Spaces:
Runtime error
Runtime error
File size: 4,327 Bytes
8fd2f2f |
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 |
import logging
from abc import ABC, abstractmethod
from typing import Dict, List, Literal, Optional, Tuple, Union
import torch
from einops import rearrange, repeat
from models.svd.sgm.util import append_dims, default
logpy = logging.getLogger(__name__)
class Guider(ABC):
@abstractmethod
def __call__(self, x: torch.Tensor, sigma: float) -> torch.Tensor:
pass
def prepare_inputs(
self, x: torch.Tensor, s: float, c: Dict, uc: Dict
) -> Tuple[torch.Tensor, float, Dict]:
pass
class VanillaCFG(Guider):
def __init__(self, scale: float):
self.scale = scale
def __call__(self, x: torch.Tensor, sigma: torch.Tensor) -> torch.Tensor:
x_u, x_c = x.chunk(2)
x_pred = x_u + self.scale * (x_c - x_u)
return x_pred
def prepare_inputs(self, x, s, c, uc):
c_out = dict()
for k in c:
if k in ["vector", "crossattn", "concat"]:
c_out[k] = torch.cat((uc[k], c[k]), 0)
else:
assert c[k] == uc[k]
c_out[k] = c[k]
return torch.cat([x] * 2), torch.cat([s] * 2), c_out
class IdentityGuider(Guider):
def __call__(self, x: torch.Tensor, sigma: float) -> torch.Tensor:
return x
def prepare_inputs(
self, x: torch.Tensor, s: float, c: Dict, uc: Dict
) -> Tuple[torch.Tensor, float, Dict]:
c_out = dict()
for k in c:
c_out[k] = c[k]
return x, s, c_out
class LinearPredictionGuider(Guider):
def __init__(
self,
max_scale: float,
num_frames: int,
min_scale: float = 1.0,
additional_cond_keys: Optional[Union[List[str], str]] = None,
):
self.min_scale = min_scale
self.max_scale = max_scale
self.num_frames = num_frames
self.scale = torch.linspace(min_scale, max_scale, num_frames).unsqueeze(0)
additional_cond_keys = default(additional_cond_keys, [])
if isinstance(additional_cond_keys, str):
additional_cond_keys = [additional_cond_keys]
self.additional_cond_keys = additional_cond_keys
def __call__(self, x: torch.Tensor, sigma: torch.Tensor) -> torch.Tensor:
x_u, x_c = x.chunk(2)
x_u = rearrange(x_u, "(b t) ... -> b t ...", t=self.num_frames)
x_c = rearrange(x_c, "(b t) ... -> b t ...", t=self.num_frames)
scale = repeat(self.scale, "1 t -> b t", b=x_u.shape[0])
scale = append_dims(scale, x_u.ndim).to(x_u.device)
scale = scale.to(x_u.dtype)
return rearrange(x_u + scale * (x_c - x_u), "b t ... -> (b t) ...")
def prepare_inputs(
self, x: torch.Tensor, s: torch.Tensor, c: dict, uc: dict
) -> Tuple[torch.Tensor, torch.Tensor, dict]:
c_out = dict()
for k in c:
if k in ["vector", "crossattn", "concat"] + self.additional_cond_keys:
c_out[k] = torch.cat((uc[k], c[k]), 0)
else:
assert c[k] == uc[k]
c_out[k] = c[k]
return torch.cat([x] * 2), torch.cat([s] * 2), c_out
class TrianglePredictionGuider(LinearPredictionGuider):
def __init__(
self,
max_scale: float,
num_frames: int,
min_scale: float = 1.0,
period: Union[float,List[float]] = 1.0,
period_fusing: Literal["mean", "multiply", "max"] = "max",
additional_cond_keys: Optional[Union[List[str], str]] = None,
):
super().__init__(max_scale, num_frames, min_scale, additional_cond_keys)
values = torch.linspace(0, 1, num_frames)
# Constructs a triangle wave
if isinstance(period, float):
period = [period]
scales = []
for p in period:
scales.append(self.triangle_wave(values, p))
if period_fusing == "mean":
scale = sum(scales) / len(period)
elif period_fusing == "multiply":
scale = torch.prod(torch.stack(scales), dim=0)
elif period_fusing == "max":
scale = torch.max(torch.stack(scales), dim=0).values
self.scale = (scale * (max_scale - min_scale) + min_scale).unsqueeze(0)
def triangle_wave(self, values: torch.Tensor, period) -> torch.Tensor:
return 2 * (values / period - torch.floor(values / period + 0.5)).abs()
|