Re-submit the kernel may help. It may not cause by your code.
", "rawMarkdown": "Re-submit the kernel may help. It may not cause by your code.", "votes": null }, { "id": "589675", "postDate": "08/01/2019 08:21:58", "content": "when I delete the TTA code, I can commit success, Addition, I try three model with TTA, unfortunately，they all fail, it make me anxious that waste my 3 times commit ,Do you meet the same problem?? thank you for your reply
", "rawMarkdown": "when I delete the TTA code, I can commit success, Addition, I try three model with TTA, unfortunately，they all fail, it make me anxious that waste my 3 times commit ,Do you meet the same problem?? thank you for your reply", "votes": null }, { "id": "589686", "postDate": "08/01/2019 08:42:31", "content": "No I haven't try TTA yet. Can you show your TTA part code here?
", "rawMarkdown": "No I haven't try TTA yet. Can you show your TTA part code here?", "votes": null }, { "id": "589708", "postDate": "08/01/2019 09:16:15", "content": "TTA_steps = 10
\n\ntest_datagen = ImageDataGenerator(\n featurewise_std_normalization= True,\n zoom_range=0.15,\n fill_mode='constant',\n cval=0., # value used for fill_mode=\"constant\"\n horizontal_flip = True,\n vertical_flip = True,\n rotation_range=360\n )
\n\npredictions = []\ntest_generator = test_datagen.flow(X_test, batch_size=16, shuffle=False)
\n\nfor i in range(TTA_steps): \n print('**TTA_{}'.format(i))\n test_generator.reset()\n preds = model.predict_generator(test_generator, steps=len(X_test)/16)\n predictions.append(preds)\ny_test = np.mean(predictions, axis=0)
\n\ny_test = np.argmax(y_test, axis=1)\ntest_df['diagnosis'] = y_test\ntest_df.to_csv('submission.csv',index=False)\nprint(test_df['diagnosis'].value_counts())\ntest_df['diagnosis'].hist()
", "rawMarkdown": "TTA_steps = 10\n\ntest_datagen = ImageDataGenerator(\n featurewise_std_normalization= True,\n zoom_range=0.15,\n fill_mode='constant',\n cval=0., # value used for fill_mode=\"constant\"\n horizontal_flip = True,\n vertical_flip = True,\n rotation_range=360\n )\n\npredictions = []\ntest_generator = test_datagen.flow(X_test, batch_size=16, shuffle=False)\n\nfor i in range(TTA_steps): \n print('********TTA_{}'.format(i))\n test_generator.reset()\n preds = model.predict_generator(test_generator, steps=len(X_test)/16)\n predictions.append(preds)\ny_test = np.mean(predictions, axis=0)\n\ny_test = np.argmax(y_test, axis=1)\ntest_df['diagnosis'] = y_test\ntest_df.to_csv('submission.csv',index=False)\nprint(test_df['diagnosis'].value_counts())\ntest_df['diagnosis'].hist()", "votes": null }, { "id": "589711", "postDate": "08/01/2019 09:17:17", "content": "thank your help me sinerely
", "rawMarkdown": "thank your help me sinerely", "votes": null }, { "id": "589742", "postDate": "08/01/2019 09:53:33", "content": "By the way, what's your commit kernel time cost?\nIn my experience, I got some submit problem when commit kernel time cost more than 400 secs.
", "rawMarkdown": "By the way, what's your commit kernel time cost?\nIn my experience, I got some submit problem when commit kernel time cost more than 400 secs.", "votes": null }, { "id": "589783", "postDate": "08/01/2019 11:18:13", "content": "I run code in kernel since this is a Kernel-only. Approximately, I spend 1 hour when I commit the kernel every times
", "rawMarkdown": "I run code in kernel since this is a Kernel-only. Approximately, I spend 1 hour when I commit the kernel every times", "votes": null }, { "id": "590068", "postDate": "08/01/2019 19:04:41", "content": "I had the exact same problem. The issue is that the private test dataset (the one used for the LB score) is ~6 times larger than the public one, that's why you can run it on your kernel but not on the private dataset. In my case this caused a memory overflow because I was using TTA with flow (which loads all the data in memory) instead of using flow_from_dataframe.
This was the code that was causing the submission error: \n```\ndef TTA(model, generator, images, n_examples=10):\n val_predictions = []\n val_generator = generator.flow(images, shuffle=False, batch_size=1)
\n\nfor i in tqdm(range(n_examples)):\n val_generator.reset()\n val_preds = model.predict_generator(\n val_generator, \n steps=images.shape[0]\n ) \n val_predictions.append(val_preds)\n```\nand this was the fix:
\n\n\ndef TTA(model, data_generator, images_df, n_examples=10):\n predictions = []\n generator = data_generator.flow_from_dataframe(\n dataframe=images_df,\n directory='../input/aptos2019-blindness-detection/test_images/',\n x_col = 'id_code',\n y_col=None,\n target_size = IMAGE_SIZE,\n shuffle=False, \n batch_size=1,\n seed=2019,\n class_mode=None)\n
I hope it helps! :)
", "rawMarkdown": "I had the exact same problem. The issue is that the private test dataset (the one used for the LB score) is ~6 times larger than the public one, that's why you can run it on your kernel but not on the private dataset. In my case this caused a memory overflow because I was using TTA with `flow` (which loads all the data in memory) instead of using `flow_from_dataframe`.\n\nThis was the code that was causing the submission error: \n```\ndef TTA(model, generator, images, n_examples=10):\n val_predictions = []\n val_generator = generator.flow(images, shuffle=False, batch_size=1)\n\n for i in tqdm(range(n_examples)):\n val_generator.reset()\n val_preds = model.predict_generator(\n val_generator, \n steps=images.shape[0]\n ) \n val_predictions.append(val_preds)\n```\nand this was the fix:\n\n```\ndef TTA(model, data_generator, images_df, n_examples=10):\n predictions = []\n generator = data_generator.flow_from_dataframe(\n dataframe=images_df,\n directory='../input/aptos2019-blindness-detection/test_images/',\n x_col = 'id_code',\n y_col=None,\n target_size = IMAGE_SIZE,\n shuffle=False, \n batch_size=1,\n seed=2019,\n class_mode=None)\n```\n\nI hope it helps! :)", "votes": null }, { "id": "590266", "postDate": "08/02/2019 01:15:56", "content": "Thank you! I will try it, hope it work!!
", "rawMarkdown": "Thank you! I will try it, hope it work!!", "votes": null }, { "id": "590301", "postDate": "08/02/2019 03:21:47", "content": "Additionlly, I have another question that the train data should be deal as same as the test data?
", "rawMarkdown": "Additionlly, I have another question that the train data should be deal as same as the test data?", "votes": null }, { "id": "590415", "postDate": "08/02/2019 07:06:18", "content": "In my case It wasn't necessary because the train set is ~3600 images and when I loaded it into an array and trained with flow it could fit in memory (my image size was 224).
as you means, I can use flow( ) function to deal train data,and use flow_from_dataframe() function deal test data , that guarantee success commit and avoid \"Submissions CSV not found\". (I am now try the flow_from_dataframe() to deal test Images to implement TTA). hope it work, and thank you sinerely, I am a beginner
", "rawMarkdown": "as you means, I can use flow( ) function to deal train data,and use flow_from_dataframe() function deal test data , that guarantee success commit and avoid \"Submissions CSV not found\". (I am now try the flow_from_dataframe() to deal test Images to implement TTA). hope it work, and thank you sinerely, I am a beginner", "votes": null } ], "comments": [ { "id": 589664, "author_name": "haqishen", "author_url": "", "post_date": "08/01/2019 08:10:18", "content": "Re-submit the kernel may help. It may not cause by your code.
", "votes": null, "replies": [ { "id": 589675, "author_name": "jiangkun2", "author_url": "", "post_date": "08/01/2019 08:21:58", "content": "when I delete the TTA code, I can commit success, Addition, I try three model with TTA, unfortunately，they all fail, it make me anxious that waste my 3 times commit ,Do you meet the same problem?? thank you for your reply
", "votes": null, "replies": [] }, { "id": 589686, "author_name": "haqishen", "author_url": "", "post_date": "08/01/2019 08:42:31", "content": "No I haven't try TTA yet. Can you show your TTA part code here?
", "votes": null, "replies": [] }, { "id": 589708, "author_name": "jiangkun2", "author_url": "", "post_date": "08/01/2019 09:16:15", "content": "TTA_steps = 10
\n\ntest_datagen = ImageDataGenerator(\n featurewise_std_normalization= True,\n zoom_range=0.15,\n fill_mode='constant',\n cval=0., # value used for fill_mode=\"constant\"\n horizontal_flip = True,\n vertical_flip = True,\n rotation_range=360\n )
\n\npredictions = []\ntest_generator = test_datagen.flow(X_test, batch_size=16, shuffle=False)
\n\nfor i in range(TTA_steps): \n print('**TTA_{}'.format(i))\n test_generator.reset()\n preds = model.predict_generator(test_generator, steps=len(X_test)/16)\n predictions.append(preds)\ny_test = np.mean(predictions, axis=0)
\n\ny_test = np.argmax(y_test, axis=1)\ntest_df['diagnosis'] = y_test\ntest_df.to_csv('submission.csv',index=False)\nprint(test_df['diagnosis'].value_counts())\ntest_df['diagnosis'].hist()
", "votes": null, "replies": [] }, { "id": 589711, "author_name": "jiangkun2", "author_url": "", "post_date": "08/01/2019 09:17:17", "content": "thank your help me sinerely
", "votes": null, "replies": [] }, { "id": 589742, "author_name": "haqishen", "author_url": "", "post_date": "08/01/2019 09:53:33", "content": "By the way, what's your commit kernel time cost?\nIn my experience, I got some submit problem when commit kernel time cost more than 400 secs.
", "votes": null, "replies": [] }, { "id": 589783, "author_name": "jiangkun2", "author_url": "", "post_date": "08/01/2019 11:18:13", "content": "I run code in kernel since this is a Kernel-only. Approximately, I spend 1 hour when I commit the kernel every times
", "votes": null, "replies": [] } ] }, { "id": 590068, "author_name": "sam1320", "author_url": "", "post_date": "08/01/2019 19:04:41", "content": "I had the exact same problem. The issue is that the private test dataset (the one used for the LB score) is ~6 times larger than the public one, that's why you can run it on your kernel but not on the private dataset. In my case this caused a memory overflow because I was using TTA with flow (which loads all the data in memory) instead of using flow_from_dataframe.
This was the code that was causing the submission error: \n```\ndef TTA(model, generator, images, n_examples=10):\n val_predictions = []\n val_generator = generator.flow(images, shuffle=False, batch_size=1)
\n\nfor i in tqdm(range(n_examples)):\n val_generator.reset()\n val_preds = model.predict_generator(\n val_generator, \n steps=images.shape[0]\n ) \n val_predictions.append(val_preds)\n```\nand this was the fix:
\n\n\ndef TTA(model, data_generator, images_df, n_examples=10):\n predictions = []\n generator = data_generator.flow_from_dataframe(\n dataframe=images_df,\n directory='../input/aptos2019-blindness-detection/test_images/',\n x_col = 'id_code',\n y_col=None,\n target_size = IMAGE_SIZE,\n shuffle=False, \n batch_size=1,\n seed=2019,\n class_mode=None)\n
I hope it helps! :)
", "votes": null, "replies": [ { "id": 590266, "author_name": "jiangkun2", "author_url": "", "post_date": "08/02/2019 01:15:56", "content": "Thank you! I will try it, hope it work!!
", "votes": null, "replies": [] }, { "id": 590301, "author_name": "jiangkun2", "author_url": "", "post_date": "08/02/2019 03:21:47", "content": "Additionlly, I have another question that the train data should be deal as same as the test data?
", "votes": null, "replies": [] }, { "id": 590415, "author_name": "sam1320", "author_url": "", "post_date": "08/02/2019 07:06:18", "content": "In my case It wasn't necessary because the train set is ~3600 images and when I loaded it into an array and trained with flow it could fit in memory (my image size was 224).
as you means, I can use flow( ) function to deal train data,and use flow_from_dataframe() function deal test data , that guarantee success commit and avoid \"Submissions CSV not found\". (I am now try the flow_from_dataframe() to deal test Images to implement TTA). hope it work, and thank you sinerely, I am a beginner
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "589627": "Batchsize:16\n- image size:256", "589664": "Re-submit the kernel may help. It may not cause by your code.", "589675": "when I delete the TTA code, I can commit success, Addition, I try three model with TTA, unfortunately，they all fail, it make me anxious that waste my 3 times commit ,Do you meet the same problem?? thank you for your reply", "589686": "No I haven't try TTA yet. Can you show your TTA part code here?", "589708": "TTA_steps = 10\n\ntest_datagen = ImageDataGenerator(\n featurewise_std_normalization= True,\n zoom_range=0.15,\n fill_mode='constant',\n cval=0., # value used for fill_mode=\"constant\"\n horizontal_flip = True,\n vertical_flip = True,\n rotation_range=360\n )\n\npredictions = []\ntest_generator = test_datagen.flow(X_test, batch_size=16, shuffle=False)\n\nfor i in range(TTA_steps): \n print('********TTA_{}'.format(i))\n test_generator.reset()\n preds = model.predict_generator(test_generator, steps=len(X_test)/16)\n predictions.append(preds)\ny_test = np.mean(predictions, axis=0)\n\ny_test = np.argmax(y_test, axis=1)\ntest_df['diagnosis'] = y_test\ntest_df.to_csv('submission.csv',index=False)\nprint(test_df['diagnosis'].value_counts())\ntest_df['diagnosis'].hist()", "589711": "thank your help me sinerely", "589742": "By the way, what's your commit kernel time cost?\nIn my experience, I got some submit problem when commit kernel time cost more than 400 secs.", "589783": "I run code in kernel since this is a Kernel-only. Approximately, I spend 1 hour when I commit the kernel every times", "590068": "I had the exact same problem. The issue is that the private test dataset (the one used for the LB score) is ~6 times larger than the public one, that's why you can run it on your kernel but not on the private dataset. In my case this caused a memory overflow because I was using TTA with `flow` (which loads all the data in memory) instead of using `flow_from_dataframe`.\n\nThis was the code that was causing the submission error: \n```\ndef TTA(model, generator, images, n_examples=10):\n val_predictions = []\n val_generator = generator.flow(images, shuffle=False, batch_size=1)\n\n for i in tqdm(range(n_examples)):\n val_generator.reset()\n val_preds = model.predict_generator(\n val_generator, \n steps=images.shape[0]\n ) \n val_predictions.append(val_preds)\n```\nand this was the fix:\n\n```\ndef TTA(model, data_generator, images_df, n_examples=10):\n predictions = []\n generator = data_generator.flow_from_dataframe(\n dataframe=images_df,\n directory='../input/aptos2019-blindness-detection/test_images/',\n x_col = 'id_code',\n y_col=None,\n target_size = IMAGE_SIZE,\n shuffle=False, \n batch_size=1,\n seed=2019,\n class_mode=None)\n```\n\nI hope it helps! :)", "590266": "Thank you! I will try it, hope it work!!", "590301": "Additionlly, I have another question that the train data should be deal as same as the test data?", "590415": "In my case It wasn't necessary because the train set is ~3600 images and when I loaded it into an array and trained with `flow` it could fit in memory (my image size was 224).", "590438": "as you means, I can use flow( ) function to deal train data,and use flow_from_dataframe() function deal test data , that guarantee success commit and avoid \"Submissions CSV not found\". (I am now try the flow_from_dataframe() to deal test Images to implement TTA). hope it work, and thank you sinerely, I am a beginner" }, "source": "meta" }