bilal-e / image-classifier

Image Classifier Command Line App using PyTorch - Udacity Nanodegree Project

Image Classifier Command Line Application

TRAINING (train.py)

PURPOSE:

Train a dataset on a new network (models to be used: vgg or densenet);

Output Training loss, Validation loss, Validation accuracy, and Best value accuracy

Saves checkpoint to file 'model_checkpoint.pth'

EXPECTED USER INPUT:

python train.py --train_dir 'path to desired directory for training' --arch 'model(vgg or densenet)'

--lr 'learning rate for training(float)' --hidden_units 'hidden units for training(int)'

--epochs 'epochs for training(int)' --gpu 'train on gpu'

EXAMPLE CALL:

python train.py --train_dir '/dir' --arch vgg --lr 0.005 --hidden_units 1024 --epochs 30 --gpu cuda

PREDICTION (predict.py)

PURPOSE:

Reads in an image and a checkpoint to predict image class

Prints the top-k probability with class name

EXPECTED USER INPUT:

python predict.py --arch 'vgg or densenet (since there is only one checkpoint file that this application generates

based on the training model; choose the same model as of training to avoid mismatch)'

--img_path 'complete path of image to be predicted' --lr 'learning rate used when training(float)'

--hidden_units 'hidden units used when training(int)' --top_k 'required no. of top K classes(int)'

--print_k '1 prints a list of top_k; 0 prints only max probability' --json_file 'complete path to category names file'

CHECKPOINT FILE:

'model_checkpoint.pth', created by train.py

EXAMPLE CALL

python predict.py --arch vgg --img_path /predict_img.jpg --lr 0.005 --hidden_units 1024 --top_k 3 --print_k 1

Tested on 102 flower categories

Model probability accuracy > 0.8