python main.py parameters
Parameter list:
| Argument | Values | Default | Description |
|---|---|---|---|
--data_path |
str | ./ |
Root of the data folder |
--model |
mlp, cnn, resnet50, resnet50_head_only, vit_head_only |
mlp |
Neural network to train |
--train |
independent, joint, continual_task, continual_online |
joint |
Training scheme |
--supervised_only |
bool | False |
Consider only supervised samples |
--augment |
bool | True |
Perform random augmentation of the training data |
--lr |
float | -0.001 |
Learning rate. If < 0 use Adam, otherwise use SGD |
--weight_decay |
float | 0.0 |
Weight decay factor |
--batch |
int | 16 |
Minibatch size |
--task_epochs |
int | 1 |
Number of epochs for each task. Incompatible with --train continual_online |
--balance |
bool | False |
Resample positives and negatives to achieve balanced training data |
--replay_buffer |
int | 0 |
Size of the replay buffer. Only with --train continual_* |
--replay_lambda |
float | 0.0 |
Weight of experience replay loss. Only with --train_continual_* |
--cem_emb_size |
int | 12 |
Embedding size for a single concept |
--hamming_margin |
int | 2 |
Margin (in bits) for the Hamming distance triplet loss |
--triplet_lambda |
float | 0.0 |
Weight of triplet loss |
--concept_lambda |
float | 0.0 |
Weight of concept loss |
--seed |
int | -1 |
Seed for random generator. If < 0 use system time |
--output_folder |
str | exp |
Output folder |
--device |
str | cpu |
Torch device for experiments |
--save_net |
bool | True |
Save network weights at the end of the experiment |
--save_results |
bool | True |
Save results at the end of the experiment |
--save_options |
bool | True |
Save options at the beginning of the experiment |
--print_every |
int | 10 |
Number of gradient steps before consecutive prints |
--wandb_project |
str | None |
W&B project name to optionally log results to |
--wandb_group |
str | None |
Group within the W&B project |