Merge models

[omid-s]

Models need to be merged in a reasonable state, we have the following options :

• run the whole data through each of the components seperately and then use the outputs as input to the final ensembler
• create a gigantic model ( not possible for the kmeans i gues ) and pass data in batches through it

what's your thoughts ?

[parya-j]

How about running each model separately and use their outputs? I think the first approach is doable ..

[raun1]

do we have the clustering and the pretrained models ready?

[raun1]

We can train a final network on top of it

[omid-s]

kmeans is ready

[raun1]

ok can u tell me the name of the file i will merge the dense net + k means and run a neural network on top of it :>

[omid-s]

kmeans http://cobweb.cs.uga.edu/~omid/models/kmeans_100_Classes.pkl

[raun1]

@parya-j do u have the outputs for the pre trained network ready?

[omid-s]

assumes 64*64 input :) to load the model see : http://scikit-learn.org/stable/modules/model_persistence.html it uses joblib libjob what ever :D

[parya-j]

What type of output you expect? Now it gives you probability of classes of imagenet

[omid-s]

I think we should stick with probabilities, it will give us mupltiple objects in a scene

[raun1]

@omid-s
We will do exactly what parya suggested ;) Clustering is unsupervised so no need to load the model. Just need the csv or some file for like

• Training each input image - cluster number (just need this to train a final network)
• Testing same thing.

I already have them for dense nets.

@parya-j yes exactly basically the class prediction for each class. Use argmax to get the class (not probability) i.e. the class will be the arg max on softmax layer. Do this for both training and testing :)

[raun1]

@omid-s for the testing set we should use the similar clusters tho :/ i guess we have to compute the shortest distance from the training centers and assign them. Let me check the source on that.

[omid-s]

I'll do that for parya's model if you do a argmax what you'll get is the one with the highest posibility! this will give us only one object in the photo ( most probably the person! ) if we stick with the probabs we can have a glimps of all the objects in the scene, like person, sea, tree, mountain! is far better than just person

[raun1]

u have to retrain the top layers of the pretrained model tho

[omid-s]

let;s talk about it tomorrow, I do not follow why is that but passing out hate 8 am class :((

[raun1]

Yea you have a point :) lets keep it all the things just give me the probability map of the softmax. my bad dont do argmax