diff --git a/models/rfd3/configs/experiment/rfd3na.yaml b/models/rfd3/configs/experiment/rfd3na.yaml deleted file mode 100644 index 8739992..0000000 --- a/models/rfd3/configs/experiment/rfd3na.yaml +++ /dev/null @@ -1,100 +0,0 @@ -# @package _global_ -# Training configuration for RFD3 - -defaults: - #- /debug/default - - override /model: rfd3_base - - override /logger: wandb - - override /datasets: design_base_rfd3na - - _self_ - -name: rfd3na_scratch_clean_test -tags: [print-model] - -#ckpt: null -#ckpt_path: /net/scratch/raktim/training/logs/train/rfd3na-fine-tune/2026-02-17_15-21_JOB_3608285/ckpt/epoch-0590.ckpt -ckpt_path: /net/scratch/raktim/training/logs/train/rfd3na_scratch_clean_test/2026-02-19_01-58_JOB_3986244/ckpt/epoch-0180.ckpt - - -model: - net: - token_initializer: - token_1d_features: - ref_motif_token_type: 3 - restype: 32 - is_dna_token: 1 - is_rna_token: 1 - is_protein_token: 1 - token_2d_features: - bp_partners: 3 # Unspecified, pair, loop - atom_1d_features: - ref_atom_name_chars: 256 - ref_element: 128 - ref_charge: 1 - ref_mask: 1 - ref_is_motif_atom_with_fixed_coord: 1 - ref_is_motif_atom_unindexed: 1 - has_zero_occupancy: 1 - ref_pos: 3 - - # Guided features - ref_atomwise_rasa: 3 - active_donor: 1 - active_acceptor: 1 - is_atom_level_hotspot: 1 - diffusion_module: - n_recycle: 2 - use_local_token_attention: True - diffusion_transformer: - n_local_tokens: 32 - n_keys: 128 - - inference_sampler: - num_timesteps: 100 - - -datasets: - diffusion_batch_size_train: 16 - crop_size: 256 - max_atoms_in_crop: 2560 # ~10x crop size. - global_transform_args: - meta_conditioning_probabilities: - p_is_nucleic_ss_example: 0.25 - p_nucleic_ss_show_partial_feats: 0.7 - p_canonical_bp_filter: 0.2 - #calculate_NA_SS: 0.3 - - association_scheme: atom23 - #add_na_pair_features: true - train_conditions: - unconditional: - frequency: 2.0 - island: - frequency: 2.0 - sequence_design: - frequency: 0.5 - tipatom: - frequency: 5.0 - ppi: - frequency: 0.0 - train: - # These are the ratios used in the preprint but we set all pdb sampling by default since not everyone might download the distillation data. - pdb: - probability: 0.6 - rna_monomer_distillation: - probability: 0.3 - monomer_distillation: - probability: 0.1 - - val: - pseudoknot: - dataset: - # eval_every_n: 10 - eval_every_n: 5 - -trainer: - #devices_per_node: 1 - #limit_train_batches: 10 - #limit_val_batches: 1 - validate_every_n_epochs: 5 - prevalidate: true diff --git a/models/rfd3/configs/experiment/rfd3na_af.yaml b/models/rfd3/configs/experiment/rfd3na_af.yaml deleted file mode 100644 index e9331b7..0000000 --- a/models/rfd3/configs/experiment/rfd3na_af.yaml +++ /dev/null @@ -1,99 +0,0 @@ -# @package _global_ -# Training configuration for RFD3 - -defaults: - - /debug/default - - override /model: rfd3_base - # - override /logger: wandb - - override /datasets: design_base_rfd3na - - _self_ - -name: rfd3na-fine-tune -tags: [print-model] -ckpt_path: /projects/ml/aa_design/models/rfd3_latest_foundry.ckpt - -model: - net: - token_initializer: - token_1d_features: - ref_motif_token_type: 3 - restype: 32 - is_dna_token: 1 - is_rna_token: 1 - is_protein_token: 1 - token_2d_features: - bp_partners: 3 # Unspecified, pair, loop - atom_1d_features: - ref_atom_name_chars: 256 - ref_element: 128 - ref_charge: 1 - ref_mask: 1 - ref_is_motif_atom_with_fixed_coord: 1 - ref_is_motif_atom_unindexed: 1 - has_zero_occupancy: 1 - ref_pos: 3 - - # Guided features - ref_atomwise_rasa: 3 - active_donor: 1 - active_acceptor: 1 - is_atom_level_hotspot: 1 - diffusion_module: - n_recycle: 2 - use_local_token_attention: True - diffusion_transformer: - n_local_tokens: 32 - n_keys: 128 - - inference_sampler: - num_timesteps: 100 - - -datasets: - diffusion_batch_size_train: 16 - crop_size: 256 - max_atoms_in_crop: 2560 # ~10x crop size. - global_transform_args: - meta_conditioning_probabilities: - # p_is_nucleic_ss_example: 0.25 - # p_nucleic_ss_show_partial_feats: 0.7 - # p_canonical_bp_filter: 0.2 - p_is_nucleic_ss_example: 1.0 - p_nucleic_ss_show_partial_feats: 0.0 - p_canonical_bp_filter: 0.0 - #calculate_NA_SS: 0.3 - - association_scheme: atom23 - #add_na_pair_features: true - train_conditions: - unconditional: - frequency: 2.0 - island: - frequency: 2.0 - sequence_design: - frequency: 0.5 - tipatom: - frequency: 5.0 - ppi: - frequency: 0.0 - train: - # These are the ratios used in the preprint but we set all pdb sampling by default since not everyone might download the distillation data. - pdb: - probability: 0.0 - rna_monomer_distillation: - probability: 1.0 - monomer_distillation: - probability: 0.0 - - val: - pseudoknot: - dataset: - # eval_every_n: 10 - eval_every_n: 5 - -trainer: - #devices_per_node: 1 - #limit_train_batches: 10 - #limit_val_batches: 1 - validate_every_n_epochs: 5 - prevalidate: true diff --git a/models/rfd3/configs/experiment/rfd3na_no_distill.yaml b/models/rfd3/configs/experiment/rfd3na_no_distill.yaml deleted file mode 100644 index abb302b..0000000 --- a/models/rfd3/configs/experiment/rfd3na_no_distill.yaml +++ /dev/null @@ -1,97 +0,0 @@ -# @package _global_ -# Training configuration for RFD3 - -defaults: - #- /debug/default - - override /model: rfd3_base - - override /logger: wandb - - override /datasets: design_base_rfd3na - - _self_ - -name: rfd3na_no_distill -tags: [print-model] -#ckpt_path: null -ckpt_path: /net/scratch/raktim/training/logs/train/rfd3na_no_distill/2026-02-19_02-16_JOB_3988845/ckpt/epoch-0170.ckpt - -model: - net: - token_initializer: - token_1d_features: - ref_motif_token_type: 3 - restype: 32 - is_dna_token: 1 - is_rna_token: 1 - is_protein_token: 1 - token_2d_features: - bp_partners: 3 # Unspecified, pair, loop - atom_1d_features: - ref_atom_name_chars: 256 - ref_element: 128 - ref_charge: 1 - ref_mask: 1 - ref_is_motif_atom_with_fixed_coord: 1 - ref_is_motif_atom_unindexed: 1 - has_zero_occupancy: 1 - ref_pos: 3 - - # Guided features - ref_atomwise_rasa: 3 - active_donor: 1 - active_acceptor: 1 - is_atom_level_hotspot: 1 - diffusion_module: - n_recycle: 2 - use_local_token_attention: True - diffusion_transformer: - n_local_tokens: 32 - n_keys: 128 - - inference_sampler: - num_timesteps: 100 - - -datasets: - diffusion_batch_size_train: 16 - crop_size: 256 - max_atoms_in_crop: 2560 # ~10x crop size. - global_transform_args: - meta_conditioning_probabilities: - p_is_nucleic_ss_example: 0.25 - p_nucleic_ss_show_partial_feats: 0.7 - p_canonical_bp_filter: 0.2 - #calculate_NA_SS: 0.3 - - association_scheme: atom23 - #add_na_pair_features: true - train_conditions: - unconditional: - frequency: 2.0 - island: - frequency: 2.0 - sequence_design: - frequency: 0.5 - tipatom: - frequency: 5.0 - ppi: - frequency: 0.0 - train: - # These are the ratios used in the preprint but we set all pdb sampling by default since not everyone might download the distillation data. - pdb: - probability: 0.7 - rna_monomer_distillation: - probability: 0.3 - monomer_distillation: - probability: 0.0 - - val: - pseudoknot: - dataset: - # eval_every_n: 10 - eval_every_n: 5 - -trainer: - #devices_per_node: 1 - #limit_train_batches: 10 - #limit_val_batches: 1 - validate_every_n_epochs: 5 - prevalidate: true diff --git a/models/rfd3/configs/experiment/rfd3na_rm.yaml b/models/rfd3/configs/experiment/rfd3na_rm.yaml deleted file mode 100644 index 1596cb1..0000000 --- a/models/rfd3/configs/experiment/rfd3na_rm.yaml +++ /dev/null @@ -1,90 +0,0 @@ -# @package _global_ -# Training configuration for RFD3 - -defaults: - #- /debug/default - - override /model: rfd3_base - - override /logger: wandb - - override /datasets: design_base_rfd3na - - _self_ - -name: rfd3na -tags: [print-model] -ckpt_path: null - -model: - net: - token_initializer: - token_1d_features: - ref_motif_token_type: 3 - restype: 32 - is_dna_token: 1 - is_rna_token: 1 - is_protein_token: 1 - #token_2d_features: - #bp_partners: 3 # Unspecified, pair, loop - atom_1d_features: - ref_atom_name_chars: 256 - ref_element: 128 - ref_charge: 1 - ref_mask: 1 - ref_is_motif_atom_with_fixed_coord: 1 - ref_is_motif_atom_unindexed: 1 - has_zero_occupancy: 1 - ref_pos: 3 - - # Guided features - ref_atomwise_rasa: 3 - active_donor: 1 - active_acceptor: 1 - is_atom_level_hotspot: 1 - diffusion_module: - n_recycle: 2 - use_local_token_attention: True - diffusion_transformer: - n_local_tokens: 32 - n_keys: 128 - - inference_sampler: - num_timesteps: 100 - - -datasets: - diffusion_batch_size_train: 16 - crop_size: 256 - max_atoms_in_crop: 2560 # ~10x crop size. - global_transform_args: - meta_conditioning_probabilities: - calculate_NA_SS: 0.0 - association_scheme: atom23 - #add_na_pair_features: true - train_conditions: - unconditional: - frequency: 2.0 - island: - frequency: 2.0 - sequence_design: - frequency: 0.5 - tipatom: - frequency: 5.0 - ppi: - frequency: 0.0 - train: - # These are the ratios used in the preprint but we set all pdb sampling by default since not everyone might download the distillation data. - pdb: - probability: 0.5 - rna_monomer_distillation: - probability: 0.5 - - val: - pseudoknot: - dataset: - # eval_every_n: 10 - eval_every_n: 5 - -trainer: - #devices_per_node: 1 - #limit_train_batches: 10 - limit_val_batches: 1 - validate_every_n_epochs: 5 - prevalidate: false