Help with Reproducing Example2_Ti-H.ipynb in WSL

[afour9961207]

Hi all,

I'm trying to reproduce Example2_Ti-H.ipynb in my WSL environment, but I encountered the following error:

Traceback (most recent call last):
  File "test.py", line 27, in <module>
    DIRECT_sampler = DIRECTSampler(
  File "/home/andy/anaconda3/envs/maml/lib/python3.8/site-packages/maml/sampling/direct.py", line 43, in __init__
    self.structure_encoder = M3GNetStructure() if structure_encoder == "M3GNet" else structure_encoder
  File "/home/andy/anaconda3/envs/maml/lib/python3.8/site-packages/maml/describers/_m3gnet.py", line 36, in __init__
    self.describer_model = M3GNet.from_dir(DEFAULT_MODEL)
  File "/home/andy/anaconda3/envs/maml/lib/python3.8/site-packages/m3gnet/models/_m3gnet.py", line 336, in from_dir
    model.load_weights(model_name)
  File "/home/andy/anaconda3/envs/maml/lib/python3.8/site-packages/keras/src/utils/traceback_utils.py", line 70, in error_handler
    raise e.with_traceback(filtered_tb) from None
  File "/home/andy/anaconda3/envs/maml/lib/python3.8/site-packages/tensorflow/python/training/py_checkpoint_reader.py", line 31, in error_translator
    raise errors_impl.NotFoundError(None, None, error_message)
tensorflow.python.framework.errors_impl.NotFoundError: Unsuccessful TensorSliceReader constructor: Failed to find any matching files for /home/andy/anaconda3/envs/maml/lib/python3.8/site-packages/maml/describers/data/m3gnet_models/matbench_mp_e_form/0/m3gnet/m3gnet

the only file in /home/andy/anaconda3/envs/maml/lib/python3.8/site-packages/maml/describers/data/m3gnet_models/matbench_mp_e_form/0/m3gnet is m3gnet.json

Here's how I set up my environment:

1. conda create --name maml python=3.8
2. conda activate maml
3. pip install ase
4. pip install m3gnet
5. pip install maml

my test.py Script

from __future__ import annotations

import pickle

import matplotlib.pyplot as plt
import matplotlib.ticker as mtick
import numpy as np

from maml.sampling.direct import BirchClustering, DIRECTSampler, SelectKFromClusters



from ase.io.trajectory import Trajectory
from pymatgen.io.ase import AseAtomsAdaptor

def load_ase_MD_traj(filename: str):
    """
    Load .traj to pymatgen structures
    """
    traj = Trajectory(filename)
    structures = [AseAtomsAdaptor.get_structure(tra) for tra in traj]
    return structures

structures_20k_MD = load_ase_MD_traj("10ps.traj")
print(structures_20k_MD[0])

DIRECT_sampler = DIRECTSampler(
    clustering=BirchClustering(n=1000, threshold_init=0.05), select_k_from_clusters=SelectKFromClusters(k=1)
)

If I understand correctly, DIRECTSampler uses the default m3gnet UP as the encoder. However, the DEFAULT_MODEL path seems to be missing weight files (e.g., .pd or .hdf5).
I suspect there's an issue with my environment setup or module dependencies. Additionally, I couldn’t find a requirements.txt file—does it contain the necessary libraries to resolve this issue?
Could someone kindly point out what might be wrong?

Sincerely,
Chang-Ti Chou

[drakeyu]

Hi,

I just tried to pip install maml. It seems that all files are included in maml/describers/data/m3gnet_models/matbench_mp_e_form/0/m3gnet. Please make sure that you have the lastest version of maml.

You can find the exact versions of all dependencies in uv.lock file. To utilize the M3GNet which is no longer being maintained, you will need to downgrade tensorflow to 2.9.0 and keras to 2.7.0. This should help you to load DIRECTSampler.

Best,
Zihan Yu

[afour9961207]

Hi Zihan,

Thanks so much for your help—it works perfectly!
I have one more quick question about saving the preprocessed data. For example, after we get the DIRECT_selection from: DIRECT_selection = DIRECT_sampler.fit_transform(structures_20k_MD) in exmaple 2

How can we save this dataset to a .pickle file so that we can later load it using:

with open("XXX.pickle", "rb") as f:
    XXX_data = pickle.load(f)

Is there a built-in function to save the preprocessed data like this?
Also, I noticed the example uses CPU—do you know if it's possible to utilize a GPU to accelerate the process?

Sincerely,
Chang-Ti Chou

[drakeyu]

Hi,

I am not sure about your question. If you want to save features of structures, they should be in your DIRECT_selection. DIRECT_selection["selected_indexes"] stores the indexes of selected structures from the originally dataset. Then you can save it to pickles. If you want to take all features, please refer to the notebook3.

Currently I do not see the advantages of using GPU, the CPU is good enough to features one structure with hundreds of atoms. If you have a large dataset, using parallel process is a more efficient way.

Bests,
Zihan Yu