keplr-io / quiver

Interactive convnet features visualization for Keras
https://keplr-io.github.io/quiver/
MIT License
1.75k stars 223 forks source link

MobilenetV3 support? #84

Open ghylander opened 3 years ago

ghylander commented 3 years ago

After bugfixing the issues in #83 and the issue in #78, feeding a mobilenetv3 model will not show the model tree on the left side of the panel, but no errors are reported. I am mamking sure of feeding a full model, as i know it is not possible to properly access collapsed keras models:

` training_model.summary()

Model: "model"


Layer (type) Output Shape Param #

input_1 (InputLayer) [(None, 224, 398, 3)] 0


sequential (Sequential) (None, 224, 398, 3) 0


model (Functional) (None, 2) 4228994

Total params: 4,228,994 Trainable params: 2,562 Non-trainable params: 4,226,432


` here the collapsed model is in the layer (model)

i create a submodel like so:

submodel = training_model.get_layer("model").get_layer("MobilenetV3large") (because of how i build the model, the actual model is nested as a collapsed model "MobilenetV3large" inside the "model" layer)

this gives the full model with the correct structure, but it is not shown in the webserver.

if that matters, the script outputs the following when the server is started:

::1 - - [2021-01-18 10:18:35] "GET /model HTTP/1.1" 200 106901 0.058173 ::1 - - [2021-01-18 10:18:35] "GET /inputs HTTP/1.1" 200 640 0.007560