{ "id": 128914, "title": "[pytorch] Tricks for Image Classification......", "url": "/competitions/bengaliai-cv19/discussion/128914", "author_name": "MachineLP", "post_date": "2020-02-04T08:01:39.728000", "votes": 61, "comment_count": 17, "views": 0, "content": "

data augmentation

\n\n
\n

auto-augment：https://github.com/DeepVoltaire/AutoAugment/blob/master/autoaugment.py\n fast-autoaugment: https://github.com/kakaobrain/fast-autoaugment\n augmix: https://github.com/google-research/augmix\n mixup/cutout: https://github.com/PistonY/torch-toolbox

\n
\n\n

model

\n\n
\n

pretrained-models.pytorch: https://github.com/Cadene/pretrained-models.pytorch

\n
\n\n

metric learning

\n\n
\n

pytorch-metric-learning: https://github.com/KevinMusgrave/pytorch-metric-learning\n top_k_optimization: https://github.com/BG2CRW/top_k_optimization

\n
\n\n

loss

\n\n
\n

Class-balanced-loss-pytorch: https://github.com/vandit15/Class-balanced-loss-pytorch

\n
\n\n

framework

\n\n
\n

pytorch_image_classification: https://github.com/hysts/pytorch_image_classification

\n
\n\n

......

", "messages": [ { "id": 736456, "postDate": "2020-02-04T08:01:39.730Z", "content": "

data augmentation

\n\n
\n

auto-augment：https://github.com/DeepVoltaire/AutoAugment/blob/master/autoaugment.py\n fast-autoaugment: https://github.com/kakaobrain/fast-autoaugment\n augmix: https://github.com/google-research/augmix\n mixup/cutout: https://github.com/PistonY/torch-toolbox

\n
\n\n

model

\n\n
\n

pretrained-models.pytorch: https://github.com/Cadene/pretrained-models.pytorch

\n
\n\n

metric learning

\n\n
\n

pytorch-metric-learning: https://github.com/KevinMusgrave/pytorch-metric-learning\n top_k_optimization: https://github.com/BG2CRW/top_k_optimization

\n
\n\n

loss

\n\n
\n

Class-balanced-loss-pytorch: https://github.com/vandit15/Class-balanced-loss-pytorch

\n
\n\n

framework

\n\n
\n

pytorch_image_classification: https://github.com/hysts/pytorch_image_classification

\n
\n\n

......

", "rawMarkdown": "## data augmentation\n> auto-augment：https://github.com/DeepVoltaire/AutoAugment/blob/master/autoaugment.py\n> fast-autoaugment: https://github.com/kakaobrain/fast-autoaugment\n> augmix: https://github.com/google-research/augmix\n> mixup/cutout: https://github.com/PistonY/torch-toolbox\n\n## model\n> pretrained-models.pytorch: https://github.com/Cadene/pretrained-models.pytorch\n\n## metric learning\n> pytorch-metric-learning: https://github.com/KevinMusgrave/pytorch-metric-learning\n> top_k_optimization: https://github.com/BG2CRW/top_k_optimization\n\n## loss\n> Class-balanced-loss-pytorch: https://github.com/vandit15/Class-balanced-loss-pytorch\n\n# framework\n> pytorch_image_classification: https://github.com/hysts/pytorch_image_classification\n\n......", "votes": 61 }, { "id": 736459, "postDate": "2020-02-04T08:06:10.357Z", "content": "

update:\n[pytorch] ohem loss implementation: https://www.kaggle.com/c/bengaliai-cv19/discussion/128637\n[pytorch] focal loss + ohem implementation: https://www.kaggle.com/c/bengaliai-cv19/discussion/128665\nmixup/cutmix with label smoothing: https://www.kaggle.com/c/bengaliai-cv19/discussion/128115\nmixup/cutmix is all you need: https://www.kaggle.com/c/bengaliai-cv19/discussion/126504

", "rawMarkdown": "update:\n[pytorch] ohem loss implementation: https://www.kaggle.com/c/bengaliai-cv19/discussion/128637\n[pytorch] focal loss + ohem implementation: https://www.kaggle.com/c/bengaliai-cv19/discussion/128665\nmixup/cutmix with label smoothing: https://www.kaggle.com/c/bengaliai-cv19/discussion/128115\nmixup/cutmix is all you need: https://www.kaggle.com/c/bengaliai-cv19/discussion/126504", "votes": 5, "replies": [ { "id": 744005, "postDate": "2020-02-12T13:19:56.270Z", "content": "

awsome

", "rawMarkdown": "awsome", "votes": 1 } ] }, { "id": 736984, "postDate": "2020-02-04T19:12:50.817Z", "content": "

One of my take-aways from this competition: I probably should get into PyTorch seriously for computer vision tasks. There seems to be a lot more resource than for Keras.

", "rawMarkdown": "One of my take-aways from this competition: I probably should get into PyTorch seriously for computer vision tasks. There seems to be a lot more resource than for Keras.", "votes": 4, "replies": [ { "id": 737129, "postDate": "2020-02-05T00:11:10.630Z", "content": "

I have a mix feelings of this. From the very beginning of this competition I see very strong community support for PyTorch. A lots of pytorch geeks share ideas and open-source implementation of various SOTA approach. So to build a quick promising prototype using pytorch and getting a really decent score is becoming really easy day to day, one evidence for this is the new competitor who just came very late and score high so quickly.

", "rawMarkdown": "I have a mix feelings of this. From the very beginning of this competition I see very strong community support for PyTorch. A lots of pytorch geeks share ideas and open-source implementation of various SOTA approach. So to build a quick promising prototype using pytorch and getting a really decent score is becoming really easy day to day, one evidence for this is the new competitor who just came very late and score high so quickly. " }, { "id": 737135, "postDate": "2020-02-05T00:27:47.170Z", "content": "

For this comp specifically pytorch is better since submitting needs to be as quick as possible.

", "rawMarkdown": "For this comp specifically pytorch is better since submitting needs to be as quick as possible.", "votes": 1 }, { "id": 737315, "postDate": "2020-02-05T07:15:56.170Z", "content": "

PyTorch seems pretty interesting to implement SOTA solutions indeed. But there are also some costs to that, having implemented a lot of my methods / classes in Keras and having to switch would take some time. Probably a great exercise though! :)

", "rawMarkdown": "PyTorch seems pretty interesting to implement SOTA solutions indeed. But there are also some costs to that, having implemented a lot of my methods / classes in Keras and having to switch would take some time. Probably a great exercise though! :)", "votes": 1 }, { "id": 740988, "postDate": "2020-02-10T03:52:41.700Z", "content": "

@maxlenormand Also I feel exactly the same. It looks like customization is harder in Keras as it has higher level of abstractions.

", "rawMarkdown": "@maxlenormand Also I feel exactly the same. It looks like customization is harder in Keras as it has higher level of abstractions.", "votes": 2 } ] }, { "id": 736894, "postDate": "2020-02-04T17:18:00.510Z", "content": "

Gridmask augmentation: https://www.kaggle.com/haqishen/gridmask

", "rawMarkdown": "Gridmask augmentation: https://www.kaggle.com/haqishen/gridmask", "votes": 1 }, { "id": 736460, "postDate": "2020-02-04T08:06:57.747Z", "content": "

did you try augmix??

", "rawMarkdown": "did you try augmix?? ", "votes": 1, "replies": [ { "id": 736469, "postDate": "2020-02-04T08:22:53.140Z", "content": "

yes

", "rawMarkdown": "yes", "votes": 3 }, { "id": 738699, "postDate": "2020-02-06T22:00:38.207Z", "content": "

@machinelp did you use augmix with JSD loss? or simply Crossentropy?

", "rawMarkdown": "@machinelp did you use augmix with JSD loss? or simply Crossentropy?" } ] }, { "id": 956846, "postDate": "2020-08-03T20:44:27.057Z", "content": "

Hello Guys. I choose ResNet for İmage Classification. I thınk you should check thıs code\nhttps://github.com/batuhan3526/ResNet50_on_Cifar_100_Without_Transfer_Learning

", "rawMarkdown": "Hello Guys. I choose ResNet for İmage Classification. I thınk you should check thıs code\nhttps://github.com/batuhan3526/ResNet50_on_Cifar_100_Without_Transfer_Learning" }, { "id": 737455, "postDate": "2020-02-05T11:15:59.320Z", "content": "

Thanks a lot @machinelp for taking the pain to compile all these resources in a single place, this is very helpful

", "rawMarkdown": "Thanks a lot @machinelp for taking the pain to compile all these resources in a single place, this is very helpful" }, { "id": 739566, "postDate": "2020-02-08T01:16:55.167Z", "rawMarkdown": "", "votes": 1, "isDeleted": true }, { "id": 1906384, "postDate": "2022-08-19T20:36:38.177Z", "content": "

Thanks for the valuable summary!

", "rawMarkdown": "Thanks for the valuable summary!", "votes": 5 }, { "id": 736890, "postDate": "2020-02-04T17:11:11.163Z", "content": "

Thank you for the nice summaries

", "rawMarkdown": "Thank you for the nice summaries", "votes": 1 }, { "id": 738326, "postDate": "2020-02-06T11:46:58.380Z", "content": "

thanks a lot

", "rawMarkdown": "thanks a lot", "votes": 2 } ], "comments": [ { "id": 736459, "author_name": "MachineLP", "author_url": "", "post_date": "2020-02-04T08:06:10.357000", "content": "

update:\n[pytorch] ohem loss implementation: https://www.kaggle.com/c/bengaliai-cv19/discussion/128637\n[pytorch] focal loss + ohem implementation: https://www.kaggle.com/c/bengaliai-cv19/discussion/128665\nmixup/cutmix with label smoothing: https://www.kaggle.com/c/bengaliai-cv19/discussion/128115\nmixup/cutmix is all you need: https://www.kaggle.com/c/bengaliai-cv19/discussion/126504

", "votes": 5, "replies": [ { "id": 744005, "author_name": "Goblin Alchemist", "author_url": "", "post_date": "2020-02-12T13:19:56.270000", "content": "

awsome

", "votes": 1, "replies": [] } ] }, { "id": 736984, "author_name": "Maxime Lenormand", "author_url": "", "post_date": "2020-02-04T19:12:50.817000", "content": "

One of my take-aways from this competition: I probably should get into PyTorch seriously for computer vision tasks. There seems to be a lot more resource than for Keras.

", "votes": 4, "replies": [ { "id": 737129, "author_name": "Innat", "author_url": "", "post_date": "2020-02-05T00:11:10.630000", "content": "

I have a mix feelings of this. From the very beginning of this competition I see very strong community support for PyTorch. A lots of pytorch geeks share ideas and open-source implementation of various SOTA approach. So to build a quick promising prototype using pytorch and getting a really decent score is becoming really easy day to day, one evidence for this is the new competitor who just came very late and score high so quickly.

", "votes": 0, "replies": [] }, { "id": 737135, "author_name": "GreatGameDota", "author_url": "", "post_date": "2020-02-05T00:27:47.170000", "content": "

For this comp specifically pytorch is better since submitting needs to be as quick as possible.

", "votes": 1, "replies": [] }, { "id": 737315, "author_name": "Maxime Lenormand", "author_url": "", "post_date": "2020-02-05T07:15:56.170000", "content": "

PyTorch seems pretty interesting to implement SOTA solutions indeed. But there are also some costs to that, having implemented a lot of my methods / classes in Keras and having to switch would take some time. Probably a great exercise though! :)

", "votes": 1, "replies": [] }, { "id": 740988, "author_name": "Kaushal Shah", "author_url": "", "post_date": "2020-02-10T03:52:41.700000", "content": "

@maxlenormand Also I feel exactly the same. It looks like customization is harder in Keras as it has higher level of abstractions.

", "votes": 2, "replies": [] } ] }, { "id": 736894, "author_name": "GreatGameDota", "author_url": "", "post_date": "2020-02-04T17:18:00.510000", "content": "

Gridmask augmentation: https://www.kaggle.com/haqishen/gridmask

", "votes": 1, "replies": [] }, { "id": 736460, "author_name": "Mobassir", "author_url": "", "post_date": "2020-02-04T08:06:57.747000", "content": "

did you try augmix??

", "votes": 1, "replies": [ { "id": 736469, "author_name": "MachineLP", "author_url": "", "post_date": "2020-02-04T08:22:53.140000", "content": "

yes

", "votes": 3, "replies": [] }, { "id": 738699, "author_name": "Yann Majewski", "author_url": "", "post_date": "2020-02-06T22:00:38.207000", "content": "

@machinelp did you use augmix with JSD loss? or simply Crossentropy?

", "votes": 0, "replies": [] } ] }, { "id": 956846, "author_name": "batu bayraktar", "author_url": "", "post_date": "2020-08-03T20:44:27.057000", "content": "

Hello Guys. I choose ResNet for İmage Classification. I thınk you should check thıs code\nhttps://github.com/batuhan3526/ResNet50_on_Cifar_100_Without_Transfer_Learning

", "votes": 0, "replies": [] }, { "id": 737455, "author_name": "Rohit Agarwal", "author_url": "", "post_date": "2020-02-05T11:15:59.320000", "content": "

Thanks a lot @machinelp for taking the pain to compile all these resources in a single place, this is very helpful

", "votes": 0, "replies": [] }, { "id": 739566, "author_name": "", "author_url": "", "post_date": "2020-02-08T01:16:55.167000", "content": "", "votes": 1, "replies": [] }, { "id": 1906384, "author_name": "Mahmud Ahad Abedin Fardin", "author_url": "", "post_date": "2022-08-19T20:36:38.177000", "content": "

Thanks for the valuable summary!

", "votes": 5, "replies": [] }, { "id": 736890, "author_name": "Soonhwan Kwon", "author_url": "", "post_date": "2020-02-04T17:11:11.163000", "content": "

Thank you for the nice summaries

", "votes": 1, "replies": [] }, { "id": 738326, "author_name": "Convophile", "author_url": "", "post_date": "2020-02-06T11:46:58.380000", "content": "

thanks a lot

", "votes": 2, "replies": [] } ], "raw_markdown_by_id": { "736456": "## data augmentation\n> auto-augment：https://github.com/DeepVoltaire/AutoAugment/blob/master/autoaugment.py\n> fast-autoaugment: https://github.com/kakaobrain/fast-autoaugment\n> augmix: https://github.com/google-research/augmix\n> mixup/cutout: https://github.com/PistonY/torch-toolbox\n\n## model\n> pretrained-models.pytorch: https://github.com/Cadene/pretrained-models.pytorch\n\n## metric learning\n> pytorch-metric-learning: https://github.com/KevinMusgrave/pytorch-metric-learning\n> top_k_optimization: https://github.com/BG2CRW/top_k_optimization\n\n## loss\n> Class-balanced-loss-pytorch: https://github.com/vandit15/Class-balanced-loss-pytorch\n\n# framework\n> pytorch_image_classification: https://github.com/hysts/pytorch_image_classification\n\n......", "736459": "update:\n[pytorch] ohem loss implementation: https://www.kaggle.com/c/bengaliai-cv19/discussion/128637\n[pytorch] focal loss + ohem implementation: https://www.kaggle.com/c/bengaliai-cv19/discussion/128665\nmixup/cutmix with label smoothing: https://www.kaggle.com/c/bengaliai-cv19/discussion/128115\nmixup/cutmix is all you need: https://www.kaggle.com/c/bengaliai-cv19/discussion/126504", "736984": "One of my take-aways from this competition: I probably should get into PyTorch seriously for computer vision tasks. There seems to be a lot more resource than for Keras.", "736894": "Gridmask augmentation: https://www.kaggle.com/haqishen/gridmask", "736460": "did you try augmix?? ", "956846": "Hello Guys. I choose ResNet for İmage Classification. I thınk you should check thıs code\nhttps://github.com/batuhan3526/ResNet50_on_Cifar_100_Without_Transfer_Learning", "737455": "Thanks a lot @machinelp for taking the pain to compile all these resources in a single place, this is very helpful", "739566": "", "1906384": "Thanks for the valuable summary!", "736890": "Thank you for the nice summaries", "738326": "thanks a lot" } }