| backbone: | |
| num_layers: 12 | |
| d_hidden: 64 | |
| d_input: ${backbone.d_hidden} # same as d hidden | |
| d_out: ${backbone.d_hidden} # same as d hidden | |
| mlp_ratio: 4 | |
| norm: rmsnorm | |
| norm_eps: 1e-8 | |
| activation: gelu | |
| num_heads: 4 | |
| max_position: 1024 | |
| dropout: 0.1 | |
| tokenizer: | |
| temporal_encoder: | |
| input_dims: 128 | |
| output_dims: 128 | |
| hidden_dims: 5 | |
| depth: 4 # Zero-index (will have 5 convolution blocks all together) | |
| kernel_size: 3 | |
| stride: 1 | |
| enable_checkpointing: False | |
| temporal_subsegment_len: 512 | |
| temporal_subsegment_step: 512 | |
| samp_frequency: 2048 | |
| num_seconds: 3 | |
| d_hidden: ${backbone.d_input} | |
| add_spatial_encoding: True | |
| spatial_grouping: destrieux # coords | destrieux | lobes | |
| embedding_max_dim: # leave empty for no normalization of embeddings | |
| embedding_init_scale: 1.0 | |