Bug: Error when loading a model with the se_a_mask descriptor

[wanghan-iapcm]

Discussed in #3924

^{Originally posted by lukasbaldauf June 28, 2024}
Hi All,
I want to evaluate a trained model using the se_a_mask descriptor, but I'm encountering an error. The training goes smoothly and I get good accuracies for my system, I'm just having trouble loading the model. I get the same error when evaluating the trained zinc_protein example system (see the error message below). It seems like something related to "dfparam" and "daparam", where the Tensors are missing.

I get the same errors for deepmd versions 2.2.7 and 2.2.10, and tensorflow versions 2.9.0 and 2.15.0.

For the zinc example, I train and freeze the model as such:

dp train zinc_se_a_mask.json --skip-neighbor-stat
dp freeze -o graph.mask.pb

The problem occurs when I want to load the model:

from deepmd.infer import DeepPot
model = DeepPot("graph.mask.pb")

---

Traceback (most recent call last):
File "", line 1, in
File "/home/lukasb/miniforge3/envs/deepmd_gpu_2.2.7/lib/python3.10/site-packages/deepmd/infer/deep_pot.py", line 156, in init
self._get_tensor(tensor_name, attr_name)
File "/home/lukasb/miniforge3/envs/deepmd_gpu_2.2.7/lib/python3.10/site-packages/deepmd/infer/deep_eval.py", line 165, in _get_tensor
tensor = self.graph.get_tensor_by_name(tensor_path)
File "/home/lukasb/miniforge3/envs/deepmd_gpu_2.2.7/lib/python3.10/site-packages/tensorflow/python/framework/ops.py", line 4128, in get_tensor_by_name
return self.as_graph_element(name, allow_tensor=True, allow_operation=False)
File "/home/lukasb/miniforge3/envs/deepmd_gpu_2.2.7/lib/python3.10/site-packages/tensorflow/python/framework/ops.py", line 3952, in as_graph_element
return self._as_graph_element_locked(obj, allow_tensor, allow_operation)
File "/home/lukasb/miniforge3/envs/deepmd_gpu_2.2.7/lib/python3.10/site-packages/tensorflow/python/framework/ops.py", line 3992, in _as_graph_element_locked
raise KeyError("The name %s refers to a Tensor which does not "
KeyError: "The name 'load/fitting_attr/dfparam:0' refers to a Tensor which does not exist. The operation, 'load/fitting_attr/dfparam', does not exist in the graph."

[njzjz]

Reproduced in v2.2.10 but failed to reproduce in devel.