File size: 7,920 Bytes
64cc34d
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
import os
import math
import numpy as np
import matplotlib.pyplot as plt
import torch


class Conceptrol:
    def __init__(self, config):
        if "name" not in config:
            raise KeyError("name has to be provided as 'conceptrol' or 'ominicontrol'")

        name = config["name"]
        if name not in ["conceptrol", "ominicontrol"]:
            raise ValueError(
                f"Name must be one of ['conceptrol', 'ominicontrol'], got {name}"
            )

        try:
            log_attn_map = config["log_attn_map"]
        except KeyError:
            log_attn_map = False

        # static
        self.NUM_BLOCKS = 19  # this is fixed for FLUX
        self.M = 512  # num of text tokens, fixed for FLUX
        self.N = 1024  # num of latent / image condtion tokens, fixed for FLUX
        self.EP = -10e6
        self.CONCEPT_BLOCK_IDX = 18

        # fixed during one generation
        self.name = name

        # variable during one generation
        self.textual_concept_mask = None
        self.forward_count = 0

        # log out for visualization
        if log_attn_map:
            self.attn_maps = {"latent_to_concept": [], "latent_to_image": []}

    def __call__(
        self,
        query: torch.FloatTensor,
        key: torch.FloatTensor,
        attention_mask: torch.Tensor,
        c_factor: float = 1.0,
    ) -> torch.Tensor:

        if not hasattr(self, "textual_concept_idx"):
            raise AttributeError(
                "textual_concept_idx must be registered before calling Conceptrol"
            )

        # Skip computation for ominicontrol
        if self.name == "ominicontrol":
            scale_factor = 1 / math.sqrt(query.size(-1))
            attention_weight = (
                query @ key.transpose(-2, -1) * scale_factor + attention_mask
            )
            attention_probs = torch.softmax(
                attention_weight, dim=-1
            )  # [B, H, M+2N, M+2N]
            return attention_probs

        if not self.textual_concept_idx[0] < self.textual_concept_idx[1]:
            raise ValueError(
                f"register_idx[0] must be less than register_idx[1], "
                f"got {self.textual_concept_idx[0]} >= {self.textual_concept_idx[1]}"
            )

        ### Reset attention mask predefined in ominicontrol
        attention_mask = torch.zeros_like(attention_mask)
        bias = torch.log(c_factor[0])
        # attention of image condition to latent
        attention_mask[-self.N :, self.M : -self.N] = bias
        # attention of latent to image condition
        attention_mask[self.M : -self.N, -self.N :] = bias

        # attention of textual concept to image condition
        attention_mask[
            self.textual_concept_idx[0] : self.textual_concept_idx[1], -self.N :
        ] = bias
        # attention of other words to image condition (set as negative inf)
        attention_mask[: self.textual_concept_idx[0], -self.N :] = self.EP
        attention_mask[self.textual_concept_idx[1] : self.M, -self.N :] = self.EP

        # If there is no textual_concept_mask, it means currently in layers previous to the first concept-specific block
        if self.textual_concept_mask is None:
            self.textual_concept_mask = (
                torch.zeros_like(attention_mask).unsqueeze(0).unsqueeze(0)
            )

        ### Compute attention
        scale_factor = 1 / math.sqrt(query.size(-1))
        attention_weight = (
            query @ key.transpose(-2, -1) * scale_factor
            + attention_mask
            + self.textual_concept_mask
        )
        # [B, H, M+2N, M+2N]
        attention_probs = torch.softmax(attention_weight, dim=-1)

        ### Extract textual concept mask if it's concept-specific block
        is_concept_block = (
            self.forward_count % self.NUM_BLOCKS == self.CONCEPT_BLOCK_IDX
        )
        if is_concept_block:
            # Shape: [B, H, N, S], where S is the token numbers of the subject
            textual_concept_mask_local = attention_probs[
                :,
                :,
                self.M : -self.N,
                self.textual_concept_idx[0] : self.textual_concept_idx[1],
            ]
            # Consider the ratio within context of text
            textual_concept_mask_local = textual_concept_mask_local / torch.sum(
                attention_probs[:, :, self.M : -self.N, : self.M], dim=-1, keepdim=True
            )
            # Average over words and head, Shape: [B, 1, N, 1]
            textual_concept_mask_local = torch.mean(
                textual_concept_mask_local, dim=(-1, 1), keepdim=True
            )
            # Normalize to average as 1
            textual_concept_mask_local = textual_concept_mask_local / torch.mean(
                textual_concept_mask_local, dim=-2, keepdim=True
            )

            self.textual_concept_mask = (
                torch.zeros_like(attention_mask).unsqueeze(0).unsqueeze(0)
            )
            # log(A) in the paper
            self.textual_concept_mask[:, :, self.M : -self.N, -self.N :] = torch.log(
                textual_concept_mask_local
            )

        self.forward_count += 1

        return attention_probs

    def register(self, textual_concept_idx):
        self.textual_concept_idx = textual_concept_idx

    def visualize_attn_map(self, config_name: str, subject: str):
        global global_concept_mask
        global forward_count

        save_dir = f"attn_maps/{config_name}/{subject}"
        if not os.path.exists(save_dir):
            os.makedirs(save_dir)
        for attn_map_name, attn_maps in self.attn_maps.items():
            if "token_to_token" in attn_map_name:
                continue
            plt.figure()

            rows, cols = 8, 19
            fig, axes = plt.subplots(
                rows, cols, figsize=(64 * cols / 100, 64 * rows / 100)
            )
            fig.subplots_adjust(
                wspace=0.1, hspace=0.1
            )  # Adjust spacing between subplots

            # Plot each array in the list on the grid
            for i, ax in enumerate(axes.flatten()):
                if i < len(attn_maps):  # Only plot existing arrays
                    attn_map = attn_maps[i] / np.amax(attn_maps[i])
                    ax.imshow(attn_map, cmap="viridis")
                    ax.axis("off")  # Turn off axes for clarity
                else:
                    ax.axis("off")  # Turn off unused subplots

            fig.set_size_inches(64 * cols / 100, 64 * rows / 100)
            save_path = os.path.join(save_dir, f"{attn_map_name}.jpg")
            plt.savefig(save_path)
            plt.close()

        for attn_map_name, attn_maps in self.attn_maps.items():
            if "token_to_token" not in attn_map_name:
                continue
            plt.figure()

            rows, cols = 8, 19
            fig, axes = plt.subplots(
                rows, cols, figsize=(2560 * cols / 100, 2560 * rows / 100)
            )
            fig.subplots_adjust(
                wspace=0.1, hspace=0.1
            )  # Adjust spacing between subplots

            # Plot each array in the list on the grid
            for i, ax in enumerate(axes.flatten()):
                if i < len(attn_maps):  # Only plot existing arrays
                    attn_map = attn_maps[i] / np.amax(attn_maps[i])
                    ax.imshow(attn_map, cmap="viridis")
                    ax.axis("off")  # Turn off axes for clarity
                else:
                    ax.axis("off")  # Turn off unused subplots

            fig.set_size_inches(64 * cols / 100, 64 * rows / 100)
            save_path = os.path.join(save_dir, f"{attn_map_name}.jpg")
            plt.savefig(save_path)
            plt.close()

        for attn_map_name in self.attn_maps.keys():
            self.attn_maps[attn_map_name] = []
        global_concept_mask = None
        forward_count = 0