utils/dift_util.py

import gc
from typing import Any, Dict, Optional, Union

import matplotlib.pyplot as plt
import numpy as np
import torch
import torch.nn as nn
from diffusers import DDIMScheduler, StableDiffusionPipeline
from diffusers.models.unet_2d_condition import UNet2DConditionModel
from PIL import Image, ImageDraw


class MyUNet2DConditionModel(UNet2DConditionModel):
    def forward(
        self,
        sample: torch.FloatTensor,
        timestep: Union[torch.Tensor, float, int],
        up_ft_indices,
        encoder_hidden_states: torch.Tensor,
        class_labels: Optional[torch.Tensor] = None,
        timestep_cond: Optional[torch.Tensor] = None,
        attention_mask: Optional[torch.Tensor] = None,
        cross_attention_kwargs: Optional[Dict[str, Any]] = None
    ):
        r"""
        Args:
            sample (`torch.FloatTensor`): (batch, channel, height, width) noisy inputs tensor
            timestep (`torch.FloatTensor` or `float` or `int`): (batch) timesteps
            encoder_hidden_states (`torch.FloatTensor`): (batch, sequence_length, feature_dim) encoder hidden states
            cross_attention_kwargs (`dict`, *optional*):
                A kwargs dictionary that if specified is passed along to the `AttnProcessor` as defined under
                `self.processor` in
                [diffusers.cross_attention](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/cross_attention.py).
        """
        # By default samples have to be AT least a multiple of the overall upsampling factor.
        # The overall upsampling factor is equal to 2 ** (# num of upsampling layears).
        # However, the upsampling interpolation output size can be forced to fit any upsampling size
        # on the fly if necessary.
        default_overall_up_factor = 2**self.num_upsamplers

        # upsample size should be forwarded when sample is not a multiple of `default_overall_up_factor`
        forward_upsample_size = False
        upsample_size = None

        if any(s % default_overall_up_factor != 0 for s in sample.shape[-2:]):
            # logger.info("Forward upsample size to force interpolation output size.")
            forward_upsample_size = True

        # prepare attention_mask
        if attention_mask is not None:
            attention_mask = (1 - attention_mask.to(sample.dtype)) * -10000.0
            attention_mask = attention_mask.unsqueeze(1)

        # 0. center input if necessary
        if self.config.center_input_sample:
            sample = 2 * sample - 1.0

        # 1. time
        timesteps = timestep
        if not torch.is_tensor(timesteps):
            # TODO: this requires sync between CPU and GPU. So try to pass timesteps as tensors if you can
            # This would be a good case for the `match` statement (Python 3.10+)
            is_mps = sample.device.type == 'mps'
            if isinstance(timestep, float):
                dtype = torch.float32 if is_mps else torch.float64
            else:
                dtype = torch.int32 if is_mps else torch.int64
            timesteps = torch.tensor([timesteps], dtype=dtype, device=sample.device)
        elif len(timesteps.shape) == 0:
            timesteps = timesteps[None].to(sample.device)

        # broadcast to batch dimension in a way that's compatible with ONNX/Core ML
        timesteps = timesteps.expand(sample.shape[0])

        t_emb = self.time_proj(timesteps)

        # timesteps does not contain any weights and will always return f32 tensors
        # but time_embedding might actually be running in fp16. so we need to cast here.
        # there might be better ways to encapsulate this.
        t_emb = t_emb.to(dtype=self.dtype)

        emb = self.time_embedding(t_emb, timestep_cond)

        if self.class_embedding is not None:
            if class_labels is None:
                raise ValueError('class_labels should be provided when num_class_embeds > 0')

            if self.config.class_embed_type == 'timestep':
                class_labels = self.time_proj(class_labels)

            class_emb = self.class_embedding(class_labels).to(dtype=self.dtype)
            emb = emb + class_emb

        # 2. pre-process
        sample = self.conv_in(sample)

        # 3. down
        down_block_res_samples = (sample,)
        for downsample_block in self.down_blocks:
            if hasattr(downsample_block, 'has_cross_attention') and downsample_block.has_cross_attention:
                sample, res_samples = downsample_block(
                    hidden_states=sample,
                    temb=emb,
                    encoder_hidden_states=encoder_hidden_states,
                    attention_mask=attention_mask,
                    cross_attention_kwargs=cross_attention_kwargs,
                )
            else:
                sample, res_samples = downsample_block(hidden_states=sample, temb=emb)

            down_block_res_samples += res_samples

        # 4. mid
        if self.mid_block is not None:
            sample = self.mid_block(
                sample,
                emb,
                encoder_hidden_states=encoder_hidden_states,
                attention_mask=attention_mask,
                cross_attention_kwargs=cross_attention_kwargs,
            )

        # 5. up
        up_ft = {}

        for i, upsample_block in enumerate(self.up_blocks):

            if i > np.max(up_ft_indices):
                break

            is_final_block = i == len(self.up_blocks) - 1

            res_samples = down_block_res_samples[-len(upsample_block.resnets):]
            down_block_res_samples = down_block_res_samples[: -len(upsample_block.resnets)]

            # if we have not reached the final block and need to forward the
            # upsample size, we do it here
            if not is_final_block and forward_upsample_size:
                upsample_size = down_block_res_samples[-1].shape[2:]

            if hasattr(upsample_block, 'has_cross_attention') and upsample_block.has_cross_attention:
                sample = upsample_block(
                    hidden_states=sample,
                    temb=emb,
                    res_hidden_states_tuple=res_samples,
                    encoder_hidden_states=encoder_hidden_states,
                    cross_attention_kwargs=cross_attention_kwargs,
                    upsample_size=upsample_size,
                    attention_mask=attention_mask,
                )
            else:
                sample = upsample_block(
                    hidden_states=sample, temb=emb, res_hidden_states_tuple=res_samples, upsample_size=upsample_size
                )

            if i in up_ft_indices:
                up_ft[i] = sample.detach()

        output = {}
        output['up_ft'] = up_ft

        return output


class OneStepSDPipeline(StableDiffusionPipeline):
    @torch.no_grad()
    def __call__(
        self,
        img_tensor,
        t,
        up_ft_indices,
        prompt_embeds: Optional[torch.FloatTensor] = None,
        cross_attention_kwargs: Optional[Dict[str, Any]] = None
    ):

        device = self._execution_device
        latents = self.vae.encode(img_tensor).latent_dist.sample() * self.vae.config.scaling_factor
        t = torch.tensor(t, dtype=torch.long, device=device)
        noise = torch.randn_like(latents).to(device)
        latents_noisy = self.scheduler.add_noise(latents, noise, t)
        unet_output = self.unet(latents_noisy, t, up_ft_indices, encoder_hidden_states=prompt_embeds, cross_attention_kwargs=cross_attention_kwargs)
        return unet_output


class SDFeaturizer:
    def __init__(self, sd_id='pretrained_models/stable-diffusion-v1-4'):
        unet = MyUNet2DConditionModel.from_pretrained(sd_id, subfolder='unet')
        onestep_pipe = OneStepSDPipeline.from_pretrained(sd_id, unet=unet, safety_checker=None)
        onestep_pipe.vae.decoder = None
        onestep_pipe.scheduler = DDIMScheduler.from_pretrained(sd_id, subfolder='scheduler')
        gc.collect()
        onestep_pipe = onestep_pipe.to('cuda')
        onestep_pipe.enable_attention_slicing()
        self.pipe = onestep_pipe

    @torch.no_grad()
    def forward(self,
                img_tensor,
                prompt,
                t=261,
                up_ft_index=0,
                ensemble_size=8):
        '''
        Args:
            img_tensor: should be a single torch tensor in the shape of [1, C, H, W] or [C, H, W]
            prompt: the prompt to use, a string
            t: the time step to use, should be an int in the range of [0, 1000]
            up_ft_index: which upsampling block of the U-Net to extract feature, you can choose [0, 1, 2, 3]
            ensemble_size: the number of repeated images used in the batch to extract features
        Return:
            unet_ft: a torch tensor in the shape of [1, c, h, w]
        '''
        img_tensor = img_tensor.repeat(ensemble_size, 1, 1, 1).cuda()  # ensem, c, h, w
        prompt_embeds = self.pipe._encode_prompt(
            prompt=prompt,
            device='cuda',
            num_images_per_prompt=1,
            do_classifier_free_guidance=False)  # [1, 77, dim]
        prompt_embeds = prompt_embeds.repeat(ensemble_size, 1, 1)
        unet_ft_all = self.pipe(
            img_tensor=img_tensor,
            t=t,
            up_ft_indices=[up_ft_index],
            prompt_embeds=prompt_embeds)
        unet_ft = unet_ft_all['up_ft'][up_ft_index]  # ensem, c, h, w
        unet_ft = unet_ft.mean(0, keepdim=True)  # 1,c,h,w
        return unet_ft


class DIFT_Demo:
    def __init__(self, source_img, source_dift, source_img_size):
        self.source_dift = source_dift  # NCHW # torch.Size([1, 1280, 28, 48])
        self.source_img = source_img
        self.source_img_size = source_img_size

    @torch.no_grad()
    def query(self, target_img, target_dift, target_img_size, query_point, target_point, visualize=False):
        num_channel = self.source_dift.size(1)
        cos = nn.CosineSimilarity(dim=1)
        source_x, source_y = int(np.round(query_point[1])), int(np.round(query_point[0]))

        src_ft = self.source_dift
        src_ft = nn.Upsample(size=self.source_img_size, mode='bilinear')(src_ft)
        src_vec = src_ft[0, :, source_y, source_x].view(1, num_channel, 1, 1)  # 1, C, 1, 1

        tgt_ft = nn.Upsample(size=target_img_size, mode='bilinear')(target_dift)
        cos_map = cos(src_vec, tgt_ft).cpu().numpy()  # N, H, W  (1, 448, 768)

        max_yx = np.unravel_index(cos_map[0].argmax(), cos_map[0].shape)
        target_x, target_y = int(np.round(target_point[1])), int(np.round(target_point[0]))

        if visualize:
            heatmap = cos_map[0]
            heatmap = (heatmap - np.min(heatmap)) / (np.max(heatmap) - np.min(heatmap))

            cmap = plt.get_cmap('viridis')
            heatmap_color = (cmap(heatmap) * 255)[..., :3].astype(np.uint8)

            alpha, radius, color = 0.5, 3, (0, 255, 0)
            blended_image = Image.blend(target_img, Image.fromarray(heatmap_color), alpha=alpha)
            draw = ImageDraw.Draw(blended_image)
            draw.ellipse((max_yx[1] - radius, max_yx[0] - radius, max_yx[1] + radius, max_yx[0] + radius), fill=color)
            draw.ellipse((target_x - radius, target_y - radius, target_x + radius, target_y + radius), fill=color)
        else:
            blended_image = None
        dift_feat, confidence = tgt_ft[0, :, target_y, target_x], cos_map[0, target_y, target_x]
        return dift_feat, confidence, blended_image