Search code examples
pythonallennlp

Trouble Finetuning Decomposable Attention Model in AllenNLP

I'm having trouble fine-tuning the decomposable-attention-elmo model. I have been able to download the model: wget https://s3-us-west-2.amazonaws.com/allennlp/models/decomposable-attention-elmo-2018.02.19.tar.gz. I'm trying to load the model and then fine-tune it on my data using the AllenNLP train command line command.

I also created a custom dataset Reader which is similar to the SNLIDatasetReader and it seems to be working well.

I created a .jsonnet file, similar to what is here, but I'm having trouble getting it to work.

When I use this version:

// Configuraiton for a textual entailment model based on:
//  Parikh, Ankur P. et al. “A Decomposable Attention Model for Natural Language Inference.” EMNLP (2016).
{
  "dataset_reader": {
    "type": "custom_reader",
    "token_indexers": {
      "elmo": {
        "type": "elmo_characters"
      }
    },
    "tokenizer": {
      "end_tokens": ["@@NULL@@"]
    }
  },
  "train_data_path": "examples_train_",
  "validation_data_path": "examples_val_",
  "model": {
    "type": "from_archive",
    "archive_file": "decomposable-attention-elmo-2018.02.19.tar.gz",
    "text_field_embedder": {
      "token_embedders": {
        "elmo": {
            "type": "elmo_token_embedder",
            "do_layer_norm": false,
            "dropout": 0.2
        }
      }
    },
   },
  "data_loader": {
    "batch_sampler": {
      "type": "bucket",
      "batch_size": 64
    }
  },
  "trainer": {
    "num_epochs": 140,
    "patience": 20,
    "grad_clipping": 5.0,
    "validation_metric": "+accuracy",
    "optimizer": {
      "type": "adagrad"
    }
  }
}

I get an error:

 File "lib/python3.6/site-packages/allennlp/common/params.py", line 423, in assert_empty
    "Extra parameters passed to {}: {}".format(class_name, self.params)
allennlp.common.checks.ConfigurationError: Extra parameters passed to Model: {'text_field_embedder': {'token_embedders': {'elmo': {'do_layer_norm': False, 'dropout': 0.2, 'type': 'elmo_token_embedder'}}}}

Then, when I take that text_field_embedder portion out, and use this version:

// Configuraiton for a textual entailment model based on:
//  Parikh, Ankur P. et al. “A Decomposable Attention Model for Natural Language Inference.” EMNLP (2016).
{
  "dataset_reader": {
    "type": "fake_news",
    "token_indexers": {
      "elmo": {
        "type": "elmo_characters"
      }
    },
    "tokenizer": {
      "end_tokens": ["@@NULL@@"]
    }
  },
  "train_data_path": "examples_train_",
  "validation_data_path": "examples_val_",
  "model": {
    "type": "from_archive",
    "archive_file": "decomposable-attention-elmo-2018.02.19.tar.gz",
   },
  "data_loader": {
    "batch_sampler": {
      "type": "bucket",
      "batch_size": 64
    }
  },
  "trainer": {
    "num_epochs": 140,
    "patience": 20,
    "grad_clipping": 5.0,
    "validation_metric": "+accuracy",
    "optimizer": {
      "type": "adagrad"
    }
  }
}

I get an error:

    raise ConfigurationError(msg)
allennlp.common.checks.ConfigurationError: key "token_embedders" is required at location "model.text_field_embedder."

The two errors seem contradictory and I'm not sure how to proceed with this fine-tuning.

Solution

  • We found out on GitHub that the problem was the old version of the model that @hockeybro was loading. The latest version right now is at https://storage.googleapis.com/allennlp-public-models/decomposable-attention-elmo-2020.04.09.tar.gz.