Deepayan137 / Basic-ML

Link to project report.

Dependencies

• Pytorch 1.4
• Numpy
• Pandas

Perceptron

Training python train_perceptron.py --name perceptron --epochs 20

• Main Arguments
  • --name: creates a directory inside the checkpoiints dir, where plots will be saved
  • epochs: No. of itertions
  • batch_size: No. of samples per batch

Results:

The perceptron converges at 15th epoch when initial weihgts are set at (1, 1)

Neural Network

Training python -m train_digit_recognizer --name digits --epochs 400 --train --mode numpy

• Main Arguments
  • --name: creates a directory inside the checkpoiints dir, where plots will be saved
  • epochs: No. of itertions
  • batch_size: No. of samples per batch
  • numpy: make sure this flag is on, else the NN code will give error
  • train: if this flag is given then the network will be trained else it performs prediction on the test set

CNN

Training

python train_chart_classifier.py --epochs 100 --name cnn --mode pytorch --csvfile train_val.csv --pretrained --train

• Main Arguments
  • --name: creates a directory inside the checkpoiints dir, where plots will be saved
  • epochs: No. of itertions
  • batch_size: No. of samples per batch
  • pretrained: loads pretrained VGG model. If this flag is not given then a 2-layer CNN will be trained from scratch
  • pytorch: make sure this flag is on, else the CNN code will give error
  • csvfile: name of the csvfile
  • dataroot: path to charts charts datafolder
  • imgdir: name of the directory; either train-val or test
  • checkpoints_dir: path to the directory containing all checkpoints