Given that the public LB score is calculated on only 15% of the data, everyone makes many different complex models and averages the result. I took a Higher LB score by tuning mloss (around -6.811) with EfficientNets & CNN and improvement it for optimizing parameters. The model has a minimalist structure, so I only varied the auxiliary parameters:
\n1) Dropout_model= [0.25, 0.32, 0.35, 0.36, 0.37, 0.38, 0.385, 0.39, 0.4],
\n2) FVC_weight and Confidence_weight = [0.15, 0.175, 0.19, 0.2, 0.21, 0.225, 0.25, 0.3, 0.35, 0.5],
\n3) GaussianNoise_stddev= [0.2, 0.25]
The analysis showed that the following options are optimal (LB_score = -6.8087):
\nUsing the library plotly, I built some 3D-interactive plots for different combinations of these parameters, which at 20 points are quite good enough to get a vision of the optimal solution.
\nOf course, you can still improve accuracy (to reduce LB score) by reducing the step, but this increases the risk of overfitting, which is not small.
\nI hope my research will be useful to someone.
\nYour comments, votes, and feedback are most welcome.
", "messages": [ { "id": 1023807, "postDate": "2020-09-23T13:15:43.530Z", "content": "Given that the public LB score is calculated on only 15% of the data, everyone makes many different complex models and averages the result. I took a Higher LB score by tuning mloss (around -6.811) with EfficientNets & CNN and improvement it for optimizing parameters. The model has a minimalist structure, so I only varied the auxiliary parameters:
\n1) Dropout_model= [0.25, 0.32, 0.35, 0.36, 0.37, 0.38, 0.385, 0.39, 0.4],
\n2) FVC_weight and Confidence_weight = [0.15, 0.175, 0.19, 0.2, 0.21, 0.225, 0.25, 0.3, 0.35, 0.5],
\n3) GaussianNoise_stddev= [0.2, 0.25]
The analysis showed that the following options are optimal (LB_score = -6.8087):
\nUsing the library plotly, I built some 3D-interactive plots for different combinations of these parameters, which at 20 points are quite good enough to get a vision of the optimal solution.
\nOf course, you can still improve accuracy (to reduce LB score) by reducing the step, but this increases the risk of overfitting, which is not small.
\nI hope my research will be useful to someone.
\nYour comments, votes, and feedback are most welcome.
", "rawMarkdown": "Given that the public LB score is calculated on only 15% of the data, everyone makes many different complex models and averages the result. I took a [Higher LB score by tuning mloss (around -6.811)](https://www.kaggle.com/reighns/higher-lb-score-by-tuning-mloss-around-6-811) with EfficientNets & CNN and **[improvement it](https://www.kaggle.com/vbmokin/higher-lb-score-by-tuning-mloss-upgrade-visual)** for optimizing parameters. The model has a minimalist structure, so I only varied the auxiliary parameters:\n\n1) **Dropout_model**= [0.25, 0.32, 0.35, 0.36, 0.37, 0.38, 0.385, 0.39, 0.4],\n2) **FVC_weight** and **Confidence_weight** = [0.15, 0.175, 0.19, 0.2, 0.21, 0.225, 0.25, 0.3, 0.35, 0.5],\n3) **GaussianNoise_stddev**= [0.2, 0.25]\n\nThe analysis showed that the following options are optimal (**LB_score = -6.8087**):\n* Dropout_model = 0.38, \n* FVC_weight = Confidence_weight = 0.2, \n* GaussianNoise_stddev = 0.2\n\nUsing the library plotly, I built [some **3D-interactive plots for different combinations of these parameters**](https://www.kaggle.com/vbmokin/higher-lb-score-by-tuning-mloss-upgrade-visual#2.3), which at 20 points are quite good enough to get a vision of the optimal solution.\n\nOf course, you can still improve accuracy (to reduce LB score) by reducing the step, but this increases the risk of overfitting, which is not small.\n\nI hope my research will be useful to someone.\n\nYour comments, votes, and feedback are most welcome.", "votes": 9 }, { "id": 1025814, "postDate": "2020-09-24T19:24:02.903Z", "content": "Changed EfficientNet with other architecture and add 3D aug and also play with parameters, got a good boost.
\nThank you for your information, it was helpful
I was planning on tuning the parameters for various EfficientNets I trained, this gives me a very good head start. Thanks a ton!
", "rawMarkdown": "I was planning on tuning the parameters for various EfficientNets I trained, this gives me a very good head start. Thanks a ton!", "votes": 1 }, { "id": 1026640, "postDate": "2020-09-25T13:15:56.950Z", "rawMarkdown": "", "votes": -1, "isDeleted": true }, { "id": 1024328, "postDate": "2020-09-23T19:00:21.003Z", "content": "Informative , Thanks for sharing
", "rawMarkdown": "Informative , Thanks for sharing", "votes": 1 }, { "id": 1024238, "postDate": "2020-09-23T17:56:46.677Z", "content": "that's really helpful thank You! upvote
", "rawMarkdown": "that's really helpful thank You! upvote" } ], "comments": [ { "id": 1025814, "author_name": "", "author_url": "", "post_date": "2020-09-24T19:24:02.903000", "content": "Changed EfficientNet with other architecture and add 3D aug and also play with parameters, got a good boost.
\nThank you for your information, it was helpful
I was planning on tuning the parameters for various EfficientNets I trained, this gives me a very good head start. Thanks a ton!
", "votes": 1, "replies": [] }, { "id": 1026640, "author_name": "", "author_url": "", "post_date": "2020-09-25T13:15:56.950000", "content": "", "votes": -1, "replies": [] }, { "id": 1024328, "author_name": "Pinaki MIshra", "author_url": "", "post_date": "2020-09-23T19:00:21.003000", "content": "Informative , Thanks for sharing
", "votes": 1, "replies": [] }, { "id": 1024238, "author_name": "Maciej Gronczynski", "author_url": "", "post_date": "2020-09-23T17:56:46.677000", "content": "that's really helpful thank You! upvote
", "votes": 0, "replies": [] } ], "raw_markdown_by_id": { "1023807": "Given that the public LB score is calculated on only 15% of the data, everyone makes many different complex models and averages the result. I took a [Higher LB score by tuning mloss (around -6.811)](https://www.kaggle.com/reighns/higher-lb-score-by-tuning-mloss-around-6-811) with EfficientNets & CNN and **[improvement it](https://www.kaggle.com/vbmokin/higher-lb-score-by-tuning-mloss-upgrade-visual)** for optimizing parameters. The model has a minimalist structure, so I only varied the auxiliary parameters:\n\n1) **Dropout_model**= [0.25, 0.32, 0.35, 0.36, 0.37, 0.38, 0.385, 0.39, 0.4],\n2) **FVC_weight** and **Confidence_weight** = [0.15, 0.175, 0.19, 0.2, 0.21, 0.225, 0.25, 0.3, 0.35, 0.5],\n3) **GaussianNoise_stddev**= [0.2, 0.25]\n\nThe analysis showed that the following options are optimal (**LB_score = -6.8087**):\n* Dropout_model = 0.38, \n* FVC_weight = Confidence_weight = 0.2, \n* GaussianNoise_stddev = 0.2\n\nUsing the library plotly, I built [some **3D-interactive plots for different combinations of these parameters**](https://www.kaggle.com/vbmokin/higher-lb-score-by-tuning-mloss-upgrade-visual#2.3), which at 20 points are quite good enough to get a vision of the optimal solution.\n\nOf course, you can still improve accuracy (to reduce LB score) by reducing the step, but this increases the risk of overfitting, which is not small.\n\nI hope my research will be useful to someone.\n\nYour comments, votes, and feedback are most welcome.", "1025814": "Changed EfficientNet with other architecture and add 3D aug and also play with parameters, got a good boost.\nThank you for your information, it was helpful ", "1024592": "I was planning on tuning the parameters for various EfficientNets I trained, this gives me a very good head start. Thanks a ton!", "1026640": "", "1024328": "Informative , Thanks for sharing", "1024238": "that's really helpful thank You! upvote" } }