Here I will try to show an example of how use deepsense.ai to compare and tune models.
\n\nFirst of all, you should register. See this. Then follow the install instructions.
\n\nTraining Data\nTo run an fast experiment in deepsense.ai I used only images from the less frequent classes. So I got 10,428 train and 2,362 validation images, and 2 files with less than 50M. Take a look at this kernel.
\n\nDirectory Structure\nCreate a folder ~/neptune/data and put the files category_names.csv, train_sample.bin and val_sample.bin. These files are created in the kernel mentioned above.
\n\nCreate a folder for you experiment: ~/neptune/my_experiment\nCreate a main.py file and paste the code from this kernel.
", "messages": [ { "id": "233057", "postDate": "10/19/2017 04:20:49", "content": "Here I will try to show an example of how use deepsense.ai to compare and tune models.
\n\nFirst of all, you should register. See this. Then follow the install instructions.
\n\nTraining Data\nTo run an fast experiment in deepsense.ai I used only images from the less frequent classes. So I got 10,428 train and 2,362 validation images, and 2 files with less than 50M. Take a look at this kernel.
\n\nDirectory Structure\nCreate a folder ~/neptune/data and put the files category_names.csv, train_sample.bin and val_sample.bin. These files are created in the kernel mentioned above.
\n\nCreate a folder for you experiment: ~/neptune/my_experiment\nCreate a main.py file and paste the code from this kernel.
", "rawMarkdown": "Here I will try to show an example of how use deepsense.ai to compare and tune models. \n\nFirst of all, you should register. [See this][1]. Then follow the install instructions. \n\n**Training Data**\nTo run an fast experiment in deepsense.ai I used only images from the less frequent classes. So I got 10,428 train and 2,362 validation images, and 2 files with less than 50M. Take a look at [this kernel][2].\n\n**Directory Structure**\nCreate a folder ~/neptune/data and put the files category_names.csv, train_sample.bin and val_sample.bin. These files are created in the kernel mentioned above.\n\nCreate a folder for you experiment: ~/neptune/my_experiment\nCreate a main.py file and paste the code from [this kernel][3]. \n\n\n [1]: https://www.kaggle.com/c/cdiscount-image-classification-challenge/discussion/41478\n [2]: https://www.kaggle.com/aloisiodn/fast-keras-generator-for-deepsense-ai\n [3]: https://www.kaggle.com/aloisiodn/deepsense-ai-experiment-example", "votes": null }, { "id": "233065", "postDate": "10/19/2017 04:53:36", "content": "Experiment Submission\nFrom your experiment folder, issue this:
\n\n>neptune send --environment keras-2.0-gpu \\
\n\n> --worker gcp-gpu-large \\
\n\n> --input train_sample.bin \\
\n\n> --input val_sample.bin \\
\n\n> --input category_names.csv
\n\nYou should see your experiment running, like this:
", "rawMarkdown": "**Experiment Submission**\nFrom your experiment folder, issue this:\n\n>neptune send --environment keras-2.0-gpu \\\n\n> --worker gcp-gpu-large \\\n\n> --input train_sample.bin \\\n\n> --input val_sample.bin \\\n\n> --input category_names.csv\n\nYou should see your experiment running, like this:", "votes": null }, { "id": "233066", "postDate": "10/19/2017 04:54:19", "content": "Training Data Upload\nIssue these commands:
\n\n>neptune data upload ../data/train_sample.bin
\n\n>neptune data upload ../data/val_sample.bin
\n\n>neptune data upload ../data/category_names.csv
", "rawMarkdown": "**Training Data Upload**\nIssue these commands:\n\n>neptune data upload ../data/train_sample.bin\n\n>neptune data upload ../data/val_sample.bin\n\n>neptune data upload ../data/category_names.csv", "votes": null }, { "id": "235262", "postDate": "10/25/2017 01:57:54", "content": "Hi Alosio, what is the PLB score I should expect from running this script?
", "rawMarkdown": "Hi Alosio, what is the PLB score I should expect from running this script?", "votes": null }, { "id": "235319", "postDate": "10/25/2017 05:55:52", "content": "Hi Aloisio. Remember that you could also use the mounted data from /public/Cdiscount if you don't want to upload.
", "rawMarkdown": "Hi Aloisio. Remember that you could also use the mounted data from /public/Cdiscount if you don't want to upload.", "votes": null }, { "id": "235501", "postDate": "10/25/2017 15:38:58", "content": "Hi Steven. I am not currently using the script to achieve high LB scores directly. I am using them to understand the effects of augmentation and other InceptionV3 traing parameters, dropout, etc . As with this dataset the NN runs pretty fast, is easy to conduct experiments.
", "rawMarkdown": "Hi Steven. I am not currently using the script to achieve high LB scores directly. I am using them to understand the effects of augmentation and other InceptionV3 traing parameters, dropout, etc . As with this dataset the NN runs pretty fast, is easy to conduct experiments.", "votes": null }, { "id": "235502", "postDate": "10/25/2017 15:41:02", "content": "I'm aware of that. I just wanted to make sure I ran everything correctly for the starter code.
", "rawMarkdown": "I'm aware of that. I just wanted to make sure I ran everything correctly for the starter code.", "votes": null }, { "id": "235503", "postDate": "10/25/2017 15:41:07", "content": "Hi Jakub, I am aware of that. But, at this moment, I am using neptune to conduct fast experiments with an smaller dataset. It has been very usefull. Thanks!
", "rawMarkdown": "Hi Jakub, I am aware of that. But, at this moment, I am using neptune to conduct fast experiments with an smaller dataset. It has been very usefull. Thanks!", "votes": null }, { "id": "235506", "postDate": "10/25/2017 15:50:04", "content": "Ok. As a matter of fact, the score should be very low, since the script gets the 400 least frequent classes. You may get better results if you change the bin file generator to get the 5 to 10 most frequent classes for example. The problem with this approach is that the NN get to 100% accuracy very fast and this is not usefull for tweaking the parameters for the final version.
", "rawMarkdown": "Ok. As a matter of fact, the score should be very low, since the script gets the 400 least frequent classes. You may get better results if you change the bin file generator to get the 5 to 10 most frequent classes for example. The problem with this approach is that the NN get to 100% accuracy very fast and this is not usefull for tweaking the parameters for the final version.", "votes": null } ], "comments": [ { "id": 233065, "author_name": "aloisiodn", "author_url": "", "post_date": "10/19/2017 04:53:36", "content": "Experiment Submission\nFrom your experiment folder, issue this:
\n\n>neptune send --environment keras-2.0-gpu \\
\n\n> --worker gcp-gpu-large \\
\n\n> --input train_sample.bin \\
\n\n> --input val_sample.bin \\
\n\n> --input category_names.csv
\n\nYou should see your experiment running, like this:
", "votes": null, "replies": [] }, { "id": 233066, "author_name": "aloisiodn", "author_url": "", "post_date": "10/19/2017 04:54:19", "content": "Training Data Upload\nIssue these commands:
\n\n>neptune data upload ../data/train_sample.bin
\n\n>neptune data upload ../data/val_sample.bin
\n\n>neptune data upload ../data/category_names.csv
", "votes": null, "replies": [ { "id": 235319, "author_name": "jakubczakon", "author_url": "", "post_date": "10/25/2017 05:55:52", "content": "Hi Aloisio. Remember that you could also use the mounted data from /public/Cdiscount if you don't want to upload.
", "votes": null, "replies": [] }, { "id": 235503, "author_name": "aloisiodn", "author_url": "", "post_date": "10/25/2017 15:41:07", "content": "Hi Jakub, I am aware of that. But, at this moment, I am using neptune to conduct fast experiments with an smaller dataset. It has been very usefull. Thanks!
", "votes": null, "replies": [] } ] }, { "id": 235262, "author_name": "stevenknguyen", "author_url": "", "post_date": "10/25/2017 01:57:54", "content": "Hi Alosio, what is the PLB score I should expect from running this script?
", "votes": null, "replies": [ { "id": 235501, "author_name": "aloisiodn", "author_url": "", "post_date": "10/25/2017 15:38:58", "content": "Hi Steven. I am not currently using the script to achieve high LB scores directly. I am using them to understand the effects of augmentation and other InceptionV3 traing parameters, dropout, etc . As with this dataset the NN runs pretty fast, is easy to conduct experiments.
", "votes": null, "replies": [] }, { "id": 235502, "author_name": "stevenknguyen", "author_url": "", "post_date": "10/25/2017 15:41:02", "content": "I'm aware of that. I just wanted to make sure I ran everything correctly for the starter code.
", "votes": null, "replies": [] }, { "id": 235506, "author_name": "aloisiodn", "author_url": "", "post_date": "10/25/2017 15:50:04", "content": "Ok. As a matter of fact, the score should be very low, since the script gets the 400 least frequent classes. You may get better results if you change the bin file generator to get the 5 to 10 most frequent classes for example. The problem with this approach is that the NN get to 100% accuracy very fast and this is not usefull for tweaking the parameters for the final version.
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "233057": "Here I will try to show an example of how use deepsense.ai to compare and tune models. \n\nFirst of all, you should register. [See this][1]. Then follow the install instructions. \n\n**Training Data**\nTo run an fast experiment in deepsense.ai I used only images from the less frequent classes. So I got 10,428 train and 2,362 validation images, and 2 files with less than 50M. Take a look at [this kernel][2].\n\n**Directory Structure**\nCreate a folder ~/neptune/data and put the files category_names.csv, train_sample.bin and val_sample.bin. These files are created in the kernel mentioned above.\n\nCreate a folder for you experiment: ~/neptune/my_experiment\nCreate a main.py file and paste the code from [this kernel][3]. \n\n\n [1]: https://www.kaggle.com/c/cdiscount-image-classification-challenge/discussion/41478\n [2]: https://www.kaggle.com/aloisiodn/fast-keras-generator-for-deepsense-ai\n [3]: https://www.kaggle.com/aloisiodn/deepsense-ai-experiment-example", "233065": "**Experiment Submission**\nFrom your experiment folder, issue this:\n\n>neptune send --environment keras-2.0-gpu \\\n\n> --worker gcp-gpu-large \\\n\n> --input train_sample.bin \\\n\n> --input val_sample.bin \\\n\n> --input category_names.csv\n\nYou should see your experiment running, like this:", "233066": "**Training Data Upload**\nIssue these commands:\n\n>neptune data upload ../data/train_sample.bin\n\n>neptune data upload ../data/val_sample.bin\n\n>neptune data upload ../data/category_names.csv", "235262": "Hi Alosio, what is the PLB score I should expect from running this script?", "235319": "Hi Aloisio. Remember that you could also use the mounted data from /public/Cdiscount if you don't want to upload.", "235501": "Hi Steven. I am not currently using the script to achieve high LB scores directly. I am using them to understand the effects of augmentation and other InceptionV3 traing parameters, dropout, etc . As with this dataset the NN runs pretty fast, is easy to conduct experiments.", "235502": "I'm aware of that. I just wanted to make sure I ran everything correctly for the starter code.", "235503": "Hi Jakub, I am aware of that. But, at this moment, I am using neptune to conduct fast experiments with an smaller dataset. It has been very usefull. Thanks!", "235506": "Ok. As a matter of fact, the score should be very low, since the script gets the 400 least frequent classes. You may get better results if you change the bin file generator to get the 5 to 10 most frequent classes for example. The problem with this approach is that the NN get to 100% accuracy very fast and this is not usefull for tweaking the parameters for the final version." }, "source": "meta" }