![\"enter](\"https://storage.googleapis.com/kaggle-forum-message-attachments/243896/7917/nasnet_scratch.png\" "\\"\\"")
https://research.googleblog.com/2017/11/automl-for-large-scale-image.html
\n\nMachined designed NASNet is is as good as hand designed SeNet, but running faster with less parameters. Wonder if this can be apply to this Cdiscount competition and future kaggle competition.
\n\nAutomation fuse CNN, RNN and RL.
", "messages": [ { "id": "242973", "postDate": "11/13/2017 04:31:55", "content": "https://research.googleblog.com/2017/11/automl-for-large-scale-image.html
\n\nMachined designed NASNet is is as good as hand designed SeNet, but running faster with less parameters. Wonder if this can be apply to this Cdiscount competition and future kaggle competition.
\n\nAutomation fuse CNN, RNN and RL.
", "rawMarkdown": "https://research.googleblog.com/2017/11/automl-for-large-scale-image.html\n\nMachined designed NASNet is is as good as hand designed SeNet, but running faster with less parameters. Wonder if this can be apply to this Cdiscount competition and future kaggle competition.\n\nAutomation fuse CNN, RNN and RL.", "votes": null }, { "id": "242987", "postDate": "11/13/2017 05:49:38", "content": "In Learning Transferable Architectures for Scalable Image Recognition they claimed:
\n\n\n\n\nIn our experiments, the pool of workers in the workqueue consisted of 500 GPUs.
\n
So maybe in the future, who have better and more devices (CPUs, Memory, GPUs etc.) will be the winners.
", "rawMarkdown": "In [Learning Transferable Architectures for Scalable Image Recognition][1] they claimed: \n\n> In our experiments, the pool of workers in the workqueue consisted of 500 GPUs.\n\n\nSo maybe in the future, who have better and more devices (CPUs, Memory, GPUs etc.) will be the winners.\n\n [1]: https://arxiv.org/pdf/1707.07012.pdf", "votes": null }, { "id": "243438", "postDate": "11/14/2017 02:42:00", "content": "hi, how about your single model performance now?
", "rawMarkdown": "hi, how about your single model performance now?", "votes": null }, { "id": "243547", "postDate": "11/14/2017 10:06:20", "content": "NASNet have good accuracy and low computation cost in flops, but it is very slow on practice https://github.com/taehoonlee/tensornets
", "rawMarkdown": "NASNet have good accuracy and low computation cost in flops, but it is very slow on practice https://github.com/taehoonlee/tensornets", "votes": null }, { "id": "243601", "postDate": "11/14/2017 14:04:44", "content": "I tried to train from scratch for a few iterations. compared to senet, nasnet is still faster and better validation (in early iterations, ... but i am not sure if it would still be better for longer iterations).
\n\nTo improve speed, i made some reduction in the layers. I have no idea if it would perform good or bad, but for now, i just give it a try.
\n\nI am interested in nasnet becuase of its dense interconnected structure, which i believe is particularly useful for this cdiscount dataset.
", "rawMarkdown": "I tried to train from scratch for a few iterations. compared to senet, nasnet is still faster and better validation (in early iterations, ... but i am not sure if it would still be better for longer iterations).\n\nTo improve speed, i made some reduction in the layers. I have no idea if it would perform good or bad, but for now, i just give it a try.\n\nI am interested in nasnet becuase of its dense interconnected structure, which i believe is particularly useful for this cdiscount dataset.", "votes": null }, { "id": "243609", "postDate": "11/14/2017 14:19:52", "content": "I change strategy. Single model performance is now irrelevant. My current results is still from the models of https://www.kaggle.com/c/cdiscount-image-classification-challenge/discussion/41021.
\n\nE.g. se-resnet image accuracy has LB 0.68939 (single 180/180 crop). Since a product has 4 images, by ensembling the \"scores from them\" correctly, i can get LB 0.749.
\n\nI am also training new models.
", "rawMarkdown": "I change strategy. Single model performance is now irrelevant. My current results is still from the models of https://www.kaggle.com/c/cdiscount-image-classification-challenge/discussion/41021.\n\nE.g. se-resnet image accuracy has LB 0.68939 (single 180/180 crop). Since a product has 4 images, by ensembling the \"scores from them\" correctly, i can get LB 0.749.\n\nI am also training new models.", "votes": null }, { "id": "243896", "postDate": "11/15/2017 05:22:20", "content": "First epoch of small nasnet training from scratch. About 18 hr per eopch. Not decided if i want to continue.
\n\nhttps://github.com/Cadene/pretrained-models.pytorch/blob/master/pretrainedmodels/nasnet.py
\n\n
Hello Heng,
\n\nI have been trying some thing around the line of ensembling the \"scores from them\" however I only get roughly from 69.5 to 71.5. Maybe I should insist more on this :)
\n\nDid you see a difference taking features from better single model? And how about ensembling features from differents models?
", "rawMarkdown": "Hello Heng,\n\nI have been trying some thing around the line of ensembling the \"scores from them\" however I only get roughly from 69.5 to 71.5. Maybe I should insist more on this :)\n\nDid you see a difference taking features from better single model? And how about ensembling features from differents models?", "votes": null }, { "id": "253123", "postDate": "12/04/2017 13:27:22", "content": "Hi, Heng. Could you please show how do you ensemble the score? Mean?
", "rawMarkdown": "Hi, Heng. Could you please show how do you ensemble the score? Mean?", "votes": null } ], "comments": [ { "id": 242987, "author_name": "dawnbreaker", "author_url": "", "post_date": "11/13/2017 05:49:38", "content": "In Learning Transferable Architectures for Scalable Image Recognition they claimed:
\n\n\n\n\nIn our experiments, the pool of workers in the workqueue consisted of 500 GPUs.
\n
So maybe in the future, who have better and more devices (CPUs, Memory, GPUs etc.) will be the winners.
", "votes": null, "replies": [] }, { "id": 243438, "author_name": "ifighting", "author_url": "", "post_date": "11/14/2017 02:42:00", "content": "hi, how about your single model performance now?
", "votes": null, "replies": [ { "id": 243609, "author_name": "hengck23", "author_url": "", "post_date": "11/14/2017 14:19:52", "content": "I change strategy. Single model performance is now irrelevant. My current results is still from the models of https://www.kaggle.com/c/cdiscount-image-classification-challenge/discussion/41021.
\n\nE.g. se-resnet image accuracy has LB 0.68939 (single 180/180 crop). Since a product has 4 images, by ensembling the \"scores from them\" correctly, i can get LB 0.749.
\n\nI am also training new models.
", "votes": null, "replies": [] }, { "id": 244254, "author_name": "lamdang", "author_url": "", "post_date": "11/15/2017 22:04:07", "content": "Hello Heng,
\n\nI have been trying some thing around the line of ensembling the \"scores from them\" however I only get roughly from 69.5 to 71.5. Maybe I should insist more on this :)
\n\nDid you see a difference taking features from better single model? And how about ensembling features from differents models?
", "votes": null, "replies": [] }, { "id": 253123, "author_name": "zhangsongwei", "author_url": "", "post_date": "12/04/2017 13:27:22", "content": "Hi, Heng. Could you please show how do you ensemble the score? Mean?
", "votes": null, "replies": [] } ] }, { "id": 243547, "author_name": "nicksergievskiy", "author_url": "", "post_date": "11/14/2017 10:06:20", "content": "NASNet have good accuracy and low computation cost in flops, but it is very slow on practice https://github.com/taehoonlee/tensornets
", "votes": null, "replies": [] }, { "id": 243601, "author_name": "hengck23", "author_url": "", "post_date": "11/14/2017 14:04:44", "content": "I tried to train from scratch for a few iterations. compared to senet, nasnet is still faster and better validation (in early iterations, ... but i am not sure if it would still be better for longer iterations).
\n\nTo improve speed, i made some reduction in the layers. I have no idea if it would perform good or bad, but for now, i just give it a try.
\n\nI am interested in nasnet becuase of its dense interconnected structure, which i believe is particularly useful for this cdiscount dataset.
", "votes": null, "replies": [] }, { "id": 243896, "author_name": "hengck23", "author_url": "", "post_date": "11/15/2017 05:22:20", "content": "First epoch of small nasnet training from scratch. About 18 hr per eopch. Not decided if i want to continue.
\n\nhttps://github.com/Cadene/pretrained-models.pytorch/blob/master/pretrainedmodels/nasnet.py
\n\n