Been doing a tour of duty using bottlneck features from Xception, Inception, VGG16, InceptionResNetV2. Anyone out there found success with MobileNet (https://github.com/keras-team/keras/blob/master/keras/applications/mobilenet.py) and/or NASNet (https://github.com/titu1994/Keras-NASNet. Using logreg to predict on MobileNet features, led to higher error than expected: Validation NASNet LogLoss 1.12359547044. NASNet higher error: Validation NASNetMobile LogLoss 4.36254491973. Welcome thoughts on successful implementation of Mobile and NAS.
", "messages": [ { "id": "282815", "postDate": "02/14/2018 16:30:32", "content": "Been doing a tour of duty using bottlneck features from Xception, Inception, VGG16, InceptionResNetV2. Anyone out there found success with MobileNet (https://github.com/keras-team/keras/blob/master/keras/applications/mobilenet.py) and/or NASNet (https://github.com/titu1994/Keras-NASNet. Using logreg to predict on MobileNet features, led to higher error than expected: Validation NASNet LogLoss 1.12359547044. NASNet higher error: Validation NASNetMobile LogLoss 4.36254491973. Welcome thoughts on successful implementation of Mobile and NAS.
", "rawMarkdown": "Been doing a tour of duty using bottlneck features from Xception, Inception, VGG16, InceptionResNetV2. Anyone out there found success with MobileNet (https://github.com/keras-team/keras/blob/master/keras/applications/mobilenet.py) and/or NASNet (https://github.com/titu1994/Keras-NASNet. Using logreg to predict on MobileNet features, led to higher error than expected: Validation NASNet LogLoss 1.12359547044. NASNet higher error: Validation NASNetMobile LogLoss 4.36254491973. Welcome thoughts on successful implementation of Mobile and NAS.", "votes": null }, { "id": "285008", "postDate": "02/19/2018 03:48:43", "content": "Nasnet pretrained on imagenet, unfreezing cell_12 to last fc. Just try it.
", "rawMarkdown": "Nasnet pretrained on imagenet, unfreezing cell_12 to last fc. Just try it.", "votes": null }, { "id": "286309", "postDate": "02/22/2018 01:04:33", "content": "My current best solution is actually a logistic regression model based on nasnet_mobile and mobilenet features, because they are small models that I can easily train on my laptop.\nI have used Tensorflow (TF slim original models) and in particular the features from this layers: final_layer/dropout/Identity (nasnet_mobile) and MobilenetV1/Logits/Dropout_1b/Identity (mobilenet). Some scripts (not documented) are here https://github.com/SlipknotTN/kaggle_dog_breed
", "rawMarkdown": "My current best solution is actually a logistic regression model based on nasnet_mobile and mobilenet features, because they are small models that I can easily train on my laptop.\nI have used Tensorflow (TF slim original models) and in particular the features from this layers: final_layer/dropout/Identity (nasnet_mobile) and MobilenetV1/Logits/Dropout_1b/Identity (mobilenet). Some scripts (not documented) are here https://github.com/SlipknotTN/kaggle_dog_breed", "votes": null }, { "id": "286791", "postDate": "02/22/2018 19:37:00", "content": "Thanks so much for the feedback, thoughts, and recommendations. So far my best model is an ensemble of two different xception models, inception_v3, and inception_resnetv2. I haven't had good success with NASMobileNet using Keras.
", "rawMarkdown": "Thanks so much for the feedback, thoughts, and recommendations. So far my best model is an ensemble of two different xception models, inception_v3, and inception_resnetv2. I haven't had good success with NASMobileNet using Keras.", "votes": null } ], "comments": [ { "id": 285008, "author_name": "yangweiioe", "author_url": "", "post_date": "02/19/2018 03:48:43", "content": "Nasnet pretrained on imagenet, unfreezing cell_12 to last fc. Just try it.
", "votes": null, "replies": [] }, { "id": 286309, "author_name": "slipknottn", "author_url": "", "post_date": "02/22/2018 01:04:33", "content": "My current best solution is actually a logistic regression model based on nasnet_mobile and mobilenet features, because they are small models that I can easily train on my laptop.\nI have used Tensorflow (TF slim original models) and in particular the features from this layers: final_layer/dropout/Identity (nasnet_mobile) and MobilenetV1/Logits/Dropout_1b/Identity (mobilenet). Some scripts (not documented) are here https://github.com/SlipknotTN/kaggle_dog_breed
", "votes": null, "replies": [] }, { "id": 286791, "author_name": "nudroc", "author_url": "", "post_date": "02/22/2018 19:37:00", "content": "Thanks so much for the feedback, thoughts, and recommendations. So far my best model is an ensemble of two different xception models, inception_v3, and inception_resnetv2. I haven't had good success with NASMobileNet using Keras.
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "282815": "Been doing a tour of duty using bottlneck features from Xception, Inception, VGG16, InceptionResNetV2. Anyone out there found success with MobileNet (https://github.com/keras-team/keras/blob/master/keras/applications/mobilenet.py) and/or NASNet (https://github.com/titu1994/Keras-NASNet. Using logreg to predict on MobileNet features, led to higher error than expected: Validation NASNet LogLoss 1.12359547044. NASNet higher error: Validation NASNetMobile LogLoss 4.36254491973. Welcome thoughts on successful implementation of Mobile and NAS.", "285008": "Nasnet pretrained on imagenet, unfreezing cell_12 to last fc. Just try it.", "286309": "My current best solution is actually a logistic regression model based on nasnet_mobile and mobilenet features, because they are small models that I can easily train on my laptop.\nI have used Tensorflow (TF slim original models) and in particular the features from this layers: final_layer/dropout/Identity (nasnet_mobile) and MobilenetV1/Logits/Dropout_1b/Identity (mobilenet). Some scripts (not documented) are here https://github.com/SlipknotTN/kaggle_dog_breed", "286791": "Thanks so much for the feedback, thoughts, and recommendations. So far my best model is an ensemble of two different xception models, inception_v3, and inception_resnetv2. I haven't had good success with NASMobileNet using Keras." }, "source": "meta" }