feat(sweep): sweep script for getting a vast hyperparam sweep

kan_gpt/VERSION

@@ -1 +1 @@
-0.2.0
+0.3.0

kan_gpt/sweep.py

@@ -0,0 +1,71 @@
+import wandb
+from kan_gpt.train import main
+
+
+def wandb_sweep():
+    run = wandb.init(resume="allow", anonymous="must")
+
+    class Args:
+        model_type = wandb.config.model_type
+        dummy_dataset = wandb.config.dummy_dataset
+        learning_rate = wandb.config.learning_rate
+        max_iters = wandb.config.max_iters
+        num_workers = wandb.config.num_workers
+        batch_size = wandb.config.batch_size
+        dataset = wandb.config.dataset
+        architecture = wandb.config.architecture
+        device = wandb.config.device
+
+    run_args = Args()
+
+    main(args=run_args, run=run)
+
+
+def sweep(args):
+    sweep_configuration = {
+        "method": "random",
+        "name": "sweep",
+        "metric": {"goal": "minimize", "name": "test_loss"},
+        "parameters": {
+            "model_type": {"values": args.model_type},
+            "batch_size": {"values": args.batch_size},
+            "dummy_dataset": {"values": args.dummy_dataset},
+            "learning_rate": {"values": args.learning_rate},
+            "max_iters": {"values": args.max_iters},
+            "num_workers": {"values": args.num_workers},
+            "dataset": {"values": args.dataset},
+            "architecture": {"values": args.architecture},
+            "device": {"values": args.device},
+        },
+    }
+
+    sweep_id = wandb.sweep(sweep_configuration, project="KAN-GPT")
+    print("sweep_id (generated)", sweep_id)
+
+    wandb.agent(sweep_id, function=wandb_sweep)
+
+
+if __name__ == "__main__":
+
+    class SweepArgs:
+        model_type = ["gpt-mini", "gpt-nano", "gpt2"]
+        dummy_dataset = [
+            False,
+        ]
+        learning_rate = [5e-3, 5e-4, 5e-5, 5e-6]
+        max_iters = [
+            1000,
+        ]
+        num_workers = [
+            0,
+        ]
+        batch_size = [1, 2, 4, 8, 12, 16]
+        dataset = [
+            "tinyshakespeare",
+        ]
+        architecture = ["MLP", "KAN"]
+        device = [
+            "auto",
+        ]
+
+    sweep(SweepArgs())

kan_gpt/train.py

@@ -62,7 +62,7 @@ def save_model(
     return save_path
 
 
-def main(args):
+def main(args, run=None):
     config = {
         "model_type": args.model_type,
         "batch_size": args.batch_size,
@@ -116,7 +116,8 @@ def main(args):
     train_config.device = args.device
     trainer = Trainer(train_config, model, train_dataset)
 
-    run = wandb.init(project="KAN-GPT", config=config)
+    if run is None:
+        run = wandb.init(project="KAN-GPT", config=config)
     wandb.watch(model)
 
     def batch_end_callback(trainer):