| name: megatron_t5_finetune_eval | |
| trainer: | |
| devices: 1 | |
| num_nodes: 1 | |
| accelerator: gpu | |
| precision: 16 | |
| logger: False # logger provided by exp_manager | |
| enable_checkpointing: False | |
| replace_sampler_ddp: False | |
| benchmark: False | |
| exp_manager: | |
| explicit_log_dir: null | |
| exp_dir: null | |
| name: megatron_t5_finetune_eval | |
| create_checkpoint_callback: False | |
| model: | |
| restore_from_path: null # Path to a trained T5 .nemo file | |
| pretrained_checkpoint: | |
| checkpoint_dir: null # Path to a folder that contains a .ckpt file | |
| checkpoint_name: null # Name of the .ckpt file within the checkpoint_dir. | |
| hparams_file: null # Path to a .yaml file that contains the hyperparameters of the checkpoint. | |
| gradient_as_bucket_view: True # Allocate gradients in a contiguous bucket to save memory (less fragmentation and buffer memory) | |
| megatron_amp_O2: False # Enable O2 optimization for megatron amp | |
| data: | |
| validation_ds: | |
| src_file_name: null # Path to the txt file corresponding to the source data. | |
| tgt_file_name: null # Path to the txt file corresponding to the target data. | |
| names: null # If src/tgt file names are ListConfigs, the corresponding label is used to log metrics. | |
| global_batch_size: 64 | |
| micro_batch_size: 64 | |
| shuffle: False | |
| num_workers: 0 | |
| pin_memory: True | |
| max_src_seq_length: 512 | |
| max_tgt_seq_length: 128 | |
| drop_last: False # TODO: Figure out if there is a way to avoid dropping last. | |
| write_predictions_to_file: False | |
| output_file_path_prefix: null # Prefix of the file to write predictions to. | |
| replace_bos_with_pad: False # Replaces bos with pad for both the encoder and decoder. This is necessary when using Google's T5 checkpoints. | |
| add_bos_to_input: False # Adds bos to the input sequence. | |
| add_eos_to_input: False # Adds eos to the input sequence. | |
| metric: | |
| name: "exact_string_match" # Name of the evaluation metric to use. | |
| average: micro # Average the metric over the dataset. Options: ['macro', 'micro']. Works only for 'F1', 'accuracy' etc. Refer to torchmetrics for metrics where this is supported. | |
| num_classes: null # Number of classes for the metric. Works only for 'F1', 'accuracy' and 'average_precision' etc. Refer to torchmetrics for metrics where this is supported. | |
| class_labels: null # If the targets in your dataset are strings and not integers/float, you need to provide a list of class labels (size = num_classes) so we can convert from strings to integer categories to compute the metric. | |
| labels_are_strings: True # NOTE: This is only required to properly handle metrics like f1, accuracy, average_precision etc. This does not affect extract_string_match. | |