{ "id": 71747, "title": "Is there anyone can help me(keras beginner)?", "url": "/competitions/quickdraw-doodle-recognition/discussion/71747", "author_name": "upup", "post_date": "2018-11-16T07:04:39.717000", "votes": 0, "comment_count": 2, "views": 0, "content": "

I'm confused about model = MobileNet(input_shape=(size, size, 1), alpha=1., weights=None, classes=NCATS) in the helpful kernel created by beluga, because the keras docs say

\n\n
\n

input_shape: optional shape tuple, only to be specified if include_top is False (otherwise the input_shape has to be (224, 224, 3) (with 'channelslast' data format) or (3, 224, 224) (with 'channelsfirst' data format). It should have exactly 3 inputs channels, and width and height should be no smaller than 32. E.g. (200, 200, 3) would be one valid value.

\n
\n\n

Why can this kernel specify the input_shape and include_top=True(default) simultaneously and run without any error？
\n(btw, how to @ someone ?)

", "messages": [ { "id": 424257, "postDate": "2018-11-19T20:02:40.603Z", "content": "", "rawMarkdown": "+ When you give `weights=None`, keras is not loading any weights, so the network need not be fixed by the weight matrix.\n+ It is just returning the architecture/network, and the input layer `input_shape=(size, size, 1)` becomes customisable .because Mobilenet/Resnet are CNNs, containing weight matrices which can be multiplied over images of any size.\n+ `classes=NCATS` can also be variable since output layer of **340 Linear Layer** is created on-the-go, by keras \n+ In recent architectures, `GlobalAveragePooling/AdaptiveAveragePooling` is used so any number of conv layers are converted to FC layer.", "votes": 1 }, { "id": 422400, "postDate": "2018-11-16T07:09:14.290Z", "content": "

Because its not trying to load pre-trained weights (weights=None) so you are free to do whatever you want with the architecture.

", "rawMarkdown": "Because its not trying to load pre-trained weights (weights=None) so you are free to do whatever you want with the architecture.", "votes": 2 }, { "id": 422398, "postDate": "2018-11-16T07:04:39.717Z", "content": "

I'm confused about model = MobileNet(input_shape=(size, size, 1), alpha=1., weights=None, classes=NCATS) in the helpful kernel created by beluga, because the keras docs say

\n\n
\n

input_shape: optional shape tuple, only to be specified if include_top is False (otherwise the input_shape has to be (224, 224, 3) (with 'channelslast' data format) or (3, 224, 224) (with 'channelsfirst' data format). It should have exactly 3 inputs channels, and width and height should be no smaller than 32. E.g. (200, 200, 3) would be one valid value.

\n
\n\n

Why can this kernel specify the input_shape and include_top=True(default) simultaneously and run without any error？
\n(btw, how to @ someone ?)

", "rawMarkdown": "I'm confused about `model = MobileNet(input_shape=(size, size, 1), alpha=1., weights=None, classes=NCATS)` in the helpful [kernel created by beluga](https://www.kaggle.com/gaborfodor/greyscale-mobilenet-lb-0-892/notebook), because the [keras docs](https://keras.io/applications/#mobilenet) say \n> `input_shape`: optional shape tuple, only to be specified if `include_top` is False (otherwise the `input_shape` has to be (224, 224, 3) (with 'channelslast' data format) or (3, 224, 224) (with 'channelsfirst' data format). It should have exactly 3 inputs channels, and width and height should be no smaller than 32. E.g. (200, 200, 3) would be one valid value.\n\nWhy can this kernel specify the `input_shape` and `include_top=True`(default) simultaneously and run without any error？ \n(btw, how to @ someone ?)" } ], "comments": [ { "id": 424257, "author_name": "remidi", "author_url": "", "post_date": "2018-11-19T20:02:40.603000", "content": "", "votes": 1, "replies": [] }, { "id": 422400, "author_name": "CVxTz", "author_url": "", "post_date": "2018-11-16T07:09:14.290000", "content": "

Because its not trying to load pre-trained weights (weights=None) so you are free to do whatever you want with the architecture.

", "votes": 2, "replies": [] } ], "raw_markdown_by_id": { "424257": "+ When you give `weights=None`, keras is not loading any weights, so the network need not be fixed by the weight matrix.\n+ It is just returning the architecture/network, and the input layer `input_shape=(size, size, 1)` becomes customisable .because Mobilenet/Resnet are CNNs, containing weight matrices which can be multiplied over images of any size.\n+ `classes=NCATS` can also be variable since output layer of **340 Linear Layer** is created on-the-go, by keras \n+ In recent architectures, `GlobalAveragePooling/AdaptiveAveragePooling` is used so any number of conv layers are converted to FC layer.", "422400": "Because its not trying to load pre-trained weights (weights=None) so you are free to do whatever you want with the architecture.", "422398": "I'm confused about `model = MobileNet(input_shape=(size, size, 1), alpha=1., weights=None, classes=NCATS)` in the helpful [kernel created by beluga](https://www.kaggle.com/gaborfodor/greyscale-mobilenet-lb-0-892/notebook), because the [keras docs](https://keras.io/applications/#mobilenet) say \n> `input_shape`: optional shape tuple, only to be specified if `include_top` is False (otherwise the `input_shape` has to be (224, 224, 3) (with 'channelslast' data format) or (3, 224, 224) (with 'channelsfirst' data format). It should have exactly 3 inputs channels, and width and height should be no smaller than 32. E.g. (200, 200, 3) would be one valid value.\n\nWhy can this kernel specify the `input_shape` and `include_top=True`(default) simultaneously and run without any error？ \n(btw, how to @ someone ?)" } }