File size: 5,894 Bytes
5fc3f91 |
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 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 |
from threading import Lock
import math
import os
import random
from diffusers import StableDiffusionPipeline
from diffusers.models.attention import get_global_heat_map, clear_heat_maps
from matplotlib import pyplot as plt
import gradio as gr
import torch
import torch.nn.functional as F
import spacy
if not os.environ.get('NO_DOWNLOAD_SPACY'):
spacy.cli.download('en_core_web_sm')
model_id = "CompVis/stable-diffusion-v1-4"
device = "cuda"
gen = torch.Generator(device='cuda')
gen.manual_seed(12758672)
orig_state = gen.get_state()
pipe = StableDiffusionPipeline.from_pretrained(model_id, use_auth_token=True).to(device)
lock = Lock()
nlp = spacy.load('en_core_web_sm')
def expand_m(m, n: int = 1, o=512, mode='bicubic'):
m = m.unsqueeze(0).unsqueeze(0) / n
m = F.interpolate(m.float().detach(), size=(o, o), mode='bicubic', align_corners=False)
m = (m - m.min()) / (m.max() - m.min() + 1e-8)
m = m.cpu().detach()
return m
@torch.no_grad()
def predict(prompt, inf_steps, threshold):
global lock
with torch.cuda.amp.autocast(), lock:
try:
plt.close('all')
except:
pass
gen.set_state(orig_state.clone())
clear_heat_maps()
out = pipe(prompt, guidance_scale=7.5, height=512, width=512, do_intermediates=False, generator=gen,
num_inference_steps=int(inf_steps))
heat_maps = get_global_heat_map()
with torch.cuda.amp.autocast(dtype=torch.float32):
m = 0
n = 0
w = ''
w_idx = 0
fig, ax = plt.subplots()
ax.imshow(out.images[0].cpu().float().detach().permute(1, 2, 0).numpy())
ax.set_xticks([])
ax.set_yticks([])
fig1, axs1 = plt.subplots(math.ceil(len(out.words) / 4), 4) # , figsize=(20, 20))
fig2, axs2 = plt.subplots(math.ceil(len(out.words) / 4), 4) # , figsize=(20, 20))
for idx in range(len(out.words) + 1):
if idx == 0:
continue
word = out.words[idx - 1]
m += heat_maps[idx]
n += 1
w += word
if '</w>' not in word:
continue
else:
mplot = expand_m(m, n)
spotlit_im = out.images[0].cpu().float().detach()
w = w.replace('</w>', '')
spotlit_im2 = torch.cat((spotlit_im, (1 - mplot.squeeze(0)).pow(1)), dim=0)
if len(out.words) <= 4:
a1 = axs1[w_idx % 4]
a2 = axs2[w_idx % 4]
else:
a1 = axs1[w_idx // 4, w_idx % 4]
a2 = axs2[w_idx // 4, w_idx % 4]
a1.set_xticks([])
a1.set_yticks([])
a1.imshow(mplot.squeeze().numpy(), cmap='jet')
a1.imshow(spotlit_im2.permute(1, 2, 0).numpy())
a1.set_title(w)
mask = torch.ones_like(mplot)
mask[mplot < threshold * mplot.max()] = 0
im2 = spotlit_im * mask.squeeze(0)
a2.set_xticks([])
a2.set_yticks([])
a2.imshow(im2.permute(1, 2, 0).numpy())
a2.set_title(w)
m = 0
n = 0
w_idx += 1
w = ''
for idx in range(w_idx, len(axs1.flatten())):
fig1.delaxes(axs1.flatten()[idx])
fig2.delaxes(axs2.flatten()[idx])
return fig, fig1, fig2
def set_prompt(prompt):
return prompt
with gr.Blocks() as demo:
md = '''# DAAM: Attention Maps for Interpreting Stable Diffusion
Check out the paper: [What the DAAM: Interpreting Stable Diffusion Using Cross Attention](http://arxiv.org/abs/2210.04885). Note that, due to server costs, this demo will transition to HuggingFace Spaces on 2022-10-20.
'''
gr.Markdown(md)
with gr.Row():
with gr.Column():
dropdown = gr.Dropdown([
'A monkey wearing a halloween costume',
'A smiling, red cat chewing gum',
# 'Doing research at Comcast Applied AI labs',
# 'Professor Jimmy Lin from the University of Waterloo',
# 'Yann Lecun teaching machine learning on a chalkboard',
# 'A cat eating cake for her birthday',
# 'Steak and dollars on a plate',
# 'A fox, a dog, and a wolf in a field'
], label='Examples', value='An angry, bald man doing research')
text = gr.Textbox(label='Prompt', value='An angry, bald man doing research')
slider1 = gr.Slider(15, 35, value=25, interactive=True, step=1, label='Inference steps')
slider2 = gr.Slider(0, 1.0, value=0.4, interactive=True, step=0.05, label='Threshold (tau)')
submit_btn = gr.Button('Submit')
with gr.Tab('Original Image'):
p0 = gr.Plot()
with gr.Tab('Soft DAAM Maps'):
p1 = gr.Plot()
with gr.Tab('Hard DAAM Maps'):
p2 = gr.Plot()
submit_btn.click(fn=predict, inputs=[text, slider1, slider2], outputs=[p0, p1, p2])
dropdown.change(set_prompt, dropdown, text)
dropdown.update()
# ADDED PART
# import portpicker
# port = portpicker.pick_unused_port()
# select_ip = "0.0.0.0:"+str(port)
# print("Port: ", port)
# from IPython.display import Javascript
# def show_port(port, height=400):
# display(Javascript("""
# (async ()=>{
# fm = document.createElement('iframe')
# fm.src = await google.colab.kernel.proxyPort(%s)
# fm.width = '95%%'
# fm.height = '%d'
# fm.frameBorder = 0
# document.body.append(fm)
# })();
# """ % (port, height)))
# get_ipython().system_raw(f'python3 -m http.server {port} &')
# show_port(port)
###
demo.launch(share=True)
# demo.launch(server_name='0.0.0.0', server_port=port)
|