Parametric input in single CNN model #163
Conversation
| @@ -0,0 +1,493 @@ | |||
| # Example config file containing all possible options for CTLearn, whether | |||
There was a problem hiding this comment.
It would be better to update the existing example config. You may want to add a TRN with parameter model to the default CTLearn models
ctlearn/default_models/single_cnn.py
Outdated
| @@ -8,7 +8,10 @@ def single_cnn_model(data, model_params): | |||
|
|
|||
| # Load neural network model | |||
| network_input_img = tf.keras.Input(shape=data.img_shape, name=f"images") | |||
There was a problem hiding this comment.
I'm not sure if we want to support MLP only without the CNNs here or in a separate file. There is no need for an additional file I guess. @nietootein What are your thoughts about this?
I ran a quick check. It is possible to request only the parameter list without the images from dl1dh. @sahilyadav27 So I guess you can just check here if images are selected or not with data.img_shape != None
There was a problem hiding this comment.
I thought having a separate MLP model would make sense because we might want to add the MLP model to other models as well like CNN_RNN and ResNet. So it would be better to have a common MLP model independent of the image model rather than having it inside the Single CNN model?
There was a problem hiding this comment.
Looks good! Can you please update the default config files? I think you would need to rename engine to engine_cnn to match your changes. At the moment we can wait with a default model for CNN + mlp till we have some solid/best-performing settings for the mlp. However, you can update the existing example config file with the parameter list and mlp settings you have been using for testing, so that new users can see how to use the mlp properly. Other than that everything is fine. We can update the docs separately.
The previous pull request was about testing a parametric model in itself and working on the suggestions received there. This pull request combines the parametric model into the single CNN model and also adds a config file showing the structure to define the parametric model's architecture (fully connected layers)