optimizer save load completed

basedformer/lm_base.py

@@ -6,7 +6,7 @@ from basedformer import gptj
 import os
 import json
 from dataclasses import dataclass
-
+from pathlib import Path
 '''
 BaseLM config dataclass:
 model_config = {
@@ -27,7 +27,6 @@ class BaseLMConfig():
     vocab_dim: int
     eps: float
 
-
 #Having common BaseLM functionality in this class instead of the torch LM itself makes sense.
 class BaseLM(nn.Module):
     def __init__(self, config=None, lm=None):

basedformer/optimizer.py

@@ -3,18 +3,29 @@ import numpy as np
 import torch
 from dotmap import DotMap
 import pickle
+import os
+from pathlib import Path
 #Based Optimizer
 
-def lr_schedule(step, warmup_steps, anneal_steps, lr, end_lr):
+def lr_schedule(step, warmup_steps, anneal_steps, lr, end_lr, cosine_warmup=False):
     warmup_percent = np.clip(step, 0, warmup_steps) / warmup_steps
     anneal_percent = np.clip(step - warmup_steps, 0, anneal_steps) / anneal_steps
-    #cosine schedule for annealing
-    return lr * warmup_percent - (lr - end_lr) * (1 - np.cos(np.pi * anneal_percent)) / 2
-    #kinda broken. doesn't start from 0
 
+    if cosine_warmup:
+        main_lr = lr * (1 - np.cos(np.pi * warmup_percent)) / 2
+    else:
+        main_lr = lr * warmup_percent
+
+    anneal_lr = (lr - end_lr) * (1 - np.cos(np.pi * anneal_percent)) / 2
+    return main_lr - anneal_lr
 
 class BasedOptimizer:
-    def __init__(self, parameters, config, optimizer):
+    def __init__(self, parameters, config, optimizer, init=True):
+        if init:
+            self.init_config(config)
+            self.init_optimizer(parameters, optimizer)
+
+    def init_config(self, config):
         defaults = {
             "lr": 6e-4,
             "end_lr": 6e-4,
@@ -27,6 +38,9 @@ class BasedOptimizer:
             "beta1": 0.9,
             "beta2": 0.95,
             "eps": 1e-4,
+            "max_lr": False,
+            "curr_step": 0,
+            "curr_lr": 0,
         }
         
         for k, v in defaults.items():
@@ -35,35 +49,34 @@ class BasedOptimizer:
         for k, v in config.items():
             setattr(self, k, v)
 
-        self.max_lr = False
-        self.curr_step = 0
-        self.curr_lr = 0
-            
-        if optimizer == "adamw":
-            self.optimizer = optim.AdamW(parameters, lr=0, weight_decay=self.weight_decay, betas=(self.beta1, self.beta2), eps=self.eps)
+    def init_optimizer(self, parameters, optimizer_name):
+        if optimizer_name == "adamw":
+            self.optimizer = optim.AdamW(self.parameters, lr=0, weight_decay=self.weight_decay, betas=(self.beta1, self.beta2), eps=self.eps)
 
-        elif optimizer == "adamw8bit":
+        elif optimizer_name == "adamw8bit":
             import bitsandbytes as bnb
-            self.optimizer = bnb.optim.Adam8bit(parameters, lr=0, weight_decay=self.weight_decay, betas=(self.beta1, self.beta2), eps=self.eps)
+            self.optimizer = bnb.optim.Adam8bit(self.parameters, lr=0, weight_decay=self.weight_decay, betas=(self.beta1, self.beta2), eps=self.eps)
 
-        elif optimizer == "adafactor":
+        elif optimizer_name == "adafactor":
             try:
                 from transformers.optimization import Adafactor
 
             except ImportError:
                 raise ImportError("Please install transformers for Adafactor")
 
-            self.optimizer = Adafactor(params=parameters)
+            self.optimizer = Adafactor(params=self.parameters)
 
-    def step(self, scaler=None):
+        
+    def step(self, dry_run=False, scaler=None):
+        if not dry_run:
             if scaler:
                 scaler.step(self.optimizer)
 
             else:
                 self.optimizer.step()
         
-        self.curr_step = self.curr_step + 1
         self.curr_lr = lr_schedule(self.curr_step, self.warmup_steps, self.anneal_steps, self.lr, self.end_lr)
+        self.curr_step = self.curr_step + 1
         
         if not self.max_lr:
             if self.curr_lr == self.end_lr:
@@ -85,15 +98,21 @@ class BasedOptimizer:
         if self.curr_step != 0:
             print(f"curr_lr: {str(self.get_current_lr())}")
 
-    def save(self, path):
-        torch.save(self.optimizer.state_dict(), path)
-        with open(path, 'wb') as f:
-            pickle.dump(self, f)
+    def save(self, path: Path):
+        path = path / "opt"
+        path.mkdir(parents=True, exist_ok=True)
+        torch.save(self.optimizer.state_dict(), path / "opt_states.pt")
+        del self.optimizer
+        metadata = self.__dict__
+        with open(path / "opt_metadata.pkl", 'wb') as f:
+            pickle.dump(metadata, f)
 
     @classmethod
-    def load(cls, path):
-        with open(path, 'rb') as f:
-            based_optimizer = pickle.load(f)
+    def load(cls, parameters, path):
+        path = path / "opt"
+        with open(path / "opt_metadata.pkl", 'rb') as f:
+            metadata = pickle.load(f)
         
-        based_optimizer.optimizer.load_state_dict(torch.load(path))
+        based_optimizer = cls(parameters, metadata, metadata["optimizer_name"])
+        based_optimizer.optimizer.load_state_dict(torch.load(path / "opt_states.pt"))
         return based_optimizer
\ No newline at end of file

hypertrain.py

@@ -146,7 +146,6 @@ class HyperNetworkSingle(nn.Module):
         return x.bfloat16()
 
 model_config = {
-    "model_class": 
     "n_layer": 28,
     "n_head": 16,
     "hidden_dim": 4096,

scripts/test_optimizer.py

+from basedformer import optimizer
+import torch
+from tqdm import tqdm
+import wandb
+import os
+from pathlib import Path
+train_config = {
+    "lr": 5e-4,
+    "end_lr": 1e-4,
+    "warmup_steps": 100,
+    "anneal_steps": 90,
+}
+
+model = torch.nn.Linear(10, 100)
+save_folder = "models/test_optimizer2"
+if not os.path.isdir(save_folder + "/opt"):
+    opt = optimizer.BasedOptimizer(model.parameters(), train_config, "adamw")
+else:
+    opt = optimizer.BasedOptimizer.load(model.parameters(), Path(save_folder))
+
+wandb.init(project="opt-test", name="test")
+
+for x in tqdm(range(opt.curr_step, 100)):
+    print(f"Step {opt.curr_step}: LR {opt.curr_lr}")
+    wandb.log({"lr": opt.curr_lr})
+    opt.step(dry_run=True)
+    #if x == 60:
+        #opt.save(Path(save_folder))