Hello everyone! This looks like a challenging classification task due to the very large dataset and large number of classes.
\n\nThis is my code so far for this competition, I hope that it helps more people to join and that it provides a decent baseline to build a robust solution on:
\n\nhttps://github.com/petrosgk/Kaggle-Cdiscount-Image-Classification-Challenge
\n\nI am still processing the .bson files so no results on the full dataset yet. These are 5 epochs on a subset of 10k samples with pre-trained VGG16 (with top dense layers removed) and image size 180x180:
\n\nEpoch 1/20\n231s - loss: 3.7500 - acc: 0.4056 - val_loss: 3.2170 - val_acc: 0.4783\nEpoch 2/20\n227s - loss: 2.8793 - acc: 0.5010 - val_loss: 3.0092 - val_acc: 0.5149\nEpoch 3/20\n228s - loss: 2.4361 - acc: 0.5570 - val_loss: 2.9464 - val_acc: 0.5323\nEpoch 4/20\n229s - loss: 2.1199 - acc: 0.5996 - val_loss: 2.9545 - val_acc: 0.5447\nEpoch 5/20\n229s - loss: 1.8476 - acc: 0.6346 - val_loss: 2.9870 - val_acc: 0.5440\nThere is overfitting with 9k training samples (90/10 train-val split) but that may change when training on the full ~7m samples.
", "messages": [ { "id": 228504, "postDate": "2017-10-06T21:45:39.263Z", "content": "Hello everyone! This looks like a challenging classification task due to the very large dataset and large number of classes.
\n\nThis is my code so far for this competition, I hope that it helps more people to join and that it provides a decent baseline to build a robust solution on:
\n\nhttps://github.com/petrosgk/Kaggle-Cdiscount-Image-Classification-Challenge
\n\nI am still processing the .bson files so no results on the full dataset yet. These are 5 epochs on a subset of 10k samples with pre-trained VGG16 (with top dense layers removed) and image size 180x180:
\n\nEpoch 1/20\n231s - loss: 3.7500 - acc: 0.4056 - val_loss: 3.2170 - val_acc: 0.4783\nEpoch 2/20\n227s - loss: 2.8793 - acc: 0.5010 - val_loss: 3.0092 - val_acc: 0.5149\nEpoch 3/20\n228s - loss: 2.4361 - acc: 0.5570 - val_loss: 2.9464 - val_acc: 0.5323\nEpoch 4/20\n229s - loss: 2.1199 - acc: 0.5996 - val_loss: 2.9545 - val_acc: 0.5447\nEpoch 5/20\n229s - loss: 1.8476 - acc: 0.6346 - val_loss: 2.9870 - val_acc: 0.5440\nThere is overfitting with 9k training samples (90/10 train-val split) but that may change when training on the full ~7m samples.
", "rawMarkdown": "Hello everyone! This looks like a challenging classification task due to the very large dataset and large number of classes. \n\nThis is my code so far for this competition, I hope that it helps more people to join and that it provides a decent baseline to build a robust solution on:\n\nhttps://github.com/petrosgk/Kaggle-Cdiscount-Image-Classification-Challenge\n\nI am still processing the .bson files so no results on the full dataset yet. These are 5 epochs on a subset of 10k samples with pre-trained VGG16 (with top dense layers removed) and image size 180x180:\n\n Epoch 1/20\n 231s - loss: 3.7500 - acc: 0.4056 - val_loss: 3.2170 - val_acc: 0.4783\n Epoch 2/20\n 227s - loss: 2.8793 - acc: 0.5010 - val_loss: 3.0092 - val_acc: 0.5149\n Epoch 3/20\n 228s - loss: 2.4361 - acc: 0.5570 - val_loss: 2.9464 - val_acc: 0.5323\n Epoch 4/20\n 229s - loss: 2.1199 - acc: 0.5996 - val_loss: 2.9545 - val_acc: 0.5447\n Epoch 5/20\n 229s - loss: 1.8476 - acc: 0.6346 - val_loss: 2.9870 - val_acc: 0.5440\n\nThere is overfitting with 9k training samples (90/10 train-val split) but that may change when training on the full ~7m samples.", "votes": 18 }, { "id": 231530, "postDate": "2017-10-15T09:07:16.680Z", "content": "I suffered from training very slow when using mxnet. I found that if random read raw images, it is very slow. Now I use the rec file (it is a file generated from all images, hence is fast) of mxnet, it improves from 140 imgs/second to 600 imgs/second.
", "rawMarkdown": "I suffered from training very slow when using mxnet. I found that if random read raw images, it is very slow. Now I use the rec file (it is a file generated from all images, hence is fast) of mxnet, it improves from 140 imgs/second to 600 imgs/second.", "votes": 1 }, { "id": 233722, "postDate": "2017-10-20T23:07:54.557Z", "content": "I took at look at your Keras starter - a few good learning points, I did a trip to Keras.io a few times, I like the 'callbacks'. \nA few observations and a question: The png files generated for the train folder are in the hundreds of gigabytes (caveat emptor) so it goes without saying -training and manipulation- can take a very very long time to those stepping into this. Don't waste time using a standard HDD, nothing less than an m.2 SSD will do. Just doing a simple ls -la can take 5 minutes. My SSD which is much quicker unfortunately is too small to hold that much data! None-the-less what variables did you tinker with to run just a 10k sample using train.py? Usually I can spot this quickly but things look a little nested. As a base line I want to replicate your result as shown. Thus far I cannot. It would help square things up.
", "rawMarkdown": "I took at look at your Keras starter - a few good learning points, I did a trip to Keras.io a few times, I like the 'callbacks'. \nA few observations and a question: The png files generated for the train folder are in the hundreds of gigabytes (caveat emptor) so it goes without saying -training and manipulation- can take a very very long time to those stepping into this. Don't waste time using a standard HDD, nothing less than an m.2 SSD will do. Just doing a simple ls -la can take 5 minutes. My SSD which is much quicker unfortunately is too small to hold that much data! None-the-less what variables did you tinker with to run just a 10k sample using train.py? Usually I can spot this quickly but things look a little nested. As a base line I want to replicate your result as shown. Thus far I cannot. It would help square things up.", "replies": [ { "id": 233796, "postDate": "2017-10-21T09:17:18.270Z", "content": "In the code at the beginning of train.py that reads the product_id's, category_id's and num of images per product_id:
\n\ndf = pd.read_csv(csv_file)\nproduct_ids, category_ids, n_pics = df['product_id'], df['category_id'], df['n_pics']\nYou can simply only use the first 10k elements of the product_ids, category_ids, n_pics arrays. So you can add something like this:
\n\n product_ids = product_ids[:10000]\n category_ids = category_ids[:10000]\n n_pics = n_pics[:10000]\nPeter - much appreciated I will give that a try and see how things play out.
", "rawMarkdown": "Peter - much appreciated I will give that a try and see how things play out." }, { "id": 233983, "postDate": "2017-10-22T00:33:02.953Z", "content": "Peter - I am getting an error on line 55 stratify=category_ids resulting in this error output:
\n\nValueError: The least populated class in y has only 1 member, which is too few. The minimum number of groups for any class cannot be less than 2.
\n\nI had a play around with no effect, but I am running out of time this week (I need to leave town). My guess is its the train/validation split not going to plan with only 10000 values? Something else needs tweaking.
", "rawMarkdown": "Peter - I am getting an error on line 55 stratify=category_ids resulting in this error output:\n\nValueError: The least populated class in y has only 1 member, which is too few. The minimum number of groups for any class cannot be less than 2.\n\nI had a play around with no effect, but I am running out of time this week (I need to leave town). My guess is its the train/validation split not going to plan with only 10000 values? Something else needs tweaking.\n\n" }, { "id": 234068, "postDate": "2017-10-22T07:57:29.200Z", "content": "Your guess is right, sklearn complains about the stratified split with 10000 values because some classes end up with just 1 product. You can do a simple split instead of stratified by removing the 'stratify=category_ids' parameter and then it'll work.
", "rawMarkdown": "Your guess is right, sklearn complains about the stratified split with 10000 values because some classes end up with just 1 product. You can do a simple split instead of stratified by removing the 'stratify=category_ids' parameter and then it'll work." }, { "id": 236759, "postDate": "2017-10-28T03:42:26.420Z", "content": "I am getting a new error on line 63: 'for pic in range(n_pics_train[prod]):'
\n\nIts a key error and is a result of the length of the data changes made above. I have had a look at the variables but so far I could not locate the issue. So supposedly its got no key for some value I assume someplace in one of the created lists?
", "rawMarkdown": "I am getting a new error on line 63: 'for pic in range(n_pics_train[prod]):'\n\nIts a key error and is a result of the length of the data changes made above. I have had a look at the variables but so far I could not locate the issue. So supposedly its got no key for some value I assume someplace in one of the created lists?" }, { "id": 236787, "postDate": "2017-10-28T07:30:23.940Z", "content": "ok this is what I needed to do, 'print' couldn't save me this time but I figured it out below. BUT the valid generator function has 'blown a gasket' so I am trying to figure that out as it runs through the test and stops just on the last iteration!
\n\nprint('Creating training data:')\nfor prod in tqdm(range(len(product_ids_train))):\n for pic in range(n_pics_train[9000]):\n product_ids_pics_train_arr.append(str(product_ids_train[9000]) + '_' + str(pic))\n category_ids_train_arr.append(category_ids_train[9000])\nproduct_ids_pics_test_arr = []\ncategory_ids_test_arr = []\nprint('Creating validation data:')\nfor prod in tqdm(range(len(product_ids_test))):\n for pic in range(1000):\n product_ids_pics_test_arr.append(str(1000) + '_' + str(pic))\n category_ids_test_arr.append(1000)\ntrain_samples = len(product_ids_pics_train_arr)\ntest_samples = len(product_ids_pics_test_arr)
", "rawMarkdown": "ok this is what I needed to do, 'print' couldn't save me this time but I figured it out below. BUT the valid generator function has 'blown a gasket' so I am trying to figure that out as it runs through the test and stops just on the last iteration!\n\nprint('Creating training data:')\nfor prod in tqdm(range(len(product_ids_train))):\n for pic in range(n_pics_train[9000]):\n product_ids_pics_train_arr.append(str(product_ids_train[9000]) + '_' + str(pic))\n category_ids_train_arr.append(category_ids_train[9000])\nproduct_ids_pics_test_arr = []\ncategory_ids_test_arr = []\nprint('Creating validation data:')\nfor prod in tqdm(range(len(product_ids_test))):\n for pic in range(1000):\n product_ids_pics_test_arr.append(str(1000) + '_' + str(pic))\n category_ids_test_arr.append(1000)\ntrain_samples = len(product_ids_pics_train_arr)\ntest_samples = len(product_ids_pics_test_arr)" } ] }, { "id": 228768, "postDate": "2017-10-07T19:39:06.710Z", "content": "Hello, thanks for sharing
\n\nHow long should a VGG16 or inception v3 epoch take?
", "rawMarkdown": "Hello, thanks for sharing\n\nHow long should a VGG16 or inception v3 epoch take?", "replies": [ { "id": 228778, "postDate": "2017-10-07T20:30:14.043Z", "content": "I'm currently training on the full dataset. 1 epoch will take ~80 hours on my GTX 1070...
", "rawMarkdown": "I'm currently training on the full dataset. 1 epoch will take ~80 hours on my GTX 1070...", "votes": 1 }, { "id": 228963, "postDate": "2017-10-08T12:23:03.650Z", "content": "It is too slow. I think that you missed something important. Do you have a look the post :https://www.kaggle.com/humananalog/keras-generator-for-reading-directly-from-bson ? It used 8 workers to speed up the reading of datas and got a better speed. On my 1080ti gpu, it will take about 22 hours per epoch.
", "rawMarkdown": "It is too slow. I think that you missed something important. Do you have a look the post :https://www.kaggle.com/humananalog/keras-generator-for-reading-directly-from-bson ? It used 8 workers to speed up the reading of datas and got a better speed. On my 1080ti gpu, it will take about 22 hours per epoch. " }, { "id": 228967, "postDate": "2017-10-08T12:30:11.730Z", "content": "Perhaps a reason it's so slow for me is because i'm reading the data from HDD as I don't have enough space on my SSD's.
", "rawMarkdown": "Perhaps a reason it's so slow for me is because i'm reading the data from HDD as I don't have enough space on my SSD's." }, { "id": 228982, "postDate": "2017-10-08T13:14:54.343Z", "content": "for your information, this is also discussed here:
\n\nhttps://www.kaggle.com/c/cdiscount-image-classification-challenge/discussion/40498:
\n\n\"Hi, Heng, I use similar way to load data as you, converting bson to files.
\n\nThe loading process is very fast at first(0.3s, CPU 30%, RAM 10% of 30G), but some iterations later, it becomes very slow(6s, CPU 5%, RAM 10% of 30G).
\n\nI set the CPU mode to performance, and try different num_workers(0,4,8,12 …), but the situation stays the same. ...\"
", "rawMarkdown": "for your information, this is also discussed here:\n\nhttps://www.kaggle.com/c/cdiscount-image-classification-challenge/discussion/40498:\n\n\"Hi, Heng, I use similar way to load data as you, converting bson to files.\n\nThe loading process is very fast at first(0.3s, CPU 30%, RAM 10% of 30G), but some iterations later, it becomes very slow(6s, CPU 5%, RAM 10% of 30G).\n\nI set the CPU mode to performance, and try different num_workers(0,4,8,12 …), but the situation stays the same. ...\"", "votes": 3 } ] }, { "id": 228529, "postDate": "2017-10-07T00:24:41.793Z", "rawMarkdown": "", "votes": -3, "isDeleted": true } ], "comments": [ { "id": 231530, "author_name": "heroxrq", "author_url": "", "post_date": "2017-10-15T09:07:16.680000", "content": "I suffered from training very slow when using mxnet. I found that if random read raw images, it is very slow. Now I use the rec file (it is a file generated from all images, hence is fast) of mxnet, it improves from 140 imgs/second to 600 imgs/second.
", "votes": 1, "replies": [] }, { "id": 233722, "author_name": "stephenl", "author_url": "", "post_date": "2017-10-20T23:07:54.557000", "content": "I took at look at your Keras starter - a few good learning points, I did a trip to Keras.io a few times, I like the 'callbacks'. \nA few observations and a question: The png files generated for the train folder are in the hundreds of gigabytes (caveat emptor) so it goes without saying -training and manipulation- can take a very very long time to those stepping into this. Don't waste time using a standard HDD, nothing less than an m.2 SSD will do. Just doing a simple ls -la can take 5 minutes. My SSD which is much quicker unfortunately is too small to hold that much data! None-the-less what variables did you tinker with to run just a 10k sample using train.py? Usually I can spot this quickly but things look a little nested. As a base line I want to replicate your result as shown. Thus far I cannot. It would help square things up.
", "votes": 0, "replies": [ { "id": 233796, "author_name": "Petros Giannakopoulos", "author_url": "", "post_date": "2017-10-21T09:17:18.270000", "content": "In the code at the beginning of train.py that reads the product_id's, category_id's and num of images per product_id:
\n\ndf = pd.read_csv(csv_file)\nproduct_ids, category_ids, n_pics = df['product_id'], df['category_id'], df['n_pics']\nYou can simply only use the first 10k elements of the product_ids, category_ids, n_pics arrays. So you can add something like this:
\n\n product_ids = product_ids[:10000]\n category_ids = category_ids[:10000]\n n_pics = n_pics[:10000]\nPeter - much appreciated I will give that a try and see how things play out.
", "votes": 0, "replies": [] }, { "id": 233983, "author_name": "stephenl", "author_url": "", "post_date": "2017-10-22T00:33:02.953000", "content": "Peter - I am getting an error on line 55 stratify=category_ids resulting in this error output:
\n\nValueError: The least populated class in y has only 1 member, which is too few. The minimum number of groups for any class cannot be less than 2.
\n\nI had a play around with no effect, but I am running out of time this week (I need to leave town). My guess is its the train/validation split not going to plan with only 10000 values? Something else needs tweaking.
", "votes": 0, "replies": [] }, { "id": 234068, "author_name": "Petros Giannakopoulos", "author_url": "", "post_date": "2017-10-22T07:57:29.200000", "content": "Your guess is right, sklearn complains about the stratified split with 10000 values because some classes end up with just 1 product. You can do a simple split instead of stratified by removing the 'stratify=category_ids' parameter and then it'll work.
", "votes": 0, "replies": [] }, { "id": 236759, "author_name": "stephenl", "author_url": "", "post_date": "2017-10-28T03:42:26.420000", "content": "I am getting a new error on line 63: 'for pic in range(n_pics_train[prod]):'
\n\nIts a key error and is a result of the length of the data changes made above. I have had a look at the variables but so far I could not locate the issue. So supposedly its got no key for some value I assume someplace in one of the created lists?
", "votes": 0, "replies": [] }, { "id": 236787, "author_name": "stephenl", "author_url": "", "post_date": "2017-10-28T07:30:23.940000", "content": "ok this is what I needed to do, 'print' couldn't save me this time but I figured it out below. BUT the valid generator function has 'blown a gasket' so I am trying to figure that out as it runs through the test and stops just on the last iteration!
\n\nprint('Creating training data:')\nfor prod in tqdm(range(len(product_ids_train))):\n for pic in range(n_pics_train[9000]):\n product_ids_pics_train_arr.append(str(product_ids_train[9000]) + '_' + str(pic))\n category_ids_train_arr.append(category_ids_train[9000])\nproduct_ids_pics_test_arr = []\ncategory_ids_test_arr = []\nprint('Creating validation data:')\nfor prod in tqdm(range(len(product_ids_test))):\n for pic in range(1000):\n product_ids_pics_test_arr.append(str(1000) + '_' + str(pic))\n category_ids_test_arr.append(1000)\ntrain_samples = len(product_ids_pics_train_arr)\ntest_samples = len(product_ids_pics_test_arr)
", "votes": 0, "replies": [] } ] }, { "id": 228768, "author_name": "Juan Pizarro", "author_url": "", "post_date": "2017-10-07T19:39:06.710000", "content": "Hello, thanks for sharing
\n\nHow long should a VGG16 or inception v3 epoch take?
", "votes": 0, "replies": [ { "id": 228778, "author_name": "Petros Giannakopoulos", "author_url": "", "post_date": "2017-10-07T20:30:14.043000", "content": "I'm currently training on the full dataset. 1 epoch will take ~80 hours on my GTX 1070...
", "votes": 1, "replies": [] }, { "id": 228963, "author_name": "huiqin", "author_url": "", "post_date": "2017-10-08T12:23:03.650000", "content": "It is too slow. I think that you missed something important. Do you have a look the post :https://www.kaggle.com/humananalog/keras-generator-for-reading-directly-from-bson ? It used 8 workers to speed up the reading of datas and got a better speed. On my 1080ti gpu, it will take about 22 hours per epoch.
", "votes": 0, "replies": [] }, { "id": 228967, "author_name": "Petros Giannakopoulos", "author_url": "", "post_date": "2017-10-08T12:30:11.730000", "content": "Perhaps a reason it's so slow for me is because i'm reading the data from HDD as I don't have enough space on my SSD's.
", "votes": 0, "replies": [] }, { "id": 228982, "author_name": "hengck23", "author_url": "", "post_date": "2017-10-08T13:14:54.343000", "content": "for your information, this is also discussed here:
\n\nhttps://www.kaggle.com/c/cdiscount-image-classification-challenge/discussion/40498:
\n\n\"Hi, Heng, I use similar way to load data as you, converting bson to files.
\n\nThe loading process is very fast at first(0.3s, CPU 30%, RAM 10% of 30G), but some iterations later, it becomes very slow(6s, CPU 5%, RAM 10% of 30G).
\n\nI set the CPU mode to performance, and try different num_workers(0,4,8,12 …), but the situation stays the same. ...\"
", "votes": 3, "replies": [] } ] }, { "id": 228529, "author_name": "", "author_url": "", "post_date": "2017-10-07T00:24:41.793000", "content": "", "votes": -3, "replies": [] } ], "raw_markdown_by_id": { "228504": "Hello everyone! This looks like a challenging classification task due to the very large dataset and large number of classes. \n\nThis is my code so far for this competition, I hope that it helps more people to join and that it provides a decent baseline to build a robust solution on:\n\nhttps://github.com/petrosgk/Kaggle-Cdiscount-Image-Classification-Challenge\n\nI am still processing the .bson files so no results on the full dataset yet. These are 5 epochs on a subset of 10k samples with pre-trained VGG16 (with top dense layers removed) and image size 180x180:\n\n Epoch 1/20\n 231s - loss: 3.7500 - acc: 0.4056 - val_loss: 3.2170 - val_acc: 0.4783\n Epoch 2/20\n 227s - loss: 2.8793 - acc: 0.5010 - val_loss: 3.0092 - val_acc: 0.5149\n Epoch 3/20\n 228s - loss: 2.4361 - acc: 0.5570 - val_loss: 2.9464 - val_acc: 0.5323\n Epoch 4/20\n 229s - loss: 2.1199 - acc: 0.5996 - val_loss: 2.9545 - val_acc: 0.5447\n Epoch 5/20\n 229s - loss: 1.8476 - acc: 0.6346 - val_loss: 2.9870 - val_acc: 0.5440\n\nThere is overfitting with 9k training samples (90/10 train-val split) but that may change when training on the full ~7m samples.", "231530": "I suffered from training very slow when using mxnet. I found that if random read raw images, it is very slow. Now I use the rec file (it is a file generated from all images, hence is fast) of mxnet, it improves from 140 imgs/second to 600 imgs/second.", "233722": "I took at look at your Keras starter - a few good learning points, I did a trip to Keras.io a few times, I like the 'callbacks'. \nA few observations and a question: The png files generated for the train folder are in the hundreds of gigabytes (caveat emptor) so it goes without saying -training and manipulation- can take a very very long time to those stepping into this. Don't waste time using a standard HDD, nothing less than an m.2 SSD will do. Just doing a simple ls -la can take 5 minutes. My SSD which is much quicker unfortunately is too small to hold that much data! None-the-less what variables did you tinker with to run just a 10k sample using train.py? Usually I can spot this quickly but things look a little nested. As a base line I want to replicate your result as shown. Thus far I cannot. It would help square things up.", "228768": "Hello, thanks for sharing\n\nHow long should a VGG16 or inception v3 epoch take?", "228529": "" } }