Needs help even to reproduce tutorial

[kimdn]

I installed TomoTwin with
pip install tomotwin-cryoet

Then embedding

TomoTwin_model="/home/kimd999/brave/DN/model/TomoTwin/model/tomotwin_latest_2024_Sep.pth"
input_tomo="tomo/tomo.mrc"
batch_size="256"

# If you run into memory issues, reduce the batchsize (-b).
CUDA_VISIBLE_DEVICES=0,1 tomotwin_embed.py tomogram -m $TomoTwin_model -v $input_tomo -b $batch_size -o out/embed/tomo/ -s 2

worked well since I didn't see an obvious error during this embedding as

  Skip version check, as it failed with the following error:
  <urlopen error [Errno 101] Network is unreachable>
  ###############################################
  Found 2 GPU(s). Start DDP + Compiling.
  reading /home/kimd999/brave/DN/model/TomoTwin/model/tomotwin_latest_2024_Sep.pth
  reading /home/kimd999/brave/DN/model/TomoTwin/model/tomotwin_latest_2024_Sep.pth
  Model config:
  {'identifier': 'SiameseNet', 'network_config': {'output_channels': 32, 'dropout': 0.2, 'repeat_layers': 0, 'norm_name': 'GroupNor\
  m', 'norm_kwargs': {'num_groups': 64, 'num_channels': 1024}, 'gem_pooling_p': 0}, 'train_config': {'loss': 'TripletLoss', 'tl_mar\
  gin': 0.539, 'miner': True, 'miner_margin': 0.734, 'learning_rate': 5.945e-05, 'optimizer': 'Adam', 'weight_decay': 0, 'batchsize\
  ': 35, 'patience': 50, 'aug_train_shift_distance': 2}, 'distance': 'COSINE', 'window_size': (37, 37, 37)}
  Embeddings have shape: (5083356, 35)
  Wrote embeddings to disk to out/embed/tomo/tomo_embeddings.temb
  Done.
  Tomogram embedding completed in 0h:36m:9s.

However, generating umap with default n_neighbors
tomotwin_tools.py umap -i $output_of_embedding -o $output_of_umap
failed as

  RuntimeError: RAFT failure at file=/opt/conda/conda-bld/work/cpp/src/umap/knn_graph/algo.cuh line=115: n_neighbors should be smaller than the graph degree computed by nn descent
  Obtained 32 stack frames

Here (umap calculation), input file ($output_of_embedding) is tomo_embeddings.temb (output from embedding).

I know that default n_neighbors for TomoTwin umap is 200.

[thorstenwagner]

Hi @kimdn ! Thanks for letting us know. There seems to be a problem with an updated version of cuml ( rapidsai/cuml#6091 ). For now, a fix to your problem is to add --neighbors=64 to your umap command. Be aware that it might affect the umap. We will provide a fix as soon as possible.

[kimdn]

Thank you for your comment @thorstenwagner

Since version of dependent library matters, I think that sharing apptainer container (e.g., https://github.com/cbmi-group/DeepETPicker?tab=readme-ov-file#option-2using-docker) might help everyone (i.e., developer, user).

Yes, I know that intentional decrease of n_neighbors (e.g., into 25~50) ran embedding part without an error.