I got some top features by matching text with LSTM and matching images as a 3d tensors(seq of images as single example vs another sequence of images and 3dConvolutions+MaxPolling). I used sigmoid and siamese netowork loss. Of course I would probably be better off by using normal features and more ensembles rather then pushing this large dataset through my GPU.
", "messages": [ { "id": "126970", "postDate": "07/12/2016 23:45:28", "content": "I got some top features by matching text with LSTM and matching images as a 3d tensors(seq of images as single example vs another sequence of images and 3dConvolutions+MaxPolling). I used sigmoid and siamese netowork loss. Of course I would probably be better off by using normal features and more ensembles rather then pushing this large dataset through my GPU.
", "rawMarkdown": "I got some top features by matching text with LSTM and matching images as a 3d tensors(seq of images as single example vs another sequence of images and 3dConvolutions+MaxPolling). I used sigmoid and siamese netowork loss. Of course I would probably be better off by using normal features and more ensembles rather then pushing this large dataset through my GPU.", "votes": null }, { "id": "126980", "postDate": "07/13/2016 00:06:02", "content": "We used a pretrained VGG-19 net for extracting 4096 deep features for each images. Then various distances between two clusters of images was computed (e.g. mean, max, min, var...).
\n\nGreat try on LSTM. Could you elaborate a bit more?
", "rawMarkdown": "We used a pretrained VGG-19 net for extracting 4096 deep features for each images. Then various distances between two clusters of images was computed (e.g. mean, max, min, var...). \r\n\r\nGreat try on LSTM. Could you elaborate a bit more?", "votes": null }, { "id": "126989", "postDate": "07/13/2016 00:29:03", "content": "I used Keras and here are the networks. \nFor text It scanned title and descriptions with same GRU(several GRU or Convolution1D+GRU) and Embedding and GRU weights are shared for title and description. \nhttps://gist.github.com/yurkor/9724b88aa602a9548b4536c5a1c8fcf7
", "rawMarkdown": "I used Keras and here are the networks. \r\nFor text It scanned title and descriptions with same GRU(several GRU or Convolution1D+GRU) and Embedding and GRU weights are shared for title and description. \r\nhttps://gist.github.com/yurkor/9724b88aa602a9548b4536c5a1c8fcf7", "votes": null }, { "id": "127041", "postDate": "07/13/2016 05:42:41", "content": "We use pretrained inception-bn in https://github.com/dmlc/mxnet-model-gallery to get feature(layer before softmax) for each image. For each item pairs, following features are used:
\n\nThese features give roughly 0.005 increase in auc.
\n\nTime of feature extraction is about 5 days with single Titan X.
", "rawMarkdown": "We use pretrained inception-bn in https://github.com/dmlc/mxnet-model-gallery to get feature(layer before softmax) for each image. For each item pairs, following features are used:\r\n\r\n1. for each item, calculate mean feature of all images; use the cosine similarity between two features.\r\n2. for item with less images, for each image within, find the most similar image in cosine similarity in another item, get this similarity; use the mean, min, max of the similarities in previous step;\r\n\r\nThese features give roughly 0.005 increase in auc.\r\n\r\nTime of feature extraction is about 5 days with single Titan X.", "votes": null } ], "comments": [ { "id": 126980, "author_name": "vinhnguyen", "author_url": "", "post_date": "07/13/2016 00:06:02", "content": "We used a pretrained VGG-19 net for extracting 4096 deep features for each images. Then various distances between two clusters of images was computed (e.g. mean, max, min, var...).
\n\nGreat try on LSTM. Could you elaborate a bit more?
", "votes": null, "replies": [] }, { "id": 126989, "author_name": "yuraka", "author_url": "", "post_date": "07/13/2016 00:29:03", "content": "I used Keras and here are the networks. \nFor text It scanned title and descriptions with same GRU(several GRU or Convolution1D+GRU) and Embedding and GRU weights are shared for title and description. \nhttps://gist.github.com/yurkor/9724b88aa602a9548b4536c5a1c8fcf7
", "votes": null, "replies": [] }, { "id": 127041, "author_name": "valueq", "author_url": "", "post_date": "07/13/2016 05:42:41", "content": "We use pretrained inception-bn in https://github.com/dmlc/mxnet-model-gallery to get feature(layer before softmax) for each image. For each item pairs, following features are used:
\n\nThese features give roughly 0.005 increase in auc.
\n\nTime of feature extraction is about 5 days with single Titan X.
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "126970": "I got some top features by matching text with LSTM and matching images as a 3d tensors(seq of images as single example vs another sequence of images and 3dConvolutions+MaxPolling). I used sigmoid and siamese netowork loss. Of course I would probably be better off by using normal features and more ensembles rather then pushing this large dataset through my GPU.", "126980": "We used a pretrained VGG-19 net for extracting 4096 deep features for each images. Then various distances between two clusters of images was computed (e.g. mean, max, min, var...). \r\n\r\nGreat try on LSTM. Could you elaborate a bit more?", "126989": "I used Keras and here are the networks. \r\nFor text It scanned title and descriptions with same GRU(several GRU or Convolution1D+GRU) and Embedding and GRU weights are shared for title and description. \r\nhttps://gist.github.com/yurkor/9724b88aa602a9548b4536c5a1c8fcf7", "127041": "We use pretrained inception-bn in https://github.com/dmlc/mxnet-model-gallery to get feature(layer before softmax) for each image. For each item pairs, following features are used:\r\n\r\n1. for each item, calculate mean feature of all images; use the cosine similarity between two features.\r\n2. for item with less images, for each image within, find the most similar image in cosine similarity in another item, get this similarity; use the mean, min, max of the similarities in previous step;\r\n\r\nThese features give roughly 0.005 increase in auc.\r\n\r\nTime of feature extraction is about 5 days with single Titan X." }, "source": "meta" }