alvanlii's picture
Update run_edit.py
c83cce1
raw
history blame
9.97 kB
import gc
import os
import io
import math
import sys
import tempfile
from PIL import Image, ImageOps
import requests
import torch
from torch import nn
from torch.nn import functional as F
from torchvision import transforms
from torchvision.transforms import functional as TF
from tqdm.notebook import tqdm
import numpy as np
from math import log2, sqrt
import argparse
import pickle
################################### mask_fusion ######################################
from util.metrics_accumulator import MetricsAccumulator
metrics_accumulator = MetricsAccumulator()
from pathlib import Path
from PIL import Image
################################### mask_fusion ######################################
import clip
import lpips
from torch.nn.functional import mse_loss
################################### CLIPseg ######################################
from torchvision import utils as vutils
import cv2
################################### CLIPseg ######################################
def str2bool(x):
return x.lower() in ('true')
USE_CPU = False
device = torch.device('cuda:0' if (torch.cuda.is_available() and not USE_CPU) else 'cpu')
def fetch(url_or_path):
if str(url_or_path).startswith('http://') or str(url_or_path).startswith('https://'):
r = requests.get(url_or_path)
r.raise_for_status()
fd = io.BytesIO()
fd.write(r.content)
fd.seek(0)
return fd
return open(url_or_path, 'rb')
class MakeCutouts(nn.Module):
def __init__(self, cut_size, cutn, cut_pow=1.):
super().__init__()
self.cut_size = cut_size
self.cutn = cutn
self.cut_pow = cut_pow
def forward(self, input):
sideY, sideX = input.shape[2:4]
max_size = min(sideX, sideY)
min_size = min(sideX, sideY, self.cut_size)
cutouts = []
for _ in range(self.cutn):
size = int(torch.rand([])**self.cut_pow * (max_size - min_size) + min_size)
offsetx = torch.randint(0, sideX - size + 1, ())
offsety = torch.randint(0, sideY - size + 1, ())
cutout = input[:, :, offsety:offsety + size, offsetx:offsetx + size]
cutouts.append(F.adaptive_avg_pool2d(cutout, self.cut_size))
return torch.cat(cutouts)
def spherical_dist_loss(x, y):
x = F.normalize(x, dim=-1)
y = F.normalize(y, dim=-1)
return (x - y).norm(dim=-1).div(2).arcsin().pow(2).mul(2)
def do_run(
arg_seed, arg_text, arg_batch_size, arg_num_batches, arg_negative, arg_cutn, arg_edit, arg_height, arg_width,
arg_edit_y, arg_edit_x, arg_edit_width, arg_edit_height, mask, arg_guidance_scale, arg_background_preservation_loss,
arg_lpips_sim_lambda, arg_l2_sim_lambda, arg_ddpm, arg_ddim, arg_enforce_background, arg_clip_guidance_scale,
arg_clip_guidance, model_params, model, diffusion, ldm, bert, clip_model
):
normalize = transforms.Normalize(mean=[0.48145466, 0.4578275, 0.40821073], std=[0.26862954, 0.26130258, 0.27577711])
if arg_seed >= 0:
torch.manual_seed(arg_seed)
text_emb = bert.encode([arg_text] * arg_batch_size).to(device).float()
text_blank = bert.encode([arg_negative] * arg_batch_size).to(device).float()
text = clip.tokenize([arg_text] * arg_batch_size, truncate=True).to(device)
text_clip_blank = clip.tokenize([arg_negative] * arg_batch_size, truncate=True).to(device)
text_emb_clip = clip_model.encode_text(text)
text_emb_clip_blank = clip_model.encode_text(text_clip_blank)
make_cutouts = MakeCutouts(clip_model.visual.input_resolution, arg_cutn)
text_emb_norm = text_emb_clip[0] / text_emb_clip[0].norm(dim=-1, keepdim=True)
image_embed = None
if arg_edit:
w = arg_edit_width if arg_edit_width else arg_width
h = arg_edit_height if arg_edit_height else arg_height
arg_edit = arg_edit.convert('RGB')
input_image_pil = arg_edit
init_image_pil = input_image_pil.resize((arg_height, arg_width), Image.Resampling.LANCZOS)
input_image_pil = ImageOps.fit(input_image_pil, (w, h))
im = transforms.ToTensor()(input_image_pil).unsqueeze(0).to(device)
init_image = (TF.to_tensor(init_image_pil).to(device).unsqueeze(0).mul(2).sub(1))
im = 2*im-1
im = ldm.encode(im).sample()
y = arg_edit_y//8
x = arg_edit_x//8
input_image = torch.zeros(1, 4, arg_height//8, arg_width//8, device=device)
ycrop = y + im.shape[2] - input_image.shape[2]
xcrop = x + im.shape[3] - input_image.shape[3]
ycrop = ycrop if ycrop > 0 else 0
xcrop = xcrop if xcrop > 0 else 0
input_image[0,:,y if y >=0 else 0:y+im.shape[2],x if x >=0 else 0:x+im.shape[3]] = im[:,:,0 if y > 0 else -y:im.shape[2]-ycrop,0 if x > 0 else -x:im.shape[3]-xcrop]
input_image_pil = ldm.decode(input_image)
input_image_pil = TF.to_pil_image(input_image_pil.squeeze(0).add(1).div(2).clamp(0, 1))
input_image *= 0.18215
new_mask = TF.resize(mask.unsqueeze(0).unsqueeze(0).to(device), (arg_width//8, arg_height//8))
mask1 = (new_mask > 0.5)
mask1 = mask1.float()
input_image *= mask1
image_embed = torch.cat(arg_batch_size*2*[input_image], dim=0).float()
elif model_params['image_condition']:
# using inpaint model but no image is provided
image_embed = torch.zeros(arg_batch_size*2, 4, arg_height//8, arg_width//8, device=device)
kwargs = {
"context": torch.cat([text_emb, text_blank], dim=0).float(),
"clip_embed": torch.cat([text_emb_clip, text_emb_clip_blank], dim=0).float() if model_params['clip_embed_dim'] else None,
"image_embed": image_embed
}
# Create a classifier-free guidance sampling function
def model_fn(x_t, ts, **kwargs):
half = x_t[: len(x_t) // 2]
combined = torch.cat([half, half], dim=0)
model_out = model(combined, ts, **kwargs)
eps, rest = model_out[:, :3], model_out[:, 3:]
cond_eps, uncond_eps = torch.split(eps, len(eps) // 2, dim=0)
half_eps = uncond_eps + arg_guidance_scale * (cond_eps - uncond_eps)
eps = torch.cat([half_eps, half_eps], dim=0)
return torch.cat([eps, rest], dim=1)
cur_t = None
@torch.no_grad()
def postprocess_fn(out, t):
if mask is not None:
background_stage_t = diffusion.q_sample(init_image, t[0])
background_stage_t = torch.tile(
background_stage_t, dims=(arg_batch_size, 1, 1, 1)
)
out["sample"] = out["sample"] * mask + background_stage_t * (1 - mask)
return out
# if arg_ddpm:
# sample_fn = diffusion.p_sample_loop_progressive
# elif arg_ddim:
# sample_fn = diffusion.ddim_sample_loop_progressive
# else:
sample_fn = diffusion.plms_sample_loop_progressive
def save_sample(i, sample):
out_ims = []
for k, image in enumerate(sample['pred_xstart'][:arg_batch_size]):
image /= 0.18215
im = image.unsqueeze(0)
out = ldm.decode(im)
metrics_accumulator.print_average_metric()
for b in range(arg_batch_size):
pred_image = sample["pred_xstart"][b]
if arg_enforce_background:
new_mask = TF.resize(mask.unsqueeze(0).unsqueeze(0).to(device), (arg_width, arg_height))
pred_image = (
init_image[0] * new_mask[0] + out * (1 - new_mask[0])
)
pred_image_pil = TF.to_pil_image(pred_image.squeeze(0).add(1).div(2).clamp(0, 1))
out_ims.append(pred_image_pil)
return out_ims
all_saved_ims = []
for i in range(arg_num_batches):
cur_t = diffusion.num_timesteps - 1
samples = sample_fn(
model_fn,
(arg_batch_size*2, 4, int(arg_height//8), int(arg_width//8)),
clip_denoised=False,
model_kwargs=kwargs,
cond_fn=None,
device=device,
progress=True,
)
for j, sample in enumerate(samples):
cur_t -= 1
if j % 5 == 0 and j != diffusion.num_timesteps - 1:
all_saved_ims += save_sample(i, sample)
all_saved_ims += save_sample(i, sample)
return all_saved_ims
def run_model(
segmodel, model, diffusion, ldm, bert, clip_model, model_params,
from_text, instruction, negative_prompt, original_img, seed, guidance_scale, clip_guidance_scale, cutn, l2_sim_lambda
):
input_image = original_img
transform = transforms.Compose([
transforms.ToTensor(),
transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]),
transforms.Resize((256, 256)),
])
img = transform(input_image).unsqueeze(0)
with torch.no_grad():
preds = segmodel(img.repeat(1,1,1,1), from_text)[0]
mask = torch.sigmoid(preds[0][0])
image = (mask.detach().cpu().numpy() * 255).astype(np.uint8) # cv2.cvtColor(image, cv2.COLOR_RGB2GRAY)
ret, thresh = cv2.threshold(image, 100, 255, cv2.THRESH_TRUNC, image)
timg = np.array(thresh)
x, y = timg.shape
for row in range(x):
for col in range(y):
if (timg[row][col]) == 100:
timg[row][col] = 255
if (timg[row][col]) < 100:
timg[row][col] = 0
fulltensor = torch.full_like(mask, fill_value=255)
bgtensor = fulltensor-timg
mask = bgtensor / 255.0
gc.collect()
use_ddim = False
use_ddpm = False
all_saved_ims = do_run(
seed, instruction, 1, 1, negative_prompt, cutn, input_image, 256, 256,
0, 0, 0, 0, mask, guidance_scale, True,
1000, l2_sim_lambda, use_ddpm, use_ddim, True, clip_guidance_scale, False,
model_params, model, diffusion, ldm, bert, clip_model
)
return all_saved_ims[-1]