{ "id": 434354, "title": "95th Bronze solutions", "url": "/competitions/asl-fingerspelling/discussion/434354", "author_name": "suguuuuu", "post_date": "2023-08-25T00:01:12.515000", "votes": 9, "comment_count": 3, "views": 0, "content": "

95th Bronze solutions

\n

Thanks to Kaggle for hosting this interesting competition!!!!
\nThis competition was diffucult for us…
\nThere were many differences from last competition which made it challenging. I enjoyed learning about transcription task.

\n

Summary

\n

We utilized the best public notebook with some modifications.

\n\n

What changes were made from the public notebook?

\n\n

Things I couldn't do

\n\n

Appendix: Model code

\n
def get:\n    inp = tf.keras.\n\n    x = tf.keras.layers.(inp)\n    x = tf.keras.layers.(x) + positional\n    x = tf.keras.layers.(x)\n\n    x = (x)\n    x = (x)\n    x = (x)\n    x = (x)\n\n    x = (x)\n    x = (x)\n    x = (x)\n    x = (x)\n\n    x = (x)\n    x = (x)\n    x = (x)\n    x = (x)\n\n    x = (x)\n    x = (x)\n    x = (x)\n    x = (x)\n\n    x = (x)\n    x = (x)\n    x = (x)\n    x_aux = (x)\n\n    x = (x_aux)\n    x = (x)\n    x = (x)\n    x = (x)\n\n    x = (x)\n    x = (x)\n    x = (x)\n    x = (x)\n\n    x = (x)\n    x = (x)\n    x = (x)\n    x = (x)\n\n    x = (x)\n    x = (x)\n    x = (x)\n    x = (x)\n\n    x = (x)\n    x = (x)\n    x = (x)\n    x = (x)\n\n    # main\n    x = tf.keras.layers.(x)\n    x = tf.keras.layers.(x)\n    x = tf.keras.layers.,name='out')(x)\n\n    #aux\n    x_aux = tf.keras.layers.(x_aux)\n    x_aux = tf.keras.layers.(x_aux)\n    x_aux = tf.keras.layers.,name='aux_out')(x_aux)    \n\n    model = tf.keras.\n\n    loss = {\n        'out':CTCLoss,\n        'aux_out':CTCLoss\n    }\n\n    # Adam Optimizer\n    optimizer = tfa.optimizers.\n    optimizer = tfa.optimizers.\n\n    model.compile(loss=loss, optimizer=optimizer)\n\n    return model\n
", "messages": [ { "id": 2407195, "postDate": "2023-08-25T00:01:12.517Z", "content": "

95th Bronze solutions

\n

Thanks to Kaggle for hosting this interesting competition!!!!
\nThis competition was diffucult for us…
\nThere were many differences from last competition which made it challenging. I enjoyed learning about transcription task.

\n

Summary

\n

We utilized the best public notebook with some modifications.

\n\n

What changes were made from the public notebook?

\n\n

Things I couldn't do

\n\n

Appendix: Model code

\n
def get:\n    inp = tf.keras.\n\n    x = tf.keras.layers.(inp)\n    x = tf.keras.layers.(x) + positional\n    x = tf.keras.layers.(x)\n\n    x = (x)\n    x = (x)\n    x = (x)\n    x = (x)\n\n    x = (x)\n    x = (x)\n    x = (x)\n    x = (x)\n\n    x = (x)\n    x = (x)\n    x = (x)\n    x = (x)\n\n    x = (x)\n    x = (x)\n    x = (x)\n    x = (x)\n\n    x = (x)\n    x = (x)\n    x = (x)\n    x_aux = (x)\n\n    x = (x_aux)\n    x = (x)\n    x = (x)\n    x = (x)\n\n    x = (x)\n    x = (x)\n    x = (x)\n    x = (x)\n\n    x = (x)\n    x = (x)\n    x = (x)\n    x = (x)\n\n    x = (x)\n    x = (x)\n    x = (x)\n    x = (x)\n\n    x = (x)\n    x = (x)\n    x = (x)\n    x = (x)\n\n    # main\n    x = tf.keras.layers.(x)\n    x = tf.keras.layers.(x)\n    x = tf.keras.layers.,name='out')(x)\n\n    #aux\n    x_aux = tf.keras.layers.(x_aux)\n    x_aux = tf.keras.layers.(x_aux)\n    x_aux = tf.keras.layers.,name='aux_out')(x_aux)    \n\n    model = tf.keras.\n\n    loss = {\n        'out':CTCLoss,\n        'aux_out':CTCLoss\n    }\n\n    # Adam Optimizer\n    optimizer = tfa.optimizers.\n    optimizer = tfa.optimizers.\n\n    model.compile(loss=loss, optimizer=optimizer)\n\n    return model\n
", "rawMarkdown": "# 95th Bronze solutions\n\nThanks to Kaggle for hosting this interesting competition!!!!\nThis competition was diffucult for us…\nThere were many differences from last competition which made it challenging. I enjoyed learning about transcription task.\n\n# Summary\nWe utilized the best public notebook with some modifications.\n- base notebook: https://www.kaggle.com/code/saidineshpola/aslfr-ctc-based-on-prev-comp-1st-place\n\n# What changes were made from the public notebook?\n* Normalization (+0.002 LB)\n\t* Adjusted all coordinates so that the center of numbers 11 and 12 is set to 0.\n![](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F2930242%2F6d424a3976cee8f769e3e90ba54a9ac2%2Finbox_2930242_fa394dbd96e33fb875aef6ed1f25c757_1.png?generation=1692922021894426&alt=media)\n* Added more features (+0.01 LB)\n\t* Motion, Shape Feature, \n\t\t* Shape is distance of this point\n![](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F2930242%2F3f7fdc56a208576b7664408fd807e422%2Finbox_2930242_b739316822c577ef9b4be204325f569e_2.png?generation=1692922063719858&alt=media)\n\n* Feature Augmentation (+0.01 LB)\n\t* Framerate, randomDrop, random padding, random crop.\n* Changed Model (+0.005 LB)\n\t* dim : 384 => 192\n\t* conv trans Block num : 3=>10\n\t* Add aux Loss\n* Pretraining with supplementay data (+0.01 LB)\n* Training params\n\t* epoch 100\n\t* batchsize 256\n\t* learningrate 1e-2\n\n# Things I couldn't do\n* Error correction Network.\n\t* I read the following paper and tried to incorporate voice recognition technology, but I couldn't make it in time.\n\t* https://www.google.com/url?sa=t&rct=j&q=&esrc=s&source=web&cd=&cad=rja&uact=8&ved=2ahUKEwjAkrKRvvaAAxWBO3AKHbDkB4MQFnoECBMQAQ&url=https%3A%2F%2Farxiv.org%2Fabs%2F2111.01690&usg=AOvVaw3B--e8pJqXTaYolWYXIBOf&opi=89978449\n* Change to SOTA OCR model.\n\t* https://arxiv.org/pdf/2205.00159.pdf\n\n# Appendix: Model code\n```\ndef get_model(dim = 192):\n inp = tf.keras.Input(INPUT_SHAPE)\n \n x = tf.keras.layers.Masking(mask_value=0.0)(inp)\n x = tf.keras.layers.Dense(dim, use_bias=False,name='stem_conv')(x) + positional_encoding(INPUT_SHAPE[0], dim)\n x = tf.keras.layers.BatchNormalization(momentum=0.95,name='stem_bn')(x)\n\n x = Conv1DBlock(dim, 11, drop_rate=0.2)(x)\n x = Conv1DBlock(dim, 5, drop_rate=0.2)(x)\n x = Conv1DBlock(dim, 3, drop_rate=0.2)(x)\n x = TransformerBlock(dim, expand=2)(x)\n\n x = Conv1DBlock(dim, 11, drop_rate=0.2)(x)\n x = Conv1DBlock(dim, 5, drop_rate=0.2)(x)\n x = Conv1DBlock(dim, 3, drop_rate=0.2)(x)\n x = TransformerBlock(dim, expand=2)(x)\n \n x = Conv1DBlock(dim, 11, drop_rate=0.2)(x)\n x = Conv1DBlock(dim, 5, drop_rate=0.2)(x)\n x = Conv1DBlock(dim, 3, drop_rate=0.2)(x)\n x = TransformerBlock(dim, expand=2)(x)\n\n x = Conv1DBlock(dim, 11, drop_rate=0.2)(x)\n x = Conv1DBlock(dim, 5, drop_rate=0.2)(x)\n x = Conv1DBlock(dim, 3, drop_rate=0.2)(x)\n x = TransformerBlock(dim, expand=2)(x)\n \n x = Conv1DBlock(dim, 11, drop_rate=0.2)(x)\n x = Conv1DBlock(dim, 5, drop_rate=0.2)(x)\n x = Conv1DBlock(dim, 3, drop_rate=0.2)(x)\n x_aux = TransformerBlock(dim, expand=2)(x)\n\n x = Conv1DBlock(dim, 11, drop_rate=0.2)(x_aux)\n x = Conv1DBlock(dim, 5, drop_rate=0.2)(x)\n x = Conv1DBlock(dim, 3, drop_rate=0.2)(x)\n x = TransformerBlock(dim, expand=2)(x)\n\n x = Conv1DBlock(dim, 11, drop_rate=0.2)(x)\n x = Conv1DBlock(dim, 5, drop_rate=0.2)(x)\n x = Conv1DBlock(dim, 3, drop_rate=0.2)(x)\n x = TransformerBlock(dim, expand=2)(x)\n \n x = Conv1DBlock(dim, 11, drop_rate=0.2)(x)\n x = Conv1DBlock(dim, 5, drop_rate=0.2)(x)\n x = Conv1DBlock(dim, 3, drop_rate=0.2)(x)\n x = TransformerBlock(dim, expand=2)(x)\n \n x = Conv1DBlock(dim, 11, drop_rate=0.2)(x)\n x = Conv1DBlock(dim, 5, drop_rate=0.2)(x)\n x = Conv1DBlock(dim, 3, drop_rate=0.2)(x)\n x = TransformerBlock(dim, expand=2)(x)\n\n x = Conv1DBlock(dim, 11, drop_rate=0.2)(x)\n x = Conv1DBlock(dim, 5, drop_rate=0.2)(x)\n x = Conv1DBlock(dim, 3, drop_rate=0.2)(x)\n x = TransformerBlock(dim, expand=2)(x)\n \n # main\n x = tf.keras.layers.Dense(dim*2,activation='relu',name='top_conv')(x)\n x = tf.keras.layers.Dropout(0.4)(x)\n x = tf.keras.layers.Dense(len(char_to_num),name='out')(x)\n\n #aux\n x_aux = tf.keras.layers.Dense(dim*2,activation='relu',name='top_aux')(x_aux)\n x_aux = tf.keras.layers.Dropout(0.4)(x_aux)\n x_aux = tf.keras.layers.Dense(len(char_to_num),name='aux_out')(x_aux) \n \n model = tf.keras.Model(inp, [x, x_aux])\n\n loss = {\n 'out':CTCLoss,\n 'aux_out':CTCLoss\n }\n \n # Adam Optimizer\n optimizer = tfa.optimizers.RectifiedAdam(sma_threshold=4)\n optimizer = tfa.optimizers.Lookahead(optimizer, sync_period=5)\n\n model.compile(loss=loss, optimizer=optimizer)\n\n return model\n```\n", "votes": 9 }, { "id": 2410713, "postDate": "2023-08-27T06:58:56.290Z", "content": "

Amazing Keep going@sugupoko

", "rawMarkdown": "Amazing Keep going@sugupoko ", "votes": 1 }, { "id": 2408364, "postDate": "2023-08-25T15:35:41.663Z", "content": "

Congratulations. Thanks for sharing the details of your solution with nice illustrations.

", "rawMarkdown": "Congratulations. Thanks for sharing the details of your solution with nice illustrations.", "votes": 1 }, { "id": 2413036, "postDate": "2023-08-28T16:34:05.037Z", "content": "

GG sir! nice one

", "rawMarkdown": "GG sir! nice one" } ], "comments": [ { "id": 2410713, "author_name": "Priyanshu1235", "author_url": "", "post_date": "2023-08-27T06:58:56.290000", "content": "

Amazing Keep going@sugupoko

", "votes": 1, "replies": [] }, { "id": 2408364, "author_name": "C R Suthikshn Kumar", "author_url": "", "post_date": "2023-08-25T15:35:41.663000", "content": "

Congratulations. Thanks for sharing the details of your solution with nice illustrations.

", "votes": 1, "replies": [] }, { "id": 2413036, "author_name": "Mystic Shadow", "author_url": "", "post_date": "2023-08-28T16:34:05.037000", "content": "

GG sir! nice one

", "votes": 0, "replies": [] } ], "raw_markdown_by_id": { "2407195": "# 95th Bronze solutions\n\nThanks to Kaggle for hosting this interesting competition!!!!\nThis competition was diffucult for us…\nThere were many differences from last competition which made it challenging. I enjoyed learning about transcription task.\n\n# Summary\nWe utilized the best public notebook with some modifications.\n- base notebook: https://www.kaggle.com/code/saidineshpola/aslfr-ctc-based-on-prev-comp-1st-place\n\n# What changes were made from the public notebook?\n* Normalization (+0.002 LB)\n\t* Adjusted all coordinates so that the center of numbers 11 and 12 is set to 0.\n![](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F2930242%2F6d424a3976cee8f769e3e90ba54a9ac2%2Finbox_2930242_fa394dbd96e33fb875aef6ed1f25c757_1.png?generation=1692922021894426&alt=media)\n* Added more features (+0.01 LB)\n\t* Motion, Shape Feature, \n\t\t* Shape is distance of this point\n![](https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F2930242%2F3f7fdc56a208576b7664408fd807e422%2Finbox_2930242_b739316822c577ef9b4be204325f569e_2.png?generation=1692922063719858&alt=media)\n\n* Feature Augmentation (+0.01 LB)\n\t* Framerate, randomDrop, random padding, random crop.\n* Changed Model (+0.005 LB)\n\t* dim : 384 => 192\n\t* conv trans Block num : 3=>10\n\t* Add aux Loss\n* Pretraining with supplementay data (+0.01 LB)\n* Training params\n\t* epoch 100\n\t* batchsize 256\n\t* learningrate 1e-2\n\n# Things I couldn't do\n* Error correction Network.\n\t* I read the following paper and tried to incorporate voice recognition technology, but I couldn't make it in time.\n\t* https://www.google.com/url?sa=t&rct=j&q=&esrc=s&source=web&cd=&cad=rja&uact=8&ved=2ahUKEwjAkrKRvvaAAxWBO3AKHbDkB4MQFnoECBMQAQ&url=https%3A%2F%2Farxiv.org%2Fabs%2F2111.01690&usg=AOvVaw3B--e8pJqXTaYolWYXIBOf&opi=89978449\n* Change to SOTA OCR model.\n\t* https://arxiv.org/pdf/2205.00159.pdf\n\n# Appendix: Model code\n```\ndef get_model(dim = 192):\n inp = tf.keras.Input(INPUT_SHAPE)\n \n x = tf.keras.layers.Masking(mask_value=0.0)(inp)\n x = tf.keras.layers.Dense(dim, use_bias=False,name='stem_conv')(x) + positional_encoding(INPUT_SHAPE[0], dim)\n x = tf.keras.layers.BatchNormalization(momentum=0.95,name='stem_bn')(x)\n\n x = Conv1DBlock(dim, 11, drop_rate=0.2)(x)\n x = Conv1DBlock(dim, 5, drop_rate=0.2)(x)\n x = Conv1DBlock(dim, 3, drop_rate=0.2)(x)\n x = TransformerBlock(dim, expand=2)(x)\n\n x = Conv1DBlock(dim, 11, drop_rate=0.2)(x)\n x = Conv1DBlock(dim, 5, drop_rate=0.2)(x)\n x = Conv1DBlock(dim, 3, drop_rate=0.2)(x)\n x = TransformerBlock(dim, expand=2)(x)\n \n x = Conv1DBlock(dim, 11, drop_rate=0.2)(x)\n x = Conv1DBlock(dim, 5, drop_rate=0.2)(x)\n x = Conv1DBlock(dim, 3, drop_rate=0.2)(x)\n x = TransformerBlock(dim, expand=2)(x)\n\n x = Conv1DBlock(dim, 11, drop_rate=0.2)(x)\n x = Conv1DBlock(dim, 5, drop_rate=0.2)(x)\n x = Conv1DBlock(dim, 3, drop_rate=0.2)(x)\n x = TransformerBlock(dim, expand=2)(x)\n \n x = Conv1DBlock(dim, 11, drop_rate=0.2)(x)\n x = Conv1DBlock(dim, 5, drop_rate=0.2)(x)\n x = Conv1DBlock(dim, 3, drop_rate=0.2)(x)\n x_aux = TransformerBlock(dim, expand=2)(x)\n\n x = Conv1DBlock(dim, 11, drop_rate=0.2)(x_aux)\n x = Conv1DBlock(dim, 5, drop_rate=0.2)(x)\n x = Conv1DBlock(dim, 3, drop_rate=0.2)(x)\n x = TransformerBlock(dim, expand=2)(x)\n\n x = Conv1DBlock(dim, 11, drop_rate=0.2)(x)\n x = Conv1DBlock(dim, 5, drop_rate=0.2)(x)\n x = Conv1DBlock(dim, 3, drop_rate=0.2)(x)\n x = TransformerBlock(dim, expand=2)(x)\n \n x = Conv1DBlock(dim, 11, drop_rate=0.2)(x)\n x = Conv1DBlock(dim, 5, drop_rate=0.2)(x)\n x = Conv1DBlock(dim, 3, drop_rate=0.2)(x)\n x = TransformerBlock(dim, expand=2)(x)\n \n x = Conv1DBlock(dim, 11, drop_rate=0.2)(x)\n x = Conv1DBlock(dim, 5, drop_rate=0.2)(x)\n x = Conv1DBlock(dim, 3, drop_rate=0.2)(x)\n x = TransformerBlock(dim, expand=2)(x)\n\n x = Conv1DBlock(dim, 11, drop_rate=0.2)(x)\n x = Conv1DBlock(dim, 5, drop_rate=0.2)(x)\n x = Conv1DBlock(dim, 3, drop_rate=0.2)(x)\n x = TransformerBlock(dim, expand=2)(x)\n \n # main\n x = tf.keras.layers.Dense(dim*2,activation='relu',name='top_conv')(x)\n x = tf.keras.layers.Dropout(0.4)(x)\n x = tf.keras.layers.Dense(len(char_to_num),name='out')(x)\n\n #aux\n x_aux = tf.keras.layers.Dense(dim*2,activation='relu',name='top_aux')(x_aux)\n x_aux = tf.keras.layers.Dropout(0.4)(x_aux)\n x_aux = tf.keras.layers.Dense(len(char_to_num),name='aux_out')(x_aux) \n \n model = tf.keras.Model(inp, [x, x_aux])\n\n loss = {\n 'out':CTCLoss,\n 'aux_out':CTCLoss\n }\n \n # Adam Optimizer\n optimizer = tfa.optimizers.RectifiedAdam(sma_threshold=4)\n optimizer = tfa.optimizers.Lookahead(optimizer, sync_period=5)\n\n model.compile(loss=loss, optimizer=optimizer)\n\n return model\n```\n", "2410713": "Amazing Keep going@sugupoko ", "2408364": "Congratulations. Thanks for sharing the details of your solution with nice illustrations.", "2413036": "GG sir! nice one" } }