Fabrice-TIERCELIN's picture
Upload 5 files
a6c349f verified
# ------------------------------------------------------------------------
#
# Ultimate VAE Tile Optimization
#
# Introducing a revolutionary new optimization designed to make
# the VAE work with giant images on limited VRAM!
# Say goodbye to the frustration of OOM and hello to seamless output!
#
# ------------------------------------------------------------------------
#
# This script is a wild hack that splits the image into tiles,
# encodes each tile separately, and merges the result back together.
#
# Advantages:
# - The VAE can now work with giant images on limited VRAM
# (~10 GB for 8K images!)
# - The merged output is completely seamless without any post-processing.
#
# Drawbacks:
# - Giant RAM needed. To store the intermediate results for a 4096x4096
# images, you need 32 GB RAM it consumes ~20GB); for 8192x8192
# you need 128 GB RAM machine (it consumes ~100 GB)
# - NaNs always appear in for 8k images when you use fp16 (half) VAE
# You must use --no-half-vae to disable half VAE for that giant image.
# - Slow speed. With default tile size, it takes around 50/200 seconds
# to encode/decode a 4096x4096 image; and 200/900 seconds to encode/decode
# a 8192x8192 image. (The speed is limited by both the GPU and the CPU.)
# - The gradient calculation is not compatible with this hack. It
# will break any backward() or torch.autograd.grad() that passes VAE.
# (But you can still use the VAE to generate training data.)
#
# How it works:
# 1) The image is split into tiles.
# - To ensure perfect results, each tile is padded with 32 pixels
# on each side.
# - Then the conv2d/silu/upsample/downsample can produce identical
# results to the original image without splitting.
# 2) The original forward is decomposed into a task queue and a task worker.
# - The task queue is a list of functions that will be executed in order.
# - The task worker is a loop that executes the tasks in the queue.
# 3) The task queue is executed for each tile.
# - Current tile is sent to GPU.
# - local operations are directly executed.
# - Group norm calculation is temporarily suspended until the mean
# and var of all tiles are calculated.
# - The residual is pre-calculated and stored and addded back later.
# - When need to go to the next tile, the current tile is send to cpu.
# 4) After all tiles are processed, tiles are merged on cpu and return.
#
# Enjoy!
#
# @author: LI YI @ Nanyang Technological University - Singapore
# @date: 2023-03-02
# @license: MIT License
#
# Please give me a star if you like this project!
#
# -------------------------------------------------------------------------
import gc
from time import time
import math
from tqdm import tqdm
import torch
import torch.version
import torch.nn.functional as F
from einops import rearrange
from diffusers.utils.import_utils import is_xformers_available
import SUPIR.utils.devices as devices
try:
import xformers
import xformers.ops
except ImportError:
pass
sd_flag = True
def get_recommend_encoder_tile_size():
if torch.cuda.is_available():
total_memory = torch.cuda.get_device_properties(
devices.device).total_memory // 2**20
if total_memory > 16*1000:
ENCODER_TILE_SIZE = 3072
elif total_memory > 12*1000:
ENCODER_TILE_SIZE = 2048
elif total_memory > 8*1000:
ENCODER_TILE_SIZE = 1536
else:
ENCODER_TILE_SIZE = 960
else:
ENCODER_TILE_SIZE = 512
return ENCODER_TILE_SIZE
def get_recommend_decoder_tile_size():
if torch.cuda.is_available():
total_memory = torch.cuda.get_device_properties(
devices.device).total_memory // 2**20
if total_memory > 30*1000:
DECODER_TILE_SIZE = 256
elif total_memory > 16*1000:
DECODER_TILE_SIZE = 192
elif total_memory > 12*1000:
DECODER_TILE_SIZE = 128
elif total_memory > 8*1000:
DECODER_TILE_SIZE = 96
else:
DECODER_TILE_SIZE = 64
else:
DECODER_TILE_SIZE = 64
return DECODER_TILE_SIZE
if 'global const':
DEFAULT_ENABLED = False
DEFAULT_MOVE_TO_GPU = False
DEFAULT_FAST_ENCODER = True
DEFAULT_FAST_DECODER = True
DEFAULT_COLOR_FIX = 0
DEFAULT_ENCODER_TILE_SIZE = get_recommend_encoder_tile_size()
DEFAULT_DECODER_TILE_SIZE = get_recommend_decoder_tile_size()
# inplace version of silu
def inplace_nonlinearity(x):
# Test: fix for Nans
return F.silu(x, inplace=True)
# extracted from ldm.modules.diffusionmodules.model
# from diffusers lib
def attn_forward_new(self, h_):
batch_size, channel, height, width = h_.shape
hidden_states = h_.view(batch_size, channel, height * width).transpose(1, 2)
attention_mask = None
encoder_hidden_states = None
batch_size, sequence_length, _ = hidden_states.shape
attention_mask = self.prepare_attention_mask(attention_mask, sequence_length, batch_size)
query = self.to_q(hidden_states)
if encoder_hidden_states is None:
encoder_hidden_states = hidden_states
elif self.norm_cross:
encoder_hidden_states = self.norm_encoder_hidden_states(encoder_hidden_states)
key = self.to_k(encoder_hidden_states)
value = self.to_v(encoder_hidden_states)
query = self.head_to_batch_dim(query)
key = self.head_to_batch_dim(key)
value = self.head_to_batch_dim(value)
attention_probs = self.get_attention_scores(query, key, attention_mask)
hidden_states = torch.bmm(attention_probs, value)
hidden_states = self.batch_to_head_dim(hidden_states)
# linear proj
hidden_states = self.to_out[0](hidden_states)
# dropout
hidden_states = self.to_out[1](hidden_states)
hidden_states = hidden_states.transpose(-1, -2).reshape(batch_size, channel, height, width)
return hidden_states
def attn_forward_new_pt2_0(self, hidden_states,):
scale = 1
attention_mask = None
encoder_hidden_states = None
input_ndim = hidden_states.ndim
if input_ndim == 4:
batch_size, channel, height, width = hidden_states.shape
hidden_states = hidden_states.view(batch_size, channel, height * width).transpose(1, 2)
batch_size, sequence_length, _ = (
hidden_states.shape if encoder_hidden_states is None else encoder_hidden_states.shape
)
if attention_mask is not None:
attention_mask = self.prepare_attention_mask(attention_mask, sequence_length, batch_size)
# scaled_dot_product_attention expects attention_mask shape to be
# (batch, heads, source_length, target_length)
attention_mask = attention_mask.view(batch_size, self.heads, -1, attention_mask.shape[-1])
if self.group_norm is not None:
hidden_states = self.group_norm(hidden_states.transpose(1, 2)).transpose(1, 2)
query = self.to_q(hidden_states, scale=scale)
if encoder_hidden_states is None:
encoder_hidden_states = hidden_states
elif self.norm_cross:
encoder_hidden_states = self.norm_encoder_hidden_states(encoder_hidden_states)
key = self.to_k(encoder_hidden_states, scale=scale)
value = self.to_v(encoder_hidden_states, scale=scale)
inner_dim = key.shape[-1]
head_dim = inner_dim // self.heads
query = query.view(batch_size, -1, self.heads, head_dim).transpose(1, 2)
key = key.view(batch_size, -1, self.heads, head_dim).transpose(1, 2)
value = value.view(batch_size, -1, self.heads, head_dim).transpose(1, 2)
# the output of sdp = (batch, num_heads, seq_len, head_dim)
# TODO: add support for attn.scale when we move to Torch 2.1
hidden_states = F.scaled_dot_product_attention(
query, key, value, attn_mask=attention_mask, dropout_p=0.0, is_causal=False
)
hidden_states = hidden_states.transpose(1, 2).reshape(batch_size, -1, self.heads * head_dim)
hidden_states = hidden_states.to(query.dtype)
# linear proj
hidden_states = self.to_out[0](hidden_states, scale=scale)
# dropout
hidden_states = self.to_out[1](hidden_states)
if input_ndim == 4:
hidden_states = hidden_states.transpose(-1, -2).reshape(batch_size, channel, height, width)
return hidden_states
def attn_forward_new_xformers(self, hidden_states):
scale = 1
attention_op = None
attention_mask = None
encoder_hidden_states = None
input_ndim = hidden_states.ndim
if input_ndim == 4:
batch_size, channel, height, width = hidden_states.shape
hidden_states = hidden_states.view(batch_size, channel, height * width).transpose(1, 2)
batch_size, key_tokens, _ = (
hidden_states.shape if encoder_hidden_states is None else encoder_hidden_states.shape
)
attention_mask = self.prepare_attention_mask(attention_mask, key_tokens, batch_size)
if attention_mask is not None:
# expand our mask's singleton query_tokens dimension:
# [batch*heads, 1, key_tokens] ->
# [batch*heads, query_tokens, key_tokens]
# so that it can be added as a bias onto the attention scores that xformers computes:
# [batch*heads, query_tokens, key_tokens]
# we do this explicitly because xformers doesn't broadcast the singleton dimension for us.
_, query_tokens, _ = hidden_states.shape
attention_mask = attention_mask.expand(-1, query_tokens, -1)
if self.group_norm is not None:
hidden_states = self.group_norm(hidden_states.transpose(1, 2)).transpose(1, 2)
query = self.to_q(hidden_states, scale=scale)
if encoder_hidden_states is None:
encoder_hidden_states = hidden_states
elif self.norm_cross:
encoder_hidden_states = self.norm_encoder_hidden_states(encoder_hidden_states)
key = self.to_k(encoder_hidden_states, scale=scale)
value = self.to_v(encoder_hidden_states, scale=scale)
query = self.head_to_batch_dim(query).contiguous()
key = self.head_to_batch_dim(key).contiguous()
value = self.head_to_batch_dim(value).contiguous()
hidden_states = xformers.ops.memory_efficient_attention(
query, key, value, attn_bias=attention_mask, op=attention_op#, scale=scale
)
hidden_states = hidden_states.to(query.dtype)
hidden_states = self.batch_to_head_dim(hidden_states)
# linear proj
hidden_states = self.to_out[0](hidden_states, scale=scale)
# dropout
hidden_states = self.to_out[1](hidden_states)
if input_ndim == 4:
hidden_states = hidden_states.transpose(-1, -2).reshape(batch_size, channel, height, width)
return hidden_states
def attn_forward(self, h_):
q = self.q(h_)
k = self.k(h_)
v = self.v(h_)
# compute attention
b, c, h, w = q.shape
q = q.reshape(b, c, h*w)
q = q.permute(0, 2, 1) # b,hw,c
k = k.reshape(b, c, h*w) # b,c,hw
w_ = torch.bmm(q, k) # b,hw,hw w[b,i,j]=sum_c q[b,i,c]k[b,c,j]
w_ = w_ * (int(c)**(-0.5))
w_ = torch.nn.functional.softmax(w_, dim=2)
# attend to values
v = v.reshape(b, c, h*w)
w_ = w_.permute(0, 2, 1) # b,hw,hw (first hw of k, second of q)
# b, c,hw (hw of q) h_[b,c,j] = sum_i v[b,c,i] w_[b,i,j]
h_ = torch.bmm(v, w_)
h_ = h_.reshape(b, c, h, w)
h_ = self.proj_out(h_)
return h_
def xformer_attn_forward(self, h_):
q = self.q(h_)
k = self.k(h_)
v = self.v(h_)
# compute attention
B, C, H, W = q.shape
q, k, v = map(lambda x: rearrange(x, 'b c h w -> b (h w) c'), (q, k, v))
q, k, v = map(
lambda t: t.unsqueeze(3)
.reshape(B, t.shape[1], 1, C)
.permute(0, 2, 1, 3)
.reshape(B * 1, t.shape[1], C)
.contiguous(),
(q, k, v),
)
out = xformers.ops.memory_efficient_attention(
q, k, v, attn_bias=None, op=self.attention_op)
out = (
out.unsqueeze(0)
.reshape(B, 1, out.shape[1], C)
.permute(0, 2, 1, 3)
.reshape(B, out.shape[1], C)
)
out = rearrange(out, 'b (h w) c -> b c h w', b=B, h=H, w=W, c=C)
out = self.proj_out(out)
return out
def attn2task(task_queue, net):
if False: #isinstance(net, AttnBlock):
task_queue.append(('store_res', lambda x: x))
task_queue.append(('pre_norm', net.norm))
task_queue.append(('attn', lambda x, net=net: attn_forward(net, x)))
task_queue.append(['add_res', None])
elif False: #isinstance(net, MemoryEfficientAttnBlock):
task_queue.append(('store_res', lambda x: x))
task_queue.append(('pre_norm', net.norm))
task_queue.append(
('attn', lambda x, net=net: xformer_attn_forward(net, x)))
task_queue.append(['add_res', None])
else:
task_queue.append(('store_res', lambda x: x))
task_queue.append(('pre_norm', net.norm))
if is_xformers_available:
# task_queue.append(('attn', lambda x, net=net: attn_forward_new_xformers(net, x)))
task_queue.append(
('attn', lambda x, net=net: xformer_attn_forward(net, x)))
elif hasattr(F, "scaled_dot_product_attention"):
task_queue.append(('attn', lambda x, net=net: attn_forward_new_pt2_0(net, x)))
else:
task_queue.append(('attn', lambda x, net=net: attn_forward_new(net, x)))
task_queue.append(['add_res', None])
def resblock2task(queue, block):
"""
Turn a ResNetBlock into a sequence of tasks and append to the task queue
@param queue: the target task queue
@param block: ResNetBlock
"""
if block.in_channels != block.out_channels:
if sd_flag:
if block.use_conv_shortcut:
queue.append(('store_res', block.conv_shortcut))
else:
queue.append(('store_res', block.nin_shortcut))
else:
if block.use_in_shortcut:
queue.append(('store_res', block.conv_shortcut))
else:
queue.append(('store_res', block.nin_shortcut))
else:
queue.append(('store_res', lambda x: x))
queue.append(('pre_norm', block.norm1))
queue.append(('silu', inplace_nonlinearity))
queue.append(('conv1', block.conv1))
queue.append(('pre_norm', block.norm2))
queue.append(('silu', inplace_nonlinearity))
queue.append(('conv2', block.conv2))
queue.append(['add_res', None])
def build_sampling(task_queue, net, is_decoder):
"""
Build the sampling part of a task queue
@param task_queue: the target task queue
@param net: the network
@param is_decoder: currently building decoder or encoder
"""
if is_decoder:
if sd_flag:
resblock2task(task_queue, net.mid.block_1)
attn2task(task_queue, net.mid.attn_1)
print(task_queue)
resblock2task(task_queue, net.mid.block_2)
resolution_iter = reversed(range(net.num_resolutions))
block_ids = net.num_res_blocks + 1
condition = 0
module = net.up
func_name = 'upsample'
else:
resblock2task(task_queue, net.mid_block.resnets[0])
attn2task(task_queue, net.mid_block.attentions[0])
resblock2task(task_queue, net.mid_block.resnets[1])
resolution_iter = (range(len(net.up_blocks))) # net.num_resolutions = 3
block_ids = 2 + 1
condition = len(net.up_blocks) - 1
module = net.up_blocks
func_name = 'upsamplers'
else:
if sd_flag:
resolution_iter = range(net.num_resolutions)
block_ids = net.num_res_blocks
condition = net.num_resolutions - 1
module = net.down
func_name = 'downsample'
else:
resolution_iter = range(len(net.down_blocks))
block_ids = 2
condition = len(net.down_blocks) - 1
module = net.down_blocks
func_name = 'downsamplers'
for i_level in resolution_iter:
for i_block in range(block_ids):
if sd_flag:
resblock2task(task_queue, module[i_level].block[i_block])
else:
resblock2task(task_queue, module[i_level].resnets[i_block])
if i_level != condition:
if sd_flag:
task_queue.append((func_name, getattr(module[i_level], func_name)))
else:
if is_decoder:
task_queue.append((func_name, module[i_level].upsamplers[0]))
else:
task_queue.append((func_name, module[i_level].downsamplers[0]))
if not is_decoder:
if sd_flag:
resblock2task(task_queue, net.mid.block_1)
attn2task(task_queue, net.mid.attn_1)
resblock2task(task_queue, net.mid.block_2)
else:
resblock2task(task_queue, net.mid_block.resnets[0])
attn2task(task_queue, net.mid_block.attentions[0])
resblock2task(task_queue, net.mid_block.resnets[1])
def build_task_queue(net, is_decoder):
"""
Build a single task queue for the encoder or decoder
@param net: the VAE decoder or encoder network
@param is_decoder: currently building decoder or encoder
@return: the task queue
"""
task_queue = []
task_queue.append(('conv_in', net.conv_in))
# construct the sampling part of the task queue
# because encoder and decoder share the same architecture, we extract the sampling part
build_sampling(task_queue, net, is_decoder)
if is_decoder and not sd_flag:
net.give_pre_end = False
net.tanh_out = False
if not is_decoder or not net.give_pre_end:
if sd_flag:
task_queue.append(('pre_norm', net.norm_out))
else:
task_queue.append(('pre_norm', net.conv_norm_out))
task_queue.append(('silu', inplace_nonlinearity))
task_queue.append(('conv_out', net.conv_out))
if is_decoder and net.tanh_out:
task_queue.append(('tanh', torch.tanh))
return task_queue
def clone_task_queue(task_queue):
"""
Clone a task queue
@param task_queue: the task queue to be cloned
@return: the cloned task queue
"""
return [[item for item in task] for task in task_queue]
def get_var_mean(input, num_groups, eps=1e-6):
"""
Get mean and var for group norm
"""
b, c = input.size(0), input.size(1)
channel_in_group = int(c/num_groups)
input_reshaped = input.contiguous().view(
1, int(b * num_groups), channel_in_group, *input.size()[2:])
var, mean = torch.var_mean(
input_reshaped, dim=[0, 2, 3, 4], unbiased=False)
return var, mean
def custom_group_norm(input, num_groups, mean, var, weight=None, bias=None, eps=1e-6):
"""
Custom group norm with fixed mean and var
@param input: input tensor
@param num_groups: number of groups. by default, num_groups = 32
@param mean: mean, must be pre-calculated by get_var_mean
@param var: var, must be pre-calculated by get_var_mean
@param weight: weight, should be fetched from the original group norm
@param bias: bias, should be fetched from the original group norm
@param eps: epsilon, by default, eps = 1e-6 to match the original group norm
@return: normalized tensor
"""
b, c = input.size(0), input.size(1)
channel_in_group = int(c/num_groups)
input_reshaped = input.contiguous().view(
1, int(b * num_groups), channel_in_group, *input.size()[2:])
out = F.batch_norm(input_reshaped, mean, var, weight=None, bias=None,
training=False, momentum=0, eps=eps)
out = out.view(b, c, *input.size()[2:])
# post affine transform
if weight is not None:
out *= weight.view(1, -1, 1, 1)
if bias is not None:
out += bias.view(1, -1, 1, 1)
return out
def crop_valid_region(x, input_bbox, target_bbox, is_decoder):
"""
Crop the valid region from the tile
@param x: input tile
@param input_bbox: original input bounding box
@param target_bbox: output bounding box
@param scale: scale factor
@return: cropped tile
"""
padded_bbox = [i * 8 if is_decoder else i//8 for i in input_bbox]
margin = [target_bbox[i] - padded_bbox[i] for i in range(4)]
return x[:, :, margin[2]:x.size(2)+margin[3], margin[0]:x.size(3)+margin[1]]
# ↓↓↓ https://github.com/Kahsolt/stable-diffusion-webui-vae-tile-infer ↓↓↓
def perfcount(fn):
def wrapper(*args, **kwargs):
ts = time()
if torch.cuda.is_available():
torch.cuda.reset_peak_memory_stats(devices.device)
devices.torch_gc()
gc.collect()
ret = fn(*args, **kwargs)
devices.torch_gc()
gc.collect()
if torch.cuda.is_available():
vram = torch.cuda.max_memory_allocated(devices.device) / 2**20
torch.cuda.reset_peak_memory_stats(devices.device)
print(
f'[Tiled VAE]: Done in {time() - ts:.3f}s, max VRAM alloc {vram:.3f} MB')
else:
print(f'[Tiled VAE]: Done in {time() - ts:.3f}s')
return ret
return wrapper
# copy end :)
class GroupNormParam:
def __init__(self):
self.var_list = []
self.mean_list = []
self.pixel_list = []
self.weight = None
self.bias = None
def add_tile(self, tile, layer):
var, mean = get_var_mean(tile, 32)
# For giant images, the variance can be larger than max float16
# In this case we create a copy to float32
if var.dtype == torch.float16 and var.isinf().any():
fp32_tile = tile.float()
var, mean = get_var_mean(fp32_tile, 32)
# ============= DEBUG: test for infinite =============
# if torch.isinf(var).any():
# print('var: ', var)
# ====================================================
self.var_list.append(var)
self.mean_list.append(mean)
self.pixel_list.append(
tile.shape[2]*tile.shape[3])
if hasattr(layer, 'weight'):
self.weight = layer.weight
self.bias = layer.bias
else:
self.weight = None
self.bias = None
def summary(self):
"""
summarize the mean and var and return a function
that apply group norm on each tile
"""
if len(self.var_list) == 0:
return None
var = torch.vstack(self.var_list)
mean = torch.vstack(self.mean_list)
max_value = max(self.pixel_list)
pixels = torch.tensor(
self.pixel_list, dtype=torch.float32, device=devices.device) / max_value
sum_pixels = torch.sum(pixels)
pixels = pixels.unsqueeze(
1) / sum_pixels
var = torch.sum(
var * pixels, dim=0)
mean = torch.sum(
mean * pixels, dim=0)
return lambda x: custom_group_norm(x, 32, mean, var, self.weight, self.bias)
@staticmethod
def from_tile(tile, norm):
"""
create a function from a single tile without summary
"""
var, mean = get_var_mean(tile, 32)
if var.dtype == torch.float16 and var.isinf().any():
fp32_tile = tile.float()
var, mean = get_var_mean(fp32_tile, 32)
# if it is a macbook, we need to convert back to float16
if var.device.type == 'mps':
# clamp to avoid overflow
var = torch.clamp(var, 0, 60000)
var = var.half()
mean = mean.half()
if hasattr(norm, 'weight'):
weight = norm.weight
bias = norm.bias
else:
weight = None
bias = None
def group_norm_func(x, mean=mean, var=var, weight=weight, bias=bias):
return custom_group_norm(x, 32, mean, var, weight, bias, 1e-6)
return group_norm_func
class VAEHook:
def __init__(self, net, tile_size, is_decoder, fast_decoder, fast_encoder, color_fix, to_gpu=False):
self.net = net # encoder | decoder
self.tile_size = tile_size
self.is_decoder = is_decoder
self.fast_mode = (fast_encoder and not is_decoder) or (
fast_decoder and is_decoder)
self.color_fix = color_fix and not is_decoder
self.to_gpu = to_gpu
self.pad = 11 if is_decoder else 32
def __call__(self, x):
B, C, H, W = x.shape
original_device = next(self.net.parameters()).device
try:
if self.to_gpu:
self.net.to(devices.get_optimal_device())
if max(H, W) <= self.pad * 2 + self.tile_size:
print("[Tiled VAE]: the input size is tiny and unnecessary to tile.")
return self.net.original_forward(x)
else:
return self.vae_tile_forward(x)
finally:
self.net.to(original_device)
def get_best_tile_size(self, lowerbound, upperbound):
"""
Get the best tile size for GPU memory
"""
divider = 32
while divider >= 2:
remainer = lowerbound % divider
if remainer == 0:
return lowerbound
candidate = lowerbound - remainer + divider
if candidate <= upperbound:
return candidate
divider //= 2
return lowerbound
def split_tiles(self, h, w):
"""
Tool function to split the image into tiles
@param h: height of the image
@param w: width of the image
@return: tile_input_bboxes, tile_output_bboxes
"""
tile_input_bboxes, tile_output_bboxes = [], []
tile_size = self.tile_size
pad = self.pad
num_height_tiles = math.ceil((h - 2 * pad) / tile_size)
num_width_tiles = math.ceil((w - 2 * pad) / tile_size)
# If any of the numbers are 0, we let it be 1
# This is to deal with long and thin images
num_height_tiles = max(num_height_tiles, 1)
num_width_tiles = max(num_width_tiles, 1)
# Suggestions from https://github.com/Kahsolt: auto shrink the tile size
real_tile_height = math.ceil((h - 2 * pad) / num_height_tiles)
real_tile_width = math.ceil((w - 2 * pad) / num_width_tiles)
real_tile_height = self.get_best_tile_size(real_tile_height, tile_size)
real_tile_width = self.get_best_tile_size(real_tile_width, tile_size)
print(f'[Tiled VAE]: split to {num_height_tiles}x{num_width_tiles} = {num_height_tiles*num_width_tiles} tiles. ' +
f'Optimal tile size {real_tile_width}x{real_tile_height}, original tile size {tile_size}x{tile_size}')
for i in range(num_height_tiles):
for j in range(num_width_tiles):
# bbox: [x1, x2, y1, y2]
# the padding is is unnessary for image borders. So we directly start from (32, 32)
input_bbox = [
pad + j * real_tile_width,
min(pad + (j + 1) * real_tile_width, w),
pad + i * real_tile_height,
min(pad + (i + 1) * real_tile_height, h),
]
# if the output bbox is close to the image boundary, we extend it to the image boundary
output_bbox = [
input_bbox[0] if input_bbox[0] > pad else 0,
input_bbox[1] if input_bbox[1] < w - pad else w,
input_bbox[2] if input_bbox[2] > pad else 0,
input_bbox[3] if input_bbox[3] < h - pad else h,
]
# scale to get the final output bbox
output_bbox = [x * 8 if self.is_decoder else x // 8 for x in output_bbox]
tile_output_bboxes.append(output_bbox)
# indistinguishable expand the input bbox by pad pixels
tile_input_bboxes.append([
max(0, input_bbox[0] - pad),
min(w, input_bbox[1] + pad),
max(0, input_bbox[2] - pad),
min(h, input_bbox[3] + pad),
])
return tile_input_bboxes, tile_output_bboxes
@torch.no_grad()
def estimate_group_norm(self, z, task_queue, color_fix):
device = z.device
tile = z
last_id = len(task_queue) - 1
while last_id >= 0 and task_queue[last_id][0] != 'pre_norm':
last_id -= 1
if last_id <= 0 or task_queue[last_id][0] != 'pre_norm':
raise ValueError('No group norm found in the task queue')
# estimate until the last group norm
for i in range(last_id + 1):
task = task_queue[i]
if task[0] == 'pre_norm':
group_norm_func = GroupNormParam.from_tile(tile, task[1])
task_queue[i] = ('apply_norm', group_norm_func)
if i == last_id:
return True
tile = group_norm_func(tile)
elif task[0] == 'store_res':
task_id = i + 1
while task_id < last_id and task_queue[task_id][0] != 'add_res':
task_id += 1
if task_id >= last_id:
continue
task_queue[task_id][1] = task[1](tile)
elif task[0] == 'add_res':
tile += task[1].to(device)
task[1] = None
elif color_fix and task[0] == 'downsample':
for j in range(i, last_id + 1):
if task_queue[j][0] == 'store_res':
task_queue[j] = ('store_res_cpu', task_queue[j][1])
return True
else:
tile = task[1](tile)
try:
devices.test_for_nans(tile, "vae")
except:
print(f'Nan detected in fast mode estimation. Fast mode disabled.')
return False
raise IndexError('Should not reach here')
@perfcount
@torch.no_grad()
def vae_tile_forward(self, z):
"""
Decode a latent vector z into an image in a tiled manner.
@param z: latent vector
@return: image
"""
device = next(self.net.parameters()).device
dtype = z.dtype
net = self.net
tile_size = self.tile_size
is_decoder = self.is_decoder
z = z.detach() # detach the input to avoid backprop
N, height, width = z.shape[0], z.shape[2], z.shape[3]
net.last_z_shape = z.shape
# Split the input into tiles and build a task queue for each tile
print(f'[Tiled VAE]: input_size: {z.shape}, tile_size: {tile_size}, padding: {self.pad}')
in_bboxes, out_bboxes = self.split_tiles(height, width)
# Prepare tiles by split the input latents
tiles = []
for input_bbox in in_bboxes:
tile = z[:, :, input_bbox[2]:input_bbox[3], input_bbox[0]:input_bbox[1]].cpu()
tiles.append(tile)
num_tiles = len(tiles)
num_completed = 0
# Build task queues
single_task_queue = build_task_queue(net, is_decoder)
#print(single_task_queue)
if self.fast_mode:
# Fast mode: downsample the input image to the tile size,
# then estimate the group norm parameters on the downsampled image
scale_factor = tile_size / max(height, width)
z = z.to(device)
downsampled_z = F.interpolate(z, scale_factor=scale_factor, mode='nearest-exact')
# use nearest-exact to keep statictics as close as possible
print(f'[Tiled VAE]: Fast mode enabled, estimating group norm parameters on {downsampled_z.shape[3]} x {downsampled_z.shape[2]} image')
# ======= Special thanks to @Kahsolt for distribution shift issue ======= #
# The downsampling will heavily distort its mean and std, so we need to recover it.
std_old, mean_old = torch.std_mean(z, dim=[0, 2, 3], keepdim=True)
std_new, mean_new = torch.std_mean(downsampled_z, dim=[0, 2, 3], keepdim=True)
downsampled_z = (downsampled_z - mean_new) / std_new * std_old + mean_old
del std_old, mean_old, std_new, mean_new
# occasionally the std_new is too small or too large, which exceeds the range of float16
# so we need to clamp it to max z's range.
downsampled_z = torch.clamp_(downsampled_z, min=z.min(), max=z.max())
estimate_task_queue = clone_task_queue(single_task_queue)
if self.estimate_group_norm(downsampled_z, estimate_task_queue, color_fix=self.color_fix):
single_task_queue = estimate_task_queue
del downsampled_z
task_queues = [clone_task_queue(single_task_queue) for _ in range(num_tiles)]
# Dummy result
result = None
result_approx = None
#try:
# with devices.autocast():
# result_approx = torch.cat([F.interpolate(cheap_approximation(x).unsqueeze(0), scale_factor=opt_f, mode='nearest-exact') for x in z], dim=0).cpu()
#except: pass
# Free memory of input latent tensor
del z
# Task queue execution
pbar = tqdm(total=num_tiles * len(task_queues[0]), desc=f"[Tiled VAE]: Executing {'Decoder' if is_decoder else 'Encoder'} Task Queue: ")
# execute the task back and forth when switch tiles so that we always
# keep one tile on the GPU to reduce unnecessary data transfer
forward = True
interrupted = False
#state.interrupted = interrupted
while True:
#if state.interrupted: interrupted = True ; break
group_norm_param = GroupNormParam()
for i in range(num_tiles) if forward else reversed(range(num_tiles)):
#if state.interrupted: interrupted = True ; break
tile = tiles[i].to(device)
input_bbox = in_bboxes[i]
task_queue = task_queues[i]
interrupted = False
while len(task_queue) > 0:
#if state.interrupted: interrupted = True ; break
# DEBUG: current task
# print('Running task: ', task_queue[0][0], ' on tile ', i, '/', num_tiles, ' with shape ', tile.shape)
task = task_queue.pop(0)
if task[0] == 'pre_norm':
group_norm_param.add_tile(tile, task[1])
break
elif task[0] == 'store_res' or task[0] == 'store_res_cpu':
task_id = 0
res = task[1](tile)
if not self.fast_mode or task[0] == 'store_res_cpu':
res = res.cpu()
while task_queue[task_id][0] != 'add_res':
task_id += 1
task_queue[task_id][1] = res
elif task[0] == 'add_res':
tile += task[1].to(device)
task[1] = None
else:
tile = task[1](tile)
#print(tiles[i].shape, tile.shape, task)
pbar.update(1)
if interrupted: break
# check for NaNs in the tile.
# If there are NaNs, we abort the process to save user's time
#devices.test_for_nans(tile, "vae")
#print(tiles[i].shape, tile.shape, i, num_tiles)
if len(task_queue) == 0:
tiles[i] = None
num_completed += 1
if result is None: # NOTE: dim C varies from different cases, can only be inited dynamically
result = torch.zeros((N, tile.shape[1], height * 8 if is_decoder else height // 8, width * 8 if is_decoder else width // 8), device=device, requires_grad=False)
result[:, :, out_bboxes[i][2]:out_bboxes[i][3], out_bboxes[i][0]:out_bboxes[i][1]] = crop_valid_region(tile, in_bboxes[i], out_bboxes[i], is_decoder)
del tile
elif i == num_tiles - 1 and forward:
forward = False
tiles[i] = tile
elif i == 0 and not forward:
forward = True
tiles[i] = tile
else:
tiles[i] = tile.cpu()
del tile
if interrupted: break
if num_completed == num_tiles: break
# insert the group norm task to the head of each task queue
group_norm_func = group_norm_param.summary()
if group_norm_func is not None:
for i in range(num_tiles):
task_queue = task_queues[i]
task_queue.insert(0, ('apply_norm', group_norm_func))
# Done!
pbar.close()
return result.to(dtype) if result is not None else result_approx.to(device)