gcp v1 large

checkpoints/best_valid.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7d1d43950a29834e7f702409bf957e9ffb75eb3cb3952074ba4c545bb4130eaf
+size 2545380820

config_gcpnet_encoder.yaml

@@ -0,0 +1,83 @@
+features:
+  module: models.gcpnet.features.factory.ProteinFeaturiser
+  kwargs:
+    representation: CA
+    scalar_node_features:
+    - amino_acid_one_hot
+    - sequence_positional_encoding
+    - alpha
+    - kappa
+    - dihedrals
+    vector_node_features:
+    - orientation
+    edge_types:
+    - knn_16
+    scalar_edge_features:
+    - edge_distance
+    vector_edge_features:
+    - edge_vectors
+task:
+  transform: null
+encoder:
+  module: models.gcpnet.models.graph_encoders.gcpnet.GCPNetModel
+  kwargs:
+    num_layers: 6
+    emb_dim: 128
+    node_s_emb_dim: 128
+    node_v_emb_dim: 16
+    edge_s_emb_dim: 32
+    edge_v_emb_dim: 4
+    r_max: 10.0
+    num_rbf: 8
+    activation: silu
+    pool: sum
+    module_cfg:
+      norm_pos_diff: true
+      scalar_gate: 0
+      vector_gate: true
+      scalar_nonlinearity: silu
+      vector_nonlinearity: silu
+      nonlinearities:
+      - silu
+      - silu
+      r_max: 10.0
+      num_rbf: 8
+      bottleneck: 4
+      vector_linear: true
+      vector_identity: true
+      default_bottleneck: 4
+      predict_node_positions: false
+      predict_node_rep: true
+      node_positions_weight: 1.0
+      update_positions_with_vector_sum: false
+      enable_e3_equivariance: false
+      pool: sum
+    model_cfg:
+      h_input_dim: 49
+      chi_input_dim: 2
+      e_input_dim: 9
+      xi_input_dim: 1
+      h_hidden_dim: 128
+      chi_hidden_dim: 16
+      e_hidden_dim: 32
+      xi_hidden_dim: 4
+      num_layers: 6
+      dropout: 0.0
+    layer_cfg:
+      pre_norm: false
+      use_gcp_norm: true
+      use_gcp_dropout: true
+      use_scalar_message_attention: true
+      num_feedforward_layers: 2
+      dropout: 0.0
+      nonlinearity_slope: 0.01
+      mp_cfg:
+        edge_encoder: false
+        edge_gate: false
+        num_message_layers: 4
+        message_residual: 0
+        message_ff_multiplier: 1
+        self_message: true
+  checkpoint_path: ./models/checkpoints/structure_denoising/gcpnet/ca_bb/last.ckpt
+top_k: 30
+num_positional_embeddings: 16

config_geometric_decoder.yaml

@@ -0,0 +1,14 @@
+dimension: 1024  # Used as dim_in and dim_out for ContinuousTransformerWrapper
+ff_mult: 4         # Multiplier for the feedforward dimension
+depth: 16      # Number of layers in the Encoder
+heads: 16      # Number of attention heads in the Encoder
+rotary_pos_emb: True
+attn_flash: True # FA-2 if installed
+attn_kv_heads: 1 # GQA
+qk_norm: True
+pre_norm: True
+residual_attn: False # Set pre_norm to False if residual_attn is True
+num_memory_tokens: 0 # Number of memory tokens, 0 means no memory tokens
+
+direction_loss_bins: 16
+pos_scale_factor: 1.0

config_vqvae.yaml

@@ -0,0 +1,129 @@
+fix_seed: 0
+checkpoints_every: 1
+tensorboard_log: True
+tqdm_progress_bar: False
+result_path: ./results/vqvae
+find_unused_parameters: True
+dispatch_batches: False
+even_batches: True
+non_blocking: False
+split_batches: False
+
+resume:
+  enabled: True
+  resume_path: results/vqvae/2025-07-17__16-36-40/checkpoints/epoch_1.pth
+  restart_optimizer: True
+  discard_decoder_weights: False
+
+model:
+  compile_model: False
+  max_length: 1280
+  decoder_output_scaling_factor: 1  # Added scaling factor for backbone prediction outputs
+  use_ndlinear: False  # Toggle for using NdLinear instead of Conv1d layers
+  encoder:
+    name: gcpnet # gcpnet
+    freeze_parameters: False
+    pretrained:
+      enabled: True
+      config_path: ./configs/pretrained/structure_denoising_pretrained_config.yaml
+      checkpoint_path: ./models/checkpoints/structure_denoising/gcpnet/ca_bb/last.ckpt # define your checkpoint directory here
+  vqvae:
+    vector_quantization:
+      enabled: True
+      freeze_parameters: False
+      dim: 256
+      decay: 0.995
+      codebook_size: 4096
+      commitment_weight: 0.05
+      orthogonal_reg_weight: 10
+      orthogonal_reg_max_codes: 512
+      orthogonal_reg_active_codes_only: True
+      rotation_trick: True
+      threshold_ema_dead_code: 2
+      kmeans_init: True
+      kmeans_iters: 10
+      alpha: 0.25
+    encoder:
+      freeze_parameters: False
+      dimension: 1024  # Used as dim_in and dim_out for ContinuousTransformerWrapper
+      ff_mult: 4         # Multiplier for the feedforward dimension
+      depth: 12      # Number of layers in the Encoder
+      heads: 12      # Number of attention heads in the Encoder
+      rotary_pos_emb: True
+      attn_flash: True # FA-2 if installed
+      attn_kv_heads: 3 # GQA
+      qk_norm: True
+      pre_norm: True
+      residual_attn: False # Set pre_norm to False if residual_attn is True
+      num_memory_tokens: 0 # Number of memory tokens, 0 means no memory tokens
+    decoder:
+      name: geometric_decoder # geometric_decoder
+      freeze_parameters: False
+
+train_settings:
+  data_path: ../../datasets/vqvae/uniref_50/
+  num_epochs: 16
+  shuffle: True
+  mixed_precision: bf16 # no, fp16, bf16, fp8
+  save_pdb_every: 1
+  batch_size: 4
+  num_workers: 24
+  grad_accumulation: 1
+  max_task_samples: 24000000
+  profile_train_loop: False
+  cutoff_augmentation:
+    enabled: False
+    probability: 0.5
+    min_length: 25
+  nan_augmentation:
+    enabled: True
+    probability: 0.05
+    max_length: 30
+  gradient_norm_logging_freq: 50  # How often to calculate and log gradient norm (in steps)
+  losses:
+    alignment_strategy: kabsch # kabsch, quaternion, no
+    mse:
+      enabled: True
+      weight: 0.005
+    backbone_distance:
+      enabled: True
+      weight: 0.01
+    backbone_direction:
+      enabled: True
+      weight: 0.05
+    binned_distance_classification:
+      enabled: False
+      weight: 0.01
+    binned_direction_classification:
+      enabled: False
+      weight: 0.01
+
+valid_settings:
+  data_path: ../../datasets/vqvae/whole_validation_2048_h5/
+  do_every: 1
+  save_pdb_every: 1
+  batch_size: 8
+  num_workers: 0
+
+visualization_settings:
+  data_path: ../../datasets/vqvae/whole_validation_2048_h5/validation_set_2048_h5
+  fasta_path: visualization/Rep_subfamily_basedon_S40pdb.fa
+  do_every: 8192
+  batch_size: 1
+  num_workers: 4
+
+optimizer:
+  name: adam
+  lr: 5e-5
+  weight_decouple: True
+  weight_decay: 1e-4
+  eps: 1e-7
+  beta_1: 0.9
+  beta_2: 0.98
+  use_8bit_adam: True
+  grad_clip_norm: 1
+  decay:
+    warmup: 16000
+    min_lr: 1e-6
+    gamma: 0.2
+    num_restarts: 1