Describe the bug
I'm using the Fisher-Callhome-Spanish as the starting point for training a Speech Translation model. When trying to use an ASR loss, the TransformerST forward_asr method is called. However, this methods expects to use the make_masks method from the parent TransformerASR class, that was removed on commit 9ca48f2, so the training fails.
Expected behaviour
The code should point to an appropriate mask method that exists.
To Reproduce
No response
Environment Details
"speechbrain": "==1.0.2"
Relevant Log Output
File "/nas/models/st/jiranzo/20250325_st_synthetic_data/recipe/apptek_dubbing/speech_translation_nemo/workflow/train_st_model_sb.py", line 588, in <mod
ule>
st_brain.fit(
File "/nas/models/st/jiranzo/20250325_st_synthetic_data/work/apptek_dubbing/common/software/py_venv/CreatePythonVEnvJob.wYSKI3ptoVz3/output/venv/lib/py
thon3.11/site-packages/speechbrain/core.py", line 1619, in fit
self._fit_train(train_set=train_set, epoch=epoch, enable=enable)
File "/nas/models/st/jiranzo/20250325_st_synthetic_data/work/apptek_dubbing/common/software/py_venv/CreatePythonVEnvJob.wYSKI3ptoVz3/output/venv/lib/py
thon3.11/site-packages/speechbrain/core.py", line 1444, in _fit_train
loss = self.fit_batch(batch)
^^^^^^^^^^^^^^^^^^^^^
File "/nas/models/st/jiranzo/20250325_st_synthetic_data/work/apptek_dubbing/common/software/py_venv/CreatePythonVEnvJob.wYSKI3ptoVz3/output/venv/lib/py
thon3.11/site-packages/speechbrain/core.py", line 1230, in fit_batch
outputs = self.compute_forward(batch, sb.Stage.TRAIN)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/nas/models/st/jiranzo/20250325_st_synthetic_data/recipe/apptek_dubbing/speech_translation_nemo/workflow/train_st_model_sb.py", line 48, in compute_forward
asr_pred = self.modules.Transformer.forward_asr(
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/nas/models/st/jiranzo/20250325_st_synthetic_data/work/apptek_dubbing/common/software/py_venv/CreatePythonVEnvJob.wYSKI3ptoVz3/output/venv/lib/python3.11/site-packages/speechbrain/lobes/models/transformer/TransformerST.py", line 238, in forward_asr
) = self.make_masks(src, tgt, wav_len, pad_idx=pad_idx)
^^^^^^^^^^^^^^^
File "/opt/conda/lib/python3.11/site-packages/torch/nn/modules/module.py", line 1931, in __getattr__
raise AttributeError(
AttributeError: 'TransformerST' object has no attribute 'make_masks'Additional Context
No response
Describe the bug
I'm using the Fisher-Callhome-Spanish as the starting point for training a Speech Translation model. When trying to use an ASR loss, the TransformerST forward_asr method is called. However, this methods expects to use the make_masks method from the parent TransformerASR class, that was removed on commit 9ca48f2, so the training fails.
Expected behaviour
The code should point to an appropriate mask method that exists.
To Reproduce
No response
Environment Details
"speechbrain": "==1.0.2"
Relevant Log Output
Additional Context
No response