Spaces:
Paused
Paused
File size: 11,214 Bytes
26557da |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 |
import torch
from einops import rearrange, repeat
class TileWorker:
def __init__(self):
pass
def mask(self, height, width, border_width):
# Create a mask with shape (height, width).
# The centre area is filled with 1, and the border line is filled with values in range (0, 1].
x = torch.arange(height).repeat(width, 1).T
y = torch.arange(width).repeat(height, 1)
mask = torch.stack([x + 1, height - x, y + 1, width - y]).min(dim=0).values
mask = (mask / border_width).clip(0, 1)
return mask
def tile(self, model_input, tile_size, tile_stride, tile_device, tile_dtype):
# Convert a tensor (b, c, h, w) to (b, c, tile_size, tile_size, tile_num)
batch_size, channel, _, _ = model_input.shape
model_input = model_input.to(device=tile_device, dtype=tile_dtype)
unfold_operator = torch.nn.Unfold(
kernel_size=(tile_size, tile_size), stride=(tile_stride, tile_stride)
)
model_input = unfold_operator(model_input)
model_input = model_input.view((batch_size, channel, tile_size, tile_size, -1))
return model_input
def tiled_inference(
self,
forward_fn,
model_input,
tile_batch_size,
inference_device,
inference_dtype,
tile_device,
tile_dtype,
):
# Call y=forward_fn(x) for each tile
tile_num = model_input.shape[-1]
model_output_stack = []
for tile_id in range(0, tile_num, tile_batch_size):
# process input
tile_id_ = min(tile_id + tile_batch_size, tile_num)
x = model_input[:, :, :, :, tile_id:tile_id_]
x = x.to(device=inference_device, dtype=inference_dtype)
x = rearrange(x, "b c h w n -> (n b) c h w")
# process output
y = forward_fn(x)
y = rearrange(y, "(n b) c h w -> b c h w n", n=tile_id_ - tile_id)
y = y.to(device=tile_device, dtype=tile_dtype)
model_output_stack.append(y)
model_output = torch.concat(model_output_stack, dim=-1)
return model_output
def io_scale(self, model_output, tile_size):
# Determine the size modification happened in forward_fn
# We only consider the same scale on height and width.
io_scale = model_output.shape[2] / tile_size
return io_scale
def untile(
self,
model_output,
height,
width,
tile_size,
tile_stride,
border_width,
tile_device,
tile_dtype,
):
# The reversed function of tile
mask = self.mask(tile_size, tile_size, border_width)
mask = mask.to(device=tile_device, dtype=tile_dtype)
mask = rearrange(mask, "h w -> 1 1 h w 1")
model_output = model_output * mask
fold_operator = torch.nn.Fold(
output_size=(height, width),
kernel_size=(tile_size, tile_size),
stride=(tile_stride, tile_stride),
)
mask = repeat(mask[0, 0, :, :, 0], "h w -> 1 (h w) n", n=model_output.shape[-1])
model_output = rearrange(model_output, "b c h w n -> b (c h w) n")
model_output = fold_operator(model_output) / fold_operator(mask)
return model_output
def tiled_forward(
self,
forward_fn,
model_input,
tile_size,
tile_stride,
tile_batch_size=1,
tile_device="cpu",
tile_dtype=torch.float32,
border_width=None,
):
# Prepare
inference_device, inference_dtype = model_input.device, model_input.dtype
height, width = model_input.shape[2], model_input.shape[3]
border_width = int(tile_stride * 0.5) if border_width is None else border_width
# tile
model_input = self.tile(
model_input, tile_size, tile_stride, tile_device, tile_dtype
)
# inference
model_output = self.tiled_inference(
forward_fn,
model_input,
tile_batch_size,
inference_device,
inference_dtype,
tile_device,
tile_dtype,
)
# resize
io_scale = self.io_scale(model_output, tile_size)
height, width = int(height * io_scale), int(width * io_scale)
tile_size, tile_stride = int(tile_size * io_scale), int(tile_stride * io_scale)
border_width = int(border_width * io_scale)
# untile
model_output = self.untile(
model_output,
height,
width,
tile_size,
tile_stride,
border_width,
tile_device,
tile_dtype,
)
# Done!
model_output = model_output.to(device=inference_device, dtype=inference_dtype)
return model_output
class FastTileWorker:
def __init__(self):
pass
def build_mask(self, data, is_bound):
_, _, H, W = data.shape
h = repeat(torch.arange(H), "H -> H W", H=H, W=W)
w = repeat(torch.arange(W), "W -> H W", H=H, W=W)
border_width = (H + W) // 4
pad = torch.ones_like(h) * border_width
mask = (
torch.stack(
[
pad if is_bound[0] else h + 1,
pad if is_bound[1] else H - h,
pad if is_bound[2] else w + 1,
pad if is_bound[3] else W - w,
]
)
.min(dim=0)
.values
)
mask = mask.clip(1, border_width)
mask = (mask / border_width).to(dtype=data.dtype, device=data.device)
mask = rearrange(mask, "H W -> 1 H W")
return mask
def tiled_forward(
self,
forward_fn,
model_input,
tile_size,
tile_stride,
tile_device="cpu",
tile_dtype=torch.float32,
border_width=None,
):
# Prepare
B, C, H, W = model_input.shape
border_width = int(tile_stride * 0.5) if border_width is None else border_width
weight = torch.zeros((1, 1, H, W), dtype=tile_dtype, device=tile_device)
values = torch.zeros((B, C, H, W), dtype=tile_dtype, device=tile_device)
# Split tasks
tasks = []
for h in range(0, H, tile_stride):
for w in range(0, W, tile_stride):
if (h - tile_stride >= 0 and h - tile_stride + tile_size >= H) or (
w - tile_stride >= 0 and w - tile_stride + tile_size >= W
):
continue
h_, w_ = h + tile_size, w + tile_size
if h_ > H:
h, h_ = H - tile_size, H
if w_ > W:
w, w_ = W - tile_size, W
tasks.append((h, h_, w, w_))
# Run
for hl, hr, wl, wr in tasks:
# Forward
hidden_states_batch = forward_fn(hl, hr, wl, wr).to(
dtype=tile_dtype, device=tile_device
)
mask = self.build_mask(
hidden_states_batch, is_bound=(hl == 0, hr >= H, wl == 0, wr >= W)
)
values[:, :, hl:hr, wl:wr] += hidden_states_batch * mask
weight[:, :, hl:hr, wl:wr] += mask
values /= weight
return values
class TileWorker2Dto3D:
"""
Process 3D tensors, but only enable TileWorker on 2D.
"""
def __init__(self):
pass
def build_mask(self, T, H, W, dtype, device, is_bound, border_width):
t = repeat(torch.arange(T), "T -> T H W", T=T, H=H, W=W)
h = repeat(torch.arange(H), "H -> T H W", T=T, H=H, W=W)
w = repeat(torch.arange(W), "W -> T H W", T=T, H=H, W=W)
border_width = (H + W) // 4 if border_width is None else border_width
pad = torch.ones_like(h) * border_width
mask = (
torch.stack(
[
pad if is_bound[0] else t + 1,
pad if is_bound[1] else T - t,
pad if is_bound[2] else h + 1,
pad if is_bound[3] else H - h,
pad if is_bound[4] else w + 1,
pad if is_bound[5] else W - w,
]
)
.min(dim=0)
.values
)
mask = mask.clip(1, border_width)
mask = (mask / border_width).to(dtype=dtype, device=device)
mask = rearrange(mask, "T H W -> 1 1 T H W")
return mask
def tiled_forward(
self,
forward_fn,
model_input,
tile_size,
tile_stride,
tile_device="cpu",
tile_dtype=torch.float32,
computation_device="cuda",
computation_dtype=torch.float32,
border_width=None,
scales=[1, 1, 1, 1],
progress_bar=lambda x: x,
):
B, C, T, H, W = model_input.shape
scale_C, scale_T, scale_H, scale_W = scales
tile_size_H, tile_size_W = tile_size
tile_stride_H, tile_stride_W = tile_stride
value = torch.zeros(
(B, int(C * scale_C), int(T * scale_T), int(H * scale_H), int(W * scale_W)),
dtype=tile_dtype,
device=tile_device,
)
weight = torch.zeros(
(1, 1, int(T * scale_T), int(H * scale_H), int(W * scale_W)),
dtype=tile_dtype,
device=tile_device,
)
# Split tasks
tasks = []
for h in range(0, H, tile_stride_H):
for w in range(0, W, tile_stride_W):
if (
h - tile_stride_H >= 0 and h - tile_stride_H + tile_size_H >= H
) or (w - tile_stride_W >= 0 and w - tile_stride_W + tile_size_W >= W):
continue
h_, w_ = h + tile_size_H, w + tile_size_W
if h_ > H:
h, h_ = max(H - tile_size_H, 0), H
if w_ > W:
w, w_ = max(W - tile_size_W, 0), W
tasks.append((h, h_, w, w_))
# Run
for hl, hr, wl, wr in progress_bar(tasks):
mask = self.build_mask(
int(T * scale_T),
int((hr - hl) * scale_H),
int((wr - wl) * scale_W),
tile_dtype,
tile_device,
is_bound=(True, True, hl == 0, hr >= H, wl == 0, wr >= W),
border_width=border_width,
)
grid_input = model_input[:, :, :, hl:hr, wl:wr].to(
dtype=computation_dtype, device=computation_device
)
grid_output = forward_fn(grid_input).to(
dtype=tile_dtype, device=tile_device
)
value[
:,
:,
:,
int(hl * scale_H) : int(hr * scale_H),
int(wl * scale_W) : int(wr * scale_W),
] += grid_output * mask
weight[
:,
:,
:,
int(hl * scale_H) : int(hr * scale_H),
int(wl * scale_W) : int(wr * scale_W),
] += mask
value = value / weight
return value
|