Hi everyone! I saw a few people asking about GPUs so I put these together:
\n\nIt uses TensorFlow (or Keras) and setup for training in the cloud using Fomoro. *Disclaimer, Fomoro is something I built.
", "messages": [ { "id": 114886, "postDate": "2016-04-14T16:21:14.853Z", "content": "Hi everyone! I saw a few people asking about GPUs so I put these together:
\n\nIt uses TensorFlow (or Keras) and setup for training in the cloud using Fomoro. *Disclaimer, Fomoro is something I built.
", "rawMarkdown": "Hi everyone! I saw a few people asking about GPUs so I put these together:\r\n\r\n- https://github.com/fomorians/distracted-drivers-tf\r\n- https://github.com/fomorians/distracted-drivers-keras\r\n\r\nIt uses TensorFlow (or Keras) and setup for training in the cloud using Fomoro. *Disclaimer, Fomoro is something I built.", "votes": 9 }, { "id": 115951, "postDate": "2016-04-20T23:58:54.743Z", "content": "@asmith26 The default format of imread is something like (WIDTH, HEIGHT, NUM_CHANNELS). Keras prefers (NUM_CHANNELS, WIDTH, HEIGHT) so we swap the axes.
I believe the problem is in prep_dataset.py. Here's a git diff showing the fix that works for me:
\n\n- paths = glob.glob(os.path.join(base, 'c{}/*.jpg'.format(j)))\n+ #paths = glob.glob(os.path.join(base, 'c{}/*.jpg'.format(j)))\n driver_ids_group = driver_imgs_grouped.get_group('c{}'.format(j))\n+ paths = base + '/c{}/'.format(j) + driver_ids_group.img\nThat's a good question! The biggest difference is that the TF version is using more folds and epochs, which will make a significant difference. Just change 3 => 8 for max folds and 10 => 20 for epochs. I'll push that change now to make them more similar.
\n\nThe models themselves are nearly identical, however, there are a couple of small differences for simplicity:
\n\nThe gain parameter of the initialization. I'm using He initialization for the weights in both models where the std deviation of the normal distribution is gain * sqrt(1 / fan_in). Keras, by default, specifies "he_normal" and no gain parameter. Ideally, you want a gain of sqrt(2) for ReLU and 1 for sigmoid.
The TF version is using a truncated normal distribution to reduce dying connections. I'm not sure what Keras is using internally.
To answer my own question above, the driver ID problem does seem to explain why my cross-validation scores were so much lower than my leaderboard scores.
\n\nAfter I applied the fix suggested by @SpammySmith, I got 1.46 from cross-validation and 1.24 on the leaderboard. Still a little puzzling why LB < CV, but maybe there's reasons.
\n\nMore notable to me is that the model appears to be badly overfitting. At least, that's my explanation for why the val_loss climbs steadily with every epoch while the training loss plummets to 0.04 or so.
\n\nAll of the above is for the Keras version. It's also still a puzzle why the TensorFlow version is substantially more accurate than Keras, when you would expect them to be about the same.
", "rawMarkdown": "To answer my own question above, the driver ID problem does seem to explain why my cross-validation scores were so much lower than my leaderboard scores.\r\n\r\nAfter I applied the fix suggested by @SpammySmith, I got 1.46 from cross-validation and 1.24 on the leaderboard. Still a little puzzling why LB < CV, but maybe there's reasons.\r\n\r\nMore notable to me is that the model appears to be badly overfitting. At least, that's my explanation for why the val_loss climbs steadily with every epoch while the training loss plummets to 0.04 or so.\r\n\r\nAll of the above is for the Keras version. It's also still a puzzle why the TensorFlow version is substantially more accurate than Keras, when you would expect them to be about the same.\r\n" }, { "id": 115735, "postDate": "2016-04-19T19:01:23.037Z", "content": "Hi @jeblist, email me with what's not working: jim [at] fomoro [dot] com
\n\nThanks @zfturbo, makes sense. That's what I get for rushing it out ;) I'll push a fix sometime today.
", "rawMarkdown": "Hi @jeblist, email me with what's not working: jim [at] fomoro [dot] com\r\n\r\nThanks @zfturbo, makes sense. That's what I get for rushing it out ;) I'll push a fix sometime today." }, { "id": 115726, "postDate": "2016-04-19T18:30:19.483Z", "content": "help setting it up !!\nI can't start
", "rawMarkdown": " help setting it up !!\r\nI can't start" }, { "id": 115671, "postDate": "2016-04-19T15:15:32.977Z", "content": "Jim Fleming, my example just to show that two images for the same driver ID has different driver on them.
\n\nI insert it in your "main.py" code after line:
\n\nfor train_index, valid_index in LabelShuffleSplit(driver_indices, n_iter=MAX_FOLDS, test_size=0.2, random_state=67):\n[quote=ZFTurbo;115616]I'm not sure, but you probably have a bug in "driver_indices" in your Keras code. Are you sure they are correct?[/quote]
\n\nI wonder if this would explain why my CV score (0.04) is so much lower than my LB score (1.26).
", "rawMarkdown": "[quote=ZFTurbo;115616]I'm not sure, but you probably have a bug in \"driver_indices\" in your Keras code. Are you sure they are correct?[/quote]\r\n\r\nI wonder if this would explain why my CV score (0.04) is so much lower than my LB score (1.26)." }, { "id": 115665, "postDate": "2016-04-19T14:54:07.473Z", "content": "@zfturbo hmm, I'll look into that. I'm not positive one way or the other. I'm not using open cv (scipy is easier to install). Where is that snippet from?
\n\n@inoryy thanks, that makes sense. I figured it was a sensible default. Part of me hoped it would be determined by the activation function. :)
", "rawMarkdown": "@zfturbo hmm, I'll look into that. I'm not positive one way or the other. I'm not using open cv (scipy is easier to install). Where is that snippet from?\r\n\r\n@inoryy thanks, that makes sense. I figured it was a sensible default. Part of me hoped it would be determined by the activation function. :)" }, { "id": 115655, "postDate": "2016-04-19T14:23:56.737Z", "content": "[quote=Jim Fleming;115005]\n- The gain parameter of the initialization. I'm using [He initialization][1] for the weights in both models where the std deviation of the normal distribution is gain * sqrt(1 / fan_in). Keras, by default, specifies "he_normal" and no gain parameter. Ideally, you want a gain of sqrt(2) for ReLU and 1 for sigmoid.\n[/quote]
FYI Keras uses sqrt(2) gain by default for "he_normal"\nhttps://github.com/fchollet/keras/blob/master/keras/initializations.py#L66
", "rawMarkdown": "[quote=Jim Fleming;115005]\r\n- The gain parameter of the initialization. I'm using [He initialization][1] for the weights in both models where the std deviation of the normal distribution is `gain * sqrt(1 / fan_in)`. Keras, by default, specifies \"he_normal\" and no gain parameter. Ideally, you want a gain of `sqrt(2)` for ReLU and `1` for sigmoid.\r\n[/quote]\r\n\r\nFYI Keras uses sqrt(2) gain by default for \"he_normal\"\r\nhttps://github.com/fchollet/keras/blob/master/keras/initializations.py#L66" }, { "id": 115616, "postDate": "2016-04-19T10:39:21.083Z", "content": "I'm not sure, but you probably have a bug in "driver_indices" in your Keras code. Are you sure they are correct?
\n\nfor i in range(len(driver_indices)):\n print(driver_indices[i])\n img = np.array(X_train_raw[i, ...]).swapaxes(0, 2)\n img = cv2.resize(img, (500, 500))\n cv2.imshow('name', img)\n cv2.waitKey(0)\n cv2.destroyAllWindows()\nThere are obviously two different drivers for the same driver ID.
", "rawMarkdown": "I'm not sure, but you probably have a bug in \"driver_indices\" in your Keras code. Are you sure they are correct?\r\n\r\n for i in range(len(driver_indices)):\r\n print(driver_indices[i])\r\n img = np.array(X_train_raw[i, ...]).swapaxes(0, 2)\r\n img = cv2.resize(img, (500, 500))\r\n cv2.imshow('name', img)\r\n cv2.waitKey(0)\r\n cv2.destroyAllWindows()\r\n\r\nThere are obviously two different drivers for the same driver ID. " }, { "id": 115326, "postDate": "2016-04-18T03:44:52.970Z", "content": "A geometric mean[0] is used to average the score and predictions from each fold. The function calc_geom calculates the mean score, while calc_geom_arr calculates the mean predictions.
[0] In mathematics, the geometric mean is a type of mean or average, which indicates the central tendency or typical value of a set of numbers by using the product of their values (as opposed to the arithmetic mean which uses their sum). (https://en.wikipedia.org/wiki/Geometric_mean)
", "rawMarkdown": "A geometric mean[0] is used to average the score and predictions from each fold. The function `calc_geom` calculates the mean score, while `calc_geom_arr` calculates the mean predictions.\r\n\r\n[0] In mathematics, the geometric mean is a type of mean or average, which indicates the central tendency or typical value of a set of numbers by using the product of their values (as opposed to the arithmetic mean which uses their sum). (https://en.wikipedia.org/wiki/Geometric_mean)" }, { "id": 115063, "postDate": "2016-04-16T00:04:16.817Z", "content": "Interesting. I'd love to know if you find the difference. I'll post here if I find it as well.
\n\nI tried to keep the implementations as similar as possible so it's likely a small detail somewhere.
", "rawMarkdown": "Interesting. I'd love to know if you find the difference. I'll post here if I find it as well.\r\n\r\nI tried to keep the implementations as similar as possible so it's likely a small detail somewhere." }, { "id": 115060, "postDate": "2016-04-15T23:35:06.927Z", "content": "I get 1.26 on the leaderboard with those changes versus 1.04 for the TF version.
\n\nI'll dig through the code some more to see if I spot anything else. These differences are a bit of a side issue of course, but I am just learning both Keras and TF and it might be instructive to understand if the differences between them are real.
", "rawMarkdown": "I get 1.26 on the leaderboard with those changes versus 1.04 for the TF version.\r\n\r\nI'll dig through the code some more to see if I spot anything else. These differences are a bit of a side issue of course, but I am just learning both Keras and TF and it might be instructive to understand if the differences between them are real." }, { "id": 114973, "postDate": "2016-04-15T06:54:23.740Z", "content": "I've run the two versions as-is without any tweaking. They give me quite different results on the leaderboard (1.04 for TensorFlow, 1. 33 for Keras).
\n\nIs that expected? Why is the Keras version worse?
", "rawMarkdown": "I've run the two versions as-is without any tweaking. They give me quite different results on the leaderboard (1.04 for TensorFlow, 1. 33 for Keras).\r\n\r\nIs that expected? Why is the Keras version worse?" }, { "id": 114950, "postDate": "2016-04-14T23:23:33.570Z", "content": "Here's the Keras version! Let me know if you need any help setting it up.
\n\nhttps://github.com/fomorians/distracted-drivers-keras
", "rawMarkdown": "Here's the Keras version! Let me know if you need any help setting it up.\r\n\r\nhttps://github.com/fomorians/distracted-drivers-keras" }, { "id": 114946, "postDate": "2016-04-14T22:34:52.423Z", "content": "Very nice, thanks!
\n\nLooking forward to the Keras version.
", "rawMarkdown": "Very nice, thanks!\r\n\r\nLooking forward to the Keras version." }, { "id": 115940, "postDate": "2016-04-20T22:38:45Z", "content": "Why do we img=img.swapaxes(2,0) in prep_dataset.py? I imagine this is because Keras requires images in a certain format, but I can't find the relevant documentation (if anyone can find it, please post here thanks).
@Jim Fleming - Thanks also for this ;-) I understand the calculations, but why do you calculate "calc_geom" and "calc_geom_arr"? Many thanks for all your help in advance!!!
", "rawMarkdown": "@Jim Fleming - Thanks also for this ;-) I understand the calculations, but why do you calculate \"calc_geom\" and \"calc_geom_arr\"? Many thanks for all your help in advance!!!", "isDeleted": true } ], "comments": [ { "id": 115951, "author_name": "Jim Fleming", "author_url": "", "post_date": "2016-04-20T23:58:54.743000", "content": "@asmith26 The default format of imread is something like (WIDTH, HEIGHT, NUM_CHANNELS). Keras prefers (NUM_CHANNELS, WIDTH, HEIGHT) so we swap the axes.
I believe the problem is in prep_dataset.py. Here's a git diff showing the fix that works for me:
\n\n- paths = glob.glob(os.path.join(base, 'c{}/*.jpg'.format(j)))\n+ #paths = glob.glob(os.path.join(base, 'c{}/*.jpg'.format(j)))\n driver_ids_group = driver_imgs_grouped.get_group('c{}'.format(j))\n+ paths = base + '/c{}/'.format(j) + driver_ids_group.img\nThat's a good question! The biggest difference is that the TF version is using more folds and epochs, which will make a significant difference. Just change 3 => 8 for max folds and 10 => 20 for epochs. I'll push that change now to make them more similar.
\n\nThe models themselves are nearly identical, however, there are a couple of small differences for simplicity:
\n\nThe gain parameter of the initialization. I'm using He initialization for the weights in both models where the std deviation of the normal distribution is gain * sqrt(1 / fan_in). Keras, by default, specifies "he_normal" and no gain parameter. Ideally, you want a gain of sqrt(2) for ReLU and 1 for sigmoid.
The TF version is using a truncated normal distribution to reduce dying connections. I'm not sure what Keras is using internally.
To answer my own question above, the driver ID problem does seem to explain why my cross-validation scores were so much lower than my leaderboard scores.
\n\nAfter I applied the fix suggested by @SpammySmith, I got 1.46 from cross-validation and 1.24 on the leaderboard. Still a little puzzling why LB < CV, but maybe there's reasons.
\n\nMore notable to me is that the model appears to be badly overfitting. At least, that's my explanation for why the val_loss climbs steadily with every epoch while the training loss plummets to 0.04 or so.
\n\nAll of the above is for the Keras version. It's also still a puzzle why the TensorFlow version is substantially more accurate than Keras, when you would expect them to be about the same.
", "votes": 0, "replies": [] }, { "id": 115735, "author_name": "Jim Fleming", "author_url": "", "post_date": "2016-04-19T19:01:23.037000", "content": "Hi @jeblist, email me with what's not working: jim [at] fomoro [dot] com
\n\nThanks @zfturbo, makes sense. That's what I get for rushing it out ;) I'll push a fix sometime today.
", "votes": 0, "replies": [] }, { "id": 115726, "author_name": "Jeblist", "author_url": "", "post_date": "2016-04-19T18:30:19.483000", "content": "help setting it up !!\nI can't start
", "votes": 0, "replies": [] }, { "id": 115671, "author_name": "ZFTurbo", "author_url": "", "post_date": "2016-04-19T15:15:32.977000", "content": "Jim Fleming, my example just to show that two images for the same driver ID has different driver on them.
\n\nI insert it in your "main.py" code after line:
\n\nfor train_index, valid_index in LabelShuffleSplit(driver_indices, n_iter=MAX_FOLDS, test_size=0.2, random_state=67):\n[quote=ZFTurbo;115616]I'm not sure, but you probably have a bug in "driver_indices" in your Keras code. Are you sure they are correct?[/quote]
\n\nI wonder if this would explain why my CV score (0.04) is so much lower than my LB score (1.26).
", "votes": 0, "replies": [] }, { "id": 115665, "author_name": "Jim Fleming", "author_url": "", "post_date": "2016-04-19T14:54:07.473000", "content": "@zfturbo hmm, I'll look into that. I'm not positive one way or the other. I'm not using open cv (scipy is easier to install). Where is that snippet from?
\n\n@inoryy thanks, that makes sense. I figured it was a sensible default. Part of me hoped it would be determined by the activation function. :)
", "votes": 0, "replies": [] }, { "id": 115655, "author_name": "Roman Ring", "author_url": "", "post_date": "2016-04-19T14:23:56.737000", "content": "[quote=Jim Fleming;115005]\n- The gain parameter of the initialization. I'm using [He initialization][1] for the weights in both models where the std deviation of the normal distribution is gain * sqrt(1 / fan_in). Keras, by default, specifies "he_normal" and no gain parameter. Ideally, you want a gain of sqrt(2) for ReLU and 1 for sigmoid.\n[/quote]
FYI Keras uses sqrt(2) gain by default for "he_normal"\nhttps://github.com/fchollet/keras/blob/master/keras/initializations.py#L66
", "votes": 0, "replies": [] }, { "id": 115616, "author_name": "ZFTurbo", "author_url": "", "post_date": "2016-04-19T10:39:21.083000", "content": "I'm not sure, but you probably have a bug in "driver_indices" in your Keras code. Are you sure they are correct?
\n\nfor i in range(len(driver_indices)):\n print(driver_indices[i])\n img = np.array(X_train_raw[i, ...]).swapaxes(0, 2)\n img = cv2.resize(img, (500, 500))\n cv2.imshow('name', img)\n cv2.waitKey(0)\n cv2.destroyAllWindows()\nThere are obviously two different drivers for the same driver ID.
", "votes": 0, "replies": [] }, { "id": 115326, "author_name": "Jim Fleming", "author_url": "", "post_date": "2016-04-18T03:44:52.970000", "content": "A geometric mean[0] is used to average the score and predictions from each fold. The function calc_geom calculates the mean score, while calc_geom_arr calculates the mean predictions.
[0] In mathematics, the geometric mean is a type of mean or average, which indicates the central tendency or typical value of a set of numbers by using the product of their values (as opposed to the arithmetic mean which uses their sum). (https://en.wikipedia.org/wiki/Geometric_mean)
", "votes": 0, "replies": [] }, { "id": 115063, "author_name": "Jim Fleming", "author_url": "", "post_date": "2016-04-16T00:04:16.817000", "content": "Interesting. I'd love to know if you find the difference. I'll post here if I find it as well.
\n\nI tried to keep the implementations as similar as possible so it's likely a small detail somewhere.
", "votes": 0, "replies": [] }, { "id": 115060, "author_name": "gauss256", "author_url": "", "post_date": "2016-04-15T23:35:06.927000", "content": "I get 1.26 on the leaderboard with those changes versus 1.04 for the TF version.
\n\nI'll dig through the code some more to see if I spot anything else. These differences are a bit of a side issue of course, but I am just learning both Keras and TF and it might be instructive to understand if the differences between them are real.
", "votes": 0, "replies": [] }, { "id": 114973, "author_name": "gauss256", "author_url": "", "post_date": "2016-04-15T06:54:23.740000", "content": "I've run the two versions as-is without any tweaking. They give me quite different results on the leaderboard (1.04 for TensorFlow, 1. 33 for Keras).
\n\nIs that expected? Why is the Keras version worse?
", "votes": 0, "replies": [] }, { "id": 114950, "author_name": "Jim Fleming", "author_url": "", "post_date": "2016-04-14T23:23:33.570000", "content": "Here's the Keras version! Let me know if you need any help setting it up.
\n\nhttps://github.com/fomorians/distracted-drivers-keras
", "votes": 0, "replies": [] }, { "id": 114946, "author_name": "gauss256", "author_url": "", "post_date": "2016-04-14T22:34:52.423000", "content": "Very nice, thanks!
\n\nLooking forward to the Keras version.
", "votes": 0, "replies": [] }, { "id": 115940, "author_name": "", "author_url": "", "post_date": "2016-04-20T22:38:45", "content": "Why do we img=img.swapaxes(2,0) in prep_dataset.py? I imagine this is because Keras requires images in a certain format, but I can't find the relevant documentation (if anyone can find it, please post here thanks).
@Jim Fleming - Thanks also for this ;-) I understand the calculations, but why do you calculate "calc_geom" and "calc_geom_arr"? Many thanks for all your help in advance!!!
", "votes": 0, "replies": [] } ], "raw_markdown_by_id": { "114886": "Hi everyone! I saw a few people asking about GPUs so I put these together:\r\n\r\n- https://github.com/fomorians/distracted-drivers-tf\r\n- https://github.com/fomorians/distracted-drivers-keras\r\n\r\nIt uses TensorFlow (or Keras) and setup for training in the cloud using Fomoro. *Disclaimer, Fomoro is something I built.", "115951": "@asmith26 The default format of `imread` is something like `(WIDTH, HEIGHT, NUM_CHANNELS)`. Keras prefers `(NUM_CHANNELS, WIDTH, HEIGHT)` so we swap the axes.", "115738": "I believe the problem is in prep_dataset.py. Here's a git diff showing the fix that works for me:\r\n\r\n - paths = glob.glob(os.path.join(base, 'c{}/*.jpg'.format(j)))\r\n + #paths = glob.glob(os.path.join(base, 'c{}/*.jpg'.format(j)))\r\n driver_ids_group = driver_imgs_grouped.get_group('c{}'.format(j))\r\n + paths = base + '/c{}/'.format(j) + driver_ids_group.img\r\n\r\n", "115005": "That's a good question! The biggest difference is that the TF version is using more folds and epochs, which will make a significant difference. Just change 3 => 8 for max folds and 10 => 20 for epochs. I'll push that change now to make them more similar.\r\n\r\nThe models themselves are nearly identical, however, there are a couple of small differences for simplicity:\r\n\r\n- The gain parameter of the initialization. I'm using [He initialization][1] for the weights in both models where the std deviation of the normal distribution is `gain * sqrt(1 / fan_in)`. Keras, by default, specifies \"he_normal\" and no gain parameter. Ideally, you want a gain of `sqrt(2)` for ReLU and `1` for sigmoid.\r\n\r\n- The TF version is using a truncated normal distribution to reduce dying connections. I'm not sure what Keras is using internally.\r\n\r\n [1]: http://arxiv.org/abs/1502.01852", "115770": "To answer my own question above, the driver ID problem does seem to explain why my cross-validation scores were so much lower than my leaderboard scores.\r\n\r\nAfter I applied the fix suggested by @SpammySmith, I got 1.46 from cross-validation and 1.24 on the leaderboard. Still a little puzzling why LB < CV, but maybe there's reasons.\r\n\r\nMore notable to me is that the model appears to be badly overfitting. At least, that's my explanation for why the val_loss climbs steadily with every epoch while the training loss plummets to 0.04 or so.\r\n\r\nAll of the above is for the Keras version. It's also still a puzzle why the TensorFlow version is substantially more accurate than Keras, when you would expect them to be about the same.\r\n", "115735": "Hi @jeblist, email me with what's not working: jim [at] fomoro [dot] com\r\n\r\nThanks @zfturbo, makes sense. That's what I get for rushing it out ;) I'll push a fix sometime today.", "115726": " help setting it up !!\r\nI can't start", "115671": "Jim Fleming, my example just to show that two images for the same driver ID has different driver on them.\r\n\r\nI insert it in your \"main.py\" code after line:\r\n\r\n for train_index, valid_index in LabelShuffleSplit(driver_indices, n_iter=MAX_FOLDS, test_size=0.2, random_state=67):", "115670": "[quote=ZFTurbo;115616]I'm not sure, but you probably have a bug in \"driver_indices\" in your Keras code. Are you sure they are correct?[/quote]\r\n\r\nI wonder if this would explain why my CV score (0.04) is so much lower than my LB score (1.26).", "115665": "@zfturbo hmm, I'll look into that. I'm not positive one way or the other. I'm not using open cv (scipy is easier to install). Where is that snippet from?\r\n\r\n@inoryy thanks, that makes sense. I figured it was a sensible default. Part of me hoped it would be determined by the activation function. :)", "115655": "[quote=Jim Fleming;115005]\r\n- The gain parameter of the initialization. I'm using [He initialization][1] for the weights in both models where the std deviation of the normal distribution is `gain * sqrt(1 / fan_in)`. Keras, by default, specifies \"he_normal\" and no gain parameter. Ideally, you want a gain of `sqrt(2)` for ReLU and `1` for sigmoid.\r\n[/quote]\r\n\r\nFYI Keras uses sqrt(2) gain by default for \"he_normal\"\r\nhttps://github.com/fchollet/keras/blob/master/keras/initializations.py#L66", "115616": "I'm not sure, but you probably have a bug in \"driver_indices\" in your Keras code. Are you sure they are correct?\r\n\r\n for i in range(len(driver_indices)):\r\n print(driver_indices[i])\r\n img = np.array(X_train_raw[i, ...]).swapaxes(0, 2)\r\n img = cv2.resize(img, (500, 500))\r\n cv2.imshow('name', img)\r\n cv2.waitKey(0)\r\n cv2.destroyAllWindows()\r\n\r\nThere are obviously two different drivers for the same driver ID. ", "115326": "A geometric mean[0] is used to average the score and predictions from each fold. The function `calc_geom` calculates the mean score, while `calc_geom_arr` calculates the mean predictions.\r\n\r\n[0] In mathematics, the geometric mean is a type of mean or average, which indicates the central tendency or typical value of a set of numbers by using the product of their values (as opposed to the arithmetic mean which uses their sum). (https://en.wikipedia.org/wiki/Geometric_mean)", "115063": "Interesting. I'd love to know if you find the difference. I'll post here if I find it as well.\r\n\r\nI tried to keep the implementations as similar as possible so it's likely a small detail somewhere.", "115060": "I get 1.26 on the leaderboard with those changes versus 1.04 for the TF version.\r\n\r\nI'll dig through the code some more to see if I spot anything else. These differences are a bit of a side issue of course, but I am just learning both Keras and TF and it might be instructive to understand if the differences between them are real.", "114973": "I've run the two versions as-is without any tweaking. They give me quite different results on the leaderboard (1.04 for TensorFlow, 1. 33 for Keras).\r\n\r\nIs that expected? Why is the Keras version worse?", "114950": "Here's the Keras version! Let me know if you need any help setting it up.\r\n\r\nhttps://github.com/fomorians/distracted-drivers-keras", "114946": "Very nice, thanks!\r\n\r\nLooking forward to the Keras version.", "115940": "Why do we **`img=img.swapaxes(2,0)`** in *prep_dataset.py*? I imagine this is because Keras requires images in a certain format, but I can't find the relevant documentation (if anyone can find it, please post here thanks).", "115187": "@Jim Fleming - Thanks also for this ;-) I understand the calculations, but why do you calculate \"calc_geom\" and \"calc_geom_arr\"? Many thanks for all your help in advance!!!" } }