perf: tiled encoding of images

The fold/unfold/split_input_params images didn't look good

README.md

@@ -300,6 +300,7 @@ docker run -it --gpus all -v $HOME/.cache/huggingface:/root/.cache/huggingface -
 
 ## ChangeLog
 
+- perf: tiled encoding of images (removes memory bottleneck)
 - perf: use Silu for performance improvement over nonlinearity
 - perf: `xformers` added as a dependency for linux and windows.  Gives a nice speed boost.
 - perf: sliced attention now runs on MacOS. A typo prevented that from happening previously.

imaginairy/feather_tile.py

@@ -7,7 +7,7 @@ import torch
 
 
 def mask_tile(tile, overlap, std_overlap, side="bottom"):
-    h, w = tile.size(2), tile.size(3)
+    b, c, h, w = tile.shape
     top_overlap, bottom_overlap, right_overlap, left_overlap = overlap
     (
         std_top_overlap,
@@ -25,7 +25,7 @@ def mask_tile(tile, overlap, std_overlap, side="bottom"):
             lin_mask_left = (
                 torch.cat([zeros_mask, lin_mask_left], 0)
                 .repeat(h, 1)
-                .repeat(3, 1, 1)
+                .repeat(c, 1, 1)
                 .unsqueeze(0)
             )
 
@@ -33,7 +33,7 @@ def mask_tile(tile, overlap, std_overlap, side="bottom"):
         lin_mask_right = (
             torch.linspace(1, 0, right_overlap, device=tile.device)
             .repeat(h, 1)
-            .repeat(3, 1, 1)
+            .repeat(c, 1, 1)
             .unsqueeze(0)
         )
     if "top" in side:
@@ -41,14 +41,14 @@ def mask_tile(tile, overlap, std_overlap, side="bottom"):
         if top_overlap > std_top_overlap:
             zeros_mask = torch.zeros(top_overlap - std_top_overlap, device=tile.device)
             lin_mask_top = torch.cat([zeros_mask, lin_mask_top], 0)
-        lin_mask_top = lin_mask_top.repeat(w, 1).rot90(3).repeat(3, 1, 1).unsqueeze(0)
+        lin_mask_top = lin_mask_top.repeat(w, 1).rot90(3).repeat(c, 1, 1).unsqueeze(0)
 
     if "bottom" in side:
         lin_mask_bottom = (
             torch.linspace(1, 0, std_bottom_overlap, device=tile.device)
             .repeat(w, 1)
             .rot90(3)
-            .repeat(3, 1, 1)
+            .repeat(c, 1, 1)
             .unsqueeze(0)
         )
 

imaginairy/modules/autoencoder.py

@@ -88,10 +88,40 @@ class AutoencoderKL(pl.LightningModule):
             self.model_ema(self)
 
     def encode(self, x):
-        h = self.encoder(x)
-        moments = self.quant_conv(h)
-        posterior = DiagonalGaussianDistribution(moments)
-        return posterior
+        return self.encode_sliced(x)
+        # h = self.encoder(x)
+        # moments = self.quant_conv(h)
+        # posterior = DiagonalGaussianDistribution(moments)
+        # return posterior.sample()
+
+    def encode_sliced(self, x, chunk_size=128 * 8):
+        """
+        encodes the image in slices.
+        """
+        b, c, h, w = x.size()
+        final_tensor = torch.zeros(
+            [1, 4, math.ceil(h / 8), math.ceil(w / 8)], device=x.device
+        )
+        for x_img in x.split(1):
+            encoded_chunks = []
+            overlap_pct = 0.5
+            chunks = tile_image(
+                x_img, tile_size=chunk_size, overlap_percent=overlap_pct
+            )
+
+            for img_chunk in chunks:
+                h = self.encoder(img_chunk)
+                moments = self.quant_conv(h)
+                posterior = DiagonalGaussianDistribution(moments)
+                encoded_chunks.append(posterior.sample())
+            final_tensor = rebuild_image(
+                encoded_chunks,
+                base_img=final_tensor,
+                tile_size=chunk_size // 8,
+                overlap_percent=overlap_pct,
+            )
+
+        return final_tensor
 
     def decode(self, z):
         try: