Upload STAR-GO checkpoint + config

README.md

@@ -0,0 +1,46 @@
+---
+title: "stargo-cc"
+tags:
+  - star-go
+  - protein
+  - gene-ontology
+  - bioinformatics
+  - pytorch
+  - lightning
+---
+
+# stargo-cc
+
+STAR-GO checkpoint published for easier discoverability. This repository stores the original Lightning `.ckpt` and the original TOML config so you can reconstruct the model as trained.
+
+## Files
+- `model.ckpt`: PyTorch Lightning checkpoint for `TrainingModel`
+- `config.toml`: training/model config (same schema as this repo's `configs/*.toml`)
+
+## Provenance
+- W&B artifact: `contempro-cc-2020-ordered-encdec-medium:best`
+
+## Usage
+This repository contains a Lightning checkpoint and the original TOML config. Load it like this:
+
+```python
+import torch
+from huggingface_hub import hf_hub_download
+
+from config import from_toml
+from model import TrainingModel, get_model_cls
+
+repo_id = "mmtf/stargo-cc"
+ckpt_path = hf_hub_download(repo_id, "model.ckpt")
+cfg_path = hf_hub_download(repo_id, "config.toml")
+
+cfg = from_toml(cfg_path)
+
+module = TrainingModel.load_from_checkpoint(
+    ckpt_path,
+    model=get_model_cls(cfg.model.name)(cfg.model),
+    training_config=cfg.train,
+)
+module = module.to("cuda" if torch.cuda.is_available() else "cpu")
+module.eval()
+```

config.toml

@@ -0,0 +1,43 @@
+[train]
+# Data paths and configuration
+data_dir = "datasets/pfresgo"
+go_embed_file = "ontology.embeddings.npy"
+protein_embed_file = "per_residue_embeddings.h5"
+subontology = "cellular_component" # overridden in train.py CLI calls
+go_release = "2020" # overridden in train.py CLI calls
+order_go_terms = true
+
+# Compute settings
+use_tpu = false
+prepare_data = false
+dm_num_workers = 0
+bf16_precision = true
+
+# Training hyperparameters
+batch_size = 8
+learning_rate = 6e-5
+weight_decay = 0.01
+max_epochs = 100
+gradient_accumulation = 4
+
+[model]
+# Model type
+name = "bert"
+decoder = true
+
+# Architecture configuration
+hidden_dim = 256
+intermediate_size = 1024
+num_encoder_layers = 6
+num_decoder_layers = 6
+num_attention_heads = 8
+
+# Input dimensions
+go_input_dim = 200
+seq_input_dim = 1024
+
+# Regularization and activation
+hidden_dropout_prob = 0.1
+attention_probs_dropout_prob = 0.1
+hidden_act = "gelu"
+layer_norm_eps = 1e-12

model.ckpt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:80c19794d909f9c745f92f439b275c8d45b52ee1f5f9f768230733a633beb129
+size 137597176