video_diffusion/prompt_attention/ptp_utils.py

'''
utils code for image visualization
'''


# 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 PIL import Image
import cv2
from typing import Optional, Union, Tuple, List, Callable, Dict

import datetime


def text_under_image(image: np.ndarray, text: str, text_color: Tuple[int, int, int] = (0, 0, 0)):
    h, w, c = image.shape
    offset = int(h * .2)
    img = np.ones((h + offset, w, c), dtype=np.uint8) * 255
    font = cv2.FONT_HERSHEY_SIMPLEX
    # font = ImageFont.truetype("/usr/share/fonts/truetype/noto/NotoMono-Regular.ttf", font_size)
    img[:h] = image
    textsize = cv2.getTextSize(text, font, 1, 2)[0]
    text_x, text_y = (w - textsize[0]) // 2, h + offset - textsize[1] // 2
    cv2.putText(img, text, (text_x, text_y ), font, 1, text_color, 2)
    return img


def view_images(images, num_rows=1, offset_ratio=0.02, save_path=None):
    if type(images) is list:
        num_empty = len(images) % num_rows
    elif images.ndim == 4:
        num_empty = images.shape[0] % num_rows
    else:
        images = [images]
        num_empty = 0

    empty_images = np.ones(images[0].shape, dtype=np.uint8) * 255
    images = [image.astype(np.uint8) for image in images] + [empty_images] * num_empty
    num_items = len(images)

    h, w, c = images[0].shape
    offset = int(h * offset_ratio)
    num_cols = num_items // num_rows
    image_ = np.ones((h * num_rows + offset * (num_rows - 1),
                      w * num_cols + offset * (num_cols - 1), 3), dtype=np.uint8) * 255
    for i in range(num_rows):
        for j in range(num_cols):
            image_[i * (h + offset): i * (h + offset) + h:, j * (w + offset): j * (w + offset) + w] = images[
                i * num_cols + j]
    
    if save_path is not None:
        pil_img = Image.fromarray(image_)
        # now = datetime.datetime.now().strftime("%Y-%m-%dT%H-%M-%S")
        pil_img.save(f'{save_path}')
        #pil_img.save(f'{save_path}/{now}.png')
    # display(pil_img)



def register_attention_control_p2p_deprecated(model, controller):
    "Original code from prompt to prompt"
    def ca_forward(self, place_in_unet):
        to_out = self.to_out
        if type(to_out) is torch.nn.modules.container.ModuleList:
            to_out = self.to_out[0]
        else:
            to_out = self.to_out

        # def forward(x, encoder_hidden_states=None, attention_mask=None):
        def forward(hidden_states, encoder_hidden_states=None, attention_mask=None, **cross_attention_kwargs):
            batch_size, sequence_length, _ = hidden_states.shape
            attention_mask = self.prepare_attention_mask(attention_mask, sequence_length, batch_size)

            query = self.to_q(hidden_states)
            query = self.head_to_batch_dim(query)

            is_cross = encoder_hidden_states is not None
            encoder_hidden_states = encoder_hidden_states if encoder_hidden_states is not None else hidden_states
            key = self.to_k(encoder_hidden_states)
            value = self.to_v(encoder_hidden_states)
            key = self.head_to_batch_dim(key)
            value = self.head_to_batch_dim(value)

            attention_probs = self.get_attention_scores(query, key, attention_mask) # [16, 4096, 4096]
            attention_probs = controller(attention_probs, is_cross, place_in_unet)
            hidden_states = torch.bmm(attention_probs, value)
            hidden_states = self.batch_to_head_dim(hidden_states)

            # linear proj
            hidden_states = self.to_out[0](hidden_states)
            # dropout
            hidden_states = self.to_out[1](hidden_states)

            return hidden_states

        return forward

    class DummyController:

        def __call__(self, *args):
            return args[0]

        def __init__(self):
            self.num_att_layers = 0

    if controller is None:
        controller = DummyController()

    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):
    # Edit the alpha map during attention map editing
    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

import omegaconf
def get_time_words_attention_alpha(prompts, num_steps,
                                   cross_replace_steps: Union[float, Dict[str, Tuple[float, float]]],
                                   tokenizer, max_num_words=77):
    # Not understand
    if (type(cross_replace_steps) is not dict) and \
        (type(cross_replace_steps) is not omegaconf.dictconfig.DictConfig):
        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)
    return alpha_time_words