old grad clip has 0d tensor problem, fix it (PaddlePaddle#3334)

luotao1 · Jun 11, 2024 · 9fbaebd · 9fbaebd
1 parent 7acf073
commit 9fbaebd
Show file tree

Hide file tree

Showing 4 changed files with 12 additions and 90 deletions.
diff --git a/paddlespeech/s2t/exps/deepspeech2/model.py b/paddlespeech/s2t/exps/deepspeech2/model.py
@@ -27,7 +27,6 @@
 from paddlespeech.s2t.io.dataloader import BatchDataLoader
 from paddlespeech.s2t.models.ds2 import DeepSpeech2InferModel
 from paddlespeech.s2t.models.ds2 import DeepSpeech2Model
-from paddlespeech.s2t.training.gradclip import ClipGradByGlobalNormWithLog
 from paddlespeech.s2t.training.reporter import report
 from paddlespeech.s2t.training.timer import Timer
 from paddlespeech.s2t.training.trainer import Trainer
@@ -148,7 +147,7 @@ def setup_model(self):
         if not self.train:
             return
 
-        grad_clip = ClipGradByGlobalNormWithLog(config.global_grad_clip)
+        grad_clip = paddle.nn.ClipGradByGlobalNorm(config.global_grad_clip)
         lr_scheduler = paddle.optimizer.lr.ExponentialDecay(
             learning_rate=config.lr, gamma=config.lr_decay, verbose=True)
         optimizer = paddle.optimizer.Adam(

diff --git a/paddlespeech/s2t/training/gradclip.py b/paddlespeech/s2t/training/gradclip.py
diff --git a/paddlespeech/s2t/training/optimizer/__init__.py b/paddlespeech/s2t/training/optimizer/__init__.py
@@ -19,7 +19,7 @@
 import paddle
 from paddle.optimizer import Optimizer
 from paddle.regularizer import L2Decay
-from paddlespeech.s2t.training.gradclip import ClipGradByGlobalNormWithLog
+
 from paddlespeech.s2t.utils.dynamic_import import dynamic_import
 from paddlespeech.s2t.utils.dynamic_import import instance_class
 from paddlespeech.s2t.utils.log import Log
@@ -100,7 +100,7 @@ def from_args(cls, name: str, args: Dict[Text, Any]):
         assert "parameters" in args, "parameters not in args."
         assert "learning_rate" in args, "learning_rate not in args."
 
-        grad_clip = ClipGradByGlobalNormWithLog(
+        grad_clip = paddle.nn.ClipGradByGlobalNorm(
             args['grad_clip']) if "grad_clip" in args else None
         weight_decay = L2Decay(
             args['weight_decay']) if "weight_decay" in args else None

diff --git a/tests/unit/tts/test_ssml.py b/tests/unit/tts/test_ssml.py
@@ -72,3 +72,12 @@
     for i, sub in enumerate(outs):
         print(i, sub)
     print()
+
+    import json
+    import xmltodict
+    text = "<speak>我们的声学模型使用了 Fast Speech Two。前浪<say-as pinyin='dao3'>倒</say-as>在沙滩上,沙滩上倒了一堆<say-as pinyin='tu3'>土</say-as>。 想象<say-as pinyin='gan1 gan1'>干干</say-as>的树干<say-as pinyin='dao3'>倒</say-as>了, 里面有个干尸，不知是被谁<say-as pinyin='gan4'>干</say-as>死的。</speak>"
+    ssml = xmltodict.parse(text)
+    print(json.dumps(ssml))
+    print(ssml['speak'].keys())
+    print(ssml['speak']['#text'])
+    print(ssml['speak']['say-as'])