Image Classification With PyTorch Deep Transfer Learning
Summary
In this project, we create a Python application for classification of flower images into 102 categories, built using transfer learning with PyTorch. The classifier model is based on one of pre-trained on ImageNet torchvision models with final fully connected classifier layers specified by user.
For application demonstration purposes (to save paid GPU time), results shown below are for early model termination prior to achievement of full saturation of training and validation accuracies after only 9 training epochs.
Note, that training data was randomly rotated, flipped, and cropped, resulting in lower training accuracy and higher training loss relative to those achieved on the validation data that has not been randomly manipulated.
Despite premature termination, model has achieved an 82.3% accuracy on the test data set.
Project Structure
|-- `Image Classifier Project.ipynb`: Code development Jupyter Notebook
|
|-- `train.py`: module for training the classifier model
|-- `predict.py`: module for using pre-trained model for image classification predicitions
|
|-- `model_utils.py`: module with helper functions for building, loading, saving,
| training and visualizing training history of a classifier model
|-- `data_utils.py`: module with helper functions for loading, processing and transforming data,
| and visualizing prediction results
|-- `workslace_utils.py`: helper functions for maintaining UDACITY active workspace during long training cycles
|-- 1cat_to_name.json1 : dictionary containing mapping of image class indices to real class names
Module Details
train.py
Description:
Trains a model with user specified classifier and saves model and optimizer parameters achieved after training epoch with the highest validation accuracy.
Options:
- data_directory : location of training, validation and test data
- -h, --help : train.py options help
- --save_path : Set model checkpoint save filename path
- --resume_checkpoint : filepath for model checkpoint for resumption of training. If resume_checkpoint is specified --arch, --learning_rate, --hidden_units, and --drop_p arguments are ignored.
- --arch : baseline torchvision model architecture from the following choices: alexnet, densenet161, resnet18, vgg16
- --learning_rate : training learning rate
- --hidden_units : list of hidden layer sizes separated by spaces
- --drop_p : dropout probability
- --epochs : number of training epochs
- --gpu : Use gpu for training, if available
Usage:
-
basic usage:
python train.py 'flowers/' -
initial training:
python train.py 'flowers/' --save_dir 'checkpoints/' --arch vgg16 --hidden_units 2048 512 --learning_rate 0.001 --epochs 5 --gpu -
resumption of training from previous checkpoint:
python train.py 'flowers/' --resume_checkpoint 'checkpoints/checkpoint1.pth' --epochs 5 --gpu
predict.py
Description:
Predicts input image class using trained model checkpoint.
Options:
- image_path : location of input image
- checkpoint: loaction of trained model checkpoint
- -h, --help : predict.py options help
- --top_k : return top K most likely classes
- --category_names: use mapping of class indices to return real class names using .json dictionary
- --gpu : Use gpu for training, if available
Usage:
-
basic usage:
python predict.py '/img/test_img.jpg' 'checkpoints/checkpoint1.pth' -
advanced usage:
python predict.py '/img/test_img.jpg' 'checkpoints/checkpoint1.pth' --top_k 5 --category_names cat_to_names.json --gpu
Requirements
Python 3.6 + with the following packages:
signal
contextlib
requests
argparse
numpy
PIL
json
matplotlib
torch
torchvision
time
copy
os
sys
erno
progress
Authors
- Alexander Manasson
Acknowledgements
- Udacity for providing template notebook code