mirror of
https://github.com/vladmandic/automatic
synced 2026-09-19 09:14:35 +02:00
major refactoring of modules
Signed-off-by: Vladimir Mandic <mandic00@live.com>
This commit is contained in:
Vladimir Mandic
@@ -0,0 +1,118 @@
|
||||
# Copyright (C) 2024 NVIDIA Corporation. All rights reserved.
|
||||
#
|
||||
# This work is licensed under the LICENSE file
|
||||
# located at the root directory.
|
||||
|
||||
import torch
|
||||
import torch.nn.functional as F
|
||||
import numpy as np
|
||||
|
||||
|
||||
## Attention Utils
|
||||
def get_dynamic_threshold(tensor):
|
||||
from skimage import filters
|
||||
return filters.threshold_otsu(tensor.float().cpu().numpy())
|
||||
|
||||
|
||||
def attn_map_to_binary(attention_map, scaler=1.):
|
||||
from skimage import filters
|
||||
attention_map_np = attention_map.float().cpu().numpy()
|
||||
threshold_value = filters.threshold_otsu(attention_map_np) * scaler
|
||||
binary_mask = (attention_map_np > threshold_value).astype(np.uint8)
|
||||
|
||||
return binary_mask
|
||||
|
||||
|
||||
## Features
|
||||
|
||||
def gaussian_smooth(input_tensor, kernel_size=3, sigma=1):
|
||||
"""
|
||||
Function to apply Gaussian smoothing on each 2D slice of a 3D tensor.
|
||||
"""
|
||||
kernel = np.fromfunction(
|
||||
lambda x, y: (1/ (2 * np.pi * sigma ** 2)) *
|
||||
np.exp(-((x - (kernel_size - 1) / 2) ** 2 + (y - (kernel_size - 1) / 2) ** 2) / (2 * sigma ** 2)),
|
||||
(kernel_size, kernel_size)
|
||||
)
|
||||
kernel = torch.Tensor(kernel / kernel.sum()).to(input_tensor.dtype).to(input_tensor.device)
|
||||
# Add batch and channel dimensions to the kernel
|
||||
kernel = kernel.unsqueeze(0).unsqueeze(0)
|
||||
# Iterate over each 2D slice and apply convolution
|
||||
smoothed_slices = []
|
||||
for i in range(input_tensor.size(0)):
|
||||
slice_tensor = input_tensor[i, :, :]
|
||||
slice_tensor = F.conv2d(slice_tensor.unsqueeze(0).unsqueeze(0), kernel, padding=kernel_size // 2)[0, 0]
|
||||
smoothed_slices.append(slice_tensor)
|
||||
# Stack the smoothed slices to get the final tensor
|
||||
smoothed_tensor = torch.stack(smoothed_slices, dim=0)
|
||||
return smoothed_tensor
|
||||
|
||||
|
||||
## Dense correspondence utils
|
||||
|
||||
def cos_dist(a, b):
|
||||
a_norm = F.normalize(a, dim=-1)
|
||||
b_norm = F.normalize(b, dim=-1)
|
||||
res = a_norm @ b_norm.T
|
||||
return 1 - res
|
||||
|
||||
|
||||
def gen_nn_map(src_features, src_mask, tgt_features, tgt_mask, device, batch_size=100, tgt_size=768):
|
||||
resized_src_features = F.interpolate(src_features.unsqueeze(0), size=tgt_size, mode='bilinear', align_corners=False).squeeze(0)
|
||||
resized_src_features = resized_src_features.permute(1,2,0).view(tgt_size**2, -1)
|
||||
resized_tgt_features = F.interpolate(tgt_features.unsqueeze(0), size=tgt_size, mode='bilinear', align_corners=False).squeeze(0)
|
||||
resized_tgt_features = resized_tgt_features.permute(1,2,0).view(tgt_size**2, -1)
|
||||
nearest_neighbor_indices = torch.zeros(tgt_size**2, dtype=torch.long, device=device)
|
||||
nearest_neighbor_distances = torch.zeros(tgt_size**2, dtype=src_features.dtype, device=device)
|
||||
if not batch_size:
|
||||
batch_size = tgt_size**2
|
||||
for i in range(0, tgt_size**2, batch_size):
|
||||
distances = cos_dist(resized_src_features, resized_tgt_features[i:i+batch_size])
|
||||
distances[~src_mask] = 2.
|
||||
min_distances, min_indices = torch.min(distances, dim=0)
|
||||
nearest_neighbor_indices[i:i+batch_size] = min_indices
|
||||
nearest_neighbor_distances[i:i+batch_size] = min_distances
|
||||
return nearest_neighbor_indices, nearest_neighbor_distances
|
||||
|
||||
|
||||
def cyclic_nn_map(features, masks, latent_resolutions, device):
|
||||
bsz = features.shape[0]
|
||||
nn_map_dict = {}
|
||||
nn_distances_dict = {}
|
||||
|
||||
for tgt_size in latent_resolutions:
|
||||
nn_map = torch.empty(bsz, bsz, tgt_size**2, dtype=torch.long, device=device)
|
||||
nn_distances = torch.full((bsz, bsz, tgt_size**2), float('inf'), dtype=features.dtype, device=device)
|
||||
|
||||
for i in range(bsz):
|
||||
for j in range(bsz):
|
||||
if i != j:
|
||||
nearest_neighbor_indices, nearest_neighbor_distances = gen_nn_map(features[j], masks[tgt_size][j], features[i], masks[tgt_size][i], device, batch_size=None, tgt_size=tgt_size)
|
||||
nn_map[i,j] = nearest_neighbor_indices
|
||||
nn_distances[i,j] = nearest_neighbor_distances
|
||||
|
||||
nn_map_dict[tgt_size] = nn_map
|
||||
nn_distances_dict[tgt_size] = nn_distances
|
||||
|
||||
return nn_map_dict, nn_distances_dict
|
||||
|
||||
|
||||
def anchor_nn_map(features, anchor_features, masks, anchor_masks, latent_resolutions, device):
|
||||
bsz = features.shape[0]
|
||||
anchor_bsz = anchor_features.shape[0]
|
||||
nn_map_dict = {}
|
||||
nn_distances_dict = {}
|
||||
|
||||
for tgt_size in latent_resolutions:
|
||||
nn_map = torch.empty(bsz, anchor_bsz, tgt_size**2, dtype=torch.long, device=device)
|
||||
nn_distances = torch.full((bsz, anchor_bsz, tgt_size**2), float('inf'), dtype=features.dtype, device=device)
|
||||
|
||||
for i in range(bsz):
|
||||
for j in range(anchor_bsz):
|
||||
nearest_neighbor_indices, nearest_neighbor_distances = gen_nn_map(anchor_features[j], anchor_masks[tgt_size][j], features[i], masks[tgt_size][i], device, batch_size=None, tgt_size=tgt_size)
|
||||
nn_map[i,j] = nearest_neighbor_indices
|
||||
nn_distances[i,j] = nearest_neighbor_distances
|
||||
nn_map_dict[tgt_size] = nn_map
|
||||
nn_distances_dict[tgt_size] = nn_distances
|
||||
|
||||
return nn_map_dict, nn_distances_dict
|
||||
@@ -0,0 +1,194 @@
|
||||
# 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.
|
||||
|
||||
# MIT License
|
||||
#
|
||||
# Copyright (c) 2023 AttendAndExcite
|
||||
#
|
||||
# Permission is hereby granted, free of charge, to any person obtaining a copy
|
||||
# of this software and associated documentation files (the "Software"), to deal
|
||||
# in the Software without restriction, including without limitation the rights
|
||||
# to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
|
||||
# copies of the Software, and to permit persons to whom the Software is
|
||||
# furnished to do so, subject to the following conditions:
|
||||
#
|
||||
# The above copyright notice and this permission notice shall be included in all
|
||||
# copies or substantial portions of the Software.
|
||||
#
|
||||
# THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
|
||||
# IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
|
||||
# FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
|
||||
# AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
|
||||
# LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
|
||||
# OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
|
||||
# SOFTWARE.
|
||||
|
||||
# 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.
|
||||
|
||||
# Not a contribution
|
||||
# Changes made by NVIDIA CORPORATION & AFFILIATES enabling ConsiStory or otherwise documented as NVIDIA-proprietary
|
||||
# are not a contribution and subject to the license under the LICENSE file located at the root directory.
|
||||
|
||||
import torch
|
||||
from collections import defaultdict
|
||||
import numpy as np
|
||||
from typing import Union, List
|
||||
from PIL import Image
|
||||
|
||||
from modules.consistory.utils.general_utils import attn_map_to_binary
|
||||
import torch.nn.functional as F
|
||||
|
||||
|
||||
class AttentionStore:
|
||||
def __init__(self, attention_store_kwargs):
|
||||
"""
|
||||
Initialize an empty AttentionStore :param step_index: used to visualize only a specific step in the diffusion
|
||||
process
|
||||
"""
|
||||
self.attn_res = attention_store_kwargs.get('attn_res', (32,32))
|
||||
self.token_indices = attention_store_kwargs['token_indices']
|
||||
bsz = self.token_indices.size(1)
|
||||
self.mask_background_query = attention_store_kwargs.get('mask_background_query', False)
|
||||
self.original_attn_masks = attention_store_kwargs.get('original_attn_masks', None)
|
||||
self.extended_mapping = attention_store_kwargs.get('extended_mapping', torch.ones(bsz, bsz).bool())
|
||||
self.mask_dropout = attention_store_kwargs.get('mask_dropout', 0.0)
|
||||
torch.manual_seed(0) # For dropout mask reproducibility
|
||||
|
||||
self.curr_iter = 0
|
||||
self.ALL_RES = [32, 64]
|
||||
self.step_store = defaultdict(list)
|
||||
self.attn_masks = {res: None for res in self.ALL_RES}
|
||||
self.last_mask = {res: None for res in self.ALL_RES}
|
||||
self.last_mask_dropout = {res: None for res in self.ALL_RES}
|
||||
|
||||
def __call__(self, attn, is_cross: bool, place_in_unet: str, attn_heads: int):
|
||||
if is_cross and attn.shape[1] == np.prod(self.attn_res):
|
||||
guidance_attention = attn[attn.size(0)//2:]
|
||||
batched_guidance_attention = guidance_attention.reshape([guidance_attention.shape[0]//attn_heads, attn_heads, *guidance_attention.shape[1:]])
|
||||
batched_guidance_attention = batched_guidance_attention.mean(dim=1)
|
||||
self.step_store[place_in_unet].append(batched_guidance_attention)
|
||||
|
||||
def reset(self):
|
||||
self.step_store = defaultdict(list)
|
||||
self.attn_masks = {res: None for res in self.ALL_RES}
|
||||
self.last_mask = {res: None for res in self.ALL_RES}
|
||||
self.last_mask_dropout = {res: None for res in self.ALL_RES}
|
||||
|
||||
torch.cuda.empty_cache()
|
||||
|
||||
def aggregate_last_steps_attention(self) -> torch.Tensor:
|
||||
"""Aggregates the attention across the different layers and heads at the specified resolution."""
|
||||
attention_maps = torch.cat([torch.stack(x[-20:]) for x in self.step_store.values()]).mean(dim=0)
|
||||
bsz, wh, _ = attention_maps.shape
|
||||
|
||||
# Create attention maps for each concept token, for each batch item
|
||||
agg_attn_maps = []
|
||||
for i in range(bsz):
|
||||
curr_prompt_indices = []
|
||||
|
||||
for concept_token_indices in self.token_indices:
|
||||
if concept_token_indices[i] != -1:
|
||||
curr_prompt_indices.append(attention_maps[i, :, concept_token_indices[i]].view(*self.attn_res))
|
||||
|
||||
agg_attn_maps.append(torch.stack(curr_prompt_indices))
|
||||
|
||||
# Upsample the attention maps to the target resolution
|
||||
# and create the attention masks, unifying masks across the different concepts
|
||||
for tgt_size in self.ALL_RES:
|
||||
pixels = tgt_size ** 2
|
||||
tgt_agg_attn_maps = [F.interpolate(x.unsqueeze(1), size=tgt_size, mode='bilinear').squeeze(1) for x in agg_attn_maps]
|
||||
|
||||
attn_masks = []
|
||||
for batch_item_map in tgt_agg_attn_maps:
|
||||
concept_attn_masks = []
|
||||
|
||||
for concept_maps in batch_item_map:
|
||||
concept_attn_masks.append(torch.from_numpy(attn_map_to_binary(concept_maps, 1.)).to(attention_maps.device).bool().view(-1))
|
||||
|
||||
concept_attn_masks = torch.stack(concept_attn_masks, dim=0).max(dim=0).values
|
||||
attn_masks.append(concept_attn_masks)
|
||||
|
||||
attn_masks = torch.stack(attn_masks)
|
||||
self.last_mask[tgt_size] = attn_masks.clone()
|
||||
|
||||
# Add mask dropout
|
||||
if self.curr_iter < 1000:
|
||||
rand_mask = (torch.rand_like(attn_masks.float()) < self.mask_dropout)
|
||||
attn_masks[rand_mask] = False
|
||||
|
||||
self.last_mask_dropout[tgt_size] = attn_masks.clone()
|
||||
|
||||
# # Create subject driven extended self attention masks
|
||||
# output_attn_mask = torch.zeros((bsz, tgt_size**2, attn_masks.view(-1).size(0)), device=attn_masks.device).bool()
|
||||
|
||||
# for i in range(bsz):
|
||||
# for j in range(bsz):
|
||||
# if i==j:
|
||||
# output_attn_mask[i, :, j*pixels:(j+1)*pixels] = 1
|
||||
# else:
|
||||
# if self.extended_mapping[i,j]:
|
||||
# if not self.mask_background_query:
|
||||
# output_attn_mask[i, :, j*pixels:(j+1)*pixels] = attn_masks[j].unsqueeze(0).expand(pixels, -1)
|
||||
# else:
|
||||
# output_attn_mask[i, attn_masks[i], j*pixels:(j+1)*pixels] = attn_masks[j].unsqueeze(0).expand(attn_masks[i].sum(), -1)
|
||||
|
||||
# self.attn_masks[tgt_size] = output_attn_mask
|
||||
|
||||
def get_attn_mask_bias(self, tgt_size, bsz=None):
|
||||
attn_mask = self.attn_masks[tgt_size] if self.original_attn_masks is None else self.original_attn_masks[tgt_size]
|
||||
|
||||
if attn_mask is None:
|
||||
return None
|
||||
|
||||
attn_bias = torch.zeros_like(attn_mask, dtype=torch.float16)
|
||||
attn_bias[~attn_mask] = float('-inf')
|
||||
|
||||
if bsz and bsz != attn_bias.shape[0]:
|
||||
attn_bias = attn_bias.repeat(bsz // attn_bias.shape[0], 1, 1)
|
||||
|
||||
return attn_bias
|
||||
|
||||
def get_extended_attn_mask_instance(self, width, i):
|
||||
attn_mask = self.last_mask_dropout[width]
|
||||
if attn_mask is None:
|
||||
return None
|
||||
|
||||
n_patches = width**2
|
||||
|
||||
|
||||
output_attn_mask = torch.zeros((attn_mask.shape[0] * attn_mask.shape[1],), device=attn_mask.device, dtype=torch.bool)
|
||||
for j in range(attn_mask.shape[0]):
|
||||
if i==j:
|
||||
output_attn_mask[j*n_patches:(j+1)*n_patches] = 1
|
||||
else:
|
||||
if self.extended_mapping[i,j]:
|
||||
if not self.mask_background_query:
|
||||
output_attn_mask[j*n_patches:(j+1)*n_patches] = attn_mask[j].unsqueeze(0) #.expand(n_patches, -1)
|
||||
else:
|
||||
raise NotImplementedError('mask_background_query is not supported anymore')
|
||||
output_attn_mask[0, attn_mask[i], k*n_patches:(k+1)*n_patches] = attn_mask[j].unsqueeze(0).expand(attn_mask[i].sum(), -1)
|
||||
|
||||
return output_attn_mask
|
||||
Reference in New Issue
Block a user