For establishing the baseline of this competition, we used a standard Inception V3 architecture with a Softmax loss function. Some standard input preprocessing and augmentation was done on the data, such as random cropping as well as small variations in brightness, hue etc. Something to note is that during training, the losses were weighted differently for each class depending on the frequency of training examples for that class. This was done to alleviate the problem of unbalanced classes. For more information on the technique used, see this paper.
", "messages": [ { "id": "777492", "postDate": "03/17/2020 17:59:59", "content": "For establishing the baseline of this competition, we used a standard Inception V3 architecture with a Softmax loss function. Some standard input preprocessing and augmentation was done on the data, such as random cropping as well as small variations in brightness, hue etc. Something to note is that during training, the losses were weighted differently for each class depending on the frequency of training examples for that class. This was done to alleviate the problem of unbalanced classes. For more information on the technique used, see this paper.
", "rawMarkdown": "For establishing the baseline of this competition, we used a standard Inception V3 architecture with a Softmax loss function. Some standard input preprocessing and augmentation was done on the data, such as random cropping as well as small variations in brightness, hue etc. Something to note is that during training, the losses were weighted differently for each class depending on the frequency of training examples for that class. This was done to alleviate the problem of unbalanced classes. For more information on the technique used, see this [paper].\n\n[paper]: https://arxiv.org/pdf/1901.05555.pdf", "votes": null }, { "id": "787167", "postDate": "03/26/2020 15:15:29", "content": "Hi @agjoka, thanks for the info. How many epochs you trained and what's the input size. Also did you use cropped images or original. Thx
", "rawMarkdown": "Hi @agjoka, thanks for the info. How many epochs you trained and what's the input size. Also did you use cropped images or original. Thx", "votes": null }, { "id": "787349", "postDate": "03/26/2020 18:14:03", "content": "We would like to keep hyperparameter information private for the sake of the competition. That being said, the main model and training decisions given above should be the most crucial pieces of reaching the baseline performance.
", "rawMarkdown": "We would like to keep hyperparameter information private for the sake of the competition. That being said, the main model and training decisions given above should be the most crucial pieces of reaching the baseline performance.", "votes": null }, { "id": "787937", "postDate": "03/27/2020 08:28:50", "content": "thank you for the prompt answer. I do not completely understand what's the point of the baseline if it is kept private, but i do respect your decision. \nThe training time and processing unit is not a hyperparameter. Would you be willing to disclose those? The reason I am asking is to understand whether I can participate with limited resources.
", "rawMarkdown": "thank you for the prompt answer. I do not completely understand what's the point of the baseline if it is kept private, but i do respect your decision. \nThe training time and processing unit is not a hyperparameter. Would you be willing to disclose those? The reason I am asking is to understand whether I can participate with limited resources.", "votes": null }, { "id": "791659", "postDate": "03/30/2020 15:26:15", "content": "Yes, I agree that those are necessary for resource management. We trained for 32 epochs on a Nvidia P100.\nHowever, we only trained on the iWildCam 2020 data. Keep in mind that this challenge also seeks to explore multimodal solutions, so there is additional data which was not trained on but could prove useful. This includes satellite imagery and high quality data from iNaturalist.
", "rawMarkdown": "Yes, I agree that those are necessary for resource management. We trained for 32 epochs on a Nvidia P100.\nHowever, we only trained on the iWildCam 2020 data. Keep in mind that this challenge also seeks to explore multimodal solutions, so there is additional data which was not trained on but could prove useful. This includes satellite imagery and high quality data from iNaturalist.", "votes": null } ], "comments": [ { "id": 787167, "author_name": "valanm", "author_url": "", "post_date": "03/26/2020 15:15:29", "content": "Hi @agjoka, thanks for the info. How many epochs you trained and what's the input size. Also did you use cropped images or original. Thx
", "votes": null, "replies": [ { "id": 787349, "author_name": "agjoka", "author_url": "", "post_date": "03/26/2020 18:14:03", "content": "We would like to keep hyperparameter information private for the sake of the competition. That being said, the main model and training decisions given above should be the most crucial pieces of reaching the baseline performance.
", "votes": null, "replies": [] }, { "id": 787937, "author_name": "valanm", "author_url": "", "post_date": "03/27/2020 08:28:50", "content": "thank you for the prompt answer. I do not completely understand what's the point of the baseline if it is kept private, but i do respect your decision. \nThe training time and processing unit is not a hyperparameter. Would you be willing to disclose those? The reason I am asking is to understand whether I can participate with limited resources.
", "votes": null, "replies": [] }, { "id": 791659, "author_name": "agjoka", "author_url": "", "post_date": "03/30/2020 15:26:15", "content": "Yes, I agree that those are necessary for resource management. We trained for 32 epochs on a Nvidia P100.\nHowever, we only trained on the iWildCam 2020 data. Keep in mind that this challenge also seeks to explore multimodal solutions, so there is additional data which was not trained on but could prove useful. This includes satellite imagery and high quality data from iNaturalist.
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "777492": "For establishing the baseline of this competition, we used a standard Inception V3 architecture with a Softmax loss function. Some standard input preprocessing and augmentation was done on the data, such as random cropping as well as small variations in brightness, hue etc. Something to note is that during training, the losses were weighted differently for each class depending on the frequency of training examples for that class. This was done to alleviate the problem of unbalanced classes. For more information on the technique used, see this [paper].\n\n[paper]: https://arxiv.org/pdf/1901.05555.pdf", "787167": "Hi @agjoka, thanks for the info. How many epochs you trained and what's the input size. Also did you use cropped images or original. Thx", "787349": "We would like to keep hyperparameter information private for the sake of the competition. That being said, the main model and training decisions given above should be the most crucial pieces of reaching the baseline performance.", "787937": "thank you for the prompt answer. I do not completely understand what's the point of the baseline if it is kept private, but i do respect your decision. \nThe training time and processing unit is not a hyperparameter. Would you be willing to disclose those? The reason I am asking is to understand whether I can participate with limited resources.", "791659": "Yes, I agree that those are necessary for resource management. We trained for 32 epochs on a Nvidia P100.\nHowever, we only trained on the iWildCam 2020 data. Keep in mind that this challenge also seeks to explore multimodal solutions, so there is additional data which was not trained on but could prove useful. This includes satellite imagery and high quality data from iNaturalist." }, "source": "meta" }