From db29c7c2a5cede4197a3eaa175d413132ba5c6e9 Mon Sep 17 00:00:00 2001 From: Leon Chen Date: Tue, 25 Oct 2016 10:36:49 -0400 Subject: [PATCH] disable xception demo for now --- demos/src/xception/xception.template.html | 6 +++--- 1 file changed, 3 insertions(+), 3 deletions(-) diff --git a/demos/src/xception/xception.template.html b/demos/src/xception/xception.template.html index c2462ef..35b4855 100644 --- a/demos/src/xception/xception.template.html +++ b/demos/src/xception/xception.template.html @@ -8,13 +8,13 @@
- Note that ~100 MB of weights must be loaded. We use the Keras architecture from here and pretrained weights from here. Enter any valid image URL as input to the network. You can also select from a list of prepopulated image URLs. The endpoint must have CORS enabled, to enable us to extract the numeric data from the canvas element, so not all URLs will work. Imgur and Flickr creative commons all work, and are good places to start. After running the network, the top-5 classes are displayed. Keep in mind also we are limited to the 1,000 classes of ImageNet. Keep in mind that this is image classification and not object detection, so the network is forced to output a single class through softmax. Best results are on images where the classification target spans a large portion of the image. All computation performed entirely in your browser. Toggling GPU on should offer significant speedups compared to CPU. Running the network may still take several seconds (optimizations to come). With "show computational flow" toggled, computation through the network will be shown in the architecture diagram (scroll down as computation is performed layer by layer). Turning this feature off will also speed up computation. + Controls currently disabled as network is very slow - SeparableConvolution2D layer implementation needs heavy optimization. Note that ~100 MB of weights must be loaded. We use the Keras architecture from here and pretrained weights from here. Enter any valid image URL as input to the network. You can also select from a list of prepopulated image URLs. The endpoint must have CORS enabled, to enable us to extract the numeric data from the canvas element, so not all URLs will work. Imgur and Flickr creative commons all work, and are good places to start. After running the network, the top-5 classes are displayed. Keep in mind also we are limited to the 1,000 classes of ImageNet. Keep in mind that this is image classification and not object detection, so the network is forced to output a single class through softmax. Best results are on images where the classification target spans a large portion of the image. All computation performed entirely in your browser. Toggling GPU on should offer significant speedups compared to CPU. Running the network may still take several seconds (optimizations to come). With "show computational flow" toggled, computation through the network will be shown in the architecture diagram (scroll down as computation is performed layer by layer). Turning this feature off will also speed up computation.
closeCLOSE
-
+