From 4d3d318a4168ef79847737cef2c0ad8a4dafd3e7 Mon Sep 17 00:00:00 2001
From: Volpeon <git@volpeon.ink>
Date: Thu, 29 Dec 2022 09:00:19 +0100
Subject: Training improvements

---
 training/lr.py           |  7 ++++---
 training/optimization.py | 43 +++++++++++++++++++++++--------------------
 2 files changed, 27 insertions(+), 23 deletions(-)

(limited to 'training')

diff --git a/training/lr.py b/training/lr.py
index c0e9b3f..0c5ce9e 100644
--- a/training/lr.py
+++ b/training/lr.py
@@ -90,6 +90,7 @@ class LRFinder():
             else:
                 if smooth_f > 0:
                     loss = smooth_f * loss + (1 - smooth_f) * losses[-1]
+                    acc = smooth_f * acc + (1 - smooth_f) * accs[-1]
                 if loss < best_loss:
                     best_loss = loss
                 if acc > best_acc:
@@ -132,9 +133,9 @@ class LRFinder():
         ax_loss.set_xlabel("Learning rate")
         ax_loss.set_ylabel("Loss")
 
-        # ax_acc = ax_loss.twinx()
-        # ax_acc.plot(lrs, accs, color='blue')
-        # ax_acc.set_ylabel("Accuracy")
+        ax_acc = ax_loss.twinx()
+        ax_acc.plot(lrs, accs, color='blue')
+        ax_acc.set_ylabel("Accuracy")
 
         print("LR suggestion: steepest gradient")
         min_grad_idx = None
diff --git a/training/optimization.py b/training/optimization.py
index a0c8673..dfee2b5 100644
--- a/training/optimization.py
+++ b/training/optimization.py
@@ -1,4 +1,7 @@
 import math
+from typing import Literal
+
+import torch
 from torch.optim.lr_scheduler import LambdaLR
 
 from diffusers.utils import logging
@@ -6,41 +9,41 @@ from diffusers.utils import logging
 logger = logging.get_logger(__name__)
 
 
-def get_one_cycle_schedule(optimizer, num_training_steps, annealing="cos", min_lr=0.04, mid_point=0.3, last_epoch=-1):
-    """
-    Create a schedule with a learning rate that decreases linearly from the initial lr set in the optimizer to 0, after
-    a warmup period during which it increases linearly from 0 to the initial lr set in the optimizer.
-    Args:
-        optimizer ([`~torch.optim.Optimizer`]):
-            The optimizer for which to schedule the learning rate.
-        num_training_steps (`int`):
-            The total number of training steps.
-        last_epoch (`int`, *optional*, defaults to -1):
-            The index of the last epoch when resuming training.
-    Return:
-        `torch.optim.lr_scheduler.LambdaLR` with the appropriate schedule.
-    """
-
+def get_one_cycle_schedule(
+    optimizer: torch.optim.Optimizer,
+    num_training_steps: int,
+    warmup: Literal["cos", "linear"] = "cos",
+    annealing: Literal["cos", "half_cos", "linear"] = "cos",
+    min_lr: int = 0.04,
+    mid_point: int = 0.3,
+    last_epoch: int = -1
+):
     def lr_lambda(current_step: int):
         thresh_up = int(num_training_steps * min(mid_point, 0.5))
 
         if current_step < thresh_up:
-            return min_lr + float(current_step) / float(max(1, thresh_up)) * (1 - min_lr)
+            progress = float(current_step) / float(max(1, thresh_up))
+
+            if warmup == "linear":
+                return min_lr + progress * (1 - min_lr)
+
+            return min_lr + 0.5 * (1.0 + math.cos(math.pi * (1 + progress)))
 
         if annealing == "linear":
             thresh_down = thresh_up * 2
 
             if current_step < thresh_down:
-                return min_lr + float(thresh_down - current_step) / float(max(1, thresh_down - thresh_up)) * (1 - min_lr)
+                progress = float(thresh_down - current_step) / float(max(1, thresh_down - thresh_up))
+                return min_lr + progress * (1 - min_lr)
 
             progress = float(num_training_steps - current_step) / float(max(1, num_training_steps - thresh_down))
-            return max(0.0, progress) * min_lr
+            return progress * min_lr
 
         progress = float(current_step - thresh_up) / float(max(1, num_training_steps - thresh_up))
 
         if annealing == "half_cos":
-            return max(0.0, 1.0 + math.cos(math.pi * (0.5 + 0.5 * progress)))
+            return 1.0 + math.cos(math.pi * (0.5 + 0.5 * progress))
 
-        return max(0.0, 0.5 * (1.0 + math.cos(math.pi * progress)))
+        return 0.5 * (1.0 + math.cos(math.pi * progress))
 
     return LambdaLR(optimizer, lr_lambda, last_epoch)
-- 
cgit v1.2.3-70-g09d2