File size: 10,071 Bytes
8d14048
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
b2d1afb
8d14048
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
# 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.

from typing import Optional, Union, Tuple, List, Callable, Dict
import torch
import torch.nn.functional as nnf
import numpy as np
import abc
from . import ptp_utils
from . import seq_aligner
from transformers import CLIPTextModel, CLIPTokenizer

pretrained_model_path = "checkpoints/CompVis/stable-diffusion-v1-4/"
ldm_stable = None
tokenizer = CLIPTokenizer.from_pretrained(pretrained_model_path, subfolder="tokenizer")

LOW_RESOURCE = False 
NUM_DIFFUSION_STEPS = 50
GUIDANCE_SCALE = 7.5
MAX_NUM_WORDS = 77
device = torch.device('cuda')


class LocalBlend:

    def __call__(self, x_t, attention_store):
        k = 1
        maps = attention_store["down_cross"][2:4] + attention_store["up_cross"][:3]
        maps = [item.reshape(self.alpha_layers.shape[0], -1, 1, 16, 16, MAX_NUM_WORDS) for item in maps]
        maps = torch.cat(maps, dim=1)
        maps = (maps * self.alpha_layers).sum(-1).mean(1)
        mask = nnf.max_pool2d(maps, (k * 2 + 1, k * 2 +1), (1, 1), padding=(k, k))
        mask = nnf.interpolate(mask, size=(x_t.shape[2:]))
        mask = mask / mask.max(2, keepdims=True)[0].max(3, keepdims=True)[0]
        mask = mask.gt(self.threshold)
        mask = (mask[:1] + mask[1:]).float()
        x_t = x_t[:1] + mask * (x_t - x_t[:1])
        return x_t
       
    # def __init__(self, prompts: List[str], words: [List[List[str]]], threshold=.3):
    def __init__(self, prompts: List[str], words: List[List[str]], threshold=.3):
        alpha_layers = torch.zeros(len(prompts),  1, 1, 1, 1, MAX_NUM_WORDS)
        for i, (prompt, words_) in enumerate(zip(prompts, words)):
            if type(words_) is str:
                words_ = [words_]
            for word in words_:
                ind = ptp_utils.get_word_inds(prompt, word, tokenizer)
                alpha_layers[i, :, :, :, :, ind] = 1
        self.alpha_layers = alpha_layers.to(device)
        self.threshold = threshold


class AttentionControl(abc.ABC):
    
    def step_callback(self, x_t):
        return x_t
    
    def between_steps(self):
        return
    
    @property
    def num_uncond_att_layers(self):
        return self.num_att_layers if LOW_RESOURCE else 0
    
    @abc.abstractmethod
    def forward (self, attn, is_cross: bool, place_in_unet: str):
        raise NotImplementedError

    def __call__(self, attn, is_cross: bool, place_in_unet: str):
        if self.cur_att_layer >= self.num_uncond_att_layers:
            if LOW_RESOURCE:
                attn = self.forward(attn, is_cross, place_in_unet)
            else:
                h = attn.shape[0]
                attn[h // 2:] = self.forward(attn[h // 2:], is_cross, place_in_unet)
        self.cur_att_layer += 1
        if self.cur_att_layer == self.num_att_layers + self.num_uncond_att_layers:
            self.cur_att_layer = 0
            self.cur_step += 1
            self.between_steps()
        return attn
    
    def reset(self):
        self.cur_step = 0
        self.cur_att_layer = 0

    def __init__(self):
        self.cur_step = 0
        self.num_att_layers = -1
        self.cur_att_layer = 0

class EmptyControl(AttentionControl):
    
    def forward (self, attn, is_cross: bool, place_in_unet: str):
        return attn
    
    
class AttentionStore(AttentionControl):

    @staticmethod
    def get_empty_store():
        return {"down_cross": [], "mid_cross": [], "up_cross": [],
                "down_self": [],  "mid_self": [],  "up_self": []}

    def forward(self, attn, is_cross: bool, place_in_unet: str):
        key = f"{place_in_unet}_{'cross' if is_cross else 'self'}"
        if attn.shape[1] <= 32 ** 2:  # avoid memory overhead
            self.step_store[key].append(attn)
        return attn

    def between_steps(self):
        if len(self.attention_store) == 0:
            self.attention_store = self.step_store
        else:
            for key in self.attention_store:
                for i in range(len(self.attention_store[key])):
                    self.attention_store[key][i] += self.step_store[key][i]
        self.step_store = self.get_empty_store()

    def get_average_attention(self):
        average_attention = {key: [item / self.cur_step for item in self.attention_store[key]] for key in self.attention_store}
        return average_attention


    def reset(self):
        super(AttentionStore, self).reset()
        self.step_store = self.get_empty_store()
        self.attention_store = {}

    def __init__(self):
        super(AttentionStore, self).__init__()
        self.step_store = self.get_empty_store()
        self.attention_store = {}

        
class AttentionControlEdit(AttentionStore, abc.ABC):
    
    def step_callback(self, x_t):
        if self.local_blend is not None:
            x_t = self.local_blend(x_t, self.attention_store)
        return x_t
        
    def replace_self_attention(self, attn_base, att_replace):
        if att_replace.shape[2] <= 16 ** 2:
            return attn_base.unsqueeze(0).expand(att_replace.shape[0], *attn_base.shape)
        else:
            return att_replace
    
    @abc.abstractmethod
    def replace_cross_attention(self, attn_base, att_replace):
        raise NotImplementedError
    
    def forward(self, attn, is_cross: bool, place_in_unet: str):
        super(AttentionControlEdit, self).forward(attn, is_cross, place_in_unet)
        if is_cross or (self.num_self_replace[0] <= self.cur_step < self.num_self_replace[1]):
            h = attn.shape[0] // (self.batch_size)
            attn = attn.reshape(self.batch_size, h, *attn.shape[1:])
            attn_base, attn_repalce = attn[0], attn[1:]
            if is_cross:
                alpha_words = self.cross_replace_alpha[self.cur_step]
                attn_repalce_new = self.replace_cross_attention(attn_base, attn_repalce) * alpha_words + (1 - alpha_words) * attn_repalce
                attn[1:] = attn_repalce_new
            else:
                attn[1:] = self.replace_self_attention(attn_base, attn_repalce)
            attn = attn.reshape(self.batch_size * h, *attn.shape[2:])
        return attn
    
    def __init__(self, prompts, num_steps: int,
                 cross_replace_steps: Union[float, Tuple[float, float], Dict[str, Tuple[float, float]]],
                 self_replace_steps: Union[float, Tuple[float, float]],
                 local_blend: Optional[LocalBlend]):
        super(AttentionControlEdit, self).__init__()
        self.batch_size = len(prompts)
        self.cross_replace_alpha = ptp_utils.get_time_words_attention_alpha(prompts, num_steps, cross_replace_steps, tokenizer).to(device)
        if type(self_replace_steps) is float:
            self_replace_steps = 0, self_replace_steps
        self.num_self_replace = int(num_steps * self_replace_steps[0]), int(num_steps * self_replace_steps[1])
        self.local_blend = local_blend


class AttentionReplace(AttentionControlEdit):

    def replace_cross_attention(self, attn_base, att_replace):
        return torch.einsum('hpw,bwn->bhpn', attn_base, self.mapper)
      
    def __init__(self, prompts, num_steps: int, cross_replace_steps: float, self_replace_steps: float,
                 local_blend: Optional[LocalBlend] = None):
        super(AttentionReplace, self).__init__(prompts, num_steps, cross_replace_steps, self_replace_steps, local_blend)
        self.mapper = seq_aligner.get_replacement_mapper(prompts, tokenizer).to(device)


class AttentionRefine(AttentionControlEdit):

    def replace_cross_attention(self, attn_base, att_replace):
        attn_base_replace = attn_base[:, :, self.mapper].permute(2, 0, 1, 3)
        attn_replace = attn_base_replace * self.alphas + att_replace * (1 - self.alphas)
        return attn_replace

    def __init__(self, prompts, num_steps: int, cross_replace_steps: float, self_replace_steps: float,
                 local_blend: Optional[LocalBlend] = None):
        super(AttentionRefine, self).__init__(prompts, num_steps, cross_replace_steps, self_replace_steps, local_blend)
        self.mapper, alphas = seq_aligner.get_refinement_mapper(prompts, tokenizer)
        self.mapper, alphas = self.mapper.to(device), alphas.to(device)
        self.alphas = alphas.reshape(alphas.shape[0], 1, 1, alphas.shape[1])


class AttentionReweight(AttentionControlEdit):

    def replace_cross_attention(self, attn_base, att_replace):
        if self.prev_controller is not None:
            attn_base = self.prev_controller.replace_cross_attention(attn_base, att_replace)
        attn_replace = attn_base[None, :, :, :] * self.equalizer[:, None, None, :]
        return attn_replace

    def __init__(self, prompts, num_steps: int, cross_replace_steps: float, self_replace_steps: float, equalizer,
                local_blend: Optional[LocalBlend] = None, controller: Optional[AttentionControlEdit] = None):
        super(AttentionReweight, self).__init__(prompts, num_steps, cross_replace_steps, self_replace_steps, local_blend)
        self.equalizer = equalizer.to(device)
        self.prev_controller = controller


def get_equalizer(text: str, word_select: Union[int, Tuple[int, ...]], values: Union[List[float],
                  Tuple[float, ...]]):
    if type(word_select) is int or type(word_select) is str:
        word_select = (word_select,)
    equalizer = torch.ones(len(values), 77)
    values = torch.tensor(values, dtype=torch.float32)
    for word in word_select:
        inds = ptp_utils.get_word_inds(text, word, tokenizer)
        equalizer[:, inds] = values
    return equalizer