Repository files navigation

Multi-Class Image Classification on Mobile Gallery Images using Transfer Learning in PyTorch.

Using the images present in your mobile gallery to train an Image Classifier using Transfer-Learning ! :D

Dataset that I have used ishttps://www.kaggle.com/n0obcoder/mobile-gallery-image-classification-dataIt has 6 classes -

• Cars
• Memes
• Mountains
• Selfies
• Trees
• Whataspp_Screenshots

A few of the sample images form the training set are shown below

Following are the transforms (ordered) applied to the images while training and testing-

• Resizing to (224, 224)
• Random Horizontal Flips (Only applied during the training phase)
• ToTensor (to convert the images into tensors)
• Normalization (using the ImageNet stats)

Architecture : Resnet34I have replaced the last linear layer of the resnet34 with another linear layer which has 6 neurons present in it (6 is the number of classes present in the Mobile Gallery Image Dataset mentioned above in STEP 1).c

Used pretrained weights of the selected architecture

We freeze the pretrained filter in the early and middle layers and train only the filters in the deep layers.

Loss Function: Cross EntropyOpimizer : Adam

• Trained 'layer 4' and 'fc' for 5 epochs.

• Then trained only 'fc' for 3 more epochs

• Test Image

• Output

  This Neural Network thinks that the given image belongs to >>> Memes <<< class with confidence of 95.21%

• Kaggle Kernel :https://www.kaggle.com/n0obcoder/mobile-gallery-image-classification-using-pytorch

• Medium Blog Post:https://towardsdatascience.com/classification-of-mobile-gallery-images-in-pytorch-8ba2d32ce2bf