Fashion_Classification

This is from a challenge from kaggle.

the labels are not that hard to interpret. see https://www.kaggle.com/am1to2/data-exploration-and-analysis

procedure

1. data analysis

   a. map images to each label and try to interpret the learnability - DONE

   b. are there enough sample images for each label, if this is the case we can directly train a cnn with y= label, x = input. but it seems the dataset is not large enough. we need to try his first -YES

   c. plot the histograms of labels . 228 labels - DID d.

2. Initial CNN training - start with first 60,000 - 50,000 - traing , 10,000 - validation 4/24/18-5/1/18

   • prepared train and validation tensors, can be fed in to tf pipeline to distort and batch - code uploded 'fashion_reads_images_createdatasetfor_TF.py'

   • read this https://www.tensorflow.org/programmers_guide/datasets

   a. x_i - different sizes - resized the images- Is this the way? PADDING?

   b. y_i - multiple labels, how to feed them

   • binary label vector - TO BE TRIED

   • can we directly feed vectors of different lenghts

   • What else - GOOGLE

   c. compare results with kaggle