Floral Classifier

A Machine Learning project by Ridhiman Dhindsa

Description

This is a command-line application developed in Python, which can be used to predict names of flowers from their images. The model has been trained on a dataset consisting of 102 flower categories, divided into training, validation and testing datasets. The directory flowers/test/ in this repository, contains only a subset of the actual testing dataset. These flower categories can be viewed in the file cat_to_name.json. Deep Learning frameworks from PyTorch, such as DenseNet-121 and DenseNet-169, have been used for building the model. The user can use the images provided in the flowers/test/ directory for prediction, or upload their own image in the project directory and run the prediction script predict.py.

Requirements

The file requirements.txt may be used to create an environment using:
conda create --name <env> --file <requirements.txt>
Platform: win-64
System requirements: Must have Python3 installed, through the Anaconda distribution. For running the prediction script on GPU, the system must have an NVIDIA GeForce RTX GPU as well as the NVIDIA CUDA-Toolkit must be installed.

If a GPU is not available, the prediction script can be run on the system CPU as well. The command-line application will automatically run on CPU.

How to Run

1. Clone the repository on your local machine and open a terminal in the project directory.
2. Type the following and press Enter: python -W ignore predict.py
3. Alternatively, you can also set optional arguments via the command line, i.e:

• python predict.py <path_to_image> <path_to_checkpoint>
  Run the prediction script on image specified by path, and using checkpoint file specified by path.
• python predict.py --top_k <k>
  Return top k most likely classes. For eg, python predict.py --top_k 3 : returns top-3 most likely classes for the image.
• python predict.py --category_names <category_to_name.json>
  Use a particular mapping of categories to real names.
• python predict.py --gpu
  Use GPU for inference.

Simply running python predict.py will set the above command line arguments to their defaults:
path_to_image = flowers/test/1/image_06743.jpg
path_to_checkpoint = checkpoint.pth
top_k = 1
category_names = cat_to_name.json
gpu = not enabled

Tech Stack

Python, PyTorch (cuda, nn, optim modules), TorchVision

Training the model

The user can train their own model as well if they wish to. For this, a system with GPU will be necessary. To get access to the dataset and train your own model, use the following Google Drive link and request access:
Flowers Dataset: Google Drive Link

Steps to run the training script:

1. Download the Flowers dataset in the directory "flowers". If you are saving the dataset in a different directory, it must be specified in the command line arguments.
2. Enable your system GPU and run the following command in terminal: python train.py --gpu.
3. Alternatively, you can also set optional arguments via the command line, i.e:

• python train.py <data_dir>
  Dataset shall be inferred from the 'data_dir' directory.
• python train.py --save_dir <save_directory>
  Set the directory to save checkpoints.
• python train.py --arch "densenet121"
  Choose which model architecture to train on. Models available: densenet121, densenet169.
  Set Hyperparameters:
• python train.py --learning_rate <0.01>
  Set the learning rate for the algortihm.
• python train.py --hidden_units [500, 256] Set the number of units in the hidden layers. Choose 2 integers between 1024 and 102 in decreasing order.
• python train.py --epochs <2>
  Set number of training epochs.
• python train.py --gpu
  Enable GPU (recommended).

Defaults:
data_dir = flowers
save_dir = root
arch = densenet121
learning_rate = 0.01
hidden_units = [500, 256]
epochs = 2
gpu = not enabled

4. This will create the file checkpoint.pth in the desired directory. Now you can run the prediction script on the model you have trained.

Further use

This command-line application can be used to train on a variety of other datasets too! Just download a dataset of your choice, along with the category-to-name mapping in a JSON file. Then choose the appropriate command-line arguments and train the PyTorch model on your dataset. The prediction script can then be run on your test images, and can even be integrated with a user interface such as an app or website, to make predictions.