Spaces:
Runtime error
Runtime error
File size: 5,326 Bytes
5b953fb 3e4185b 5b953fb |
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 |
# Copyright 2022 Google LLC
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import numpy as np
import torch
from typing import Optional, Union, Tuple, Dict
def register_attention_control(model, controller):
def ca_forward(self, place_in_unet):
def forward(x, context=None, mask=None):
batch_size, sequence_length, dim = x.shape
h = self.heads
q = self.to_q(x)
is_cross = context is not None
context = context if is_cross else x
k = self.to_k(context)
v = self.to_v(context)
q = self.reshape_heads_to_batch_dim(q)
k = self.reshape_heads_to_batch_dim(k)
v = self.reshape_heads_to_batch_dim(v)
sim = torch.einsum("b i d, b j d -> b i j", q, k) * self.scale
if mask is not None:
mask = mask.reshape(batch_size, -1)
max_neg_value = -torch.finfo(sim.dtype).max
mask = mask[:, None, :].repeat(h, 1, 1)
sim.masked_fill_(~mask, max_neg_value)
# attention, what we cannot get enough of
attn = sim.softmax(dim=-1)
attn = controller(attn, is_cross, place_in_unet)
out = torch.einsum("b i j, b j d -> b i d", attn, v)
out = self.reshape_batch_dim_to_heads(out)
# TODO: Chen (new version of diffusers)
# return self.to_out(out)
# linear proj
out = self.to_out[0](out)
# dropout
out = self.to_out[1](out)
return out
return forward
def register_recr(net_, count, place_in_unet):
if net_.__class__.__name__ == 'CrossAttention':
net_.forward = ca_forward(net_, place_in_unet)
return count + 1
elif hasattr(net_, 'children'):
for net__ in net_.children():
count = register_recr(net__, count, place_in_unet)
return count
cross_att_count = 0
sub_nets = model.unet.named_children()
for net in sub_nets:
if "down" in net[0]:
cross_att_count += register_recr(net[1], 0, "down")
elif "up" in net[0]:
cross_att_count += register_recr(net[1], 0, "up")
elif "mid" in net[0]:
cross_att_count += register_recr(net[1], 0, "mid")
controller.num_att_layers = cross_att_count
def get_word_inds(text: str, word_place: int, tokenizer):
split_text = text.split(" ")
if type(word_place) is str:
word_place = [i for i, word in enumerate(split_text) if word_place == word]
elif type(word_place) is int:
word_place = [word_place]
out = []
if len(word_place) > 0:
words_encode = [tokenizer.decode([item]).strip("#") for item in tokenizer.encode(text)][1:-1]
cur_len, ptr = 0, 0
for i in range(len(words_encode)):
cur_len += len(words_encode[i])
if ptr in word_place:
out.append(i + 1)
if cur_len >= len(split_text[ptr]):
ptr += 1
cur_len = 0
return np.array(out)
def update_alpha_time_word(alpha, bounds: Union[float, Tuple[float, float]], prompt_ind: int, word_inds: Optional[torch.Tensor]=None):
if type(bounds) is float:
bounds = 0, bounds
start, end = int(bounds[0] * alpha.shape[0]), int(bounds[1] * alpha.shape[0])
if word_inds is None:
word_inds = torch.arange(alpha.shape[2])
alpha[: start, prompt_ind, word_inds] = 0
alpha[start: end, prompt_ind, word_inds] = 1
alpha[end:, prompt_ind, word_inds] = 0
return alpha
def get_time_words_attention_alpha(prompts, num_steps, cross_replace_steps: Union[float, Tuple[float, float], Dict[str, Tuple[float, float]]],
tokenizer, max_num_words=77):
if type(cross_replace_steps) is not dict:
cross_replace_steps = {"default_": cross_replace_steps}
if "default_" not in cross_replace_steps:
cross_replace_steps["default_"] = (0., 1.)
alpha_time_words = torch.zeros(num_steps + 1, len(prompts) - 1, max_num_words)
for i in range(len(prompts) - 1):
alpha_time_words = update_alpha_time_word(alpha_time_words, cross_replace_steps["default_"],
i)
for key, item in cross_replace_steps.items():
if key != "default_":
inds = [get_word_inds(prompts[i], key, tokenizer) for i in range(1, len(prompts))]
for i, ind in enumerate(inds):
if len(ind) > 0:
alpha_time_words = update_alpha_time_word(alpha_time_words, item, i, ind)
alpha_time_words = alpha_time_words.reshape(num_steps + 1, len(prompts) - 1, 1, 1, max_num_words) # time, batch, heads, pixels, words
return alpha_time_words
|