Regarding training with schnet architecture

[SoumyaCYZ]

Hi developers,
When I'm trying to train with schnet architecture model, I'm getting the following error. When I used graph-network I received the same error. How can I train schnet architecture using torchmd-net?
torchmd-train --conf t.yaml
Seed set to 1
Number of files: 1
Combined dataset size 1868861
Preloading Custom dataset (of size 427.75 MB)
train 1618159, val 93443, test 157259
Traceback (most recent call last):
File "/home/chemistry/phd/cyz218385/mambaforge/envs/torchmd-net/bin/torchmd-train", line 11, in
sys.exit(main())
File "/home/chemistry/phd/cyz218385/mambaforge/envs/torchmd-net/lib/python3.10/site-packages/torchmdnet/scripts/train.py", line 202, in main
model = LNNP(args, prior_model=prior_models, mean=data.mean, std=data.std)
File "/home/chemistry/phd/cyz218385/mambaforge/envs/torchmd-net/lib/python3.10/site-packages/torchmdnet/module.py", line 89, in init
self.model = create_model(self.hparams, prior_model, mean, std)
File "/home/chemistry/phd/cyz218385/mambaforge/envs/torchmd-net/lib/python3.10/site-packages/torchmdnet/models/model.py", line 110, in create_model
raise ValueError(f'Unknown architecture: {args["model"]}')
ValueError: Unknown architecture: torchmd_gn
Here is my input file
activation: tanh
attn_activation: silu
dataset: Custom
coord_files: "coords.npy"
force_files: "deltaforces.npy"
embed_files: "embeddings.npy"
log_dir: /scratch/chemistry/phd/cyz218385/chig_cgschnet
num_epochs: 200
batch_size: 256
model: torchmd_gn
lr: 0.0001
lr_patience: 4
lr_min: 1.e-06
lr_factor: 0.8
ngpus: 2
num_nodes: 1
num_workers: 8
seed: 1
lr_patience: 1
save_interval: 1
val_size: 0.05
test_size: 157259
#hidden_channels: 128
#num_filters: 128
#num_gaussians: 128
#num_interactions: 3
#max_z: 100
#cutoff: 9

[SoumyaCYZ]

@RaulPPelaez @giadefa can you please help?

[stefdoerr]

It has been renamed to graph-network

[SoumyaCYZ]

Thanks. @stefdoerr sir, but the following syntax (for schnet) are not working with graph-network in torchmd-net.
num_filters: 128
num_gaussians: 128
num_interactions: 3
max_z: 100
cutoff: 9

[stefdoerr]

embedding_dimension: 128
num_rbf: 128
num_layers: 3
max_z: 100
cutoff_upper: 9

[SoumyaCYZ]

@stefdoerr thanks. This is my input file attached below
activation: tanh
attn_activation: tanh
dataset: Custom
coord_files: "deca_ca_coords.npy"
force_files: "deca_ca_deltaforces.npy"
embed_files: "deca_ca_embeddings.npy"
log_dir: /scratch/chemistry/phd/cyz218385/chig_cgschnet
batch_size: 256
model: graph-network
lr: 0.0001
lr_patience: 4
lr_min: 1.e-06
lr_factor: 0.8
num_workers: 2
seed: 1
lr_patience: 1
save_interval: 1
num_rbf: 128
num_layers: 3
max_z: 100
cutoff_upper: 9
lr_metric: val_total_mse_loss
test_size: 0.1
train_size: null
val_size: 0.05
num_epochs: 100
num_heads: 8
num_layers: 1
num_nodes: 1
ngpus: -1
precision: 32
[rank0]: return Variable._execution_engine.run_backward( # Calls into the C++ engine to run the backward pass
[rank0]: RuntimeError: element 0 of tensors does not require grad and does not have a grad_fn
Epoch 0: 0%| | 0/3103 [00:01<?, ?it/s]
I have a total of 1868861 frames in my dataset. How should I fix train_size, test_size, and val_size of the 1868861 dataset? It'll be great help if you help me to solve this error.

[stefdoerr]

this looks like coarse graining. You should report this to https://github.com/torchmd/torchmd-cg

[SoumyaCYZ]

Sir, I think https://github.com/torchmd/torchmd-cg is dead now. After 2023, there has been no discussion in the torchmd_cg repo.