Remove warnings (open-mmlab#2402)

* replace iteritems with items

* replace iteritems with items

* remove seqeval warning

* Update metric.py

* set num_workers to 0 by default

Some huggingface checkpoints such as deberta-v3, mdeberta-v3, roberta, flan-t5-xl do not work with ddp_spawn when setting num_workers=2

* update ner presets

* add fixme

* Update presets.py

multimodal/src/autogluon/multimodal/configs/environment/default.yaml

@@ -6,7 +6,7 @@ env:
   eval_batch_size_ratio: 4  # per_gpu_batch_size_evaluation = per_gpu_batch_size * eval_batch_size_ratio
   per_gpu_batch_size_evaluation:  # This is deprecated. Use eval_batch_size_ratio instead.
   precision: 16  # training precision. Refer to https://pytorch-lightning.readthedocs.io/en/latest/advanced/mixed_precision.html
-  num_workers: 2  # pytorch training dataloader workers
+  num_workers: 0  # pytorch training dataloader workers. FIXME: setting it to 0 to avoid ProcessExitedException for deberta-v3, flan-t5-xl, etc.
   num_workers_evaluation: 2  # pytorch prediction/test dataloader workers
   fast_dev_run: False
   deterministic: False

multimodal/src/autogluon/multimodal/data/infer_types.py

@@ -571,7 +571,7 @@ def infer_problem_type_output_shape(
         elif provided_problem_type == NER:
             unique_entity_groups = [
                 annot[ENTITY_GROUP]
-                for annotation in data[label_column].iteritems()
+                for annotation in data[label_column].items()
                 for annot in json.loads(annotation[-1])
             ]
             return provided_problem_type, len(set(unique_entity_groups)) + 2

multimodal/src/autogluon/multimodal/data/labelencoder_ner.py

@@ -71,7 +71,7 @@ def extract_ner_annotations(self, y: pd.Series):
         }
         all_annotations = []
         all_entity_groups = []
-        for ner_annotations in y.iteritems():
+        for ner_annotations in y.items():
             json_ner_annotations = json.loads(ner_annotations[-1])  # load the json annotations
             try:
                 jsonschema.validate(json_ner_annotations, schema)  # verify the json schema
@@ -142,7 +142,7 @@ def transform_label_for_metric(self, y: pd.Series, x: pd.Series, tokenizer):
         """
         all_annotations, _ = self.extract_ner_annotations(y)
         transformed_y = []
-        for annotation, text_snippet in zip(all_annotations, x.iteritems()):
+        for annotation, text_snippet in zip(all_annotations, x.items()):
             word_label, _, _, _ = process_ner_annotations(annotation, text_snippet[-1], tokenizer, is_eval=True)
             word_label_invers = []
             for l in word_label:

multimodal/src/autogluon/multimodal/presets.py

@@ -395,25 +395,23 @@ def medium_quality_faster_inference_image_text_similarity():
 def best_quality_ner():
     return {
         "model.names": ["ner_text"],
-        "model.ner_text.checkpoint_name": "google/flan-t5-base",
-        "env.precision": "bf16",
+        "model.ner_text.checkpoint_name": "microsoft/deberta-v3-base",
     }
 
 
 @automm_presets.register()
 def medium_quality_faster_inference_ner():
     return {
         "model.names": ["ner_text"],
-        "model.ner_text.checkpoint_name": "google/flan-t5-small",
-        "env.precision": "bf16",
+        "model.ner_text.checkpoint_name": "google/electra-small-discriminator",
     }
 
 
 @automm_presets.register()
 def high_quality_fast_inference_ner():
     return {
         "model.names": ["ner_text"],
-        "model.ner_text.checkpoint_name": "bert-base-cased",
+        "model.ner_text.checkpoint_name": "microsoft/deberta-v3-base",
     }
 
 

multimodal/src/autogluon/multimodal/utils/metric.py

@@ -171,6 +171,7 @@ def compute_score(
     """
     if metric_name == OVERALL_ACCURACY:
         metric = evaluate.load("seqeval")
+        warnings.filterwarnings("ignore")
         return metric.compute(references=metric_data[Y_TRUE], predictions=metric_data[Y_PRED])
 
     metric = get_metric(metric_name)