Added package versions; Added date tag on output CSV file; Fixed(?) number of steps;

logger.py

@@ -2,8 +2,11 @@ import torch
 import psutil
 import csv
 import math
+import datetime
+
+NOW_STRING = datetime.datetime.now().strftime("%Y-%m-%d_%H-%M-%S")
+CSV_FILE = f"output_{NOW_STRING}.csv"
 
-CSV_FILE = "output.csv"
 
 def init_csv():
     """Initialize CSV file with headers."""
@@ -13,6 +16,7 @@ def init_csv():
                          "peak_memory_usage_allocated_GB", "peak_memory_usage_reserved_GB", 
                          "loss", "perplexity"])
 
+
 def measure_memory():
     """Measure memory usage from CUDA or CPU."""
     if torch.cuda.is_available():
@@ -28,6 +32,7 @@ def measure_memory():
 
     return peak_history, max_allocated, max_reserved
 
+
 def log_to_csv(epoch, step, compute_time, loss):
     """Log training metrics to CSV file."""
     peak_history, max_allocated, max_reserved = measure_memory()

main.py

@@ -72,13 +72,6 @@ if __name__ == "__main__":
 
     dataset = load_data(args, tokenizer)
 
-    optimizer, model = get_optimizer(args, model)
-
-    num_steps = ceil(args.num_epochs * len(dataset) / args.batch_size)
-    scheduler = get_scheduler(
-        optimizer, args.lr_scheduler, args.warm_up_fraction, num_steps, args.lr, args.lr_min
-    )
-
     shuffle = True if args.shuffle == "true" else False
     if args.mode == "pretraining":
         dataloader = DataLoader(dataset, batch_size=args.batch_size, shuffle=shuffle)
@@ -87,6 +80,13 @@ if __name__ == "__main__":
     else:
         raise ValueError("Invalid mode. Choose 'pretraining' or 'finetuning'")
 
+    optimizer, model = get_optimizer(args, model)
+
+    num_steps = ceil(args.num_epochs * len(dataloader))
+    scheduler = get_scheduler(
+        optimizer, args.lr_scheduler, args.warm_up_fraction, num_steps, args.lr, args.lr_min
+    )
+
     trained_model = train(device, accelerator, scheduler, model, optimizer, dataloader, num_epochs=args.num_epochs)
 
     file_name = f"{args.model}_{args.optimizer}_pretrained" if args.mode == "pretraining" else f"{args.model}_{args.optimizer}_finetuned"

requirements.txt

-galore-torch
-datasets
-transformers
-torch
-accelerate
-psutil
-peft
-argparse
\ No newline at end of file
+galore-torch==1.0
+datasets==3.1.0
+transformers==4.46.3
+torch==2.4.1
+accelerate==1.0.1
+psutil==7.0.0
+peft==0.13.2
+argparse==1.4.0
\ No newline at end of file