Spaces:
Running
on
Zero
Running
on
Zero
File size: 2,504 Bytes
7f2690b |
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 |
import torch
import torch.nn.functional as F
import numpy as np
from PIL import Image
attn_maps = {}
def hook_fn(name):
def forward_hook(module, input, output):
if hasattr(module.processor, "attn_map"):
attn_maps[name] = module.processor.attn_map
del module.processor.attn_map
return forward_hook
def register_cross_attention_hook(unet):
for name, module in unet.named_modules():
if name.split('.')[-1].startswith('attn2'):
module.register_forward_hook(hook_fn(name))
return unet
def upscale(attn_map, target_size):
attn_map = torch.mean(attn_map, dim=0)
attn_map = attn_map.permute(1,0)
temp_size = None
for i in range(0,5):
scale = 2 ** i
if ( target_size[0] // scale ) * ( target_size[1] // scale) == attn_map.shape[1]*64:
temp_size = (target_size[0]//(scale*8), target_size[1]//(scale*8))
break
assert temp_size is not None, "temp_size cannot is None"
attn_map = attn_map.view(attn_map.shape[0], *temp_size)
attn_map = F.interpolate(
attn_map.unsqueeze(0).to(dtype=torch.float32),
size=target_size,
mode='bilinear',
align_corners=False
)[0]
attn_map = torch.softmax(attn_map, dim=0)
return attn_map
def get_net_attn_map(image_size, batch_size=2, instance_or_negative=False, detach=True):
idx = 0 if instance_or_negative else 1
net_attn_maps = []
for name, attn_map in attn_maps.items():
attn_map = attn_map.cpu() if detach else attn_map
attn_map = torch.chunk(attn_map, batch_size)[idx].squeeze()
attn_map = upscale(attn_map, image_size)
net_attn_maps.append(attn_map)
net_attn_maps = torch.mean(torch.stack(net_attn_maps,dim=0),dim=0)
return net_attn_maps
def attnmaps2images(net_attn_maps):
#total_attn_scores = 0
images = []
for attn_map in net_attn_maps:
attn_map = attn_map.cpu().numpy()
#total_attn_scores += attn_map.mean().item()
normalized_attn_map = (attn_map - np.min(attn_map)) / (np.max(attn_map) - np.min(attn_map)) * 255
normalized_attn_map = normalized_attn_map.astype(np.uint8)
#print("norm: ", normalized_attn_map.shape)
image = Image.fromarray(normalized_attn_map)
#image = fix_save_attn_map(attn_map)
images.append(image)
#print(total_attn_scores)
return images
def is_torch2_available():
return hasattr(F, "scaled_dot_product_attention") |