inatVisionTraining

Prerequisites

The following python modules are required:

• numpy
• pandas
• yaml
• tensorflow
• nets (pip only)
• tqdm
• json

Configuration

Specify various parameters for training, validating, and testing the model in a YAML file. An example is provided in config.yml.example. This example has been modified from the original to better reflect the state of the code.

The following tables describe parameters that must be set. There are no defaults within the code.

Storage Parameters

Parameters for the storage of log files, checkpoints, etc.

Name	Function
TENSORBOARD_LOG_DIR	Directory for saving log information
CHECKPOINT_DIR	Directory for storing checkpoints
FINAL_SAVE_DIR	Directory for saving the final training data
BACKUP_DIR	Directory for storing backup data; this is an experimental Keras feature

Data Parameters

Parameters for specifying dataset location and metadata. Datasets should be stored in JSON format and are read in using Pandas.

Name	Function
TRAINING_DATA	JSON file with training data
VAL_DATA	JSON file with validation data
TEST_DATA	JSON file with test data
NUM_CLASSES	Number of output classes
LABEL_COLUMN_NAME	Name of the column containing class labels

Model Parameters

Parameters specifying the neural network.

Name	Function
IMAGE_SIZE	2D list of integers specifying the size of each input image.
DROPOUT_PCT	Dropout percentage for the layer between "pool" and "logits"
PRETRAINED_MODEL	"None" (for random initialization), "imagenet" (if pre-training is hosted there), or path to a weights file

Training Parameters

Parameters for training the model. The training framework is hardcoded to use the "rmsprop" optimizer.

Name	Function
TRAIN_MIXED_PRECISION	Set to "True" only if your GPU supports CUDA >= 7.0
MULTIGPU	Boolean; use more than one GPU
BATCH_SIZE	Size of batch per GPU
NUM_EPOCHS	Number of training epochs
DO_LABEL_SMOOTH	Enable label smoothing in the categorical cross entropy loss function
LABEL_SMOOTH_MODE	Distribution; the only currently supported mode is "flat"
LABEL_SMOOTH_PCT	Percentage by which to smooth the labels
INITIAL_LEARNING_RATE	Initial learning rate
LR_DECAY_FACTOR	Learning rate decay factor
EPOCHS_PER_LR_DECAY	Number of epochs over which to decay by
RMSPROP_RHO	Parameter "rho" for the "rmsprop" optimizer
RMSPROP_MOMENTUM	Parameter "momentum" for the "rmsprop" optimizer
RMSPROP_EPSILON	Parameter "epsilon" for the "rmsprop" optimizer

Training and Evaluating the Model

Training

After constructing a configuration file, the model can be trained by running python train.py --config_file <config.yml>.

Model Augmentation

Some basic augmentation is performed on the training data. This augmentation includes:

• Random cropping (image is resized to IMAGE_SIZE)
• Flip 50% of the time
• Recolor 30% of the time (random hue, saturation, brightness, contrast)

Model Evaluation

A trained model can be evaluated using python eval.py --config_file <config.yml>. This is the one required argument.

There are also the following optional arguments:

Argument	Function
--use_checkpoint	If present, will evaluate using a checkpoint rather than the final export
--should_save_results	If yes, will save evaluation results to --save_file
--save_file	File in which to save evaluation results. Will be .npz.