[WIP] Define finetuning tasks in command-line hparams

configure_finetuning.py

@@ -128,6 +128,7 @@ def __init__(self, model_name, data_dir, **kwargs):
         "{:}_{:}_{:}_predictions.pkl").format
 
     # update defaults with passed-in hyperparameters
+    self.tasks = {}
     self.update(kwargs)
 
     # default hyperparameters for single-task models

finetune/classification/classification_tasks.py

@@ -437,3 +437,23 @@ def _create_examples(self, lines, split):
       examples += self._load_glue(
           lines, split, -3, -2, -1, True, len(examples), True)
     return examples
+
+class StandardTSV(ClassificationTask):
+  def __init__(self, config: configure_finetuning.FinetuningConfig,
+               task_name: str, task_config: dict, tokenizer):
+    super(StandardTSV, self).__init__(config, task_name, tokenizer,
+                               task_config["labels"])
+    self.task_config = task_config
+
+  def get_examples(self, split):
+    return self._create_examples(read_tsv(
+        os.path.join(self.config.raw_data_dir(self.name), split + ".tsv"),
+        quotechar="\"",
+        max_lines=100 if self.config.debug else None), split)
+
+  def _create_examples(self, lines, split):
+    text_column_2 = self.task_config.get("text_column_2", None)
+    header = self.task_config.get("header", False)
+    return self._load_glue(lines, split, self.task_config["text_column"],
+                           text_column_2, self.task_config["label_column"],
+                           skip_first_line=header)

finetune/task_builder.py

@@ -66,5 +66,12 @@ def get_task(config: configure_finetuning.FinetuningConfig, task_name,
     return qa_tasks.SearchQA(config, tokenizer)
   elif task_name == "chunk":
     return tagging_tasks.Chunking(config, tokenizer)
+  elif (task_name in config.tasks):
+    if config.tasks[task_name]["type"] == "classification":
+      return classification_tasks.StandardTSV(config, task_name,
+                                              config.tasks[task_name],
+                                              tokenizer)
+    else:
+      raise ValueError("Unknown task type: " + config.tasks[task_name]["type"])
   else:
     raise ValueError("Unknown task " + task_name)

run_finetuning.py

@@ -309,11 +309,21 @@ def main():
                       help="The name of the model being fine-tuned.")
   parser.add_argument("--hparams", default="{}",
                       help="JSON dict of model hyperparameters.")
+  parser.add_argument("--task-config", default="{}",
+                      help="JSON dict of custom fine-tuning task parameters")
   args = parser.parse_args()
   if args.hparams.endswith(".json"):
     hparams = utils.load_json(args.hparams)
   else:
     hparams = json.loads(args.hparams)
+
+  if args.task_config.endswith(".json"):
+    task_config = utils.load_json(args.task_config)
+  else:
+    task_config = json.loads(args.task_config)
+  if len(task_config.keys()) > 0:
+    hparams["tasks"] = task_config
+
   tf.logging.set_verbosity(tf.logging.ERROR)
   run_finetuning(configure_finetuning.FinetuningConfig(
       args.model_name, args.data_dir, **hparams))