Home

3D CNN

Environment Variables Table

1. Training & Model Behavior

Variable Name	Description	Suggested Default
SHOULD_TRAIN	Boolean to control whether the training process should be performed.	True
LOAD_FROM_CHECKPOINTS	Boolean to load model weights from a saved checkpoint.	False
CHECKPOINTS_PATH	Path to the checkpoint file to load model weights from.	None

---

2. Dataset Settings

Variable Name	Description	Suggested Default
PATH_TO_DATA	Path to the folder containing the dataset.	/Cubes32
CLASSES_NAME	Name of the label class in dataset.	voxel
TEST_HAS_LABELS	Whether the test data contains labels. Useful for evaluating or inference-only setup.	False
BATCH_SIZE	Batch size for the dataloader.	2
NUM_WORKERS	Number of CPU workers for the dataloader.	2

---

3. Model & Group Settings

Variable Name	Description	Suggested Default
IN_CHANNELS	Number of input channels for the model (e.g., 1 for grayscale images).	1
OUT_CHANNELS	Output channels of model. For autoencoders it's usually 1, but could vary (e.g., for MLP or NdLinear head).	1
NONLIN	Non-linearity activation function. Options: relu, leaky-relu, or elu.	leaky-relu
NORMALIZATION	Type of normalization layer, e.g., bn (Batch Norm) or in (Instance Norm).	bn
MODEL_DEPTH	Depth of the U-Net model.	3
DIVIDER	Channel divisor used to scale down encoder/decoder width.	4
DROPOUT	Dropout rate.	0.0

---

4. Logs & Saving

Variable Name	Description	Suggested Default
LOGS_DIR	Path to the directory where TensorBoard logs will be saved.	/TensorBoard
LOG_NAME	Name prefix for this specific run in TensorBoard and results.	Test

---

5. Loss Function & Optimizer

Variable Name	Description	Suggested Default
LEARNING_RATE	The learning rate for the optimizer.	0.0001

---

6. Trainer Settings

Variable Name	Description	Suggested Default
MAX_EPOCHS	Maximum number of epochs to train for.	10
GPUS	Number or identifier of the GPU(s) to use.	1
PRECISION	GPU precision to use. Options: 16, 32, 64.	16

---

7. Data Normalization

Variable Name	Description	Suggested Default
INTENSITY_MEAN	Global mean intensity value of all training data, used for normalization.	38.90965
INTENSITY_STD	Global standard deviation of intensity, used for normalization.	45.17005

---

---

CubeNet

Environment Variables Table

1. Training & Model Behavior

Variable Name	Description	Suggested Default
SHOULD_TRAIN	Boolean to control whether the training process should be performed.	True
LOAD_FROM_CHECKPOINTS	Boolean to load model weights from a saved checkpoint.	False
CHECKPOINTS_PATH	Path to the checkpoint file to load model weights from.	None

---

2. Dataset Settings

Variable Name	Description	Suggested Default
PATH_TO_DATA	Path to the folder containing the dataset.	./data
BATCH_SIZE	Batch size for the dataloader.	16
NUM_WORKERS	Number of CPU workers for the dataloader.	4
NUM_CELLS	Number of cells in a worm. Only needed for LAP; max possible is 558.	20
SEED	Random seed for train/validation splits to ensure reproducibility.	42

---

3. Model & Group Settings

Variable Name	Description	Suggested Default
GROUP	Name of the group for Group Equivariant CNNs (G-CNNs), e.g., S4 or T4. Remove if using standard CNN.	None
GROUP_DIM	Dimension of the group for G-CNNs. 24 for S4, 12 for T4.	None
IN_CHANNELS	Number of input channels (e.g., 1 for grayscale images).	1
OUT_CHANNELS	Number of output channels. Equal to classes for classifiers; 1 for Autoencoder; None for LAP.	None
NONLIN	Non-linearity activation function: relu, leaky-relu, or elu.	leaky-relu
NORMALIZATION	Normalization layer type: bn (Batch Norm) or in (Instance Norm).	bn
DIVIDER	Integer divisor to reduce channels in each layer; controls model size.	4
MODEL_DEPTH	Depth of the U-Net model.	4
DROPOUT	Dropout rate.	0.1

---

4. Logs & Saving

Variable Name	Description	Suggested Default
LOGS_DIR	Directory path for saving Tensorboard logs.	./logs
LOG_NAME	Name prefix for this run in Tensorboard and results folders.	default_run

---

5. Loss Function & Optimizer

Variable Name	Description	Suggested Default
LEARNING_RATE	Learning rate for the optimizer.	0.001
LR_PATIENCE	Patience for LR scheduler before reducing LR (ReduceLROnPlateau).	5
LR_FACTOR	Factor to multiply LR when reducing (new_lr = lr * factor).	0.1
LR_MIN	Minimum allowable learning rate.	1e-6
DISTANCE_TYPE	Loss function distance metric: MSE (L2 Loss) or L1.	MSE
LAMBDA	Parameter from paper Differentiation of Blackbox Combinatorial Solvers.	15

---

6. Trainer Settings

Variable Name	Description	Suggested Default
EARLY_STOPPING	Enable/disable Early Stopping callback.	True
EARLY_STOPPING_PATIENCE	Patience for Early Stopping (epochs without improvement).	10
GPUS	Number or identifier of GPUs to use.	1
PRECISION	GPU precision: 16 (or 16-mixed), 32, 64.	32
MAX_EPOCHS	Maximum number of epochs to train.	50
VAL_CHECK_INTERVAL	Frequency of validation checks (1.0 = once per epoch).	1.0
LOG_EVERY_N_STEPS	Log metrics every N steps.	50
PROGRESS_BAR_REFRESH_RATE	Progress bar refresh rate.	20

---

7. Data Normalization

Variable Name	Description	Suggested Default
INTENSITY_MEAN	Global mean intensity for normalization (must be computed from dataset).	None
INTENSITY_STD	Global standard deviation for normalization (must be computed from dataset).	None

---

---

Graph Neural Network

Environment Variables Table

1. Dataset Settings

Variable Name	Description	Suggested Default
PATH_TO_DATA	Path to the folder containing the dataset.	./data
GRAPH_DATA_FOLDER	Path to the folder that contains the graph data.	./Graph_Data/DoubleConv/Graph_data_MLP512_16_3Layers_L1_R45
BATCH_SIZE	Batch size for the dataloader.	4
NUM_WORKERS	Number of CPU workers for the dataloader.	2

2. Model & Group Settings

Variable Name	Description	Suggested Default
NODE_FEATURE_DIM	Input dimension of the node feature vector. 8 for only Geo features, 520 for full features.	520
GNN_HIDDEN_DIM	Hidden layer dimension in GNN.	256
GNN_OUTPUT_DIM	Output dimension of GNN encoder.	128
GAT_HEADS	Number of attention heads for GAT (Graph Attention Network).	8
DROPOUT	Dropout rate.	0.0

3. Logs & Saving

Variable Name	Description	Suggested Default
LOG_NAME	Name prefix for this specific run in Tensorboard and results folders.	MLP512_16_3Layers_L1_R45_200EpochES_Lambda200.00_LR0.0005

4. Loss Function & Optimizer

Variable Name	Description	Suggested Default
LEARNING_RATE	The learning rate for the optimizer.	0.0005
DISTANCE_TYPE	The distance metric used for the loss function.	MSE
LAMBDA	Fixed lambda value. Used when lambda scheduler is disabled.	200.0

5. Trainer Settings

Variable Name	Description	Suggested Default
MAX_EPOCHS	Maximum number of epochs to train for.	1000
LOG_EVERY_N_STEPS	How often to log metrics every N steps.	50
PROGRESS_BAR_REFRESH_RATE	Refresh rate for the progress bar.	50

6. Lambda Scheduling (Conditional)

Variable Name	Description	Suggested Default
USE_LAMBDA_SCHEDULER	Whether to use dynamic lambda scheduling (annealing).	False
LAMBDA_START	Starting value of lambda if using annealing.	50.0
LAMBDA_END	Ending value of lambda if using annealing.	200.0
LAMBDA_WARMUP_EPOCHS	Number of epochs to warm-up the lambda value during annealing.	10