Fix text NER evaluation

.gitignore

@@ -1,2 +1,7 @@
-
+# general
 .DS_Store
+dataset/
+manifest/
+save/
+slue_toolkit.egg-info/
+__pycache__/

baselines/ner/nlp_scripts/eval-deberta.sh

@@ -8,4 +8,5 @@ python slue_toolkit/text_ner/ner_deberta.py eval \
 --model_type $model_type \
 --eval_asr False \
 --eval_subset $eval_set \
---eval_label $eval_label
+--eval_label $eval_label \
+--save_results True

baselines/sentiment/README.md

@@ -31,6 +31,6 @@ To evaluate the fine-tuned nlp model, run following command or run `baselines/se
 
 First, ASR transcription need to be prepared in manifest dir, and then evalution can be done using the same evaluation script with nlp topline.
 ```sh
-python slue_toolkit/prepare/prepare_voxceleb_asr_pred.py --data manifest/slue-voxceleb --pred-data datasets/slue-voxceleb/preds/vc1/w2v2-large-lv60k-ft-slue-vc1-12h-lr1e-5-s1-mt800000-8gpu-update280000
+python slue_toolkit/prepare/prepare_voxceleb_asr_pred.py --data manifest/slue-voxceleb --pred-data dataset/slue-voxceleb/preds/vc1/w2v2-large-lv60k-ft-slue-vc1-12h-lr1e-5-s1-mt800000-8gpu-update280000
 python slue_toolkit/eval/eval_nlp_sentiment.py --save-dir save/sentiment/nlp_topline_bert-base-cased --data manifest/slue-voxceleb --subset test.asr-pred
 ```

baselines/sentiment/pipeline_scripts/eval.sh

@@ -1,11 +1,11 @@
 #!/bin/bash
 
-python3 slue_toolkit/prepare/prepare_voxceleb_asr_pred.py --data manifest/slue-voxceleb --pred-data datasets/slue-voxceleb/preds/vc1/w2v2-large-lv60k-ft-slue-vc1-12h-lr1e-5-s1-mt800000-8gpu-update280000
+python3 slue_toolkit/prepare/prepare_voxceleb_asr_pred.py --data manifest/slue-voxceleb --pred-data dataset/slue-voxceleb/preds/vc1/w2v2-large-lv60k-ft-slue-vc1-12h-lr1e-5-s1-mt800000-8gpu-update280000
 
 python3 slue_toolkit/eval/eval_nlp_sentiment.py \
         --data manifest/slue-voxceleb \
         --subset test.asr-pred \
         --save-dir save/sentiment/nlp_topline_bert-base-cased \
         --use-gpu \
         --eval \
-
+

scripts/download_datasets.sh

@@ -1,12 +1,12 @@
 #!/bin/bash
 
 #1. Download
-wget https://papers-slue.awsdev.asapp.com/slue-voxceleb_blind.tar.gz -P datasets/
-wget https://papers-slue.awsdev.asapp.com/slue-voxpopuli_blind.tar.gz -P datasets/
+wget https://papers-slue.awsdev.asapp.com/slue-voxceleb_blind.tar.gz -P dataset/
+wget https://papers-slue.awsdev.asapp.com/slue-voxpopuli_blind.tar.gz -P dataset/
 
 #2. Extract
-tar -xzvf datasets/slue-voxceleb_blind.tar.gz -C datasets/
-tar -xzvf datasets/slue-voxpopuli_blind.tar.gz -C datasets/
+tar -xzvf dataset/slue-voxceleb_blind.tar.gz -C dataset/
+tar -xzvf dataset/slue-voxpopuli_blind.tar.gz -C dataset/
 
 #3. preprocess
 

setup.py

@@ -23,6 +23,9 @@
         "fire",
         "editdistance",
         "soundfile",
+        "transformers",
+        "datasets",
+        "seqeval",
     ],
     entry_points={},
     include_package_data=True,

slue_toolkit/prepare/prepare_voxceleb.py

@@ -106,7 +106,7 @@ def create_split(
 
 
 def create_manifest(
-    data_dir="datasets/slue-voxceleb",
+    data_dir="dataset/slue-voxceleb",
     manifest_dir="manifest/slue-voxceleb",
     is_blind=True,
 ):

slue_toolkit/prepare/prepare_voxceleb_asr_pred.py

@@ -18,7 +18,7 @@ def main():
         "--pred-data",
         type=str,
         required=True,
-        default="datasets/slue-voxceleb/preds/vc1/w2v2-large-lv60k-ft-slue-vc1-12h-lr1e-5-s1-mt800000-8gpu-update280000",
+        default="dataset/slue-voxceleb/preds/vc1/w2v2-large-lv60k-ft-slue-vc1-12h-lr1e-5-s1-mt800000-8gpu-update280000",
         help="Root directory containing voxceleb1_slue data files,"
         "This dir should contain audio/ voxceleb1_slue_{finetune,dev,test} folders ",
     )

slue_toolkit/prepare/prepare_voxpopuli.py

@@ -31,7 +31,7 @@ def create_split(
 
 
 def create_manifest(
-    data_dir="datasets/slue-voxpopuli",
+    data_dir="dataset/slue-voxpopuli",
     manifest_dir="manifest/slue-voxpopuli",
     is_blind=True,
 ):

slue_toolkit/text_ner/ner_deberta.py

@@ -38,8 +38,8 @@ def eval(
     label_list = read_lst(os.path.join(data_dir, f"{train_label}_tag_lst_ordered"))
     if save_results:
         ner_results_dir = os.path.join(log_dir, "error_analysis")
+        os.makedirs(ner_results_dir, exist_ok=True)
     os.makedirs(log_dir, exist_ok=True)
-    os.makedirs(ner_results_dir, exist_ok=True)
 
     data_obj = NDM.DataSetup(data_dir, model_type)
     _ = data_obj.prep_data(
@@ -49,17 +49,20 @@ def eval(
         "fine-tune", "combined", get_map_files=True
     )  # prepare tag-id mapping files
 
+    # TODO: please verify this function
     if "combined" in eval_label:
         tag_lst = read_lst(os.path.join(data_dir, "combined_tag_lst_ordered"))
 
-    val_texts, val_tags, _, _, _, _ = data_obj.prep_data(eval_subset, "raw")
+    val_texts, val_tags, _, _, _ = data_obj.prep_data(eval_subset)
     if eval_asr:
         asr_val_texts, _, _, _, val_dataset = data_obj.prep_data(
             f"{eval_subset}-{asr_model_type}-asr-{lm}", "raw"
         )
     else:
-        asr_val_texts = None
-    eval_obj = NDM.Eval(model_dir, model_type, label_list, eval_label, eval_asr)
+        asr_val_texts, asr_val_dataset = None, None
+
+    label_list = read_lst(os.path.join(data_dir, f"{eval_label}_tag_lst_ordered"))
+    eval_obj = NDM.Eval(data_dir, model_dir, model_type, label_list, eval_label, eval_asr)
     for score_type in ["standard", "label"]:
         if eval_asr:
             res_fn = "-".join(

slue_toolkit/text_ner/ner_deberta_modules.py

@@ -30,6 +30,7 @@
 from slue_toolkit.eval import eval_utils
 
 
+
 class VPDataset(torch.utils.data.Dataset):
     def __init__(self, encodings, labels):
         self.encodings = encodings
@@ -76,12 +77,14 @@ def read_data(self, file_path):
 
     def align_labels(self, tag2id, tags, encodings, label_all_tokens=False):
         """
-        Align labels with appropriate padding labels for sub-tokens
+		Align labels with appropriate padding labels for sub-tokens
+
+		label_all_tokens: Whether to put the label for one word on all tokens of generated by that word or just on the
+            one (in which case the other tokens will have a padding index).
+		"""
+        # TODO : Check this line
+        labels = [[tag2id[tag] if tag in tag2id else tag2id['O'] for tag in doc] for doc in tags]
 
-        label_all_tokens: Whether to put the label for one word on all tokens of generated by that word or just on the
-                         one (in which case the other tokens will have a padding index).
-        """
-        labels = [[tag2id[tag] for tag in doc] for doc in tags]
         encoded_labels = []
         for idx, doc_labels in enumerate(labels):
             word_ids = encodings.word_ids(batch_index=idx)
@@ -242,7 +245,7 @@ def compute_metrics(p, return_entity_level_metrics=True):
         model=model,  # the instantiated 🤗 Transformers model to be trained
         args=training_args,  # training arguments, defined above
         train_dataset=train_dataset,  # training dataset
-        eval_dataset=eval_dataset,  # evaluation dataset
+        eval_dataset=val_dataset,  # evaluation dataset
         compute_metrics=compute_metrics,
     )
 
@@ -288,7 +291,7 @@ def compute_metrics(p, return_entity_level_metrics=True):
     if training_args.do_eval:
         logger.info("*** Evaluate ***")
         metrics = trainer.evaluate()
-        metrics["eval_samples"] = len(eval_dataset)
+        metrics["eval_samples"] = len(val_dataset)
         trainer.log_metrics("eval", metrics)
         trainer.save_metrics("eval", metrics)
 
@@ -298,11 +301,11 @@ def __init__(
         self, data_dir, model_dir, model_type, label_list, eval_label, eval_asr=False
     ):
         """
-        Inference with batch size = 1
-        """
+		Inference with batch size = 1
+		"""
         self.data_dir = data_dir
         self.model_dir = model_dir
-        best_model_ckpt_dir = os.path.join(self.model_dir, "best-checkpoint")
+        best_model_ckpt_dir = os.path.join(self.model_dir)
         self.model = DebertaForTokenClassification.from_pretrained(
             best_model_ckpt_dir, output_loading_info=False
         )