IkeaProductClassification

Project to identify products from IKEA product categories.

Requirements

• All dependencies can be installed by running

  pip install -r requirements.txt

Data

• This project has data scraped from IKEA on four different classes:

  • Chairs

  • Beds

  • Candle Holders

  • Wardrobes

• There are total 2,421 product images in data/model_data directory. However, Images scrapped alone from IKEA website (using src/FetchData.py) for each product category is not enough to train a multiclass classifier using a pretrained model.
• Hence additional data scraped from google images has been added to raise number of samples per class at least to 400 images(Images from google images were downloaded using Download all Images chrome extension).
• Scrapped data is split into train and validation datasets where validation data has atleast 40 samples per each class

Data preprocessing

• Images are standardized by using ImageNet data mean and standard deviations.

Data Augmentations used

• random_flip: with p=0.5
• rotate: between -10 to 10 degrees
• zoom
• lightning variations
• wrapping

Training a multi-class classifier

Why ResNext_50?

• In general ResNet based architectures are found to generalise better on custom datasets.
• ResNext based architecture has a higher model capacity at parameter cost almost equivalent to ResNet based equivalent architecture

Model specifications

• ResNext_50(R50) pretrained is used as a classifier.
• Model is defined and trained using src/ResNext_50.ipynb
• For transfer learning, the ImageNet specific FC layer of the R50 is removed and replaced with the following layers
  • AdaptiveAvgPool2d
  • AdaptiveMaxPool2d
  • Flatten
  • BatchNorm1d
  • Dropout(p=0.25)
  • FC (512 units)
  • ReLU
  • BatchNorm1d
  • Dropout(p=0.5)
  • FC(4 units)
• R50 is trained for 20 epochs using one cycle learning rate policy

Train model on IKEA data

• To train the model use src/ResNext_50.ipynb notebook
• If you are unable to see the notebook in github, you can open it in colab here

Training and Validation loss

Confusion matrix

Interpreting model predictions

As it can be clearly seen the first and second are wrong predictions.

• However, in the first example even though the actual label is Bed model predicts it as charis because this product looks like a sofa, convertable to bed and it also shares lot of common features with chairs class.

• In second prediction, model is predicting as wardrobe because of the background in the image which do have a mini shelf structure similar to wardrobe.

Model weights

Model weights are saved to data/model_data/models dir when model.save(‘model_name’’) is performed in src/ResNext_50.ipynb