Skip to content

Record parameters/number of trainable parameters and dataset sizes in experiment directory #22

Description

@ikanher

Currently, the number of model parameters or the number of trainable parameters for PEFT, is not recorded to the experiment directory. This can be annoying to pull from logs and should be saved to the experiment directory.

Same applies for dataset (or dataset subset) size.

Let's actually do bit more fine-grained recording of the parameter count:

P_backbone: number of parameters in the backbone

P_frozen: number of parameters that remain frozen during training

P_model: all instantiated parameters (e.g. for LoRA this will include the backbone plus LoRA weights and head count)

d_peft: number of trainable (PEFT) parameters

d_head: number of parameters in head

d_train:  all trainable parameters (d_peft +  d_head)

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Fields

    No fields configured for issues without a type.

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions