In the rules of the competition it is stated that \"External data not permitted, but whitelisted pre-trained models are permitted. \". Which are these pre-trained models that are allowed to be used? And how do I get them into the kernel if internet is blocked?
\n\nThanks.
", "messages": [ { "id": "443046", "postDate": "12/20/2018 23:29:41", "content": "In the rules of the competition it is stated that \"External data not permitted, but whitelisted pre-trained models are permitted. \". Which are these pre-trained models that are allowed to be used? And how do I get them into the kernel if internet is blocked?
\n\nThanks.
", "rawMarkdown": "In the rules of the competition it is stated that \"External data not permitted, but whitelisted pre-trained models are permitted. \". Which are these pre-trained models that are allowed to be used? And how do I get them into the kernel if internet is blocked?\n\nThanks.", "votes": null }, { "id": "443087", "postDate": "12/21/2018 02:04:16", "content": "i have exactly same question can someone please answer this.I was planning to Fast.ai pretrained model but i asssume it is Trained on a huge corpus of wiki data so it might not be allowed.
", "rawMarkdown": "i have exactly same question can someone please answer this.I was planning to Fast.ai pretrained model but i asssume it is Trained on a huge corpus of wiki data so it might not be allowed.", "votes": null }, { "id": "443159", "postDate": "12/21/2018 05:36:31", "content": "I think the rules mean we can use the pre-trained word vectors which are provided in the dataset
", "rawMarkdown": "I think the rules mean we can use the pre-trained word vectors which are provided in the dataset", "votes": null }, { "id": "443194", "postDate": "12/21/2018 07:41:30", "content": "Everything that works out of the box within the docker image should be allowed I think.
", "rawMarkdown": "Everything that works out of the box within the docker image should be allowed I think.", "votes": null }, { "id": "443199", "postDate": "12/21/2018 07:46:15", "content": "Check the data tab and scroll all the way to the bottom. Under \"Embeddings\", you will find a list.
", "rawMarkdown": "Check the data tab and scroll all the way to the bottom. Under \"Embeddings\", you will find a list.", "votes": null }, { "id": "443519", "postDate": "12/21/2018 19:50:14", "content": "What I can find under embeddings are, well, embeddings :) They are results of pretrained models, but definitely not pretrained models themselves... What I wanted to use is actual models, which can do inference on the fly.
", "rawMarkdown": "What I can find under embeddings are, well, embeddings :) They are results of pretrained models, but definitely not pretrained models themselves... What I wanted to use is actual models, which can do inference on the fly.", "votes": null }, { "id": "443522", "postDate": "12/21/2018 19:56:12", "content": "Sorry, that's the only pretrained anything you get :) I guess their wording was ambiguous, but there's nothing but embeddings.
\n\nAlso, embeddings are not the result of models. They are the models. One embedding matrix is basically a giant weight matrix of a neural network.
", "rawMarkdown": "Sorry, that's the only pretrained anything you get :) I guess their wording was ambiguous, but there's nothing but embeddings.\n\nAlso, embeddings are not the result of models. They *are* the models. One embedding matrix is basically a giant weight matrix of a neural network.", "votes": null }, { "id": "443556", "postDate": "12/21/2018 21:16:30", "content": "Well, technically speaking, you are right. But this factor is diminished by the fact that you pass one-hot encoded vector as input - it is the same as just picking a vector from embedding as is where the value of input data is 1. The rest will be multiplied with zeros.
", "rawMarkdown": "Well, technically speaking, you are right. But this factor is diminished by the fact that you pass one-hot encoded vector as input - it is the same as just picking a vector from embedding as is where the value of input data is 1. The rest will be multiplied with zeros.", "votes": null }, { "id": "443566", "postDate": "12/21/2018 21:55:32", "content": "Well, my main point was: There's only the embeddings. No other pretrained models.
", "rawMarkdown": "Well, my main point was: There's only the embeddings. No other pretrained models.", "votes": null } ], "comments": [ { "id": 443087, "author_name": "chansbest93", "author_url": "", "post_date": "12/21/2018 02:04:16", "content": "i have exactly same question can someone please answer this.I was planning to Fast.ai pretrained model but i asssume it is Trained on a huge corpus of wiki data so it might not be allowed.
", "votes": null, "replies": [] }, { "id": 443159, "author_name": "wujianhong", "author_url": "", "post_date": "12/21/2018 05:36:31", "content": "I think the rules mean we can use the pre-trained word vectors which are provided in the dataset
", "votes": null, "replies": [] }, { "id": 443194, "author_name": "philippsinger", "author_url": "", "post_date": "12/21/2018 07:41:30", "content": "Everything that works out of the box within the docker image should be allowed I think.
", "votes": null, "replies": [] }, { "id": 443199, "author_name": "mschumacher", "author_url": "", "post_date": "12/21/2018 07:46:15", "content": "Check the data tab and scroll all the way to the bottom. Under \"Embeddings\", you will find a list.
", "votes": null, "replies": [ { "id": 443519, "author_name": "ishitori", "author_url": "", "post_date": "12/21/2018 19:50:14", "content": "What I can find under embeddings are, well, embeddings :) They are results of pretrained models, but definitely not pretrained models themselves... What I wanted to use is actual models, which can do inference on the fly.
", "votes": null, "replies": [] }, { "id": 443522, "author_name": "mschumacher", "author_url": "", "post_date": "12/21/2018 19:56:12", "content": "Sorry, that's the only pretrained anything you get :) I guess their wording was ambiguous, but there's nothing but embeddings.
\n\nAlso, embeddings are not the result of models. They are the models. One embedding matrix is basically a giant weight matrix of a neural network.
", "votes": null, "replies": [] }, { "id": 443556, "author_name": "ishitori", "author_url": "", "post_date": "12/21/2018 21:16:30", "content": "Well, technically speaking, you are right. But this factor is diminished by the fact that you pass one-hot encoded vector as input - it is the same as just picking a vector from embedding as is where the value of input data is 1. The rest will be multiplied with zeros.
", "votes": null, "replies": [] }, { "id": 443566, "author_name": "mschumacher", "author_url": "", "post_date": "12/21/2018 21:55:32", "content": "Well, my main point was: There's only the embeddings. No other pretrained models.
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "443046": "In the rules of the competition it is stated that \"External data not permitted, but whitelisted pre-trained models are permitted. \". Which are these pre-trained models that are allowed to be used? And how do I get them into the kernel if internet is blocked?\n\nThanks.", "443087": "i have exactly same question can someone please answer this.I was planning to Fast.ai pretrained model but i asssume it is Trained on a huge corpus of wiki data so it might not be allowed.", "443159": "I think the rules mean we can use the pre-trained word vectors which are provided in the dataset", "443194": "Everything that works out of the box within the docker image should be allowed I think.", "443199": "Check the data tab and scroll all the way to the bottom. Under \"Embeddings\", you will find a list.", "443519": "What I can find under embeddings are, well, embeddings :) They are results of pretrained models, but definitely not pretrained models themselves... What I wanted to use is actual models, which can do inference on the fly.", "443522": "Sorry, that's the only pretrained anything you get :) I guess their wording was ambiguous, but there's nothing but embeddings.\n\nAlso, embeddings are not the result of models. They *are* the models. One embedding matrix is basically a giant weight matrix of a neural network.", "443556": "Well, technically speaking, you are right. But this factor is diminished by the fact that you pass one-hot encoded vector as input - it is the same as just picking a vector from embedding as is where the value of input data is 1. The rest will be multiplied with zeros.", "443566": "Well, my main point was: There's only the embeddings. No other pretrained models." }, "source": "meta" }