utils.py

import numpy as np
import torch
from PIL import Image, ImageDraw, ImageFont
import cv2
from typing import Optional, Union, Tuple, List, Callable, Dict
from IPython.display import display
from tqdm.notebook import tqdm
from einops import rearrange, repeat


class ScoreParams:

    def __init__(self, gap, match, mismatch):
        self.gap = gap
        self.match = match
        self.mismatch = mismatch

    def mis_match_char(self, x, y):
        if x != y:
            return self.mismatch
        else:
            return self.match


def get_matrix(size_x, size_y, gap):
    matrix = np.zeros((size_x + 1, size_y + 1), dtype=np.int32)
    matrix[0, 1:] = (np.arange(size_y) + 1) * gap
    matrix[1:, 0] = (np.arange(size_x) + 1) * gap
    return matrix


def get_traceback_matrix(size_x, size_y):
    matrix = np.zeros((size_x + 1, size_y +1), dtype=np.int32)
    matrix[0, 1:] = 1
    matrix[1:, 0] = 2
    matrix[0, 0] = 4
    return matrix


def global_align(x, y, score):
    matrix = get_matrix(len(x), len(y), score.gap)
    trace_back = get_traceback_matrix(len(x), len(y))
    for i in range(1, len(x) + 1):
        for j in range(1, len(y) + 1):
            left = matrix[i, j - 1] + score.gap
            up = matrix[i - 1, j] + score.gap
            diag = matrix[i - 1, j - 1] + score.mis_match_char(x[i - 1], y[j - 1])
            matrix[i, j] = max(left, up, diag)
            if matrix[i, j] == left:
                trace_back[i, j] = 1
            elif matrix[i, j] == up:
                trace_back[i, j] = 2
            else:
                trace_back[i, j] = 3
    return matrix, trace_back


def get_aligned_sequences(x, y, trace_back):
    x_seq = []
    y_seq = []
    i = len(x)
    j = len(y)
    mapper_y_to_x = []
    while i > 0 or j > 0:
        if trace_back[i, j] == 3:
            x_seq.append(x[i-1])
            y_seq.append(y[j-1])
            i = i-1
            j = j-1
            mapper_y_to_x.append((j, i))
        elif trace_back[i][j] == 1:
            x_seq.append('-')
            y_seq.append(y[j-1])
            j = j-1
            mapper_y_to_x.append((j, -1))
        elif trace_back[i][j] == 2:
            x_seq.append(x[i-1])
            y_seq.append('-')
            i = i-1
        elif trace_back[i][j] == 4:
            break
    mapper_y_to_x.reverse()
    return x_seq, y_seq, torch.tensor(mapper_y_to_x, dtype=torch.int64)


def get_mapper(x: str, y: str, tokenizer, max_len=77):
    x_seq = tokenizer.encode(x)
    y_seq = tokenizer.encode(y)
    score = ScoreParams(0, 1, -1)
    matrix, trace_back = global_align(x_seq, y_seq, score)
    mapper_base = get_aligned_sequences(x_seq, y_seq, trace_back)[-1]
    alphas = torch.ones(max_len)
    alphas[: mapper_base.shape[0]] = mapper_base[:, 1].ne(-1).float()
    mapper = torch.zeros(max_len, dtype=torch.int64)
    mapper[:mapper_base.shape[0]] = mapper_base[:, 1]
    mapper[mapper_base.shape[0]:] = len(y_seq) + torch.arange(max_len - len(y_seq))
    return mapper, alphas


def get_refinement_mapper(prompts, tokenizer, max_len=77):
    x_seq = prompts[0]
    mappers, alphas = [], []
    for i in range(1, len(prompts)):
        mapper, alpha = get_mapper(x_seq, prompts[i], tokenizer, max_len)
        mappers.append(mapper)
        alphas.append(alpha)
    return torch.stack(mappers), torch.stack(alphas)


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 diffusion_step(model, controller, latents, context, t, guidance_scale, low_resource=False):
    if low_resource:
        noise_pred_uncond = model.unet(latents, t, encoder_hidden_states=context[0])["sample"]
        noise_prediction_text = model.unet(latents, t, encoder_hidden_states=context[1])["sample"]
    else:
        latents_input = torch.cat([latents] * 2)
        noise_pred = model.unet(latents_input, t, encoder_hidden_states=context)["sample"]
        noise_pred_uncond, noise_prediction_text = noise_pred.chunk(2)
    noise_pred = noise_pred_uncond + guidance_scale * (noise_prediction_text - noise_pred_uncond)
    latents = model.scheduler.step(noise_pred, t, latents)["prev_sample"]

    import pdb
    pdb.set_trace()
    latents = controller.step_callback(latents)
    return latents


def latent2image(vae, latents):
    latents = 1 / 0.18215 * latents
    image = vae.decode(latents)['sample']
    image = (image / 2 + 0.5).clamp(0, 1)
    image = image.cpu().permute(0, 2, 3, 1).numpy()
    image = (image * 255).astype(np.uint8)
    return image


def init_latent(latent, model, height, width, generator, batch_size):
    if latent is None:
        latent = torch.randn(
            (1, model.unet.in_channels, height // 8, width // 8),
            generator=generator,
        )
    breakpoint()
    latents = latent.expand(batch_size,  model.unet.in_channels, height // 8, width // 8).to(model.device)
    return latent, latents


@torch.no_grad()
def text2image(
    model,
    prompt:  List[str],
    controller,
    num_inference_steps: int = 50,
    guidance_scale: Optional[float] = 7.5,
    generator: Optional[torch.Generator] = None,
    latent: Optional[torch.FloatTensor] = None,
    uncond_embeddings=None,
    start_time=50,
    return_type='image'
):
    batch_size = len(prompt)
    register_attention_control(model, controller)
    height = width = 512
    
    text_input = model.tokenizer(
        prompt,
        padding="max_length",
        max_length=model.tokenizer.model_max_length,
        truncation=True,
        return_tensors="pt",
    )
    text_embeddings = model.text_encoder(text_input.input_ids.to(model.device))[0]
    max_length = text_input.input_ids.shape[-1]
    if uncond_embeddings is None:
        uncond_input = model.tokenizer(
            [""] * batch_size, padding="max_length", max_length=max_length, return_tensors="pt"
        )
        uncond_embeddings_ = model.text_encoder(uncond_input.input_ids.to(model.device))[0]
    else:
        uncond_embeddings_ = None

    latent, latents = init_latent(latent, model, height, width, generator, batch_size)
    model.scheduler.set_timesteps(num_inference_steps)
    for i, t in enumerate(tqdm(model.scheduler.timesteps[-start_time:])):
        if uncond_embeddings_ is None:
            context = torch.cat([uncond_embeddings[i].expand(*text_embeddings.shape), text_embeddings])
        else:
            context = torch.cat([uncond_embeddings_, text_embeddings])
        latents = diffusion_step(model, controller, latents, context, t, guidance_scale, low_resource=False)
        
    if return_type == 'image':
        image = latent2image(model.vae, latents)
    else:
        image = latents
    return image, latent


@torch.no_grad()
def text2image_controlnet(
    model,
    prompt:  List[str],
    control_factor,
    controller,
    num_inference_steps: int = 50,
    guidance_scale: Optional[float] = 7.5,
    generator: Optional[torch.Generator] = None,
    latent: Optional[torch.FloatTensor] = None,
    uncond_embeddings=None,
    start_time=50,
    return_type='image'
):
    batch_size = len(prompt)
    register_attention_control(model, controller)
    height = width = 512
    
    text_input = model.tokenizer(
        prompt,
        padding="max_length",
        max_length=model.tokenizer.model_max_length,
        truncation=True,
        return_tensors="pt",
    )
    text_embeddings = model.text_encoder(text_input.input_ids.to(model.device))[0]
    max_length = text_input.input_ids.shape[-1]
    if uncond_embeddings is None:
        uncond_input = model.tokenizer(
            [""] * batch_size, padding="max_length", max_length=max_length, return_tensors="pt"
        )
        uncond_embeddings_ = model.text_encoder(uncond_input.input_ids.to(model.device))[0]
    else:
        uncond_embeddings_ = None

    latent, latents = init_latent(latent, model, height, width, generator, batch_size)
    model.scheduler.set_timesteps(num_inference_steps)
    for i, t in enumerate(tqdm(model.scheduler.timesteps[-start_time:])):
        if uncond_embeddings_ is None:
            context = torch.cat([uncond_embeddings[i].expand(*text_embeddings.shape), text_embeddings])
        else:
            context = torch.cat([uncond_embeddings_, text_embeddings])
        latents = diffusion_step(model, controller, latents, context, t, guidance_scale, low_resource=False)
        
    if return_type == 'image':
        image = latent2image(model.vae, latents)
    else:
        image = latents
    return image, latent



def register_attention_control(model, controller):
    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(hidden_states, encoder_hidden_states=None, attention_mask=None,temb=None,):
            is_cross = encoder_hidden_states is not None
            
            residual = hidden_states

            if self.spatial_norm is not None:
                hidden_states = self.spatial_norm(hidden_states, temb)

            input_ndim = hidden_states.ndim

            if input_ndim == 4:
                batch_size, channel, height, width = hidden_states.shape
                hidden_states = hidden_states.view(batch_size, channel, height * width).transpose(1, 2)

            batch_size, sequence_length, _ = (
                hidden_states.shape if encoder_hidden_states is None else encoder_hidden_states.shape
            )
            attention_mask = self.prepare_attention_mask(attention_mask, sequence_length, batch_size)

            if self.group_norm is not None:
                hidden_states = self.group_norm(hidden_states.transpose(1, 2)).transpose(1, 2)

            query = self.to_q(hidden_states)

            if encoder_hidden_states is None:
                encoder_hidden_states = hidden_states
            elif self.norm_cross:
                encoder_hidden_states = self.norm_encoder_hidden_states(encoder_hidden_states)

            key = self.to_k(encoder_hidden_states)
            value = self.to_v(encoder_hidden_states)

            query = self.head_to_batch_dim(query)
            key = self.head_to_batch_dim(key)
            value = self.head_to_batch_dim(value)

            attention_probs = self.get_attention_scores(query, key, attention_mask)
            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 = to_out(hidden_states)

            if input_ndim == 4:
                hidden_states = hidden_states.transpose(-1, -2).reshape(batch_size, channel, height, width)

            if self.residual_connection:
                hidden_states = hidden_states + residual

            hidden_states = hidden_states / self.rescale_output_factor

            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':
        if net_.__class__.__name__ == 'Attention':
            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, 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)
    return alpha_time_words