This is in requirement of BT5153 Applied Machine Learning Term Paper. Dataset used : - https://www.kaggle.com/paramaggarwal/fashion-product-images-small
This package includes two solutions:-
- Image Classification using Transfer Learning :-VGG19 and Inception V3
- Image Search using a custom CNN based AutoEncoder and Resnet50
- Three schemes to conduct product classification for e-commerce -by gender+masterCategory -by subCategory -by articleType
-
MODEL ARCHITECTURE VGG-19 has 19 deep layers with unified kernel size of 3×3
-
LAYERS TUNED Only the last 5 layers are re-trained with newly add-in 3 dense layers in order to predict the correct number of class for current study
-
HYPER-PARAMETERS TUNED Optimizer using adam with learning rate of 1e-5 and decay of 1e-6, new add-in layer with 512 hidden units with drop out rate of 0.5
-
REGULARIZATION Early stopping, reduce learning rate on plateau, drop-out, data augmentation (ONLY during training) are applied, weight sharing is done implicitly through convolution
-
MODEL ARCHITECTURE Inception V3 has 48 dense layers and stacks 11 inception modules, each consisting of pooling layers and convolutional filters
-
LAYERS TUNED All the layers are re-trained with newly add-in 3 dense layers in order to predict the correct number of class for current study
-
HYPER-PARAMETERS TUNED Optimizer using rmsprop with learning rate of 1e-4 and decay of 1e-6, new add-in layer with 128 hidden units with drop out rate of 0.4
-
REGULARIZATION Early stopping, reduce learning rate on plateau, drop-out, data augmentation (ONLY during training) are applied, weight sharing is done implicitly through convolution
Accuracy,Precision and Recall
-
MODEL ARCHITECTURE CNN architecture is used to build AntoEncoder for image search. It comprises of both an Encoder and a Decoder. Embedding features is generated bythe Encoder.
-
LAYERS DESIGN Encoder consists of the first 7 layers Decoder consists of the last 7 layers
-
HYPER-PARAMETERS TUNED Optimizer using adam with learning rate of 1e-5 and decay of 1e-6
-
TRAINING Autoencoders are self-supervised learning model with input and output being the image itself
The names of the creators of these workbooks and their LinkedIn Profiles are given below
Name | |
---|---|
Li Fengzi | https://www.linkedin.com/in/fengzi-li-14b65946/ |
Shashi Kant | https://www.linkedin.com/in/shashi-kant-bksc/ |
Shunichi Araki | https://www.linkedin.com/in/shunichi-araki-76a68398/ |
Sumer Bangera | https://www.linkedin.com/in/sumerbangera/ |
Swapna Samir Shukla | https://www.linkedin.com/in/swapna-samir-shukla-079491a0/ |
- Fine Tune Inception V3
- Improve Image Search
NUS
All authors are graduate students of MSBA program Class of2020 at NUS Business School. .BT5153 Applied Machine Learning Term Paper, MSBA @ NUS,2020. Copyright 2020 by the author(s).