avans06's picture
feat(dataops): Proactively split large tiles in auto_split_upscale to prevent CUDA OOM errors.
aa7cfca
raw
history blame
7.64 kB
#!/usr/bin/env python3
# -*- coding: utf-8 -*-
# The file source is from the [ESRGAN](https://github.com/xinntao/ESRGAN) project
# forked by authors [joeyballentine](https://github.com/joeyballentine/ESRGAN) and [BlueAmulet](https://github.com/BlueAmulet/ESRGAN).
import gc
import numpy as np
import torch
def bgr_to_rgb(image: torch.Tensor) -> torch.Tensor:
# flip image channels
# https://github.com/pytorch/pytorch/issues/229
out: torch.Tensor = image.flip(-3)
# out: torch.Tensor = image[[2, 1, 0], :, :] #RGB to BGR #may be faster
return out
def rgb_to_bgr(image: torch.Tensor) -> torch.Tensor:
# same operation as bgr_to_rgb(), flip image channels
return bgr_to_rgb(image)
def bgra_to_rgba(image: torch.Tensor) -> torch.Tensor:
out: torch.Tensor = image[[2, 1, 0, 3], :, :]
return out
def rgba_to_bgra(image: torch.Tensor) -> torch.Tensor:
# same operation as bgra_to_rgba(), flip image channels
return bgra_to_rgba(image)
def auto_split_upscale(
lr_img: np.ndarray,
upscale_function,
scale: int = 4,
overlap: int = 32,
# A heuristic to proactively split tiles that are too large, avoiding a CUDA error.
# The default (2048*2048) is a conservative value for moderate VRAM (e.g., 8-12GB).
# Adjust this based on your GPU and model's memory footprint.
max_tile_pixels: int = 4194304, # Default: 2048 * 2048 pixels
# Internal parameters for recursion state. Do not set these manually.
known_max_depth: int = None,
current_depth: int = 1,
current_tile: int = 1, # Tracks the current tile being processed
total_tiles: int = 1, # Total number of tiles at this depth level
):
# --- Step 0: Handle CPU-only environment ---
# The entire splitting logic is designed to overcome GPU VRAM limitations.
# If no CUDA-enabled GPU is present, this logic is unnecessary and adds overhead.
# Therefore, we process the image in one go on the CPU.
if not torch.cuda.is_available():
# Note: This assumes the image fits into system RAM, which is usually the case.
result, _ = upscale_function(lr_img, scale)
# The conceptual depth is 1 since no splitting was performed.
return result, 1
"""
Automatically splits an image into tiles for upscaling to avoid CUDA out-of-memory errors.
It uses a combination of a pixel-count heuristic and reactive error handling to find the
optimal processing depth, then applies this depth to all subsequent tiles.
"""
input_h, input_w, input_c = lr_img.shape
# --- Step 1: Decide if we should ATTEMPT to upscale or MUST split ---
# We must split if:
# A) The tile is too large based on our heuristic, and we don't have a known working depth yet.
# B) We have a known working depth from a sibling tile, but we haven't recursed deep enough to reach it yet.
must_split = (known_max_depth is None and (input_h * input_w) > max_tile_pixels) or \
(known_max_depth is not None and current_depth < known_max_depth)
if not must_split:
# If we are not forced to split, let's try to upscale the current tile.
try:
print(f"auto_split_upscale depth: {current_depth}", end=" ", flush=True)
result, _ = upscale_function(lr_img, scale)
# SUCCESS! The upscale worked at this depth.
print(f"progress: {current_tile}/{total_tiles}")
# Return the result and the current depth, which is now the "known_max_depth".
return result, current_depth
except RuntimeError as e:
# Check to see if its actually the CUDA out of memory error
if "CUDA" in str(e):
# OOM ERROR. Our heuristic was too optimistic. This depth is not viable.
print("RuntimeError: CUDA out of memory...")
# Clean up VRAM and proceed to the splitting logic below.
torch.cuda.empty_cache()
gc.collect()
else:
# A different runtime error occurred, so we should not suppress it.
raise RuntimeError(e)
# If an OOM error occurred, flow continues to the splitting section.
# --- Step 2: If we reached here, we MUST split the image ---
# Safety break to prevent infinite recursion if something goes wrong.
if current_depth > 10:
raise RuntimeError("Maximum recursion depth exceeded. Check max_tile_pixels or model requirements.")
# Prepare parameters for the next level of recursion.
next_depth = current_depth + 1
new_total_tiles = total_tiles * 4
base_tile_for_next_level = (current_tile - 1) * 4
# Announce the split only when it's happening.
print(f"Splitting tile at depth {current_depth} into 4 tiles for depth {next_depth}.")
# Split the image into 4 quadrants with overlap.
top_left = lr_img[: input_h // 2 + overlap, : input_w // 2 + overlap, :]
top_right = lr_img[: input_h // 2 + overlap, input_w // 2 - overlap :, :]
bottom_left = lr_img[input_h // 2 - overlap :, : input_w // 2 + overlap, :]
bottom_right = lr_img[input_h // 2 - overlap :, input_w // 2 - overlap :, :]
# Recursively process each quadrant.
# Process the first quadrant to discover the safe depth.
# The first quadrant (top_left) will "discover" the correct processing depth.
# Pass the current `known_max_depth` down.
top_left_rlt, discovered_depth = auto_split_upscale(
top_left, upscale_function, scale=scale, overlap=overlap,
max_tile_pixels=max_tile_pixels,
known_max_depth=known_max_depth,
current_depth=next_depth,
current_tile=base_tile_for_next_level + 1,
total_tiles=new_total_tiles,
)
# Once the depth is discovered, pass it to the other quadrants to avoid redundant checks.
top_right_rlt, _ = auto_split_upscale(
top_right, upscale_function, scale=scale, overlap=overlap,
max_tile_pixels=max_tile_pixels,
known_max_depth=discovered_depth,
current_depth=next_depth,
current_tile=base_tile_for_next_level + 2,
total_tiles=new_total_tiles,
)
bottom_left_rlt, _ = auto_split_upscale(
bottom_left, upscale_function, scale=scale, overlap=overlap,
max_tile_pixels=max_tile_pixels,
known_max_depth=discovered_depth,
current_depth=next_depth,
current_tile=base_tile_for_next_level + 3,
total_tiles=new_total_tiles,
)
bottom_right_rlt, _ = auto_split_upscale(
bottom_right, upscale_function, scale=scale, overlap=overlap,
max_tile_pixels=max_tile_pixels,
known_max_depth=discovered_depth,
current_depth=next_depth,
current_tile=base_tile_for_next_level + 4,
total_tiles=new_total_tiles,
)
# --- Step 3: Stitch the results back together ---
# Reassemble the upscaled quadrants into a single image.
out_h = input_h * scale
out_w = input_w * scale
# Create an empty output image
output_img = np.zeros((out_h, out_w, input_c), np.uint8)
# Fill the output image, removing the overlap regions to prevent artifacts
output_img[: out_h // 2, : out_w // 2, :] = top_left_rlt[: out_h // 2, : out_w // 2, :]
output_img[: out_h // 2, -out_w // 2 :, :] = top_right_rlt[: out_h // 2, -out_w // 2 :, :]
output_img[-out_h // 2 :, : out_w // 2, :] = bottom_left_rlt[-out_h // 2 :, : out_w // 2, :]
output_img[-out_h // 2 :, -out_w // 2 :, :] = bottom_right_rlt[-out_h // 2 :, -out_w // 2 :, :]
return output_img, discovered_depth