udacity-dog-breed-classification

My way to implement and train CNN classifiers both from scratch and use PyTorch's implementations of popular CNNs for transfer learning. Inspired from task from Udacity Deep Learning Nano Degree Program (see original task in original_task/dog_app_orig.ipynb, the original repo where you can find this task and more).

The purpose of this work

It is my own learning process of training convolution neural networks using PyTorch framework, from very beginning

What interesting and useful can be found in this repo

• The code is well structured, scripts for training and evaluation are provided with wide range of input parameters
• A series of experiments are provided, well documented, see experiments_part_1.ipynb
• The tensorboard logs are provided, for each experiment you can see plots of train loss, validation loss, test loss and test accuracy. The logs are in tensorboard_logs directory
• You can view step-by-step on model's improvements and finetuning
• You can repeat all experiments

How to use

1. The original task you can read in original_task/dog_app_orig.ipynb. You can also get links to dog breed dataset there
2. To see experiments description (from 1 to 25), run experiments_part_1.ipynb. For next serie of experiments (from 26 to 51) see experiments_part_2.ipynb. Also, you can find conclusions and future work in there notebooks.
3. To run specific experiment, go to train_run.sh bash script, find experiment you are interested in, copy command and run in terminal (to see available input parameters for train.py, run python3 train.py -h)
4. To evaluate model from specific experiment, go to evaluate_run.sh or evaluate_run_2.sh bash script, find experiment you are interested in, copy command and run in terminal

Dataset

The dog breed dataset is used. Summary:

• 133 classes
• 6680 train images
• 835 valid images
• 836 test images

Checkpoints

Checkpoints are not provided in this repo, but you can run all experiments using bash scripts:

• train_run.sh - bash script for running experiments 1-25
• train_run_p2.sh - bash script for running experiments 26-50

Tensorboard logs

Tensorboard logs for all experiments are provided. To run tensorboard locally, run:

tensorboard -logdir tensorboard_logs/ 

or

tensorboard -logdir tensorboard_logs_p2/ 

To run tensorboard to get access from remote server, run (change [port] to your value):

tensorboard -logdir tensorboard_logs/ --port [port] --host 0.0.0.0

Also, you can merge all logs to same directory to view all experiments on same plot

Models description

Models are described in experiments_part_1.ipynb and experiments_part_2.ipynb. Brief info:

• Three implementations of AlexNet from original paper (Base_fix and Base_1_fix and Base_2), which are training from scratch
• AlexNet from torchvision's implementation, which is training both from scratch and using transfer learning
• VGG16 from torchvision's implementation, which is training both from scratch and using transfer learning
• VGG16 with batch normalization from torchvision's implementation, which is training both from scratch and using transfer learning

Train models from scratch

A lot of experiments are dedicated to train models from scratch. More detail information see in experiments_part_1.ipynb and experiments_part_2.ipynb

Transfer learning

Transfer learning are implemented for pretrained torchvision's models AlexNet and VGG16, with unfreezing from 1 to 3 last fully connected layers (unfreezing of convolution layers will be implemented in future)

Mean and std for dog breed dataset

To calculate mean and std for dog breed dataset, run utils/mean_std_calculate.py

Evaluation models from checkpoints

See examples in evaluate_run.sh and evaluate_run_2.sh

Feel free to use this work and have a good time with it!

If you have any ideas, or noticed any bugs, feel free to open issues, I'll be happy to improve this work!