flower-image-classifier

This repository contains the console example of training a convolutional neural network and recognize different species of flowers for choosen image. A neural network will be train on this dataset with 102 categories. The program was created during Udacity's AI Programming with Python Nanodegree Program.

Dependencies

The next points are necessary for correct execution:

• Python 3.6 or later
• Torch 0.4.0
• Pillow 5.2.0
• torchvision 0.2.1

Please, check your environment. Also this repository contains workspace_utils.py, which is important for long-running work, for example, training a network. So you need install requests package on an environment.

Quickstart

To train and use pretrained model you need run train.py and predict.py scripts. Firstly, train.py has the one required argument - data_dir, which takes in the path to a directory with training images. The repository contains directory flowers already, so you can train a network this way:

python train.py flowers/

By defaulf script creates a VGG19 model with 512 hidden units, trains it on 10 epochs with learning rate - 0.01, dropout - 0.2 in CPU mode (it could takes a lot of time) and saves the trained model to checkpoints/checkpoint.pth. The checkpoint could be a quite large file. Usually 130 Mb or more.

After that you can test this trained neural network with a flower image. predict.py need the path to a choosen image and the path to checkpoint as required arguments.

python predict.py flowers/valid/3/image_06631.jpg checkpoints/

By default script uses flowers_name.json as flower labels, prints top 5 predictions and runs in CPU mode (it could takes a lot of time). When the scripts is completed you'll see results like this:

Traning a neural network

train.py takes in the next optional arguments:

• --save_directory - the directory for saving checkpoints (default=checkpoints/)
• --arch - the architecture of a model (default=VGG19). The full list you can see here
• --learning_rate - the learning rate for training (default=0.01). Try to find the most suitable one.
• --hidden_units - the number of hidden units in a model (default=512). More is better.
• --epochs - the number of training epochs (default=10). In general 30 is OK to train a network completely.
• --dropout - the value of dropout during training (default=0.2). The default value is common.
• --gpu - The running mode (default=CPU). Training could takes a lot of time, so GPU is the great idea (if your hardware supports GPU). There is the "how to check" instruction here.

Using a pretrained model for prediction

predict.py takes in 3 optional arguments:

• --top_k - The number of top K predictions to print as results (default=5).
• --category_names - A json file with species of flowers (default=flowers_name.json).
• --gpu- The running mode (default=CPU). The prediction is quite short process, so GPU isn't nesessary.

Issues

• The scripts work with JPEG images

License

The contents of this repository are covered under the MIT License.