\n```
\n\nSuccessful Code\n```\nclass GraphemeDataset(Dataset):\n def init(self, fname, transform):\n self.df = pd.read_parquet(fname)\n self.data = self.df.iloc[:, 1:].values.reshape(-1, HEIGHT, WIDTH).astype(np.uint8)\n self.transform = transform
\n\ndef __len__(self):\n return len(self.data)\n\ndef __getitem__(self, idx):\n name = self.df.iloc[idx,0]\n try:\n img = 255 - self.data[idx].astype(np.uint8)\n img = self.transform(image=img)['image']\n except:\n img = 255 - self.data[idx].astype(np.uint8)\n img = backup_tsfm(image=img)['image']\n return img, name\n```
\n\nStill investigating why crop-resize fails, though. Any hypotheses?\nHere is the code for my crop-resize and transformation\n```\ndef bbox(img):\n rows = np.any(img, axis=1)\n cols = np.any(img, axis=0)\n rmin, rmax = np.where(rows)[0][[0, -1]]\n cmin, cmax = np.where(cols)[0][[0, -1]]\n return rmin, rmax, cmin, cmax
\n\ndef crop_resize(img0, size=SIZE, pad=8, cols=None, rows=None, force_apply=False):\n #crop a box around pixels large than the threshold \n #some images contain line at the sides\n ymin,ymax,xmin,xmax = bbox(img0[5:-5,5:-5] > 80)\n #cropping may cut too much, so we need to add it back\n xmin = xmin - 13 if (xmin > 13) else 0\n ymin = ymin - 10 if (ymin > 10) else 0\n xmax = xmax + 13 if (xmax < WIDTH - 13) else WIDTH\n ymax = ymax + 10 if (ymax < HEIGHT - 10) else HEIGHT\n img = img0[ymin:ymax,xmin:xmax]\n #remove lo intensity pixels as noise\n img[img < 28] = 0\n lx, ly = xmax-xmin,ymax-ymin\n l = max(lx,ly) + pad\n #make sure that the aspect ratio is kept in rescaling\n img = np.pad(img, [((l-ly)//2,), ((l-lx)//2,)], mode='constant')\n return cv2.resize(img,(size,size))
\n\ntsfm = Compose([\n Lambda(image=crop_resize),\n Resize(SIZE, SIZE),\n Normalize((.0855,),(.2,)),\n ToTensor(),\n ])\n```
", "rawMarkdown": "Hi! I solved my submission error by doing 9 controlled experiments (yes, I failed 8 times) and 29 previously failed submissions👋 . Seems like the first challenge in the competition is indeed making a successful submission :) Here is what worked for me, hope it helps.\n I found that the crop_resize function (which I edited a little) and transformation led to failure. **I added a backup transform without crop_resize with try:...except:... and solved the problem.** Single model which achieves .972 CV leads to my current position.\n\nOriginal Code (fails)\n```\nclass GraphemeDataset(Dataset):\n def __init__(self, fname, transform):\n self.df = pd.read_parquet(fname)\n self.data = 255 - self.df.iloc[:, 1:].values.reshape(-1, HEIGHT, WIDTH).astype(np.uint8)\n self.transform = transform\n\n def __len__(self):\n return len(self.data)\n\n def __getitem__(self, idx):\n name = self.df.iloc[idx,0]\n img = self.data[idx].astype(np.uint8)\n img = self.transform(image=img)['image']\n return img, name\n```\n\nSuccessful Code\n```\nclass GraphemeDataset(Dataset):\n def __init__(self, fname, transform):\n self.df = pd.read_parquet(fname)\n self.data = self.df.iloc[:, 1:].values.reshape(-1, HEIGHT, WIDTH).astype(np.uint8)\n self.transform = transform\n\n def __len__(self):\n return len(self.data)\n\n def __getitem__(self, idx):\n name = self.df.iloc[idx,0]\n try:\n img = 255 - self.data[idx].astype(np.uint8)\n img = self.transform(image=img)['image']\n except:\n img = 255 - self.data[idx].astype(np.uint8)\n img = backup_tsfm(image=img)['image']\n return img, name\n```\n\nStill investigating why crop-resize fails, though. Any hypotheses?\nHere is the code for my crop-resize and transformation\n```\ndef bbox(img):\n rows = np.any(img, axis=1)\n cols = np.any(img, axis=0)\n rmin, rmax = np.where(rows)[0][[0, -1]]\n cmin, cmax = np.where(cols)[0][[0, -1]]\n return rmin, rmax, cmin, cmax\n\ndef crop_resize(img0, size=SIZE, pad=8, cols=None, rows=None, force_apply=False):\n #crop a box around pixels large than the threshold \n #some images contain line at the sides\n ymin,ymax,xmin,xmax = bbox(img0[5:-5,5:-5] > 80)\n #cropping may cut too much, so we need to add it back\n xmin = xmin - 13 if (xmin > 13) else 0\n ymin = ymin - 10 if (ymin > 10) else 0\n xmax = xmax + 13 if (xmax < WIDTH - 13) else WIDTH\n ymax = ymax + 10 if (ymax < HEIGHT - 10) else HEIGHT\n img = img0[ymin:ymax,xmin:xmax]\n #remove lo intensity pixels as noise\n img[img < 28] = 0\n lx, ly = xmax-xmin,ymax-ymin\n l = max(lx,ly) + pad\n #make sure that the aspect ratio is kept in rescaling\n img = np.pad(img, [((l-ly)//2,), ((l-lx)//2,)], mode='constant')\n return cv2.resize(img,(size,size))\n\ntsfm = Compose([\n Lambda(image=crop_resize),\n Resize(SIZE, SIZE),\n Normalize((.0855,),(.2,)),\n ToTensor(),\n ])\n```", "votes": null }, { "id": "724963", "postDate": "01/21/2020 16:41:18", "content": "it is really time consuming to debug uninformative errors like these given in the competition. we are data scientists not alchemists
", "rawMarkdown": "it is really time consuming to debug uninformative errors like these given in the competition. we are data scientists not alchemists", "votes": null }, { "id": "725297", "postDate": "01/22/2020 01:25:44", "content": "@ma7555 My suggestion is that you start from an easy-to-understand successful inference kernel then change it to your own submission kernel step-by-step. I did this for 8 times and finally succeeded. In retrospect it was only 2 days, though (8 submissions).
", "rawMarkdown": "ma7555 My suggestion is that you start from an easy-to-understand successful inference kernel then change it to your own submission kernel step-by-step. I did this for 8 times and finally succeeded. In retrospect it was only 2 days, though (8 submissions).", "votes": null }, { "id": "725714", "postDate": "01/22/2020 12:28:19", "content": "Same case for me as well but with an additional error. Not sure what I did wrong as commit only take about 400-500s.\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F2116526%2Fc10a4c43fbda19be97406fd98a2b7c00%2FScreenshot%202020-01-21%20at%2012.38.23%20PM.png?generation=1579696072557804&alt=media\")
I ran into this myself -- my best guess is there must be some blank or extremely faint image in the test set, which causes the indexing in bbox to fail, since both rows and cols would be empty. I seem to have solved this by adding simple checks in bbox. Again frustrating to fix, but that's what I get for copying over code without thinking hard enough :P
welcome board.✌️
", "rawMarkdown": "welcome board.✌️", "votes": null }, { "id": "730084", "postDate": "01/27/2020 05:04:05", "content": "@roshanpadaki but how could inference succeed here?\nhttps://www.kaggle.com/iafoss/grapheme-fast-ai-starter-inference
", "rawMarkdown": "roshanpadaki but how could inference succeed here?\nhttps://www.kaggle.com/iafoss/grapheme-fast-ai-starter-inference", "votes": null }, { "id": "730133", "postDate": "01/27/2020 06:17:13", "content": "I am wondering about the same thing, as fixing this problem could potentially help LB. Maybe because I changed some parameters? Did you change any? @roshanpadaki
", "rawMarkdown": "I am wondering about the same thing, as fixing this problem could potentially help LB. Maybe because I changed some parameters? Did you change any? @roshanpadaki", "votes": null }, { "id": "732689", "postDate": "01/30/2020 05:38:22", "content": "Could I know what you did on backup_tsfm?\nJust resize image and do data augmentation?\nLooking forward to your reply
", "rawMarkdown": "Could I know what you did on backup_tsfm?\nJust resize image and do data augmentation?\nLooking forward to your reply", "votes": null }, { "id": "735742", "postDate": "02/03/2020 12:10:07", "content": "@roguekk007 I think I understand now why his inference kernel works. simply because he normalized the image before calling the crop_resize so this fixed the too faint image problem...
", "rawMarkdown": "roguekk007 I think I understand now why his inference kernel works. simply because he normalized the image before calling the crop_resize so this fixed the *too faint image* problem...", "votes": null }, { "id": "744275", "postDate": "02/12/2020 17:35:48", "content": "I've been having huge issues with this and I couldn't even submit 1's. So far I found that when I go to reshape the data, that seems to be the breaking point for me. I created a public notebook related to my findings.\nhttps://www.kaggle.com/yeayates21/cant-submit-1s-whats-wrong-please-help
", "rawMarkdown": "I've been having huge issues with this and I couldn't even submit 1's. So far I found that when I go to reshape the data, that seems to be the breaking point for me. I created a public notebook related to my findings.\nhttps://www.kaggle.com/yeayates21/cant-submit-1s-whats-wrong-please-help", "votes": null }, { "id": "744280", "postDate": "02/12/2020 17:40:01", "content": "I think I came to the same conclusion that reshaping the parquet data causes submission to fail. I've been having huge issues with this and I couldn't even submit 1's (so no model at all). So far I found that when I go to reshape the data, that seems to be the breaking point for me. I created a public notebook related to my findings.\nhttps://www.kaggle.com/yeayates21/cant-submit-1s-whats-wrong-please-help
", "rawMarkdown": "I think I came to the same conclusion that reshaping the parquet data causes submission to fail. I've been having huge issues with this and I couldn't even submit 1's (so no model at all). So far I found that when I go to reshape the data, that seems to be the breaking point for me. I created a public notebook related to my findings.\nhttps://www.kaggle.com/yeayates21/cant-submit-1s-whats-wrong-please-help", "votes": null } ], "comments": [ { "id": 722247, "author_name": "pestipeti", "author_url": "", "post_date": "01/18/2020 10:04:54", "content": "This code is not optimized yet. You probably ran out of memory or time. Try a test commit with the train parquet files to see where is the issue.
\n\nYou can find my optimization tips here. (Implementation notebook included)
", "votes": null, "replies": [ { "id": 722259, "author_name": "roguekk007", "author_url": "", "post_date": "01/18/2020 10:23:31", "content": "Hi Peter, thanks for your answer! I was trying to keep the code minimal. It is probably not a timeout error I have received an explicit \"Notebook timeout error\" before. The \"Did not find CSV\" error is really confusing. I have read your optimization tips and found them very useful :)
", "votes": null, "replies": [] }, { "id": 723558, "author_name": "anirbank", "author_url": "", "post_date": "01/20/2020 06:42:05", "content": "In order to avoid memory issues, i first saved the model as a h5 file. Now in my kernel i am loading this model making a prediction and submitting, even that results in Error, and no log is provided. Can someone help pls.
", "votes": null, "replies": [] } ] }, { "id": 722354, "author_name": "roguekk007", "author_url": "", "post_date": "01/18/2020 13:01:34", "content": "There, I also managed to get a submission scoring error...I think I just obtained a collection of all make-able kaggle submission errors in my log now :)
", "votes": null, "replies": [ { "id": 724963, "author_name": "ma7555", "author_url": "", "post_date": "01/21/2020 16:41:18", "content": "it is really time consuming to debug uninformative errors like these given in the competition. we are data scientists not alchemists
", "votes": null, "replies": [] }, { "id": 725297, "author_name": "roguekk007", "author_url": "", "post_date": "01/22/2020 01:25:44", "content": "@ma7555 My suggestion is that you start from an easy-to-understand successful inference kernel then change it to your own submission kernel step-by-step. I did this for 8 times and finally succeeded. In retrospect it was only 2 days, though (8 submissions).
", "votes": null, "replies": [] } ] }, { "id": 722872, "author_name": "nroman", "author_url": "", "post_date": "01/19/2020 08:30:28", "content": "This competition has two challenges. First of them is to make a submission :)
\n\n
There is no way to find out what exact exception was raised during submission. And I cannot reproduce it, since kernel's commit is successful. It would be really helpful to show the stack trace in this case, since there is no way to debug this issue.
\n\n", "votes": null, "replies": [] }, { "id": 723428, "author_name": "roguekk007", "author_url": "", "post_date": "01/20/2020 02:52:31", "content": "Exactly the same here! My commit kernel produces a valid submission but I got a \"submission CSV not found\" error. Debugging the submission process is really frustrating.\nLooking forward to solving this problem.\n@addisonhoward @sohier
\n\nHere is the tragic submission log for me\n
Hi! I solved my submission error by doing 9 controlled experiments (yes, I failed 8 times) and 29 previously failed submissions👋 . Seems like the first challenge in the competition is indeed making a successful submission :) Here is what worked for me, hope it helps.\n I found that the crop_resize function (which I edited a little) and transformation led to failure. I added a backup transform without crop_resize with try:...except:... and solved the problem. Single model which achieves .972 CV leads to my current position.
\n\nOriginal Code (fails)\n```\nclass GraphemeDataset(Dataset):\n def init(self, fname, transform):\n self.df = pd.read_parquet(fname)\n self.data = 255 - self.df.iloc[:, 1:].values.reshape(-1, HEIGHT, WIDTH).astype(np.uint8)\n self.transform = transform
\n\ndef __len__(self):\n return len(self.data)\n\ndef __getitem__(self, idx):\n name = self.df.iloc[idx,0]\n img = self.data[idx].astype(np.uint8)\n img = self.transform(image=img)['image']\n return img, name\n```
\n\nSuccessful Code\n```\nclass GraphemeDataset(Dataset):\n def init(self, fname, transform):\n self.df = pd.read_parquet(fname)\n self.data = self.df.iloc[:, 1:].values.reshape(-1, HEIGHT, WIDTH).astype(np.uint8)\n self.transform = transform
\n\ndef __len__(self):\n return len(self.data)\n\ndef __getitem__(self, idx):\n name = self.df.iloc[idx,0]\n try:\n img = 255 - self.data[idx].astype(np.uint8)\n img = self.transform(image=img)['image']\n except:\n img = 255 - self.data[idx].astype(np.uint8)\n img = backup_tsfm(image=img)['image']\n return img, name\n```
\n\nStill investigating why crop-resize fails, though. Any hypotheses?\nHere is the code for my crop-resize and transformation\n```\ndef bbox(img):\n rows = np.any(img, axis=1)\n cols = np.any(img, axis=0)\n rmin, rmax = np.where(rows)[0][[0, -1]]\n cmin, cmax = np.where(cols)[0][[0, -1]]\n return rmin, rmax, cmin, cmax
\n\ndef crop_resize(img0, size=SIZE, pad=8, cols=None, rows=None, force_apply=False):\n #crop a box around pixels large than the threshold \n #some images contain line at the sides\n ymin,ymax,xmin,xmax = bbox(img0[5:-5,5:-5] > 80)\n #cropping may cut too much, so we need to add it back\n xmin = xmin - 13 if (xmin > 13) else 0\n ymin = ymin - 10 if (ymin > 10) else 0\n xmax = xmax + 13 if (xmax < WIDTH - 13) else WIDTH\n ymax = ymax + 10 if (ymax < HEIGHT - 10) else HEIGHT\n img = img0[ymin:ymax,xmin:xmax]\n #remove lo intensity pixels as noise\n img[img < 28] = 0\n lx, ly = xmax-xmin,ymax-ymin\n l = max(lx,ly) + pad\n #make sure that the aspect ratio is kept in rescaling\n img = np.pad(img, [((l-ly)//2,), ((l-lx)//2,)], mode='constant')\n return cv2.resize(img,(size,size))
\n\ntsfm = Compose([\n Lambda(image=crop_resize),\n Resize(SIZE, SIZE),\n Normalize((.0855,),(.2,)),\n ToTensor(),\n ])\n```
", "votes": null, "replies": [ { "id": 727842, "author_name": "roshanpadaki", "author_url": "", "post_date": "01/24/2020 05:33:55", "content": "I ran into this myself -- my best guess is there must be some blank or extremely faint image in the test set, which causes the indexing in bbox to fail, since both rows and cols would be empty. I seem to have solved this by adding simple checks in bbox. Again frustrating to fix, but that's what I get for copying over code without thinking hard enough :P
@roshanpadaki but how could inference succeed here?\nhttps://www.kaggle.com/iafoss/grapheme-fast-ai-starter-inference
", "votes": null, "replies": [] }, { "id": 730133, "author_name": "roguekk007", "author_url": "", "post_date": "01/27/2020 06:17:13", "content": "I am wondering about the same thing, as fixing this problem could potentially help LB. Maybe because I changed some parameters? Did you change any? @roshanpadaki
", "votes": null, "replies": [] }, { "id": 732689, "author_name": "kasim0226", "author_url": "", "post_date": "01/30/2020 05:38:22", "content": "Could I know what you did on backup_tsfm?\nJust resize image and do data augmentation?\nLooking forward to your reply
", "votes": null, "replies": [] }, { "id": 744280, "author_name": "yeayates21", "author_url": "", "post_date": "02/12/2020 17:40:01", "content": "I think I came to the same conclusion that reshaping the parquet data causes submission to fail. I've been having huge issues with this and I couldn't even submit 1's (so no model at all). So far I found that when I go to reshape the data, that seems to be the breaking point for me. I created a public notebook related to my findings.\nhttps://www.kaggle.com/yeayates21/cant-submit-1s-whats-wrong-please-help
", "votes": null, "replies": [] } ] }, { "id": 725714, "author_name": "gaur128", "author_url": "", "post_date": "01/22/2020 12:28:19", "content": "Same case for me as well but with an additional error. Not sure what I did wrong as commit only take about 400-500s.\n
I've been having huge issues with this and I couldn't even submit 1's. So far I found that when I go to reshape the data, that seems to be the breaking point for me. I created a public notebook related to my findings.\nhttps://www.kaggle.com/yeayates21/cant-submit-1s-whats-wrong-please-help
", "votes": null, "replies": [] } ] }, { "id": 728727, "author_name": "tiandaye", "author_url": "", "post_date": "01/25/2020 06:43:48", "content": "welcome board.✌️
", "votes": null, "replies": [] }, { "id": 735742, "author_name": "ma7555", "author_url": "", "post_date": "02/03/2020 12:10:07", "content": "@roguekk007 I think I understand now why his inference kernel works. simply because he normalized the image before calling the crop_resize so this fixed the too faint image problem...
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "722241": "Hi, everyone! Been struggling to make a successful submission. There are the timeouts, kaggle error, etc. So I decide to give the discussion a try.\n\nHere is a kernel I forked and edited. \nhttps://www.kaggle.com/roguekk007/another-stab\nThe resnet34 achieves .970 CV locally, but when running it I got **submission CSV not found**. I am running at 224x224 resolution which I found helps. \n\nCommiting the kernel gives me a submission.csv. Why is there this error?\nAny help is appreciated, thanks👍", "722247": "This code is not optimized yet. You probably ran out of memory or time. Try a test commit with the train parquet files to see where is the issue. \n\nYou can find my optimization tips [here](https://www.kaggle.com/c/bengaliai-cv19/discussion/126054). (Implementation notebook included)", "722259": "Hi Peter, thanks for your answer! I was trying to keep the code minimal. It is probably not a timeout error I have received an explicit \"Notebook timeout error\" before. The \"Did not find CSV\" error is really confusing. I have read your optimization tips and found them very useful :)", "722354": "There, I also managed to get a submission scoring error...I think I just obtained a collection of all make-able kaggle submission errors in my log now :)", "722872": "This competition has two challenges. First of them is to make a submission :)\n\n", "723141": "There is no way to find out what exact exception was raised during submission. And I cannot reproduce it, since kernel's commit is successful. It would be really helpful to show the stack trace in this case, since there is no way to debug this issue.\n\n@addisonhoward @sohier", "723428": "Exactly the same here! My commit kernel produces a valid submission but I got a \"submission CSV not found\" error. Debugging the submission process is really frustrating.\nLooking forward to solving this problem.\n@addisonhoward @sohier \n\nHere is the tragic submission log for me\n", "723558": "In order to avoid memory issues, i first saved the model as a h5 file. Now in my kernel i am loading this model making a prediction and submitting, even that results in Error, and no log is provided. Can someone help pls.", "724318": "Hi! I solved my submission error by doing 9 controlled experiments (yes, I failed 8 times) and 29 previously failed submissions👋 . Seems like the first challenge in the competition is indeed making a successful submission :) Here is what worked for me, hope it helps.\n I found that the crop_resize function (which I edited a little) and transformation led to failure. **I added a backup transform without crop_resize with try:...except:... and solved the problem.** Single model which achieves .972 CV leads to my current position.\n\nOriginal Code (fails)\n```\nclass GraphemeDataset(Dataset):\n def __init__(self, fname, transform):\n self.df = pd.read_parquet(fname)\n self.data = 255 - self.df.iloc[:, 1:].values.reshape(-1, HEIGHT, WIDTH).astype(np.uint8)\n self.transform = transform\n\n def __len__(self):\n return len(self.data)\n\n def __getitem__(self, idx):\n name = self.df.iloc[idx,0]\n img = self.data[idx].astype(np.uint8)\n img = self.transform(image=img)['image']\n return img, name\n```\n\nSuccessful Code\n```\nclass GraphemeDataset(Dataset):\n def __init__(self, fname, transform):\n self.df = pd.read_parquet(fname)\n self.data = self.df.iloc[:, 1:].values.reshape(-1, HEIGHT, WIDTH).astype(np.uint8)\n self.transform = transform\n\n def __len__(self):\n return len(self.data)\n\n def __getitem__(self, idx):\n name = self.df.iloc[idx,0]\n try:\n img = 255 - self.data[idx].astype(np.uint8)\n img = self.transform(image=img)['image']\n except:\n img = 255 - self.data[idx].astype(np.uint8)\n img = backup_tsfm(image=img)['image']\n return img, name\n```\n\nStill investigating why crop-resize fails, though. Any hypotheses?\nHere is the code for my crop-resize and transformation\n```\ndef bbox(img):\n rows = np.any(img, axis=1)\n cols = np.any(img, axis=0)\n rmin, rmax = np.where(rows)[0][[0, -1]]\n cmin, cmax = np.where(cols)[0][[0, -1]]\n return rmin, rmax, cmin, cmax\n\ndef crop_resize(img0, size=SIZE, pad=8, cols=None, rows=None, force_apply=False):\n #crop a box around pixels large than the threshold \n #some images contain line at the sides\n ymin,ymax,xmin,xmax = bbox(img0[5:-5,5:-5] > 80)\n #cropping may cut too much, so we need to add it back\n xmin = xmin - 13 if (xmin > 13) else 0\n ymin = ymin - 10 if (ymin > 10) else 0\n xmax = xmax + 13 if (xmax < WIDTH - 13) else WIDTH\n ymax = ymax + 10 if (ymax < HEIGHT - 10) else HEIGHT\n img = img0[ymin:ymax,xmin:xmax]\n #remove lo intensity pixels as noise\n img[img < 28] = 0\n lx, ly = xmax-xmin,ymax-ymin\n l = max(lx,ly) + pad\n #make sure that the aspect ratio is kept in rescaling\n img = np.pad(img, [((l-ly)//2,), ((l-lx)//2,)], mode='constant')\n return cv2.resize(img,(size,size))\n\ntsfm = Compose([\n Lambda(image=crop_resize),\n Resize(SIZE, SIZE),\n Normalize((.0855,),(.2,)),\n ToTensor(),\n ])\n```", "724963": "it is really time consuming to debug uninformative errors like these given in the competition. we are data scientists not alchemists", "725297": "ma7555 My suggestion is that you start from an easy-to-understand successful inference kernel then change it to your own submission kernel step-by-step. I did this for 8 times and finally succeeded. In retrospect it was only 2 days, though (8 submissions).", "725714": "Same case for me as well but with an additional error. Not sure what I did wrong as commit only take about 400-500s.\n", "727842": "I ran into this myself -- my best guess is there must be some blank or extremely faint image in the test set, which causes the indexing in `bbox` to fail, since both `rows` and `cols` would be empty. I seem to have solved this by adding simple checks in `bbox`. Again frustrating to fix, but that's what I get for copying over code without thinking hard enough :P", "728727": "welcome board.✌️", "730084": "roshanpadaki but how could inference succeed here?\nhttps://www.kaggle.com/iafoss/grapheme-fast-ai-starter-inference", "730133": "I am wondering about the same thing, as fixing this problem could potentially help LB. Maybe because I changed some parameters? Did you change any? @roshanpadaki", "732689": "Could I know what you did on backup_tsfm?\nJust resize image and do data augmentation?\nLooking forward to your reply", "735742": "roguekk007 I think I understand now why his inference kernel works. simply because he normalized the image before calling the crop_resize so this fixed the *too faint image* problem...", "744275": "I've been having huge issues with this and I couldn't even submit 1's. So far I found that when I go to reshape the data, that seems to be the breaking point for me. I created a public notebook related to my findings.\nhttps://www.kaggle.com/yeayates21/cant-submit-1s-whats-wrong-please-help", "744280": "I think I came to the same conclusion that reshaping the parquet data causes submission to fail. I've been having huge issues with this and I couldn't even submit 1's (so no model at all). So far I found that when I go to reshape the data, that seems to be the breaking point for me. I created a public notebook related to my findings.\nhttps://www.kaggle.com/yeayates21/cant-submit-1s-whats-wrong-please-help" }, "source": "meta" }