Improved aspect ratio bucketing

1 files changed, 19 insertions, 3 deletions

diff --git a/data/csv.py b/data/csv.py
index 7527b7d..55a1988 100644
--- a/data/csv.py
+++ b/data/csv.py
@@ -44,18 +44,25 @@ def generate_buckets(
     items: list[str],
     base_size: int,
     step_size: int = 64,
+    max_pixels: Optional[int] = None,
     num_buckets: int = 4,
     progressive_buckets: bool = False,
     return_tensor: bool = True
 ):
+    if max_pixels is None:
+        max_pixels = (base_size + step_size) ** 2
+
+    max_pixels = max(max_pixels, base_size * base_size)
+
     bucket_items: list[int] = []
     bucket_assignments: list[int] = []
     buckets = [1.0]
 
     for i in range(1, num_buckets + 1):
-        s = base_size + i * step_size
-        buckets.append(s / base_size)
-        buckets.append(base_size / s)
+        long_side = base_size + i * step_size
+        short_side = min(base_size - math.ceil((base_size - max_pixels / long_side) / step_size) * step_size, base_size)
+        buckets.append(long_side / short_side)
+        buckets.append(short_side / long_side)
 
     buckets = torch.tensor(buckets)
     bucket_indices = torch.arange(len(buckets))
@@ -110,6 +117,8 @@ class VlpnDataModule():
         num_class_images: int = 1,
         size: int = 768,
         num_buckets: int = 0,
+        bucket_step_size: int = 64,
+        max_pixels_per_bucket: Optional[int] = None,
         progressive_buckets: bool = False,
         dropout: float = 0,
         interpolation: str = "bicubic",
@@ -135,6 +144,8 @@ class VlpnDataModule():
         self.prompt_processor = prompt_processor
         self.size = size
         self.num_buckets = num_buckets
+        self.bucket_step_size = bucket_step_size
+        self.max_pixels_per_bucket = max_pixels_per_bucket
         self.progressive_buckets = progressive_buckets
         self.dropout = dropout
         self.template_key = template_key
@@ -223,6 +234,7 @@ class VlpnDataModule():
         train_dataset = VlpnDataset(
             self.data_train, self.prompt_processor,
             num_buckets=self.num_buckets, progressive_buckets=self.progressive_buckets,
+            bucket_step_size=self.bucket_step_size, max_pixels_per_bucket=self.max_pixels_per_bucket,
             batch_size=self.batch_size, generator=generator,
             size=self.size, interpolation=self.interpolation,
             num_class_images=self.num_class_images, dropout=self.dropout, shuffle=True,
@@ -251,6 +263,8 @@ class VlpnDataset(IterableDataset):
         items: list[VlpnDataItem],
         prompt_processor: PromptProcessor,
         num_buckets: int = 1,
+        bucket_step_size: int = 64,
+        max_pixels_per_bucket: Optional[int] = None,
         progressive_buckets: bool = False,
         batch_size: int = 1,
         num_class_images: int = 0,
@@ -274,7 +288,9 @@ class VlpnDataset(IterableDataset):
         self.buckets, self.bucket_items, self.bucket_assignments = generate_buckets(
             [item.instance_image_path for item in items],
             base_size=size,
+            step_size=bucket_step_size,
             num_buckets=num_buckets,
+            max_pixels=max_pixels_per_bucket,
             progressive_buckets=progressive_buckets,
         )