[Feature Request] support `eval_descriptor` usage in PT frozen model
Closed this issue · 3 comments
QuantumMisaka commented
Summary
The DeepEval.eval_descriptor() function should support frozen model with extension name .pth
Detailed Description
Now the DeepEval.eval_descriptor() function can only support PT model with extension .pt, when use frozen model with extension pth, error orrurs:
In [22]: DPf = DeepPot('model-64.pth')
You can use the environment variable DP_INFER_BATCH_SIZE tocontrol the inference batch size (nframes * natoms). The default value is 1024.
In [23]: anodes2 = descriptor_from_model(onedata, DPf)
---------------------------------------------------------------------------
AttributeError Traceback (most recent call last)
Cell In[23], line 1
----> 1 anodes2 = descriptor_from_model(onedata, DPf)
Cell In[7], line 7, in descriptor_from_model(sys, model)
5 type_trans = np.array([model_type_map.index(i) for i in sys.data['atom_names']])
6 atypes = list(type_trans[sys.data['atom_types']])
----> 7 predict = model.eval_descriptor(coords, cells, atypes)
8 return predict
File /data/softwares/miniconda3/envs/deepmd-devel/lib/python3.12/site-packages/deepmd/infer/deep_eval.py:445, in DeepEval.eval_descriptor(self, coords, cells, atom_types, fparam, aparam, mixed_type, **kwargs)
398 """Evaluate descriptors by using this DP.
399
400 Parameters
(...)
434 Descriptors.
435 """
436 (
437 coords,
438 cells,
(...)
443 natoms,
444 ) = self._standard_input(coords, cells, atom_types, fparam, aparam, mixed_type)
--> 445 descriptor = self.deep_eval.eval_descriptor(
446 coords,
447 cells,
448 atom_types,
449 fparam=fparam,
450 aparam=aparam,
451 **kwargs,
452 )
453 return descriptor
File /data/softwares/miniconda3/envs/deepmd-devel/lib/python3.12/site-packages/deepmd/pt/infer/deep_eval.py:648, in DeepEval.eval_descriptor(self, coords, cells, atom_types, fparam, aparam, **kwargs)
616 """Evaluate descriptors by using this DP.
617
618 Parameters
(...)
645 Descriptors.
646 """
647 model = self.dp.model["Default"]
--> 648 model.set_eval_descriptor_hook(True)
649 self.eval(
650 coords,
651 cells,
(...)
656 **kwargs,
657 )
658 descriptor = model.eval_descriptor()
File /data/softwares/miniconda3/envs/deepmd-devel/lib/python3.12/site-packages/torch/jit/_script.py:826, in RecursiveScriptModule.__getattr__(self, attr)
823 self.__dict__[attr] = script_method
824 return script_method
--> 826 return super().__getattr__(attr)
File /data/softwares/miniconda3/envs/deepmd-devel/lib/python3.12/site-packages/torch/jit/_script.py:533, in ScriptModule.__getattr__(self, attr)
531 def __getattr__(self, attr):
532 if "_actual_script_module" not in self.__dict__:
--> 533 return super().__getattr__(attr)
534 return getattr(self._actual_script_module, attr)
File /data/softwares/miniconda3/envs/deepmd-devel/lib/python3.12/site-packages/torch/nn/modules/module.py:1931, in Module.__getattr__(self, name)
1929 if name in modules:
1930 return modules[name]
-> 1931 raise AttributeError(
1932 f"'{type(self).__name__}' object has no attribute '{name}'"
1933 )
AttributeError: 'RecursiveScriptModule' object has no attribute 'set_eval_descriptor_hook'
Further Information, Files, and Links
No response
njzjz commented
It does support. In the unit tests, only .pth has been tested
deepmd-kit/source/tests/infer/test_models.py
Lines 159 to 160 in 85e5e20
QuantumMisaka commented
The frozen process is done with 3.0.0b4 code, but the eval_descriptor usage is by using repo https://github.com/iProzd/deepmd-kit/tree/devel_compat_b4. The problem may be from there, but .pt model can be used normally