def __init__(self, config, metadata,
text_embedder: EmbedderInterface.Config):
super().__init__(channels=[
C.TensorBoardChannel(SummaryWriter(config.output_path)),
C.ConsoleChannel()
])
self.text_embedder = EmbedderInterface.from_config(text_embedder)
self._reset()
def __init__(self, embedder_cfg: EmbedderInterface.Config, all_responses: bool = False,
fixed_n_turns: bool = False):
# Common setup in the process for it's lifetime
self.text_embedder = EmbedderInterface.from_config(embedder_cfg)
self.pad_token_idx = self.text_embedder.pad_idx
self.unk_token_idx = self.text_embedder.unk_idx
self.bos_token_idx = self.text_embedder.bos_idx
self.eos_token_idx = self.text_embedder.eos_idx
self.fixed_n_turns = fixed_n_turns
self.all_responses = all_responses
class Config(ConfigBase):
model: RetrievalModel.Config = RetrievalModel.Config()
trainer: RetrievalTrainer.Config = RetrievalTrainer.Config(
report_train_metrics=True,
save_modules_checkpoint=True,
modules_save_dir="exp/retrieval"
)
featurizer: TokenIdFeaturizer.Config = TokenIdFeaturizer.Config()
features: ModelInputConfig = ModelInputConfig()
labels: ModelOutputConfig = ModelOutputConfig() # was: WordLabelConfig
metric_reporter: MetaLearnMetricReporter.Config = MetaLearnMetricReporter.Config()
text_embedder: EmbedderInterface.Config = EmbedderInterface.Config()
# Maybe we could just have a single instance, would need the nested batch iterator
data_handler: MetaDataHandler.Config = MetaDataHandler.Config()
model_needs_meta_training: bool = True
def from_config(cls, config: Config,
feature_config: ModelInputConfig,
target_config: ModelOutputConfig,
text_embedder_config: EmbedderInterface.Config,
**kwargs):
text_embedder: EmbedderInterface = EmbedderInterface.from_config(text_embedder_config)
features: Dict[str, Field] = {
ModelInput.SEQ: BPEField(text_embedder)
}
assert len(features)
targets: Dict[str, Field] = {
ModelOutputConfig._name: BPEField(text_embedder, is_target=True, all_responses=config.all_responses),
}
extra_fields = {
RAW_TEXT: RawField(),
ModelInput.DLG_LEN: RawField(),
ModelInput.DLG_ID: RawField(),
ModelInput.DOMAIN_ID: RawField(),
ModelInput.TASK_ID: RawField()
}
kwargs.update(config.items())
self = cls(
raw_columns=[], # ignored in our read function
features=features,
labels=targets,
extra_fields=extra_fields,
**kwargs,
)
self.max_turns = config.max_turns
self.text_embedder_cfg = text_embedder_config
self.all_responses = config.all_responses
self.preproc_chunksize = config.preproc_chunksize
self.train_domains = config.train_domains
self.eval_domains = config.eval_domains
self.featurized_cache_dir = config.featurized_cache_dir
self.test_domains = config.test_domains
self.text_embedder = text_embedder
self.seed = config.seed
return self
def __init__(self, config: Config, feature_config, text_embedder_config):
self.text_embedder = EmbedderInterface.from_config(
text_embedder_config)