Training Parameters

Iterations

The training process is iterative, that is, carried out in cycles. Each cycle is called Season. In each epoch all examples are presented for the neural network, so that the patterns are learned. The training algorithm will then measure the error (Loss) and adjust the neural network to minimize it.

It is common to think that with more times the network will learn more, but in practice it is not exactly like that. Depending on the quantity / quality of the examples, a point is reached where the error no longer decreases and learning stagnates. Another common occurrence is that the error (Loss) continues to decrease, but the Val Lossstarts to increase. This phenomenon is known as* Overfitting * and it means that the neural network has become addicted to the training examples and is no longer able to generalize to new examples.

The recommendation is to set 5 training periods in the beginning while the dataset has less than 100 examples, and then go up.

In the sequence, we have several other parameters that govern the training process and all can influence positively or negatively on the final result of the network learning. Do not be frightened by the quantity, nor by the complexity of them, it is natural to take a long time to acquire mastery over the whole process.

Rest assured, Eyeflow.AI has a great set of defaults for the parameters that solve the needs of most of the needs. In addition, our team is available to answer questions and give tips in our Forum.

Train Parameters

Parameters for Neural Network Training

Parameter	Values	Default	Description
Epochs	int [1 - 200]	5	Number of epochs for training
Steps per Epoch	int [50 - 2000]	100	Number of Steps for training in each Epoch
Batch Size	int [1 - 64]	10	Number of examples in each step
Val Size	number [0.01 - 0.9]	0.1	Percent of examples selected for Validation
Test Size	number [0.01 - 0.9]	0.1	Percent of examples selected for Final Test
Confidence Threshold	number [0.05 - 1.0]	0.6	Minimum confidence threshold for valid detection
IoU Detection Threshold	number [0.05 - 1.0]	0.45	Minimum threshold for IoU detection
Maximum Boxes	int [1 - 300]	30	Maximum number of boxes in detection
Expand Boxes	number [0 - 2]	0	Percent to expand size of boxes in detection

Input Resolution

Input image dimensions

Parameter	Values	Default	Description
Minimum Side	int [20 - 800]	50	Size of the smaller side
Maximum Side	int [20 - 1000]	80	Size of the bigger side
Channels	choice [1, 3]	1	Color channels

Optimizer Parameters

Parameters for Train Optimizer

Parameter	Values	Default	Description
beta_2	number [0.1 - 1.0]	0.999	Beta 2
beta_1	number [0.1 - 1.0]	0.9	Beta 1
Learning Rate	number [1e-06 - 0.1]	0.001	Optimizer Learning Rate
amsgrad	bool [True - False]	False	AMSGrad

Early Stopping

Early stopping for training

Parameter	Values	Default	Description
Patience	int [1 - ]	5	Num of epochs to wait for progress
Monitor variable	choice [‘val_loss’, ‘loss’, ‘categorical_accuracy’, ‘val_categorical_accuracy’]	val_loss	Variable to monitor progress
Minimum Delta	number [0 - ]	0.01	The minimum variantion in variable
Evaluation mode	choice [‘min’, ‘max’, ‘auto’]	min	Monitor decrement or increment of variable value

Reduce LR on plateau

Reduce Learning Rate on plateau

Parameter	Values	Default	Description
Patience	int [1 - ]	4	Num of epochs to wait for progress
Monitor variable	choice [‘val_loss’, ‘loss’, ‘categorical_accuracy’, ‘val_categorical_accuracy’]	val_loss	Variable to monitor for progress
Minimum Delta	number [0 - ]	0.01	The minimum variantion in variable
Reducing factor	number [0.1 - 0.9]	0.5	Ammount to reduce
Cooldown	number [0 - ]	0	Cool Down

Save Checkpoint

Trigger to save model training progress

Parameter	Values	Default	Description
Monitor variable	choice [‘val_loss’, ‘loss’, ‘categorical_accuracy’, ‘val_categorical_accuracy’]	val_loss	Variable to monitor for saving
Evaluation mode	choice [‘min’, ‘max’, ‘auto’]	min	Save on decrement or increment of variable value

Where should I go now?

• Training Parameters
• Parameters of Neural Networks
• Data Expansion Parameters