Attention-refocusing / gligen /task_grounded_generation.py
Quα»³nh PhΓΉng
update
589b7f1
raw
history blame
13.1 kB
import argparse
from PIL import Image, ImageDraw
from evaluator import Evaluator
from omegaconf import OmegaConf
from ldm.models.diffusion.ddim import DDIMSampler
from ldm.models.diffusion.plms import PLMSSampler
import os
from transformers import CLIPProcessor, CLIPModel
from copy import deepcopy
import torch
from ldm.util import instantiate_from_config
from trainer import read_official_ckpt, batch_to_device
from evaluator import set_alpha_scale, save_images, draw_masks_from_boxes
import numpy as np
import clip
from functools import partial
import torchvision.transforms.functional as F
import random
device = "cuda"
def alpha_generator(length, type=[1,0,0]):
"""
length is total timestpes needed for sampling.
type should be a list containing three values which sum should be 1
It means the percentage of three stages:
alpha=1 stage
linear deacy stage
alpha=0 stage.
For example if length=100, type=[0.8,0.1,0.1]
then the first 800 stpes, alpha will be 1, and then linearly decay to 0 in the next 100 steps,
and the last 100 stpes are 0.
"""
assert len(type)==3
assert type[0] + type[1] + type[2] == 1
stage0_length = int(type[0]*length)
stage1_length = int(type[1]*length)
stage2_length = length - stage0_length - stage1_length
if stage1_length != 0:
decay_alphas = np.arange(start=0, stop=1, step=1/stage1_length)[::-1]
decay_alphas = list(decay_alphas)
else:
decay_alphas = []
alphas = [1]*stage0_length + decay_alphas + [0]*stage2_length
assert len(alphas) == length
return alphas
def draw_box(img, locations):
colors = ["red", "green", "blue", "olive", "orange", "brown", "cyan", "purple"]
draw = ImageDraw.Draw(img)
WW,HH = img.size
for bid, box in enumerate(locations):
draw.rectangle([box[0]*WW, box[1]*HH, box[2]*WW, box[3]*HH], outline =colors[bid % len(colors)], width=5)
return img
def load_common_ckpt(config, common_ckpt):
autoencoder = instantiate_from_config(config.autoencoder).to(device).eval()
text_encoder = instantiate_from_config(config.text_encoder).to(device).eval()
diffusion = instantiate_from_config(config.diffusion).to(device)
autoencoder.load_state_dict( common_ckpt["autoencoder"] )
text_encoder.load_state_dict( common_ckpt["text_encoder"] )
diffusion.load_state_dict( common_ckpt["diffusion"] )
return [autoencoder, text_encoder, diffusion]
def load_ckpt(config, state_dict, common_instances):
model = instantiate_from_config(config.model).to(device)
model.load_state_dict(state_dict['model'])
set_alpha_scale(model, config.alpha_scale)
print("ckpt is loaded")
return [model] + common_instances
def project(x, projection_matrix):
"""
x (Batch*768) should be the penultimate feature of CLIP (before projection)
projection_matrix (768*768) is the CLIP projection matrix, which should be weight.data of Linear layer
defined in CLIP (out_dim, in_dim), thus we need to apply transpose below.
this function will return the CLIP feature (without normalziation)
"""
return x@torch.transpose(projection_matrix, 0, 1)
@torch.no_grad()
def get_clip_feature(model, processor, input, is_image=False):
feature_type = ['before','after_reproject'] # text feature, image feature
if is_image:
image = input #Image.open(input).convert("RGB")
inputs = processor(images=[image], return_tensors="pt", padding=True)
inputs['pixel_values'] = inputs['pixel_values'].cuda() # we use our own preprocessing without center_crop
inputs['input_ids'] = torch.tensor([[0,1,2,3]]).cuda() # placeholder
outputs = model(**inputs)
feature = outputs.image_embeds
if feature_type[1] == 'after_renorm':
feature = feature*28.7
if feature_type[1] == 'after_reproject':
feature = project( feature, torch.load('gligen/projection_matrix.pth').cuda().T ).squeeze(0)
feature = ( feature / feature.norm() ) * 28.7
feature = feature.unsqueeze(0)
else:
inputs = processor(text=input, return_tensors="pt", padding=True)
inputs['input_ids'] = inputs['input_ids'].cuda()
inputs['pixel_values'] = torch.ones(1,3,224,224).cuda() # placeholder
inputs['attention_mask'] = inputs['attention_mask'].cuda()
outputs = model(**inputs)
feature = outputs.text_embeds if feature_type[0] == 'after' else outputs.text_model_output.pooler_output
return feature
def complete_mask(has_mask, max_objs):
mask = torch.ones(1,max_objs)
if type(has_mask) == int or type(has_mask) == float:
return mask * has_mask
else:
for idx, value in enumerate(has_mask):
mask[0,idx] = value
return mask
@torch.no_grad()
def fire_clip(text_encoder, meta, batch=1, max_objs=30, clip_model=None):
# import pdb; pdb.set_trace()
phrases = meta["phrases"]
images = meta["images"]
if clip_model is None:
version = "openai/clip-vit-large-patch14"
model = CLIPModel.from_pretrained(version).cuda()
processor = CLIPProcessor.from_pretrained(version)
else:
version = "openai/clip-vit-large-patch14"
assert clip_model['version'] == version
model = clip_model['model']
processor = clip_model['processor']
boxes = torch.zeros(max_objs, 4)
masks = torch.zeros(max_objs)
text_embeddings = torch.zeros(max_objs, 768)
image_embeddings = torch.zeros(max_objs, 768)
text_features = []
image_features = []
for phrase, image in zip(phrases,images):
text_features.append( get_clip_feature(model, processor, phrase, is_image=False) )
image_features.append( get_clip_feature(model, processor, image, is_image=True) )
if len(text_features) > 0:
text_features = torch.cat(text_features, dim=0)
image_features = torch.cat(image_features, dim=0)
for idx, (box, text_feature, image_feature) in enumerate(zip( meta['locations'], text_features, image_features)):
boxes[idx] = torch.tensor(box)
masks[idx] = 1
text_embeddings[idx] = text_feature
image_embeddings[idx] = image_feature
out = {
"boxes" : boxes.unsqueeze(0).repeat(batch,1,1),
"masks" : masks.unsqueeze(0).repeat(batch,1),
"text_masks" : masks.unsqueeze(0).repeat(batch,1)*complete_mask( meta["has_text_mask"], max_objs ),
"image_masks" : masks.unsqueeze(0).repeat(batch,1)*complete_mask( meta["has_image_mask"], max_objs ),
"text_embeddings" : text_embeddings.unsqueeze(0).repeat(batch,1,1),
"image_embeddings" : image_embeddings.unsqueeze(0).repeat(batch,1,1)
}
return batch_to_device(out, device)
def remove_numbers(text):
result = ''.join([char for char in text if not char.isdigit()])
return result
def process_box_phrase(names, bboxes):
d = {}
for i, phrase in enumerate(names):
phrase = phrase.replace('_',' ')
list_noun = phrase.split(' ')
for n in list_noun:
n = remove_numbers(n)
if not n in d.keys():
d.update({n:[np.array(bboxes[i])]})
else:
d[n].append(np.array(bboxes[i]))
return d
def Pharse2idx_2(prompt, name_box):
prompt = prompt.replace('.','')
prompt = prompt.replace(',','')
prompt_list = prompt.strip('.').split(' ')
object_positions = []
bbox_to_self_att = []
for obj in name_box.keys():
obj_position = []
in_prompt = False
for word in obj.split(' '):
if word in prompt_list:
obj_first_index = prompt_list.index(word) + 1
obj_position.append(obj_first_index)
in_prompt = True
elif word +'s' in prompt_list:
obj_first_index = prompt_list.index(word+'s') + 1
obj_position.append(obj_first_index)
in_prompt = True
elif word +'es' in prompt_list:
obj_first_index = prompt_list.index(word+'es') + 1
obj_position.append(obj_first_index)
in_prompt = True
if in_prompt :
bbox_to_self_att.append(np.array(name_box[obj]))
object_positions.append(obj_position)
return object_positions, bbox_to_self_att
# @torch.no_grad()
def grounded_generation_box(loaded_model_list, instruction, *args, **kwargs):
# -------------- prepare model and misc --------------- #
model, autoencoder, text_encoder, diffusion = loaded_model_list
batch_size = instruction["batch_size"]
is_inpaint = True if "input_image" in instruction else False
save_folder = os.path.join("create_samples", instruction["save_folder_name"])
# -------------- set seed if required --------------- #
if instruction.get('fix_seed', False):
random_seed = instruction['rand_seed']
random.seed(random_seed)
np.random.seed(random_seed)
torch.manual_seed(random_seed)
# ------------- prepare input for the model ------------- #
with torch.no_grad():
batch = fire_clip(text_encoder, instruction, batch_size, clip_model=kwargs.get('clip_model', None))
context = text_encoder.encode( [instruction["prompt"]]*batch_size )
uc = text_encoder.encode( batch_size*[""] )
name_box = process_box_phrase(instruction['phrases'], instruction['locations'])
position, box_att = Pharse2idx_2(instruction['prompt'],name_box )
input = dict(x = None,
timesteps = None,
context = context,
boxes = batch['boxes'],
masks = batch['masks'],
text_masks = batch['text_masks'],
image_masks = batch['image_masks'],
text_embeddings = batch["text_embeddings"],
image_embeddings = batch["image_embeddings"],
boxes_att=box_att,
object_position = position )
inpainting_mask = x0 = None # used for inpainting
if is_inpaint:
input_image = F.pil_to_tensor( instruction["input_image"] )
input_image = ( input_image.float().unsqueeze(0).cuda() / 255 - 0.5 ) / 0.5
x0 = autoencoder.encode( input_image )
if instruction["actual_mask"] is not None:
inpainting_mask = instruction["actual_mask"][None, None].expand(batch['boxes'].shape[0], -1, -1, -1).cuda()
else:
actual_boxes = [instruction['inpainting_boxes_nodrop'] for _ in range(batch['boxes'].shape[0])]
inpainting_mask = draw_masks_from_boxes(actual_boxes, (x0.shape[-2], x0.shape[-1]) ).cuda()
masked_x0 = x0*inpainting_mask
inpainting_extra_input = torch.cat([masked_x0,inpainting_mask], dim=1)
input["inpainting_extra_input"] = inpainting_extra_input
# ------------- prepare sampler ------------- #
alpha_generator_func = partial(alpha_generator, type=instruction["alpha_type"])
if False:
sampler = DDIMSampler(diffusion, model, alpha_generator_func=alpha_generator_func, set_alpha_scale=set_alpha_scale)
steps = 250
else:
sampler = PLMSSampler(diffusion, model, alpha_generator_func=alpha_generator_func, set_alpha_scale=set_alpha_scale)
steps = 50
# ------------- run sampler ... ------------- #
shape = (batch_size, model.in_channels, model.image_size, model.image_size)
samples_fake = sampler.sample(S=steps, shape=shape, input=input, uc=uc, guidance_scale=instruction['guidance_scale'], mask=inpainting_mask, x0=x0)
with torch.no_grad():
samples_fake = autoencoder.decode(samples_fake)
# ------------- other logistics ------------- #
sample_list = []
for sample in samples_fake:
sample = torch.clamp(sample, min=-1, max=1) * 0.5 + 0.5
sample = sample.cpu().numpy().transpose(1,2,0) * 255
sample = Image.fromarray(sample.astype(np.uint8))
sample_list.append(sample)
return sample_list, None
# if __name__ == "__main__":
# parser = argparse.ArgumentParser()
# parser.add_argument("--folder", type=str, default="create_samples", help="path to OUTPUT")
# parser.add_argument("--official_ckpt", type=str, default='../../../data/sd-v1-4.ckpt', help="")
# parser.add_argument("--batch_size", type=int, default=10, help="This will overwrite the one in yaml.")
# parser.add_argument("--no_plms", action='store_true')
# parser.add_argument("--guidance_scale", type=float, default=5, help="")
# parser.add_argument("--alpha_scale", type=float, default=1, help="scale tanh(alpha). If 0, the behaviour is same as original model")
# args = parser.parse_args()
# assert "sd-v1-4.ckpt" in args.official_ckpt, "only support for stable-diffusion model"
# grounded_generation(args)