This is a demonstration of fine-tuning the ResNet 50 V2 feature vector module to predict in a multi-label classification problem. The final layers of the model has been fine-tuned along with the customly appended dense layers.
Download the data from Hackerearth's predicting attribute of animal, and place the extracted contents from it in Data folder if you would like to work with default parameter settings.
ResNet V2 module has the necessity of having the input image size of 224 x 224. Hence, we will have to preprocess the input images to the fixed size of 224 x 224. In Preprocess.py file, pass the downloaded train and test folder paths in resize_paths variable. Run the below code to start with the preprocessing:
python3 Preprocess.py
To begin with the training procedure, configure the below parameters in Main.py file.
EPOCHS: Number of epochs you would like to run the code
BATCH_SIZE: Batch size during training
LEARNING_RATE: Starting Learning rate
N_CLASS: The number of classes present in the dataset
DIVIDE_LEARNING_RATE_AT: At which epochs, learning rate should be divided by 10. Epoch count starts from 0.
TRAIN_PATH: Folder path where your train folder has been resized to.
TEST_PATH: Folder path where your test folder has been resized to.
TRAIN_VAL_RATIO: The train-validation ratio to be performed on your train dataset.
DATA_LABELS: File path where the labels of train folder are present.
To start the training, run the below code:
python3 Main.py