File size: 2,752 Bytes
0c48771 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 | """transformers-compatible config for the RadGraph-IT DyGIE++ joint NER + relation model.
Mirrors the hyperparameters of `training_v2/src/dygie/model.py`'s `DyGIEModel` plus the
label vocabulary (`training_v2/src/dygie/vocab.py`'s `Vocabulary`), so a `RadgraphModel` can
be reconstructed from `config.json` alone, matching the checkpoint trained by
`run_medbit_full_cv.sh` (single train run on the full split, not actual cross-validation --
see that script's own header comment).
"""
from transformers import PretrainedConfig
class RadgraphConfig(PretrainedConfig):
model_type = "radgraph_it"
def __init__(
self,
encoder_name: str = "IVN-RIN/medBIT-r3-plus",
max_length: int = 512,
max_span_width: int = 12,
feature_size: int = 20,
feedforward_params: dict = None,
loss_weights: dict = None,
relation_spans_per_word: float = 0.5,
train_encoder: bool = True,
span_pooling: bool = False,
transformer_params: dict = None,
relation_context: bool = False,
relation_feedforward_params: dict = None,
dataset: str = "radgraph-it",
ner_labels: dict = None,
relation_labels: dict = None,
**kwargs,
):
self.encoder_name = encoder_name
self.max_length = max_length
self.max_span_width = max_span_width
self.feature_size = feature_size
self.feedforward_params = feedforward_params or {"hidden_dims": [150, 150], "dropout": 0.4}
self.loss_weights = loss_weights or {"ner": 0.2, "relation": 1.0}
self.relation_spans_per_word = relation_spans_per_word
self.train_encoder = train_encoder
self.span_pooling = span_pooling
self.transformer_params = transformer_params
self.relation_context = relation_context
self.relation_feedforward_params = relation_feedforward_params
self.dataset = dataset
# Namespace-unqualified label -> index maps, null label "" pinned to 0. Namespaced as
# f"{dataset}__ner_labels" / f"{dataset}__relation_labels" when rebuilt into a Vocabulary
# (see modeling_radgraph.py), matching training_v2/src/dygie/vocab.py exactly.
self.ner_labels = ner_labels or {
"": 0,
"Anatomy::definitely present": 1,
"Observation::definitely present": 2,
"Observation::definitely absent": 3,
"Observation::uncertain": 4,
"Anatomy::definitely absent": 5,
"Anatomy::uncertain": 6,
}
self.relation_labels = relation_labels or {
"": 0,
"modify": 1,
"located_at": 2,
"suggestive_of": 3,
}
super().__init__(**kwargs)
|