khumairraj / Visualisation-Keras-models

Code to visualize how different layers view the input when the output is changed. Also visualize the salient features as seen by the input image

Understanding the output of Keras Models

Outline of the Process

• The GCAM method helps to analyse which part the model is focussing while making a prediction
• The other method include the following: Saliency Map, Epsilon-LRP and Deeplift. These give an idea of the saliency of the model
• The model is working on mobilenetv2 now. Provisions are made to analyse any model.

Analyse Results

Layer view analysis

The initial layer focus on textures and features

As we move forward in deep layers the focus starts shifting

Finally it starts seeing high level features in the image. It is to observe how the model sees at different location for 
- dog and 
- lion

Saliency Analysis

This are the salient features in the input image

Requirements

• The code has been written for python 3.6
• Run the following command

    pip install -r requirements.txt

• Run the following command to produce the layernames which can be used in the params. Default is set though.

    python produce_layernames.py

• Run the following command to produce the GCAM outputs

    python produce_GCAM.py

• Run the following command to produce the other outputs

    python produce_others.py

Parameters

The params contain the following keys

   "sourceclasslabel" : default = "imagenet" else provide the path for the classnames.json

   "sourcemodel": 
                        default = "imagenet". Else change it to None and save the model in
                        the model folder.

    "readtypeofmodel" : 
                        default : "plain". Specify the type in which model is saved. Other 
                        options are : "json" and "yaml". Please make sure the model is 
                        saved in this format in the model folder.

    "whethermodelmodified" : 
                        default = 0. Specify if the model has been modified before the
                        previous run. This is to reduce the time spent in removing the
                        softmax in final layer

    "imgfolderpath":    default = ".\\sample"

    "savepath" : 		default = ".\\analyse_results"

    "penultimate_layer_names" :
                        default = ["block_2_expand", "block_9_expand", "Conv_1"]
                        Specify the layers to study from the network. To get the layernames
                        run the file producelayernames.py

    "topNclass" :       default = 2. Set the number of outputs of the model to study.