Upload generate.py
Browse files- generate.py +990 -0
generate.py
ADDED
@@ -0,0 +1,990 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# Originally made by Katherine Crowson (https://github.com/crowsonkb, https://twitter.com/RiversHaveWings)
|
2 |
+
# The original BigGAN+CLIP method was by https://twitter.com/advadnoun
|
3 |
+
|
4 |
+
import argparse
|
5 |
+
import math
|
6 |
+
import random
|
7 |
+
# from email.policy import default
|
8 |
+
from urllib.request import urlopen
|
9 |
+
from tqdm import tqdm
|
10 |
+
import sys
|
11 |
+
import os
|
12 |
+
|
13 |
+
# pip install taming-transformers doesn't work with Gumbel, but does not yet work with coco etc
|
14 |
+
# appending the path does work with Gumbel, but gives ModuleNotFoundError: No module named 'transformers' for coco etc
|
15 |
+
sys.path.append('taming-transformers')
|
16 |
+
|
17 |
+
from omegaconf import OmegaConf
|
18 |
+
from taming.models import cond_transformer, vqgan
|
19 |
+
#import taming.modules
|
20 |
+
|
21 |
+
import torch
|
22 |
+
from torch import nn, optim
|
23 |
+
from torch.nn import functional as F
|
24 |
+
from torchvision import transforms
|
25 |
+
from torchvision.transforms import functional as TF
|
26 |
+
from torch.cuda import get_device_properties
|
27 |
+
torch.backends.cudnn.benchmark = False # NR: True is a bit faster, but can lead to OOM. False is more deterministic.
|
28 |
+
#torch.use_deterministic_algorithms(True) # NR: grid_sampler_2d_backward_cuda does not have a deterministic implementation
|
29 |
+
|
30 |
+
from torch_optimizer import DiffGrad, AdamP
|
31 |
+
|
32 |
+
from CLIP import clip
|
33 |
+
import kornia.augmentation as K
|
34 |
+
import numpy as np
|
35 |
+
import imageio
|
36 |
+
|
37 |
+
from PIL import ImageFile, Image, PngImagePlugin, ImageChops
|
38 |
+
ImageFile.LOAD_TRUNCATED_IMAGES = True
|
39 |
+
|
40 |
+
from subprocess import Popen, PIPE
|
41 |
+
import re
|
42 |
+
|
43 |
+
# Supress warnings
|
44 |
+
import warnings
|
45 |
+
warnings.filterwarnings('ignore')
|
46 |
+
|
47 |
+
|
48 |
+
# Check for GPU and reduce the default image size if low VRAM
|
49 |
+
default_image_size = 512 # >8GB VRAM
|
50 |
+
if not torch.cuda.is_available():
|
51 |
+
default_image_size = 256 # no GPU found
|
52 |
+
elif get_device_properties(0).total_memory <= 2 ** 33: # 2 ** 33 = 8,589,934,592 bytes = 8 GB
|
53 |
+
default_image_size = 304 # <8GB VRAM
|
54 |
+
|
55 |
+
# Create the parser
|
56 |
+
vq_parser = argparse.ArgumentParser(description='Image generation using VQGAN+CLIP')
|
57 |
+
|
58 |
+
# Add the arguments
|
59 |
+
vq_parser.add_argument("-p", "--prompts", type=str, help="Text prompts", default=None, dest='prompts')
|
60 |
+
vq_parser.add_argument("-ip", "--image_prompts", type=str, help="Image prompts / target image", default=[], dest='image_prompts')
|
61 |
+
vq_parser.add_argument("-i", "--iterations", type=int, help="Number of iterations", default=500, dest='max_iterations')
|
62 |
+
vq_parser.add_argument("-se", "--save_every", type=int, help="Save image iterations", default=50, dest='display_freq')
|
63 |
+
vq_parser.add_argument("-s", "--size", nargs=2, type=int, help="Image size (width height) (default: %(default)s)", default=[default_image_size,default_image_size], dest='size')
|
64 |
+
vq_parser.add_argument("-ii", "--init_image", type=str, help="Initial image", default=None, dest='init_image')
|
65 |
+
vq_parser.add_argument("-in", "--init_noise", type=str, help="Initial noise image (pixels or gradient)", default=None, dest='init_noise')
|
66 |
+
vq_parser.add_argument("-iw", "--init_weight", type=float, help="Initial weight", default=0., dest='init_weight')
|
67 |
+
vq_parser.add_argument("-m", "--clip_model", type=str, help="CLIP model (e.g. ViT-B/32, ViT-B/16)", default='ViT-B/32', dest='clip_model')
|
68 |
+
vq_parser.add_argument("-conf", "--vqgan_config", type=str, help="VQGAN config", default=f'checkpoints/vqgan_imagenet_f16_16384.yaml', dest='vqgan_config')
|
69 |
+
vq_parser.add_argument("-ckpt", "--vqgan_checkpoint", type=str, help="VQGAN checkpoint", default=f'checkpoints/vqgan_imagenet_f16_16384.ckpt', dest='vqgan_checkpoint')
|
70 |
+
vq_parser.add_argument("-nps", "--noise_prompt_seeds", nargs="*", type=int, help="Noise prompt seeds", default=[], dest='noise_prompt_seeds')
|
71 |
+
vq_parser.add_argument("-npw", "--noise_prompt_weights", nargs="*", type=float, help="Noise prompt weights", default=[], dest='noise_prompt_weights')
|
72 |
+
vq_parser.add_argument("-lr", "--learning_rate", type=float, help="Learning rate", default=0.1, dest='step_size')
|
73 |
+
vq_parser.add_argument("-cutm", "--cut_method", type=str, help="Cut method", choices=['original','updated','nrupdated','updatedpooling','latest'], default='latest', dest='cut_method')
|
74 |
+
vq_parser.add_argument("-cuts", "--num_cuts", type=int, help="Number of cuts", default=32, dest='cutn')
|
75 |
+
vq_parser.add_argument("-cutp", "--cut_power", type=float, help="Cut power", default=1., dest='cut_pow')
|
76 |
+
vq_parser.add_argument("-sd", "--seed", type=int, help="Seed", default=None, dest='seed')
|
77 |
+
vq_parser.add_argument("-opt", "--optimiser", type=str, help="Optimiser", choices=['Adam','AdamW','Adagrad','Adamax','DiffGrad','AdamP','RAdam','RMSprop'], default='Adam', dest='optimiser')
|
78 |
+
vq_parser.add_argument("-o", "--output", type=str, help="Output image filename", default="output.png", dest='output')
|
79 |
+
vq_parser.add_argument("-vid", "--video", action='store_true', help="Create video frames?", dest='make_video')
|
80 |
+
vq_parser.add_argument("-zvid", "--zoom_video", action='store_true', help="Create zoom video?", dest='make_zoom_video')
|
81 |
+
vq_parser.add_argument("-zs", "--zoom_start", type=int, help="Zoom start iteration", default=0, dest='zoom_start')
|
82 |
+
vq_parser.add_argument("-zse", "--zoom_save_every", type=int, help="Save zoom image iterations", default=10, dest='zoom_frequency')
|
83 |
+
vq_parser.add_argument("-zsc", "--zoom_scale", type=float, help="Zoom scale %%", default=0.99, dest='zoom_scale')
|
84 |
+
vq_parser.add_argument("-zsx", "--zoom_shift_x", type=int, help="Zoom shift x (left/right) amount in pixels", default=0, dest='zoom_shift_x')
|
85 |
+
vq_parser.add_argument("-zsy", "--zoom_shift_y", type=int, help="Zoom shift y (up/down) amount in pixels", default=0, dest='zoom_shift_y')
|
86 |
+
vq_parser.add_argument("-cpe", "--change_prompt_every", type=int, help="Prompt change frequency", default=0, dest='prompt_frequency')
|
87 |
+
vq_parser.add_argument("-vl", "--video_length", type=float, help="Video length in seconds (not interpolated)", default=10, dest='video_length')
|
88 |
+
vq_parser.add_argument("-ofps", "--output_video_fps", type=float, help="Create an interpolated video (Nvidia GPU only) with this fps (min 10. best set to 30 or 60)", default=0, dest='output_video_fps')
|
89 |
+
vq_parser.add_argument("-ifps", "--input_video_fps", type=float, help="When creating an interpolated video, use this as the input fps to interpolate from (>0 & <ofps)", default=15, dest='input_video_fps')
|
90 |
+
vq_parser.add_argument("-d", "--deterministic", action='store_true', help="Enable cudnn.deterministic?", dest='cudnn_determinism')
|
91 |
+
vq_parser.add_argument("-aug", "--augments", nargs='+', action='append', type=str, choices=['Ji','Sh','Gn','Pe','Ro','Af','Et','Ts','Cr','Er','Re'], help="Enabled augments (latest vut method only)", default=[], dest='augments')
|
92 |
+
vq_parser.add_argument("-vsd", "--video_style_dir", type=str, help="Directory with video frames to style", default=None, dest='video_style_dir')
|
93 |
+
vq_parser.add_argument("-cd", "--cuda_device", type=str, help="Cuda device to use", default="cuda:0", dest='cuda_device')
|
94 |
+
|
95 |
+
|
96 |
+
# Execute the parse_args() method
|
97 |
+
args = vq_parser.parse_args()
|
98 |
+
|
99 |
+
if not args.prompts and not args.image_prompts:
|
100 |
+
args.prompts = "A cute, smiling, Nerdy Rodent"
|
101 |
+
|
102 |
+
if args.cudnn_determinism:
|
103 |
+
torch.backends.cudnn.deterministic = True
|
104 |
+
|
105 |
+
if not args.augments:
|
106 |
+
args.augments = [['Af', 'Pe', 'Ji', 'Er']]
|
107 |
+
|
108 |
+
# Split text prompts using the pipe character (weights are split later)
|
109 |
+
if args.prompts:
|
110 |
+
# For stories, there will be many phrases
|
111 |
+
story_phrases = [phrase.strip() for phrase in args.prompts.split("^")]
|
112 |
+
|
113 |
+
# Make a list of all phrases
|
114 |
+
all_phrases = []
|
115 |
+
for phrase in story_phrases:
|
116 |
+
all_phrases.append(phrase.split("|"))
|
117 |
+
|
118 |
+
# First phrase
|
119 |
+
args.prompts = all_phrases[0]
|
120 |
+
|
121 |
+
# Split target images using the pipe character (weights are split later)
|
122 |
+
if args.image_prompts:
|
123 |
+
args.image_prompts = args.image_prompts.split("|")
|
124 |
+
args.image_prompts = [image.strip() for image in args.image_prompts]
|
125 |
+
|
126 |
+
if args.make_video and args.make_zoom_video:
|
127 |
+
print("Warning: Make video and make zoom video are mutually exclusive.")
|
128 |
+
args.make_video = False
|
129 |
+
|
130 |
+
# Make video steps directory
|
131 |
+
if args.make_video or args.make_zoom_video:
|
132 |
+
if not os.path.exists('steps'):
|
133 |
+
os.mkdir('steps')
|
134 |
+
|
135 |
+
# Fallback to CPU if CUDA is not found and make sure GPU video rendering is also disabled
|
136 |
+
# NB. May not work for AMD cards?
|
137 |
+
if not args.cuda_device == 'cpu' and not torch.cuda.is_available():
|
138 |
+
args.cuda_device = 'cpu'
|
139 |
+
args.video_fps = 0
|
140 |
+
print("Warning: No GPU found! Using the CPU instead. The iterations will be slow.")
|
141 |
+
print("Perhaps CUDA/ROCm or the right pytorch version is not properly installed?")
|
142 |
+
|
143 |
+
# If a video_style_dir has been, then create a list of all the images
|
144 |
+
if args.video_style_dir:
|
145 |
+
print("Locating video frames...")
|
146 |
+
video_frame_list = []
|
147 |
+
for entry in os.scandir(args.video_style_dir):
|
148 |
+
if (entry.path.endswith(".jpg")
|
149 |
+
or entry.path.endswith(".png")) and entry.is_file():
|
150 |
+
video_frame_list.append(entry.path)
|
151 |
+
|
152 |
+
# Reset a few options - same filename, different directory
|
153 |
+
if not os.path.exists('steps'):
|
154 |
+
os.mkdir('steps')
|
155 |
+
|
156 |
+
args.init_image = video_frame_list[0]
|
157 |
+
filename = os.path.basename(args.init_image)
|
158 |
+
cwd = os.getcwd()
|
159 |
+
args.output = os.path.join(cwd, "steps", filename)
|
160 |
+
num_video_frames = len(video_frame_list) # for video styling
|
161 |
+
|
162 |
+
|
163 |
+
# Various functions and classes
|
164 |
+
def sinc(x):
|
165 |
+
return torch.where(x != 0, torch.sin(math.pi * x) / (math.pi * x), x.new_ones([]))
|
166 |
+
|
167 |
+
|
168 |
+
def lanczos(x, a):
|
169 |
+
cond = torch.logical_and(-a < x, x < a)
|
170 |
+
out = torch.where(cond, sinc(x) * sinc(x/a), x.new_zeros([]))
|
171 |
+
return out / out.sum()
|
172 |
+
|
173 |
+
|
174 |
+
def ramp(ratio, width):
|
175 |
+
n = math.ceil(width / ratio + 1)
|
176 |
+
out = torch.empty([n])
|
177 |
+
cur = 0
|
178 |
+
for i in range(out.shape[0]):
|
179 |
+
out[i] = cur
|
180 |
+
cur += ratio
|
181 |
+
return torch.cat([-out[1:].flip([0]), out])[1:-1]
|
182 |
+
|
183 |
+
|
184 |
+
# For zoom video
|
185 |
+
def zoom_at(img, x, y, zoom):
|
186 |
+
w, h = img.size
|
187 |
+
zoom2 = zoom * 2
|
188 |
+
img = img.crop((x - w / zoom2, y - h / zoom2,
|
189 |
+
x + w / zoom2, y + h / zoom2))
|
190 |
+
return img.resize((w, h), Image.LANCZOS)
|
191 |
+
|
192 |
+
|
193 |
+
# NR: Testing with different intital images
|
194 |
+
def random_noise_image(w,h):
|
195 |
+
random_image = Image.fromarray(np.random.randint(0,255,(w,h,3),dtype=np.dtype('uint8')))
|
196 |
+
return random_image
|
197 |
+
|
198 |
+
|
199 |
+
# create initial gradient image
|
200 |
+
def gradient_2d(start, stop, width, height, is_horizontal):
|
201 |
+
if is_horizontal:
|
202 |
+
return np.tile(np.linspace(start, stop, width), (height, 1))
|
203 |
+
else:
|
204 |
+
return np.tile(np.linspace(start, stop, height), (width, 1)).T
|
205 |
+
|
206 |
+
|
207 |
+
def gradient_3d(width, height, start_list, stop_list, is_horizontal_list):
|
208 |
+
result = np.zeros((height, width, len(start_list)), dtype=float)
|
209 |
+
|
210 |
+
for i, (start, stop, is_horizontal) in enumerate(zip(start_list, stop_list, is_horizontal_list)):
|
211 |
+
result[:, :, i] = gradient_2d(start, stop, width, height, is_horizontal)
|
212 |
+
|
213 |
+
return result
|
214 |
+
|
215 |
+
|
216 |
+
def random_gradient_image(w,h):
|
217 |
+
array = gradient_3d(w, h, (0, 0, np.random.randint(0,255)), (np.random.randint(1,255), np.random.randint(2,255), np.random.randint(3,128)), (True, False, False))
|
218 |
+
random_image = Image.fromarray(np.uint8(array))
|
219 |
+
return random_image
|
220 |
+
|
221 |
+
|
222 |
+
# Used in older MakeCutouts
|
223 |
+
def resample(input, size, align_corners=True):
|
224 |
+
n, c, h, w = input.shape
|
225 |
+
dh, dw = size
|
226 |
+
|
227 |
+
input = input.view([n * c, 1, h, w])
|
228 |
+
|
229 |
+
if dh < h:
|
230 |
+
kernel_h = lanczos(ramp(dh / h, 2), 2).to(input.device, input.dtype)
|
231 |
+
pad_h = (kernel_h.shape[0] - 1) // 2
|
232 |
+
input = F.pad(input, (0, 0, pad_h, pad_h), 'reflect')
|
233 |
+
input = F.conv2d(input, kernel_h[None, None, :, None])
|
234 |
+
|
235 |
+
if dw < w:
|
236 |
+
kernel_w = lanczos(ramp(dw / w, 2), 2).to(input.device, input.dtype)
|
237 |
+
pad_w = (kernel_w.shape[0] - 1) // 2
|
238 |
+
input = F.pad(input, (pad_w, pad_w, 0, 0), 'reflect')
|
239 |
+
input = F.conv2d(input, kernel_w[None, None, None, :])
|
240 |
+
|
241 |
+
input = input.view([n, c, h, w])
|
242 |
+
return F.interpolate(input, size, mode='bicubic', align_corners=align_corners)
|
243 |
+
|
244 |
+
|
245 |
+
class ReplaceGrad(torch.autograd.Function):
|
246 |
+
@staticmethod
|
247 |
+
def forward(ctx, x_forward, x_backward):
|
248 |
+
ctx.shape = x_backward.shape
|
249 |
+
return x_forward
|
250 |
+
|
251 |
+
@staticmethod
|
252 |
+
def backward(ctx, grad_in):
|
253 |
+
return None, grad_in.sum_to_size(ctx.shape)
|
254 |
+
|
255 |
+
replace_grad = ReplaceGrad.apply
|
256 |
+
|
257 |
+
|
258 |
+
class ClampWithGrad(torch.autograd.Function):
|
259 |
+
@staticmethod
|
260 |
+
def forward(ctx, input, min, max):
|
261 |
+
ctx.min = min
|
262 |
+
ctx.max = max
|
263 |
+
ctx.save_for_backward(input)
|
264 |
+
return input.clamp(min, max)
|
265 |
+
|
266 |
+
@staticmethod
|
267 |
+
def backward(ctx, grad_in):
|
268 |
+
input, = ctx.saved_tensors
|
269 |
+
return grad_in * (grad_in * (input - input.clamp(ctx.min, ctx.max)) >= 0), None, None
|
270 |
+
|
271 |
+
clamp_with_grad = ClampWithGrad.apply
|
272 |
+
|
273 |
+
|
274 |
+
def vector_quantize(x, codebook):
|
275 |
+
d = x.pow(2).sum(dim=-1, keepdim=True) + codebook.pow(2).sum(dim=1) - 2 * x @ codebook.T
|
276 |
+
indices = d.argmin(-1)
|
277 |
+
x_q = F.one_hot(indices, codebook.shape[0]).to(d.dtype) @ codebook
|
278 |
+
return replace_grad(x_q, x)
|
279 |
+
|
280 |
+
|
281 |
+
class Prompt(nn.Module):
|
282 |
+
def __init__(self, embed, weight=1., stop=float('-inf')):
|
283 |
+
super().__init__()
|
284 |
+
self.register_buffer('embed', embed)
|
285 |
+
self.register_buffer('weight', torch.as_tensor(weight))
|
286 |
+
self.register_buffer('stop', torch.as_tensor(stop))
|
287 |
+
|
288 |
+
def forward(self, input):
|
289 |
+
input_normed = F.normalize(input.unsqueeze(1), dim=2)
|
290 |
+
embed_normed = F.normalize(self.embed.unsqueeze(0), dim=2)
|
291 |
+
dists = input_normed.sub(embed_normed).norm(dim=2).div(2).arcsin().pow(2).mul(2)
|
292 |
+
dists = dists * self.weight.sign()
|
293 |
+
return self.weight.abs() * replace_grad(dists, torch.maximum(dists, self.stop)).mean()
|
294 |
+
|
295 |
+
|
296 |
+
#NR: Split prompts and weights
|
297 |
+
def split_prompt(prompt):
|
298 |
+
vals = prompt.rsplit(':', 2)
|
299 |
+
vals = vals + ['', '1', '-inf'][len(vals):]
|
300 |
+
return vals[0], float(vals[1]), float(vals[2])
|
301 |
+
|
302 |
+
|
303 |
+
class MakeCutouts(nn.Module):
|
304 |
+
def __init__(self, cut_size, cutn, cut_pow=1.):
|
305 |
+
super().__init__()
|
306 |
+
self.cut_size = cut_size
|
307 |
+
self.cutn = cutn
|
308 |
+
self.cut_pow = cut_pow # not used with pooling
|
309 |
+
|
310 |
+
# Pick your own augments & their order
|
311 |
+
augment_list = []
|
312 |
+
for item in args.augments[0]:
|
313 |
+
if item == 'Ji':
|
314 |
+
augment_list.append(K.ColorJitter(brightness=0.1, contrast=0.1, saturation=0.1, hue=0.1, p=0.7))
|
315 |
+
elif item == 'Sh':
|
316 |
+
augment_list.append(K.RandomSharpness(sharpness=0.3, p=0.5))
|
317 |
+
elif item == 'Gn':
|
318 |
+
augment_list.append(K.RandomGaussianNoise(mean=0.0, std=1., p=0.5))
|
319 |
+
elif item == 'Pe':
|
320 |
+
augment_list.append(K.RandomPerspective(distortion_scale=0.7, p=0.7))
|
321 |
+
elif item == 'Ro':
|
322 |
+
augment_list.append(K.RandomRotation(degrees=15, p=0.7))
|
323 |
+
elif item == 'Af':
|
324 |
+
augment_list.append(K.RandomAffine(degrees=15, translate=0.1, shear=5, p=0.7, padding_mode='zeros', keepdim=True)) # border, reflection, zeros
|
325 |
+
elif item == 'Et':
|
326 |
+
augment_list.append(K.RandomElasticTransform(p=0.7))
|
327 |
+
elif item == 'Ts':
|
328 |
+
augment_list.append(K.RandomThinPlateSpline(scale=0.8, same_on_batch=True, p=0.7))
|
329 |
+
elif item == 'Cr':
|
330 |
+
augment_list.append(K.RandomCrop(size=(self.cut_size,self.cut_size), pad_if_needed=True, padding_mode='reflect', p=0.5))
|
331 |
+
elif item == 'Er':
|
332 |
+
augment_list.append(K.RandomErasing(scale=(.1, .4), ratio=(.3, 1/.3), same_on_batch=True, p=0.7))
|
333 |
+
elif item == 'Re':
|
334 |
+
augment_list.append(K.RandomResizedCrop(size=(self.cut_size,self.cut_size), scale=(0.1,1), ratio=(0.75,1.333), cropping_mode='resample', p=0.5))
|
335 |
+
|
336 |
+
self.augs = nn.Sequential(*augment_list)
|
337 |
+
self.noise_fac = 0.1
|
338 |
+
# self.noise_fac = False
|
339 |
+
|
340 |
+
# Uncomment if you like seeing the list ;)
|
341 |
+
# print(augment_list)
|
342 |
+
|
343 |
+
# Pooling
|
344 |
+
self.av_pool = nn.AdaptiveAvgPool2d((self.cut_size, self.cut_size))
|
345 |
+
self.max_pool = nn.AdaptiveMaxPool2d((self.cut_size, self.cut_size))
|
346 |
+
|
347 |
+
def forward(self, input):
|
348 |
+
cutouts = []
|
349 |
+
|
350 |
+
for _ in range(self.cutn):
|
351 |
+
# Use Pooling
|
352 |
+
cutout = (self.av_pool(input) + self.max_pool(input))/2
|
353 |
+
cutouts.append(cutout)
|
354 |
+
|
355 |
+
batch = self.augs(torch.cat(cutouts, dim=0))
|
356 |
+
|
357 |
+
if self.noise_fac:
|
358 |
+
facs = batch.new_empty([self.cutn, 1, 1, 1]).uniform_(0, self.noise_fac)
|
359 |
+
batch = batch + facs * torch.randn_like(batch)
|
360 |
+
return batch
|
361 |
+
|
362 |
+
|
363 |
+
# An updated version with Kornia augments and pooling (where my version started):
|
364 |
+
class MakeCutoutsPoolingUpdate(nn.Module):
|
365 |
+
def __init__(self, cut_size, cutn, cut_pow=1.):
|
366 |
+
super().__init__()
|
367 |
+
self.cut_size = cut_size
|
368 |
+
self.cutn = cutn
|
369 |
+
self.cut_pow = cut_pow # Not used with pooling
|
370 |
+
|
371 |
+
self.augs = nn.Sequential(
|
372 |
+
K.RandomAffine(degrees=15, translate=0.1, p=0.7, padding_mode='border'),
|
373 |
+
K.RandomPerspective(0.7,p=0.7),
|
374 |
+
K.ColorJitter(hue=0.1, saturation=0.1, p=0.7),
|
375 |
+
K.RandomErasing((.1, .4), (.3, 1/.3), same_on_batch=True, p=0.7),
|
376 |
+
)
|
377 |
+
|
378 |
+
self.noise_fac = 0.1
|
379 |
+
self.av_pool = nn.AdaptiveAvgPool2d((self.cut_size, self.cut_size))
|
380 |
+
self.max_pool = nn.AdaptiveMaxPool2d((self.cut_size, self.cut_size))
|
381 |
+
|
382 |
+
def forward(self, input):
|
383 |
+
sideY, sideX = input.shape[2:4]
|
384 |
+
max_size = min(sideX, sideY)
|
385 |
+
min_size = min(sideX, sideY, self.cut_size)
|
386 |
+
cutouts = []
|
387 |
+
|
388 |
+
for _ in range(self.cutn):
|
389 |
+
cutout = (self.av_pool(input) + self.max_pool(input))/2
|
390 |
+
cutouts.append(cutout)
|
391 |
+
|
392 |
+
batch = self.augs(torch.cat(cutouts, dim=0))
|
393 |
+
|
394 |
+
if self.noise_fac:
|
395 |
+
facs = batch.new_empty([self.cutn, 1, 1, 1]).uniform_(0, self.noise_fac)
|
396 |
+
batch = batch + facs * torch.randn_like(batch)
|
397 |
+
return batch
|
398 |
+
|
399 |
+
|
400 |
+
# An Nerdy updated version with selectable Kornia augments, but no pooling:
|
401 |
+
class MakeCutoutsNRUpdate(nn.Module):
|
402 |
+
def __init__(self, cut_size, cutn, cut_pow=1.):
|
403 |
+
super().__init__()
|
404 |
+
self.cut_size = cut_size
|
405 |
+
self.cutn = cutn
|
406 |
+
self.cut_pow = cut_pow
|
407 |
+
self.noise_fac = 0.1
|
408 |
+
|
409 |
+
# Pick your own augments & their order
|
410 |
+
augment_list = []
|
411 |
+
for item in args.augments[0]:
|
412 |
+
if item == 'Ji':
|
413 |
+
augment_list.append(K.ColorJitter(brightness=0.1, contrast=0.1, saturation=0.1, hue=0.1, p=0.7))
|
414 |
+
elif item == 'Sh':
|
415 |
+
augment_list.append(K.RandomSharpness(sharpness=0.3, p=0.5))
|
416 |
+
elif item == 'Gn':
|
417 |
+
augment_list.append(K.RandomGaussianNoise(mean=0.0, std=1., p=0.5))
|
418 |
+
elif item == 'Pe':
|
419 |
+
augment_list.append(K.RandomPerspective(distortion_scale=0.5, p=0.7))
|
420 |
+
elif item == 'Ro':
|
421 |
+
augment_list.append(K.RandomRotation(degrees=15, p=0.7))
|
422 |
+
elif item == 'Af':
|
423 |
+
augment_list.append(K.RandomAffine(degrees=30, translate=0.1, shear=5, p=0.7, padding_mode='zeros', keepdim=True)) # border, reflection, zeros
|
424 |
+
elif item == 'Et':
|
425 |
+
augment_list.append(K.RandomElasticTransform(p=0.7))
|
426 |
+
elif item == 'Ts':
|
427 |
+
augment_list.append(K.RandomThinPlateSpline(scale=0.8, same_on_batch=True, p=0.7))
|
428 |
+
elif item == 'Cr':
|
429 |
+
augment_list.append(K.RandomCrop(size=(self.cut_size,self.cut_size), pad_if_needed=True, padding_mode='reflect', p=0.5))
|
430 |
+
elif item == 'Er':
|
431 |
+
augment_list.append(K.RandomErasing(scale=(.1, .4), ratio=(.3, 1/.3), same_on_batch=True, p=0.7))
|
432 |
+
elif item == 'Re':
|
433 |
+
augment_list.append(K.RandomResizedCrop(size=(self.cut_size,self.cut_size), scale=(0.1,1), ratio=(0.75,1.333), cropping_mode='resample', p=0.5))
|
434 |
+
|
435 |
+
self.augs = nn.Sequential(*augment_list)
|
436 |
+
|
437 |
+
|
438 |
+
def forward(self, input):
|
439 |
+
sideY, sideX = input.shape[2:4]
|
440 |
+
max_size = min(sideX, sideY)
|
441 |
+
min_size = min(sideX, sideY, self.cut_size)
|
442 |
+
cutouts = []
|
443 |
+
for _ in range(self.cutn):
|
444 |
+
size = int(torch.rand([])**self.cut_pow * (max_size - min_size) + min_size)
|
445 |
+
offsetx = torch.randint(0, sideX - size + 1, ())
|
446 |
+
offsety = torch.randint(0, sideY - size + 1, ())
|
447 |
+
cutout = input[:, :, offsety:offsety + size, offsetx:offsetx + size]
|
448 |
+
cutouts.append(resample(cutout, (self.cut_size, self.cut_size)))
|
449 |
+
batch = self.augs(torch.cat(cutouts, dim=0))
|
450 |
+
if self.noise_fac:
|
451 |
+
facs = batch.new_empty([self.cutn, 1, 1, 1]).uniform_(0, self.noise_fac)
|
452 |
+
batch = batch + facs * torch.randn_like(batch)
|
453 |
+
return batch
|
454 |
+
|
455 |
+
|
456 |
+
# An updated version with Kornia augments, but no pooling:
|
457 |
+
class MakeCutoutsUpdate(nn.Module):
|
458 |
+
def __init__(self, cut_size, cutn, cut_pow=1.):
|
459 |
+
super().__init__()
|
460 |
+
self.cut_size = cut_size
|
461 |
+
self.cutn = cutn
|
462 |
+
self.cut_pow = cut_pow
|
463 |
+
self.augs = nn.Sequential(
|
464 |
+
K.RandomHorizontalFlip(p=0.5),
|
465 |
+
K.ColorJitter(hue=0.01, saturation=0.01, p=0.7),
|
466 |
+
# K.RandomSolarize(0.01, 0.01, p=0.7),
|
467 |
+
K.RandomSharpness(0.3,p=0.4),
|
468 |
+
K.RandomAffine(degrees=30, translate=0.1, p=0.8, padding_mode='border'),
|
469 |
+
K.RandomPerspective(0.2,p=0.4),)
|
470 |
+
self.noise_fac = 0.1
|
471 |
+
|
472 |
+
|
473 |
+
def forward(self, input):
|
474 |
+
sideY, sideX = input.shape[2:4]
|
475 |
+
max_size = min(sideX, sideY)
|
476 |
+
min_size = min(sideX, sideY, self.cut_size)
|
477 |
+
cutouts = []
|
478 |
+
for _ in range(self.cutn):
|
479 |
+
size = int(torch.rand([])**self.cut_pow * (max_size - min_size) + min_size)
|
480 |
+
offsetx = torch.randint(0, sideX - size + 1, ())
|
481 |
+
offsety = torch.randint(0, sideY - size + 1, ())
|
482 |
+
cutout = input[:, :, offsety:offsety + size, offsetx:offsetx + size]
|
483 |
+
cutouts.append(resample(cutout, (self.cut_size, self.cut_size)))
|
484 |
+
batch = self.augs(torch.cat(cutouts, dim=0))
|
485 |
+
if self.noise_fac:
|
486 |
+
facs = batch.new_empty([self.cutn, 1, 1, 1]).uniform_(0, self.noise_fac)
|
487 |
+
batch = batch + facs * torch.randn_like(batch)
|
488 |
+
return batch
|
489 |
+
|
490 |
+
|
491 |
+
# This is the original version (No pooling)
|
492 |
+
class MakeCutoutsOrig(nn.Module):
|
493 |
+
def __init__(self, cut_size, cutn, cut_pow=1.):
|
494 |
+
super().__init__()
|
495 |
+
self.cut_size = cut_size
|
496 |
+
self.cutn = cutn
|
497 |
+
self.cut_pow = cut_pow
|
498 |
+
|
499 |
+
def forward(self, input):
|
500 |
+
sideY, sideX = input.shape[2:4]
|
501 |
+
max_size = min(sideX, sideY)
|
502 |
+
min_size = min(sideX, sideY, self.cut_size)
|
503 |
+
cutouts = []
|
504 |
+
for _ in range(self.cutn):
|
505 |
+
size = int(torch.rand([])**self.cut_pow * (max_size - min_size) + min_size)
|
506 |
+
offsetx = torch.randint(0, sideX - size + 1, ())
|
507 |
+
offsety = torch.randint(0, sideY - size + 1, ())
|
508 |
+
cutout = input[:, :, offsety:offsety + size, offsetx:offsetx + size]
|
509 |
+
cutouts.append(resample(cutout, (self.cut_size, self.cut_size)))
|
510 |
+
return clamp_with_grad(torch.cat(cutouts, dim=0), 0, 1)
|
511 |
+
|
512 |
+
|
513 |
+
def load_vqgan_model(config_path, checkpoint_path):
|
514 |
+
global gumbel
|
515 |
+
gumbel = False
|
516 |
+
config = OmegaConf.load(config_path)
|
517 |
+
if config.model.target == 'taming.models.vqgan.VQModel':
|
518 |
+
model = vqgan.VQModel(**config.model.params)
|
519 |
+
model.eval().requires_grad_(False)
|
520 |
+
model.init_from_ckpt(checkpoint_path)
|
521 |
+
elif config.model.target == 'taming.models.vqgan.GumbelVQ':
|
522 |
+
model = vqgan.GumbelVQ(**config.model.params)
|
523 |
+
model.eval().requires_grad_(False)
|
524 |
+
model.init_from_ckpt(checkpoint_path)
|
525 |
+
gumbel = True
|
526 |
+
elif config.model.target == 'taming.models.cond_transformer.Net2NetTransformer':
|
527 |
+
parent_model = cond_transformer.Net2NetTransformer(**config.model.params)
|
528 |
+
parent_model.eval().requires_grad_(False)
|
529 |
+
parent_model.init_from_ckpt(checkpoint_path)
|
530 |
+
model = parent_model.first_stage_model
|
531 |
+
else:
|
532 |
+
raise ValueError(f'unknown model type: {config.model.target}')
|
533 |
+
del model.loss
|
534 |
+
return model
|
535 |
+
|
536 |
+
|
537 |
+
def resize_image(image, out_size):
|
538 |
+
ratio = image.size[0] / image.size[1]
|
539 |
+
area = min(image.size[0] * image.size[1], out_size[0] * out_size[1])
|
540 |
+
size = round((area * ratio)**0.5), round((area / ratio)**0.5)
|
541 |
+
return image.resize(size, Image.LANCZOS)
|
542 |
+
|
543 |
+
|
544 |
+
# Do it
|
545 |
+
device = torch.device(args.cuda_device)
|
546 |
+
model = load_vqgan_model(args.vqgan_config, args.vqgan_checkpoint).to(device)
|
547 |
+
jit = True if "1.7.1" in torch.__version__ else False
|
548 |
+
perceptor = clip.load(args.clip_model, jit=jit)[0].eval().requires_grad_(False).to(device)
|
549 |
+
|
550 |
+
# clock=deepcopy(perceptor.visual.positional_embedding.data)
|
551 |
+
# perceptor.visual.positional_embedding.data = clock/clock.max()
|
552 |
+
# perceptor.visual.positional_embedding.data=clamp_with_grad(clock,0,1)
|
553 |
+
|
554 |
+
cut_size = perceptor.visual.input_resolution
|
555 |
+
f = 2**(model.decoder.num_resolutions - 1)
|
556 |
+
|
557 |
+
# Cutout class options:
|
558 |
+
# 'latest','original','updated' or 'updatedpooling'
|
559 |
+
if args.cut_method == 'latest':
|
560 |
+
make_cutouts = MakeCutouts(cut_size, args.cutn, cut_pow=args.cut_pow)
|
561 |
+
elif args.cut_method == 'original':
|
562 |
+
make_cutouts = MakeCutoutsOrig(cut_size, args.cutn, cut_pow=args.cut_pow)
|
563 |
+
elif args.cut_method == 'updated':
|
564 |
+
make_cutouts = MakeCutoutsUpdate(cut_size, args.cutn, cut_pow=args.cut_pow)
|
565 |
+
elif args.cut_method == 'nrupdated':
|
566 |
+
make_cutouts = MakeCutoutsNRUpdate(cut_size, args.cutn, cut_pow=args.cut_pow)
|
567 |
+
else:
|
568 |
+
make_cutouts = MakeCutoutsPoolingUpdate(cut_size, args.cutn, cut_pow=args.cut_pow)
|
569 |
+
|
570 |
+
toksX, toksY = args.size[0] // f, args.size[1] // f
|
571 |
+
sideX, sideY = toksX * f, toksY * f
|
572 |
+
|
573 |
+
# Gumbel or not?
|
574 |
+
if gumbel:
|
575 |
+
e_dim = 256
|
576 |
+
n_toks = model.quantize.n_embed
|
577 |
+
z_min = model.quantize.embed.weight.min(dim=0).values[None, :, None, None]
|
578 |
+
z_max = model.quantize.embed.weight.max(dim=0).values[None, :, None, None]
|
579 |
+
else:
|
580 |
+
e_dim = model.quantize.e_dim
|
581 |
+
n_toks = model.quantize.n_e
|
582 |
+
z_min = model.quantize.embedding.weight.min(dim=0).values[None, :, None, None]
|
583 |
+
z_max = model.quantize.embedding.weight.max(dim=0).values[None, :, None, None]
|
584 |
+
|
585 |
+
|
586 |
+
if args.init_image:
|
587 |
+
if 'http' in args.init_image:
|
588 |
+
img = Image.open(urlopen(args.init_image))
|
589 |
+
else:
|
590 |
+
img = Image.open(args.init_image)
|
591 |
+
pil_image = img.convert('RGB')
|
592 |
+
pil_image = pil_image.resize((sideX, sideY), Image.LANCZOS)
|
593 |
+
pil_tensor = TF.to_tensor(pil_image)
|
594 |
+
z, *_ = model.encode(pil_tensor.to(device).unsqueeze(0) * 2 - 1)
|
595 |
+
elif args.init_noise == 'pixels':
|
596 |
+
img = random_noise_image(args.size[0], args.size[1])
|
597 |
+
pil_image = img.convert('RGB')
|
598 |
+
pil_image = pil_image.resize((sideX, sideY), Image.LANCZOS)
|
599 |
+
pil_tensor = TF.to_tensor(pil_image)
|
600 |
+
z, *_ = model.encode(pil_tensor.to(device).unsqueeze(0) * 2 - 1)
|
601 |
+
elif args.init_noise == 'gradient':
|
602 |
+
img = random_gradient_image(args.size[0], args.size[1])
|
603 |
+
pil_image = img.convert('RGB')
|
604 |
+
pil_image = pil_image.resize((sideX, sideY), Image.LANCZOS)
|
605 |
+
pil_tensor = TF.to_tensor(pil_image)
|
606 |
+
z, *_ = model.encode(pil_tensor.to(device).unsqueeze(0) * 2 - 1)
|
607 |
+
else:
|
608 |
+
one_hot = F.one_hot(torch.randint(n_toks, [toksY * toksX], device=device), n_toks).float()
|
609 |
+
# z = one_hot @ model.quantize.embedding.weight
|
610 |
+
if gumbel:
|
611 |
+
z = one_hot @ model.quantize.embed.weight
|
612 |
+
else:
|
613 |
+
z = one_hot @ model.quantize.embedding.weight
|
614 |
+
|
615 |
+
z = z.view([-1, toksY, toksX, e_dim]).permute(0, 3, 1, 2)
|
616 |
+
#z = torch.rand_like(z)*2 # NR: check
|
617 |
+
|
618 |
+
z_orig = z.clone()
|
619 |
+
z.requires_grad_(True)
|
620 |
+
|
621 |
+
pMs = []
|
622 |
+
normalize = transforms.Normalize(mean=[0.48145466, 0.4578275, 0.40821073],
|
623 |
+
std=[0.26862954, 0.26130258, 0.27577711])
|
624 |
+
|
625 |
+
# From imagenet - Which is better?
|
626 |
+
#normalize = transforms.Normalize(mean=[0.485, 0.456, 0.406],
|
627 |
+
# std=[0.229, 0.224, 0.225])
|
628 |
+
|
629 |
+
# CLIP tokenize/encode
|
630 |
+
if args.prompts:
|
631 |
+
for prompt in args.prompts:
|
632 |
+
txt, weight, stop = split_prompt(prompt)
|
633 |
+
embed = perceptor.encode_text(clip.tokenize(txt).to(device)).float()
|
634 |
+
pMs.append(Prompt(embed, weight, stop).to(device))
|
635 |
+
|
636 |
+
for prompt in args.image_prompts:
|
637 |
+
path, weight, stop = split_prompt(prompt)
|
638 |
+
img = Image.open(path)
|
639 |
+
pil_image = img.convert('RGB')
|
640 |
+
img = resize_image(pil_image, (sideX, sideY))
|
641 |
+
batch = make_cutouts(TF.to_tensor(img).unsqueeze(0).to(device))
|
642 |
+
embed = perceptor.encode_image(normalize(batch)).float()
|
643 |
+
pMs.append(Prompt(embed, weight, stop).to(device))
|
644 |
+
|
645 |
+
for seed, weight in zip(args.noise_prompt_seeds, args.noise_prompt_weights):
|
646 |
+
gen = torch.Generator().manual_seed(seed)
|
647 |
+
embed = torch.empty([1, perceptor.visual.output_dim]).normal_(generator=gen)
|
648 |
+
pMs.append(Prompt(embed, weight).to(device))
|
649 |
+
|
650 |
+
|
651 |
+
# Set the optimiser
|
652 |
+
def get_opt(opt_name, opt_lr):
|
653 |
+
if opt_name == "Adam":
|
654 |
+
opt = optim.Adam([z], lr=opt_lr) # LR=0.1 (Default)
|
655 |
+
elif opt_name == "AdamW":
|
656 |
+
opt = optim.AdamW([z], lr=opt_lr)
|
657 |
+
elif opt_name == "Adagrad":
|
658 |
+
opt = optim.Adagrad([z], lr=opt_lr)
|
659 |
+
elif opt_name == "Adamax":
|
660 |
+
opt = optim.Adamax([z], lr=opt_lr)
|
661 |
+
elif opt_name == "DiffGrad":
|
662 |
+
opt = DiffGrad([z], lr=opt_lr, eps=1e-9, weight_decay=1e-9) # NR: Playing for reasons
|
663 |
+
elif opt_name == "AdamP":
|
664 |
+
opt = AdamP([z], lr=opt_lr)
|
665 |
+
elif opt_name == "RAdam":
|
666 |
+
opt = optim.RAdam([z], lr=opt_lr)
|
667 |
+
elif opt_name == "RMSprop":
|
668 |
+
opt = optim.RMSprop([z], lr=opt_lr)
|
669 |
+
else:
|
670 |
+
print("Unknown optimiser. Are choices broken?")
|
671 |
+
opt = optim.Adam([z], lr=opt_lr)
|
672 |
+
return opt
|
673 |
+
|
674 |
+
opt = get_opt(args.optimiser, args.step_size)
|
675 |
+
|
676 |
+
|
677 |
+
# Output for the user
|
678 |
+
print('Using device:', device)
|
679 |
+
print('Optimising using:', args.optimiser)
|
680 |
+
|
681 |
+
if args.prompts:
|
682 |
+
print('Using text prompts:', args.prompts)
|
683 |
+
if args.image_prompts:
|
684 |
+
print('Using image prompts:', args.image_prompts)
|
685 |
+
if args.init_image:
|
686 |
+
print('Using initial image:', args.init_image)
|
687 |
+
if args.noise_prompt_weights:
|
688 |
+
print('Noise prompt weights:', args.noise_prompt_weights)
|
689 |
+
|
690 |
+
|
691 |
+
if args.seed is None:
|
692 |
+
seed = torch.seed()
|
693 |
+
else:
|
694 |
+
seed = args.seed
|
695 |
+
torch.manual_seed(seed)
|
696 |
+
print('Using seed:', seed)
|
697 |
+
|
698 |
+
|
699 |
+
# Vector quantize
|
700 |
+
def synth(z):
|
701 |
+
if gumbel:
|
702 |
+
z_q = vector_quantize(z.movedim(1, 3), model.quantize.embed.weight).movedim(3, 1)
|
703 |
+
else:
|
704 |
+
z_q = vector_quantize(z.movedim(1, 3), model.quantize.embedding.weight).movedim(3, 1)
|
705 |
+
return clamp_with_grad(model.decode(z_q).add(1).div(2), 0, 1)
|
706 |
+
|
707 |
+
|
708 |
+
#@torch.no_grad()
|
709 |
+
@torch.inference_mode()
|
710 |
+
def checkin(i, losses):
|
711 |
+
losses_str = ', '.join(f'{loss.item():g}' for loss in losses)
|
712 |
+
tqdm.write(f'i: {i}, loss: {sum(losses).item():g}, losses: {losses_str}')
|
713 |
+
out = synth(z)
|
714 |
+
info = PngImagePlugin.PngInfo()
|
715 |
+
info.add_text('comment', f'{args.prompts}')
|
716 |
+
TF.to_pil_image(out[0].cpu()).save(args.output, pnginfo=info)
|
717 |
+
|
718 |
+
|
719 |
+
def ascend_txt():
|
720 |
+
global i
|
721 |
+
out = synth(z)
|
722 |
+
iii = perceptor.encode_image(normalize(make_cutouts(out))).float()
|
723 |
+
|
724 |
+
result = []
|
725 |
+
|
726 |
+
if args.init_weight:
|
727 |
+
# result.append(F.mse_loss(z, z_orig) * args.init_weight / 2)
|
728 |
+
result.append(F.mse_loss(z, torch.zeros_like(z_orig)) * ((1/torch.tensor(i*2 + 1))*args.init_weight) / 2)
|
729 |
+
|
730 |
+
for prompt in pMs:
|
731 |
+
result.append(prompt(iii))
|
732 |
+
|
733 |
+
if args.make_video:
|
734 |
+
img = np.array(out.mul(255).clamp(0, 255)[0].cpu().detach().numpy().astype(np.uint8))[:,:,:]
|
735 |
+
img = np.transpose(img, (1, 2, 0))
|
736 |
+
imageio.imwrite('./steps/' + str(i) + '.png', np.array(img))
|
737 |
+
|
738 |
+
return result # return loss
|
739 |
+
|
740 |
+
|
741 |
+
def train(i):
|
742 |
+
opt.zero_grad(set_to_none=True)
|
743 |
+
lossAll = ascend_txt()
|
744 |
+
|
745 |
+
if i % args.display_freq == 0:
|
746 |
+
checkin(i, lossAll)
|
747 |
+
|
748 |
+
loss = sum(lossAll)
|
749 |
+
loss.backward()
|
750 |
+
opt.step()
|
751 |
+
|
752 |
+
#with torch.no_grad():
|
753 |
+
with torch.inference_mode():
|
754 |
+
z.copy_(z.maximum(z_min).minimum(z_max))
|
755 |
+
|
756 |
+
|
757 |
+
|
758 |
+
i = 0 # Iteration counter
|
759 |
+
j = 0 # Zoom video frame counter
|
760 |
+
p = 1 # Phrase counter
|
761 |
+
smoother = 0 # Smoother counter
|
762 |
+
this_video_frame = 0 # for video styling
|
763 |
+
|
764 |
+
# Messing with learning rate / optimisers
|
765 |
+
#variable_lr = args.step_size
|
766 |
+
#optimiser_list = [['Adam',0.075],['AdamW',0.125],['Adagrad',0.2],['Adamax',0.125],['DiffGrad',0.075],['RAdam',0.125],['RMSprop',0.02]]
|
767 |
+
|
768 |
+
# Do it
|
769 |
+
try:
|
770 |
+
with tqdm() as pbar:
|
771 |
+
while True:
|
772 |
+
# Change generated image
|
773 |
+
if args.make_zoom_video:
|
774 |
+
if i % args.zoom_frequency == 0:
|
775 |
+
out = synth(z)
|
776 |
+
|
777 |
+
# Save image
|
778 |
+
img = np.array(out.mul(255).clamp(0, 255)[0].cpu().detach().numpy().astype(np.uint8))[:,:,:]
|
779 |
+
img = np.transpose(img, (1, 2, 0))
|
780 |
+
imageio.imwrite('./steps/' + str(j) + '.png', np.array(img))
|
781 |
+
|
782 |
+
# Time to start zooming?
|
783 |
+
if args.zoom_start <= i:
|
784 |
+
# Convert z back into a Pil image
|
785 |
+
#pil_image = TF.to_pil_image(out[0].cpu())
|
786 |
+
|
787 |
+
# Convert NP to Pil image
|
788 |
+
pil_image = Image.fromarray(np.array(img).astype('uint8'), 'RGB')
|
789 |
+
|
790 |
+
# Zoom
|
791 |
+
if args.zoom_scale != 1:
|
792 |
+
pil_image_zoom = zoom_at(pil_image, sideX/2, sideY/2, args.zoom_scale)
|
793 |
+
else:
|
794 |
+
pil_image_zoom = pil_image
|
795 |
+
|
796 |
+
# Shift - https://pillow.readthedocs.io/en/latest/reference/ImageChops.html
|
797 |
+
if args.zoom_shift_x or args.zoom_shift_y:
|
798 |
+
# This one wraps the image
|
799 |
+
pil_image_zoom = ImageChops.offset(pil_image_zoom, args.zoom_shift_x, args.zoom_shift_y)
|
800 |
+
|
801 |
+
# Convert image back to a tensor again
|
802 |
+
pil_tensor = TF.to_tensor(pil_image_zoom)
|
803 |
+
|
804 |
+
# Re-encode
|
805 |
+
z, *_ = model.encode(pil_tensor.to(device).unsqueeze(0) * 2 - 1)
|
806 |
+
z_orig = z.clone()
|
807 |
+
z.requires_grad_(True)
|
808 |
+
|
809 |
+
# Re-create optimiser
|
810 |
+
opt = get_opt(args.optimiser, args.step_size)
|
811 |
+
|
812 |
+
# Next
|
813 |
+
j += 1
|
814 |
+
|
815 |
+
# Change text prompt
|
816 |
+
if args.prompt_frequency > 0:
|
817 |
+
if i % args.prompt_frequency == 0 and i > 0:
|
818 |
+
# In case there aren't enough phrases, just loop
|
819 |
+
if p >= len(all_phrases):
|
820 |
+
p = 0
|
821 |
+
|
822 |
+
pMs = []
|
823 |
+
args.prompts = all_phrases[p]
|
824 |
+
|
825 |
+
# Show user we're changing prompt
|
826 |
+
print(args.prompts)
|
827 |
+
|
828 |
+
for prompt in args.prompts:
|
829 |
+
txt, weight, stop = split_prompt(prompt)
|
830 |
+
embed = perceptor.encode_text(clip.tokenize(txt).to(device)).float()
|
831 |
+
pMs.append(Prompt(embed, weight, stop).to(device))
|
832 |
+
|
833 |
+
'''
|
834 |
+
# Smooth test
|
835 |
+
smoother = args.zoom_frequency * 15 # smoothing over x frames
|
836 |
+
variable_lr = args.step_size * 0.25
|
837 |
+
opt = get_opt(args.optimiser, variable_lr)
|
838 |
+
'''
|
839 |
+
|
840 |
+
p += 1
|
841 |
+
|
842 |
+
'''
|
843 |
+
if smoother > 0:
|
844 |
+
if smoother == 1:
|
845 |
+
opt = get_opt(args.optimiser, args.step_size)
|
846 |
+
smoother -= 1
|
847 |
+
'''
|
848 |
+
|
849 |
+
'''
|
850 |
+
# Messing with learning rate / optimisers
|
851 |
+
if i % 225 == 0 and i > 0:
|
852 |
+
variable_optimiser_item = random.choice(optimiser_list)
|
853 |
+
variable_optimiser = variable_optimiser_item[0]
|
854 |
+
variable_lr = variable_optimiser_item[1]
|
855 |
+
|
856 |
+
opt = get_opt(variable_optimiser, variable_lr)
|
857 |
+
print("New opt: %s, lr= %f" %(variable_optimiser,variable_lr))
|
858 |
+
'''
|
859 |
+
|
860 |
+
|
861 |
+
# Training time
|
862 |
+
train(i)
|
863 |
+
|
864 |
+
# Ready to stop yet?
|
865 |
+
if i == args.max_iterations:
|
866 |
+
if not args.video_style_dir:
|
867 |
+
# we're done
|
868 |
+
break
|
869 |
+
else:
|
870 |
+
if this_video_frame == (num_video_frames - 1):
|
871 |
+
# we're done
|
872 |
+
make_styled_video = True
|
873 |
+
break
|
874 |
+
else:
|
875 |
+
# Next video frame
|
876 |
+
this_video_frame += 1
|
877 |
+
|
878 |
+
# Reset the iteration count
|
879 |
+
i = -1
|
880 |
+
pbar.reset()
|
881 |
+
|
882 |
+
# Load the next frame, reset a few options - same filename, different directory
|
883 |
+
args.init_image = video_frame_list[this_video_frame]
|
884 |
+
print("Next frame: ", args.init_image)
|
885 |
+
|
886 |
+
if args.seed is None:
|
887 |
+
seed = torch.seed()
|
888 |
+
else:
|
889 |
+
seed = args.seed
|
890 |
+
torch.manual_seed(seed)
|
891 |
+
print("Seed: ", seed)
|
892 |
+
|
893 |
+
filename = os.path.basename(args.init_image)
|
894 |
+
args.output = os.path.join(cwd, "steps", filename)
|
895 |
+
|
896 |
+
# Load and resize image
|
897 |
+
img = Image.open(args.init_image)
|
898 |
+
pil_image = img.convert('RGB')
|
899 |
+
pil_image = pil_image.resize((sideX, sideY), Image.LANCZOS)
|
900 |
+
pil_tensor = TF.to_tensor(pil_image)
|
901 |
+
|
902 |
+
# Re-encode
|
903 |
+
z, *_ = model.encode(pil_tensor.to(device).unsqueeze(0) * 2 - 1)
|
904 |
+
z_orig = z.clone()
|
905 |
+
z.requires_grad_(True)
|
906 |
+
|
907 |
+
# Re-create optimiser
|
908 |
+
opt = get_opt(args.optimiser, args.step_size)
|
909 |
+
|
910 |
+
i += 1
|
911 |
+
pbar.update()
|
912 |
+
except KeyboardInterrupt:
|
913 |
+
pass
|
914 |
+
|
915 |
+
# All done :)
|
916 |
+
|
917 |
+
# Video generation
|
918 |
+
if args.make_video or args.make_zoom_video:
|
919 |
+
init_frame = 1 # Initial video frame
|
920 |
+
if args.make_zoom_video:
|
921 |
+
last_frame = j
|
922 |
+
else:
|
923 |
+
last_frame = i # This will raise an error if that number of frames does not exist.
|
924 |
+
|
925 |
+
length = args.video_length # Desired time of the video in seconds
|
926 |
+
|
927 |
+
min_fps = 10
|
928 |
+
max_fps = 60
|
929 |
+
|
930 |
+
total_frames = last_frame-init_frame
|
931 |
+
|
932 |
+
frames = []
|
933 |
+
tqdm.write('Generating video...')
|
934 |
+
for i in range(init_frame,last_frame):
|
935 |
+
temp = Image.open("./steps/"+ str(i) +'.png')
|
936 |
+
keep = temp.copy()
|
937 |
+
frames.append(keep)
|
938 |
+
temp.close()
|
939 |
+
|
940 |
+
if args.output_video_fps > 9:
|
941 |
+
# Hardware encoding and video frame interpolation
|
942 |
+
print("Creating interpolated frames...")
|
943 |
+
ffmpeg_filter = f"minterpolate='mi_mode=mci:me=hexbs:me_mode=bidir:mc_mode=aobmc:vsbmc=1:mb_size=8:search_param=32:fps={args.output_video_fps}'"
|
944 |
+
output_file = re.compile('\.png$').sub('.mp4', args.output)
|
945 |
+
try:
|
946 |
+
p = Popen(['ffmpeg',
|
947 |
+
'-y',
|
948 |
+
'-f', 'image2pipe',
|
949 |
+
'-vcodec', 'png',
|
950 |
+
'-r', str(args.input_video_fps),
|
951 |
+
'-i',
|
952 |
+
'-',
|
953 |
+
'-b:v', '10M',
|
954 |
+
'-vcodec', 'h264_nvenc',
|
955 |
+
'-pix_fmt', 'yuv420p',
|
956 |
+
'-strict', '-2',
|
957 |
+
'-filter:v', f'{ffmpeg_filter}',
|
958 |
+
'-metadata', f'comment={args.prompts}',
|
959 |
+
output_file], stdin=PIPE)
|
960 |
+
except FileNotFoundError:
|
961 |
+
print("ffmpeg command failed - check your installation")
|
962 |
+
for im in tqdm(frames):
|
963 |
+
im.save(p.stdin, 'PNG')
|
964 |
+
p.stdin.close()
|
965 |
+
p.wait()
|
966 |
+
else:
|
967 |
+
# CPU
|
968 |
+
fps = np.clip(total_frames/length,min_fps,max_fps)
|
969 |
+
output_file = re.compile('\.png$').sub('.mp4', args.output)
|
970 |
+
try:
|
971 |
+
p = Popen(['ffmpeg',
|
972 |
+
'-y',
|
973 |
+
'-f', 'image2pipe',
|
974 |
+
'-vcodec', 'png',
|
975 |
+
'-r', str(fps),
|
976 |
+
'-i',
|
977 |
+
'-',
|
978 |
+
'-vcodec', 'libx264',
|
979 |
+
'-r', str(fps),
|
980 |
+
'-pix_fmt', 'yuv420p',
|
981 |
+
'-crf', '17',
|
982 |
+
'-preset', 'veryslow',
|
983 |
+
'-metadata', f'comment={args.prompts}',
|
984 |
+
output_file], stdin=PIPE)
|
985 |
+
except FileNotFoundError:
|
986 |
+
print("ffmpeg command failed - check your installation")
|
987 |
+
for im in tqdm(frames):
|
988 |
+
im.save(p.stdin, 'PNG')
|
989 |
+
p.stdin.close()
|
990 |
+
p.wait()
|