From 9926ad685b75f6673a9f924dbebbce7b5b52bf17 Mon Sep 17 00:00:00 2001
From: zsl01670416 <zsl01670416@alibaba-inc.com>
Date: Tue, 25 Jul 2023 19:02:32 +0800
Subject: [PATCH] support getting labels from dataset in sbert text
 classification and building dataset from file in chatglm-6b

1.Add getting labels from dataset in "text_classificationfinetune_text_classification.py" to simplify user's operation in flex training. Parameters "--num_labels" and "--labels" were removed in "run_train.sh".
2.In "chatglm6b / finetune.py", building dataset from file  is necessary to support flex training.
Link: https://code.alibaba-inc.com/Ali-MaaS/MaaS-lib/codereview/13382745
* support getting labels from dataset in sbert text classification and building dataset from file in chatglm-6b

* support getting labels from dataset in sbert text classification and building dataset from file in chatglm-6b

* remove repetitive labels in a concise manner of using set

* reserve parameter labels in finetune_text_classification

* Merge branch 'master' of http://gitlab.alibaba-inc.com/Ali-MaaS/MaaS-lib

reserve parameter labels in finetune_text_classification

* Merge branch 'support_text_cls_labels_chatglm_json'
reserve parameter labels in finetune_text_classification
---
 examples/pytorch/chatglm6b/finetune.py        | 24 ++++++++++++-------
 .../finetune_text_classification.py           | 17 ++++++++++---
 .../pytorch/text_classification/run_train.sh  |  2 --
 3 files changed, 29 insertions(+), 14 deletions(-)

diff --git a/examples/pytorch/chatglm6b/finetune.py b/examples/pytorch/chatglm6b/finetune.py
index 3fa73ba0..40eb8720 100644
--- a/examples/pytorch/chatglm6b/finetune.py
+++ b/examples/pytorch/chatglm6b/finetune.py
@@ -7,7 +7,7 @@ from chatglm_trainer import Seq2SeqTrainer
 from text_generation_metric import TextGenerationMetric
 from transformers import DataCollatorForSeq2Seq
 
-from modelscope import snapshot_download
+from modelscope import build_dataset_from_file, snapshot_download
 from modelscope.metainfo import Models
 from modelscope.models import Model
 from modelscope.msdatasets import MsDataset
@@ -172,14 +172,20 @@ def cfg_modify_fn(cfg):
     return cfg
 
 
-train_dataset = MsDataset.load(
-    args.train_dataset_name,
-    subset_name=args.train_subset_name,
-    split=args.train_split)
-validation_dataset = MsDataset.load(
-    args.val_dataset_name,
-    subset_name=args.val_subset_name,
-    split=args.val_split)
+if args.dataset_json_file is None:
+    train_dataset = MsDataset.load(
+        args.train_dataset_name,
+        subset_name=args.train_subset_name,
+        split=args.train_split,
+        namespace=args.train_dataset_namespace)
+    validation_dataset = MsDataset.load(
+        args.val_dataset_name,
+        subset_name=args.val_subset_name,
+        split=args.val_split,
+        namespace=args.val_dataset_namespace)
+else:
+    train_dataset, validation_dataset = build_dataset_from_file(
+        args.dataset_json_file)
 
 model_dir = snapshot_download(args.model)
 model_config = read_config(model_dir)
diff --git a/examples/pytorch/text_classification/finetune_text_classification.py b/examples/pytorch/text_classification/finetune_text_classification.py
index e9c6eb27..111243f6 100644
--- a/examples/pytorch/text_classification/finetune_text_classification.py
+++ b/examples/pytorch/text_classification/finetune_text_classification.py
@@ -8,8 +8,14 @@ from modelscope.trainers import build_trainer
 
 def set_labels(labels):
     if isinstance(labels, str):
-        labels = labels.split(',')
-    return {label: id for id, label in enumerate(labels)}
+        label_list = labels.split(',')
+    else:
+        unique_labels = set(labels)
+        label_list = list(unique_labels)
+        label_list.sort()
+        label_list = list(
+            map(lambda x: x if isinstance(x, str) else str(x), label_list))
+    return {label: id for id, label in enumerate(label_list)}
 
 
 @dataclass(init=False)
@@ -52,7 +58,8 @@ class TextClassificationArguments(TrainingArgs):
         })
 
 
-config, args = TextClassificationArguments().parse_cli().to_config()
+training_args = TextClassificationArguments().parse_cli()
+config, args = training_args.to_config()
 
 print(config, args)
 
@@ -62,6 +69,10 @@ def cfg_modify_fn(cfg):
         cfg.merge_from_dict(config)
     else:
         cfg = config
+    if training_args.labels is None:
+        labels = train_dataset[training_args.label] + validation_dataset[
+            training_args.label]
+        cfg.merge_from_dict({'preprocessor.label2id': set_labels(labels)})
     cfg.model['num_labels'] = len(cfg.preprocessor.label2id)
     if cfg.evaluation.period.eval_strategy == 'by_epoch':
         cfg.evaluation.period.by_epoch = True
diff --git a/examples/pytorch/text_classification/run_train.sh b/examples/pytorch/text_classification/run_train.sh
index e148655c..e05c71bc 100644
--- a/examples/pytorch/text_classification/run_train.sh
+++ b/examples/pytorch/text_classification/run_train.sh
@@ -9,8 +9,6 @@ PYTHONPATH=. python examples/pytorch/text_classification/finetune_text_classific
     --val_split 'validation' \
     --first_sequence 'sentence' \
     --label label \
-    --num_labels 15 \
-    --labels '0,1,2,3,4,5,6,7,8,9,10,11,12,13,14' \
     --preprocessor 'sen-cls-tokenizer' \
     --use_model_config True \
     --max_epochs 1 \