Skip to content

Latest commit

 

History

History
34 lines (21 loc) · 1.72 KB

README.md

File metadata and controls

34 lines (21 loc) · 1.72 KB

Pytorch Classifier - male / female, long sleeves. short sleeves - Experiments

predictions-2

predictions-1

classifier_pytorch

pytorch classifier - male/female gender . Long sleeves/short sleeves clothing

1) Code for downloading images from Myntra..

Explanation - I used chromedriver which loads page 1 to page 10 at a time. Each page has 50 images. I downloaded 500 images for 4 categories [Men Long sleeves, Men short sleeves, women long sleeves, women short sleeves]

Challenge faced - usually jumping to the end of a document was not loading all images on pages.Myntra uses lazy loading. I used code to scroll part by part on screen so that all images get a chance to load.

2) classifier_model_pytorch (poor 25-35% accuracy)

Explanation - for each category I have 500 images. I used 80% as training images (400) and 100 images as tst dataset. So for 4 categories in all 2000 = 1600 (train images) + 400 (test images).

Challenge Faced - For inference testing. I took an image from amazon (unseen image to model) Accuracy is very poor. Inference is wrong. Prediction Probabilities for all 4 categories were ~25%. Not good.

3) classifier_model_transfer_learning (85% accuracy)

Explanation - I used Transfer learning by using pretrained model efficientnet_b0. Base model was put in freeze. Model was fine tuned on 2000 images (1600 training + 400 test)

Result - For inference testing. I took an image from amazon (unseen image to model). Accuracy , Prediction Probabilities on all 4 categories is good.