Add CNN support

[nidhaloff]

@Prasanna28Devadiga Let's discuss how we will add CNN support. Anyone is free to join and share ideas.

I want to keep it very simple. My suggestion would be just to write CNN in the algorithm field in order to use a CNN model. Here is an example to illustrate what I mean:

model:
     type: classification
     algorithm: CNN # here just CNN insteead of NeuralNetClassifier. Everyone will recognize this because sklearn supports no CNN models.

Example of yaml configuration for CNN:

model:
    type: classification
    algorithm: CNN
    arguments:
        conv_layers:
            -
                num_filters: 32
                filter_size: 3
                pool_size: 2
                pool_stride: 2
            -
                num_filters: 64
                filter_size: 3
                pool_size: 2
                pool_stride: 2
                dropout: 0.4
        fc_layers:
            -
                fc_size: 128
                dropout: 0.4

[GouravWadhwa]

Hi @nidhaloff, For using CNN we might need the information of whether to use a pre-trained model or not and number of layers required. Will this be provided by the user?

[nidhaloff]

@GouravWadhwa yes eventually we will need that, thanks. I would suggest either writing the pre-trained model name in the algorithm field (e.g algorithm: AlexNet) or adding a new field/parameter (e.g use-pre-trained: AlexNet) below the algorithm field. Would you like to work on this feature?

[GouravWadhwa]

Yes I am trying to work on this feature. Further I am thinking of using tensorflow/keras for implementing the CNN models.
Currently I will just focus on the Pre-Trained models on classification problems after that we can maybe focus on random user designed models, and on the regression problem too.
Please let me know if I should add something else

[nidhaloff]

@GouravWadhwa sounds good ;) I have nothing to add at the moment

[Prasanna28Devadiga]

@GouravWadhwa this sounds great.
@nidhaloff I have started working on implementing user-defined models. The way I see it, besides just coding up a CNN , we also need to find a way to work with image datasets. Keras can be used to perform pre-processing on the image dataset once its parsed by the CLI. Do you have any ideas about we might want to go about giving image inputs to the CLI, considering right now it only takes text inputs( csv , excel etc.)

[Prasanna28Devadiga]

I'm sorry for not being very clear. I meant to say the function takes Dense(64) as an input. Cant we parse the yaml file in a way such that Dense(64) is being sent to the function?

[nidhaloff]

@Prasanna28Devadiga of course, it should be pretty straightforward to do.

[Prasanna28Devadiga]

Then that wont be a problem I think. I'll try to fix the dataset function and try to setup the yaml part too asap.

[nidhaloff]

@Prasanna28Devadiga which problem are you having with the dataset? If you followed the steps I post, then you should already have a dataset example to start with

[Prasanna28Devadiga]

I did follow those steps and was able to get a csv file . However keras' flow_from_dataframe method is unable to recognize the images. The error i'm getting is "invalid filename" which on looking around online, seems to have a common fix i.e. providing the absolute paths in the csv file.
However i dont think that would be the case each time and hence am stuck here.

[GouravWadhwa]

Hi @nidhaloff, in my experience pre-trained models are much more powerful than the models trained from scratch because of which they are used in general. Further, I also support the idea of adding customisable models in this library because of which I think we can give the option of using pre-trained model in yaml file. I the user does not wants to use a pre-trained model then they have to specify the model parameters.

[nidhaloff]

@GouravWadhwa What do you mean by much more powerful? Can you provide which experience you had with this?

If you mean that pre-trained models always achieve better performance, then you are wrong. Sometimes adding accidental complexity will make your model worse. Furthermore, the network will need much more time and energy to train. So why would a user need that if he/she has a problem that a smaller network can solve?

Moreover, if you are using a pre-trained model, then you would have no idea what is happening, why your model is performing badly? what is the issue? On the other side, building a network step by step and seeing the differences in metrics for each step may help you as a user to figure out what is the best architecture for your problem.

If you are working on adding support for pre-trained models, then you would have pre-defined network architecture, so this would not be a problem for pre-trained models. However, we need to implement this for customizable models (this is what @Prasanna28Devadiga is working on)

[GouravWadhwa]

@nidhaloff I was just saying if we have 2 models with same parameters then the pre-trained model generally performs better(I have worked on cross domain adaption and you would find all the state of the art works uses pre-trained feature extractors as these models have performed very nicely on some of the toughest datasets.).
Also I suggested that in your proposed yaml file we can add a pertained_model argument in the model parameter which gives the pertained model we want to use if any otherwise the user can opt for there own configured model too. I was just suggesting of giving the user an extra choice of using pre-trained too.

[Shrimadh]

Hey @nidhaloff, as @Prasanna28Devadiga is working on User-defined models using Tensorflow/Keras.
I was wondering if I could work on the problem of user-defined models using Pytorch and use the custom dataset feature in pytorch to load up the images for training once the csv file is constructed using the gist mentioned above.