Skip to content

Commit e0be682

Browse files
shunk031shunk031
committed
update
1 parent 728b911 commit e0be682

File tree

2 files changed

+383
-0
lines changed

2 files changed

+383
-0
lines changed

‎scripts/1-2_text-to-image-generation.py

Lines changed: 277 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,277 @@
1+
# ---
2+
# jupyter:
3+
# jupytext:
4+
# text_representation:
5+
# extension: .py
6+
# format_name: percent
7+
# format_version: '1.3'
8+
# jupytext_version: 1.16.7
9+
# kernelspec:
10+
# display_name: Python 3 (ipykernel)
11+
# language: python
12+
# name: python3
13+
# ---
14+
15+
# %% [markdown]
16+
# # Text-to-Image 手法の実践
17+
18+
# %% [markdown]
19+
# [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/py-img-gen/python-image-generation/blob/main/notebooks/1-2_text-to-image-generation.ipynb)
20+
21+
# %% [markdown]
22+
# ## 準備
23+
24+
# %%
25+
# !pip install -qq py-img-gen
26+
27+
# %%
28+
import torch
29+
30+
# GPU が使用できる場合は GPU (= cuda) を指定し、
31+
# そうでない場合は CPU を指定
32+
device = torch.device(
33+
"cuda" if torch.cuda.is_available() else "cpu"
34+
)
35+
# 通常は単精度 (float32) を使用するが、
36+
# メモリ使用量削減のため半精度 (float16)を使用
37+
dtype = torch.float16
38+
# 生成結果の再現性を確保するためにシード値を設定
39+
seed = 42
40+
41+
# %%
42+
import logging
43+
44+
# error ログのみを表示する
45+
logger_name = "diffusers.pipelines.pipeline_utils"
46+
logging.getLogger(logger_name).setLevel(logging.ERROR)
47+
48+
# %% [markdown]
49+
# ## Stable Diffusion を扱うパイプラインの構築
50+
51+
# %%
52+
from diffusers import StableDiffusionPipeline
53+
54+
model_id = "stable-diffusion-v1-5/stable-diffusion-v1-5"
55+
56+
pipe = StableDiffusionPipeline.from_pretrained(
57+
model_id, torch_dtype=dtype
58+
)
59+
60+
# %%
61+
print(f"Move pipeline to {device}")
62+
63+
pipe = pipe.to(device)
64+
65+
# %%
66+
# 訳: 宇宙飛行士が馬に乗っている写真
67+
text = "a photograph of an astronaut riding a horse"
68+
69+
# 画像を生成
70+
output = pipe(prompt=text)
71+
72+
# ここで image は
73+
# pillow (https://pillow.readthedocs.io/en/stable) 形式
74+
image = output.images[0]
75+
76+
# 画像を表示
77+
image
78+
79+
# %%
80+
# 乱数生成器に指定されたシード値を設定
81+
generator = torch.manual_seed(seed)
82+
83+
# pipe の引数である generator に上記の乱数生成器を渡して画像を生成
84+
output = pipe(prompt=text, generator=generator)
85+
image = output.images[0]
86+
image # 画像を表示
87+
88+
# %%
89+
generator = torch.manual_seed(seed)
90+
91+
# 推論時のステップ数である num_inference_steps を 15 に設定 (デフォルトは 50)
92+
output = pipe(
93+
prompt=text, generator=generator, num_inference_steps=15
94+
)
95+
image = output.images[0]
96+
image # 画像を表示
97+
98+
# %%
99+
from diffusers.utils import make_image_grid
100+
101+
text = "a photograph of an astronaut riding a horse"
102+
103+
num_rows, num_cols = 4, 3 # 行数・列数を指定
104+
num_images = num_rows * num_cols # 生成画像数
105+
106+
output = pipe(prompt=text, num_images_per_prompt=num_images)
107+
108+
# make_image_grid 関数を使用してグリッド上に複数生成画像を表示
109+
make_image_grid(
110+
images=output.images, rows=num_rows, cols=num_cols
111+
)
112+
113+
# %% [markdown]
114+
# ## Stable Diffusion v1 による画像生成
115+
116+
# %%
117+
# 吾輩は猫である。名前はまだ無い。（夏目漱石「吾輩は猫である」冒頭より）
118+
text = "I am a cat. As yet I have no name."
119+
120+
# シード値を固定して画像を生成
121+
output = pipe(
122+
prompt=text, generator=torch.manual_seed(seed)
123+
)
124+
image = output.images[0]
125+
image # 画像を表示
126+
127+
# %%
128+
# 国境の長いトンネルを抜けると雪国であった。（川端康成「雪国」冒頭より）
129+
text = "The train came out of the long tunnel into the snow country."
130+
131+
# シード値を固定して画像を生成
132+
output = pipe(
133+
prompt=text, generator=torch.manual_seed(seed)
134+
)
135+
image = output.images[0]
136+
image # 画像を表示
137+
138+
# %%
139+
# 春はあけぼの、ようよう白く成りゆく山際、少し明かりて、紫だちたる雲の細くたなびきたる。
140+
# （清少納言「枕草子」冒頭より）
141+
text = "In the dawn of spring, the mountains are turning white, and the purple clouds are trailing thinly with a little light"
142+
143+
output = pipe(
144+
prompt=text, generator=torch.manual_seed(seed)
145+
)
146+
image = output.images[0]
147+
image # 画像を表示
148+
149+
# %%
150+
import gc
151+
152+
pipe = pipe.to("cpu")
153+
del pipe
154+
gc.collect()
155+
torch.cuda.empty_cache()
156+
157+
158+
# %% [markdown]
159+
# ## Stable Diffusion v2 による画像生成
160+
161+
# %%
162+
model_id = "stabilityai/stable-diffusion-2"
163+
164+
pipe = StableDiffusionPipeline.from_pretrained(
165+
model_id, torch_dtype=dtype
166+
)
167+
pipe = pipe.to(device)
168+
169+
# %%
170+
pipe.enable_attention_slicing()
171+
172+
# %%
173+
text = "a photograph of an astronaut riding a horse"
174+
175+
generator = torch.manual_seed(seed)
176+
output = pipe(prompt=text, generator=generator)
177+
image = output.images[0]
178+
image # 画像を表示
179+
180+
# %%
181+
# 吾輩は猫である。名前はまだ無い。
182+
text = "I am a cat. As yet I have no name."
183+
184+
output = pipe(
185+
prompt=text, generator=torch.manual_seed(seed)
186+
)
187+
image = output.images[0]
188+
image # 画像を表示
189+
190+
# %%
191+
# 国境の長いトンネルを抜けると雪国であった。
192+
text = "The train came out of the long tunnel into the snow country."
193+
194+
output = pipe(
195+
prompt=text, generator=torch.manual_seed(seed)
196+
)
197+
image = output.images[0]
198+
image # 画像を表示
199+
200+
# %%
201+
# 春はあけぼの、ようよう白く成りゆく山際、少し明かりて、紫だちたる雲の細くたなびきたる。
202+
text = "In the dawn of spring, the mountains are turning white, and the purple clouds are trailing thinly with a little light"
203+
204+
output = pipe(
205+
prompt=text, generator=torch.manual_seed(seed)
206+
)
207+
image = output.images[0]
208+
image # 画像を表示
209+
210+
# %%
211+
pipe = pipe.to("cpu")
212+
del pipe
213+
gc.collect()
214+
torch.cuda.empty_cache()
215+
216+
# %% [markdown]
217+
# ## waifu-diffusion による画像生成
218+
219+
# %%
220+
pipe = StableDiffusionPipeline.from_pretrained(
221+
"hakurei/waifu-diffusion", torch_dtype=dtype
222+
)
223+
pipe = pipe.to(device)
224+
225+
# %%
226+
text = "1girl, aqua eyes, baseball cap, blonde hair, closed mouth, earrings, green background, hat, hoop earrings, jewelry, looking at viewer, shirt, short hair, simple background, solo, upper body, yellow shirt"
227+
228+
num_rows, num_cols = 4, 3
229+
num_images = num_rows * num_cols
230+
231+
generator = torch.manual_seed(seed)
232+
output = pipe(
233+
prompt=text,
234+
generator=generator,
235+
num_images_per_prompt=num_images,
236+
)
237+
238+
make_image_grid(output.images, rows=num_rows, cols=num_cols)
239+
240+
# %%
241+
pipe = pipe.to("cpu")
242+
del pipe
243+
gc.collect()
244+
torch.cuda.empty_cache()
245+
246+
247+
# %% [markdown]
248+
# ## nitro-diffusion による画像生成
249+
250+
# %%
251+
pipe = StableDiffusionPipeline.from_pretrained(
252+
"nitrosocke/nitro-diffusion", torch_dtype=dtype
253+
)
254+
pipe = pipe.to(device)
255+
256+
# %%
257+
text = (
258+
"archer arcane style magical princess with golden hair"
259+
)
260+
261+
num_rows, num_cols = 4, 3
262+
num_images = num_rows * num_cols
263+
264+
generator = torch.manual_seed(seed)
265+
output = pipe(
266+
prompt=text,
267+
generator=generator,
268+
num_images_per_prompt=num_images,
269+
)
270+
271+
make_image_grid(output.images, rows=num_rows, cols=num_cols)
272+
273+
# %%
274+
pipe = pipe.to("cpu")
275+
del pipe
276+
gc.collect()
277+
torch.cuda.empty_cache()

‎scripts/2-2_activation-function.py

Lines changed: 106 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,106 @@
1+
# ---
2+
# jupyter:
3+
# jupytext:
4+
# text_representation:
5+
# extension: .py
6+
# format_name: percent
7+
# format_version: '1.3'
8+
# jupytext_version: 1.16.7
9+
# kernelspec:
10+
# display_name: Python 3 (ipykernel)
11+
# language: python
12+
# name: python3
13+
# ---
14+
15+
# %% [markdown]
16+
# # 活性化関数の可視化
17+
18+
# %% [markdown]
19+
# [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/py-img-gen/python-image-generation/blob/main/notebooks/2-2_activation-function.ipynb)
20+
21+
# %% [markdown]
22+
# ## 準備
23+
24+
# %%
25+
# !pip install -qq py-img-gen
26+
27+
# %%
28+
import pathlib
29+
30+
current_dir = pathlib.Path.cwd()
31+
project_dir = current_dir / "data" / "activation-function"
32+
project_dir.mkdir(parents=True, exist_ok=True)
33+
34+
# %%
35+
import matplotlib.pyplot as plt
36+
import numpy as np
37+
38+
plt.style.use("ggplot")
39+
plt.rcParams["font.size"] = 20
40+
41+
42+
def visualize(
43+
ax,
44+
x: np.ndarray,
45+
y: np.ndarray,
46+
bottom: float,
47+
top: float,
48+
):
49+
ax.set_xlim(-10, 10)
50+
ax.set_ylim(bottom, top)
51+
ax.set_xlabel(r"$x$")
52+
ax.set_ylabel(r"$y$")
53+
ax.plot(x, y, linewidth=4)
54+
return ax
55+
56+
57+
# %%
58+
x = np.arange(-10, 10, 0.01)
59+
x
60+
61+
62+
# %% [markdown]
63+
# ## シグモイド関数
64+
65+
# %%
66+
def sigmoid_function(x):
67+
return 1.0 / (1.0 + np.exp(-x))
68+
69+
70+
y = sigmoid_function(x)
71+
72+
fig, ax = plt.subplots(figsize=(5, 4))
73+
ax = visualize(ax, x, y, bottom=-0.02, top=1.02)
74+
fig.savefig(project_dir / "sigmoid.pdf")
75+
76+
77+
# %% [markdown]
78+
# ## tanh 関数
79+
80+
# %%
81+
def tanh_function(x):
82+
return (np.exp(x) - np.exp(-x)) / (
83+
np.exp(x) + np.exp(-x)
84+
)
85+
86+
87+
y = tanh_function(x)
88+
89+
fig, ax = plt.subplots(figsize=(5, 4))
90+
ax = visualize(ax, x, y, bottom=-1.05, top=1.05)
91+
fig.savefig(project_dir / "tanh.pdf")
92+
93+
94+
# %% [markdown]
95+
# ## ReLU 関数
96+
97+
# %%
98+
def relu_function(x):
99+
return np.maximum(0, x)
100+
101+
102+
y = relu_function(x)
103+
104+
fig, ax = plt.subplots(figsize=(5, 4))
105+
ax = visualize(ax, x, y, bottom=-0.2, top=10.2)
106+
fig.savefig(project_dir / "relu.pdf")

0 commit comments

Comments
 (0)