configs/train_params/default.yaml

# Random Seed
seed: 0
# Maximum number of images per GPU (changes based on available GPU memory)
max_num_of_imgs_per_gpu: 48
# Accumulate gradient iterations (for increasing the effective batch size under memory constraints)
accum_iter: 1
# Maximum number of epochs for the scheduler
epochs: 100
## Default Optimizer parameters
# Learning rate (absolute lr)
lr: 0.0001
# Lower lr bound for cyclic schedulers that hit 0
min_lr: 1e-06
# Epochs to warmup LR
warmup_epochs: 10
# Weight decay
weight_decay: 0.05
# LR schedule type
schedule_type: "linear_warmup_half_cycle_cosine_decay"
# Warn if model params are not in the below submodule_configs
warn_not_in_submodule: False
# Optimizer parameters specific to submodules
submodule_configs: {}
# Use Automatic Mixed Precision for pretraining
amp: 1
# Floating point type to use for mixed precision training
amp_dtype: "bf16"
# Disable CUDNN Benchmark (Disable for variable resolution & number of view training)
disable_cudnn_benchmark: true
# Freeze the validation samples across all epochs
freeze_val_samples_across_all_epochs: true
# Test loss evaluation frequency
eval_freq: 1
# Frequency (number of epochs) to save checkpoint in checkpoint-last.pth
save_freq: 1
# Frequency (number of epochs) to save checkpoint in checkpoint-%d.pth
keep_freq: 10
# Frequence (number of iterations) to print infos while training (includes tensorboard logging)
print_freq: 20
# Resume Training from last checkpoint
resume: True