!pip install diffusers transformers accelerate
from diffusers import DiffusionPipeline
savename = 'drive/MyDrive/mj1.png'
id = 'midjourney-community/midjourney-mini'
device = 'cuda'
pipe = DiffusionPipeline.from_pretrained(id)
pipe = pipe.to(device)
prompt = 'rabbit, Studio Ghibli --ar 16:9'
image = pipe(prompt).images[0]
image.save(savename)
アップロードするファイルは こちらから ダウンロードできます。
[
{ "times": 3,
"prompt": "good quality photography of a singular red apple on simple background for e-commerce site --stop 75"
},
{ "times": 1,
"prompts": [
"Detailing oil painting of The great white castle on deep forest landscape ",
"by CASPAR DAVID FRIEDRICH and CLAUDE LORRAIN,perfect lighting, ",
"golden hour, taken with Canon 5D Mk4 --ar 16:9"
]
},
{ "times": 2,
"prompt": "(Vintage Victorian Design Style, Elsa Beskow, Colored Pencils and Watercolors, Softly Lit Lanterns)"
}
]
from diffusers import DiffusionPipeline
class ImgGenAI:
def __init__(self, id, device = 'cuda'):
self.pipe = DiffusionPipeline.from_pretrained(id, use_safetensors=False)
self.pipe = self.pipe.to(device)
def generate(self, prompt):
image = self.pipe(prompt).images[0]
return image
import json
import sys
class Orders:
def __init__(self, filename):
with open(filename, 'r') as f:
jsonData = json.load(f)
self.orders = []
for order in jsonData:
if 'times' in order:
t = int(order['times'])
else:
keyword = 'times'
print(f'必須キー[ {keyword} ]が存在しません!')
sys.exit(-1)
if 'prompt' in order:
prompt = order['prompt']
elif 'prompts' in order:
prompt = ''
for p in order['prompts']:
prompt = prompt + p
if prompt[-1] != ' ':
prompt = prompt + ' '
else:
keywords = ['prompt', 'prompts']
print(f'必須キー[ {keywords[0]} ] もしくは [ {keywords[1]} ]が存在しません!')
sys.exit(-1)
for i in range(t):
self.orders.append(prompt)
def __len__(self):
return len(self.orders)
def __getitem__(self, index):
return self.orders[index]
!pip install diffusers transformers accelerate
import sys
sys.path.append('drive/MyDrive/libAI')
from Orders import Orders
from ImgGenAI import ImgGenAI
import datetime
import pytz
import os
def createFolderName():
dt_now = datetime.datetime.now(pytz.timezone('Asia/Tokyo'))
s = dt_now.strftime('%Y%m%d_%H%M%S')
return s
target_json = 'drive/MyDrive/gen1.json'
id = 'midjourney-community/midjourney-mini'
folder_name = 'drive/MyDrive/result/' + createFolderName()
os.makedirs(folder_name)
ai = ImgGenAI(id)
order = Orders(target_json)
n = len(order)
for i in range(n):
save_name = os.path.join(folder_name, f'mj{i+1}.png')
img = ai.generate(order[i])
img.save(save_name)
import torch
x = torch.ones(2, 3, requires_grad=True)
print(x)
y = x + 2
print(y)
print(y.grad_fn)
z = y * 3
print(z)
out = z.mean()
print(out)
a = torch.tensor([1.0], requires_grad=True)
b = a * 2 # bの変化量はaの2倍
b.backward() # 逆伝播
print(a.grad) # aの勾配（aの変化に対するbの変化の割合）
def calc(a):
b = a*2 + 1
c = b*b
d = c/(c + 2)
e = d.mean()
return e
x = [1.0, 2.0, 3.0]
x = torch.tensor(x, requires_grad=True)
y = calc(x)
y.backward()
print(x.grad) # xの勾配（xの各値の変化に対するyの変化の割合）
delta = 0.001 #xの微小変化
x = [1.0, 2.0, 3.0]
x = torch.tensor(x)
y = calc(x)
x_1 = [1.0+delta, 2.0, 3.0]
x_1 = torch.tensor(x_1)
y_1 = calc(x_1)
x_2 = [1.0, 2.0+delta, 3.0]
x_2 = torch.tensor(x_2)
y_2 = calc(x_2)
x_3 = [1.0, 2.0, 3.0+delta]
x_3 = torch.tensor(x_3)
y_3 = calc(x_3)
# 勾配の計算 (yの微小変化)/(xの微小変化)
grad_1 = (y_1 - y) / delta
grad_2 = (y_2 - y) / delta
grad_3 = (y_3 - y) / delta
grads = torch.stack((grad_1, grad_2, grad_3)) # Tensorを結合
print(grads)
from torchvision.datasets import MNIST
from torchvision import transforms
img_size = 28 # 画像の高さと幅
# 訓練データを取得
mnist_train = MNIST("./data",
train=True, download=True,
transform=transforms.ToTensor()) # Tensorに変換
# テストデータの取得
mnist_test = MNIST("./data",
train=False, download=True,
transform=transforms.ToTensor()) # Tensorに変換
print("訓練データの数:", len(mnist_train), "テストデータの数:", len(mnist_test))
from torch.utils.data import DataLoader
# DataLoaderの設定
batch_size = 256 # バッチサイズ
train_loader = DataLoader(mnist_train,
batch_size=batch_size,
shuffle=True)
test_loader = DataLoader(mnist_test,
batch_size=batch_size,
shuffle=False)
import torch.nn as nn
class Net(nn.Module):
def __init__(self):
super().__init__()
self.fc1 = nn.Linear(img_size*img_size, 1024) # 全結合層
self.fc2 = nn.Linear(1024, 512)
self.fc3 = nn.Linear(512, 10)
self.relu = nn.ReLU() # ReLU 学習するパラメータがないので使い回しできる
def forward(self, x):
x = x.view(-1, img_size*img_size) # (バッチサイズ, 入力の数): 画像を1次元に変換
x = self.relu(self.fc1(x))
x = self.relu(self.fc2(x))
x = self.fc3(x)
return x
net = Net()
net.cuda() # GPU対応
print(net)
from torch import optim
# 交差エントロピー誤差関数
loss_fnc = nn.CrossEntropyLoss()
# SGD
optimizer = optim.SGD(net.parameters(), lr=0.01)
# 損失のログ
record_loss_train = []
record_loss_test = []
# 学習
for i in range(10): # 10エポック学習
net.train() # 訓練モード
loss_train = 0
for j, (x, t) in enumerate(train_loader): # ミニバッチ（x, t）を取り出す
x, t = x.cuda(), t.cuda() # GPU対応
y = net(x)
loss = loss_fnc(y, t)
loss_train += loss.item()
optimizer.zero_grad()
loss.backward()
optimizer.step()
loss_train /= j+1
record_loss_train.append(loss_train)
net.eval() # 評価モード
loss_test = 0
for j, (x, t) in enumerate(test_loader): # ミニバッチ（x, t）を取り出す
x, t = x.cuda(), t.cuda() # GPU対応
y = net(x)
loss = loss_fnc(y, t)
loss_test += loss.item()
loss_test /= j+1
record_loss_test.append(loss_test)
if i%1 == 0:
print("Epoch:", i, "Loss_Train:", loss_train, "Loss_Test:", loss_test)
import matplotlib.pyplot as plt
plt.plot(range(len(record_loss_train)), record_loss_train, label="Train")
plt.plot(range(len(record_loss_test)), record_loss_test, label="Test")
plt.legend()
plt.xlabel("Epochs")
plt.ylabel("Error")
plt.show()
correct = 0
total = 0
net.eval() # 評価モード
for i, (x, t) in enumerate(test_loader):
x, t = x.cuda(), t.cuda() # GPU対応
y = net(x)
correct += (y.argmax(1) == t).sum().item()
total += len(x)
print("正解率:", str(correct/total*100) + "%")
from torchvision.datasets import CIFAR10
import torchvision.transforms as transforms
from torch.utils.data import DataLoader
import matplotlib.pyplot as plt
cifar10_data = CIFAR10(root="./data",
train=False,download=True,
transform=transforms.ToTensor())
cifar10_classes = ["airplane", "automobile", "bird", "cat", "deer",
"dog", "frog", "horse", "ship", "truck"]
print("データの数:", len(cifar10_data))
n_image = 25 # 表示する画像の数
cifar10_loader = DataLoader(cifar10_data, batch_size=n_image, shuffle=True)
dataiter = iter(cifar10_loader) # イテレータ: 要素を順番に取り出せるようにする
images, labels = next(dataiter) # 最初のバッチを取り出す
plt.figure(figsize=(10,10)) # 画像の表示サイズ
for i in range(n_image):
ax = plt.subplot(5,5,i+1)
ax.imshow(images[i].permute(1, 2, 0)) # チャンネルを一番後の次元に
label = cifar10_classes[labels[i]]
ax.set_title(label)
ax.get_xaxis().set_visible(False) # 軸を非表示に
ax.get_yaxis().set_visible(False)
plt.show()
transform = transforms.Compose([transforms.RandomAffine((-45, 45), scale=(0.5, 1.5)), # 回転とリサイズ
transforms.ToTensor()])
cifar10_data = CIFAR10(root="./data",
train=False,download=True,
transform=transform)
cifar10_loader = DataLoader(cifar10_data, batch_size=n_image, shuffle=True)
dataiter = iter(cifar10_loader)
images, labels = next(dataiter)
plt.figure(figsize=(10,10)) # 画像の表示サイズ
for i in range(n_image):
ax = plt.subplot(5,5,i+1)
ax.imshow(images[i].permute(1, 2, 0))
label = cifar10_classes[labels[i]]
ax.set_title(label)
ax.get_xaxis().set_visible(False)
ax.get_yaxis().set_visible(False)
plt.show()
transform = transforms.Compose([transforms.RandomAffine((0, 0), translate=(0.5, 0.5)), # 上下左右へのシフト
transforms.ToTensor()])
cifar10_data = CIFAR10(root="./data",
train=False,download=True,
transform=transform)
cifar10_loader = DataLoader(cifar10_data, batch_size=n_image, shuffle=True)
dataiter = iter(cifar10_loader)
images, labels = next(dataiter)
plt.figure(figsize=(10,10)) # 画像の表示サイズ
for i in range(n_image):
ax = plt.subplot(5,5,i+1)
ax.imshow(images[i].permute(1, 2, 0))
label = cifar10_classes[labels[i]]
ax.set_title(label)
ax.get_xaxis().set_visible(False)
ax.get_yaxis().set_visible(False)
plt.show()
transform = transforms.Compose([transforms.RandomHorizontalFlip(p=0.5), # 左右反転
transforms.RandomVerticalFlip(p=0.5), # 上下反転
transforms.ToTensor()])
cifar10_data = CIFAR10(root="./data",
train=False,download=True,
transform=transform)
cifar10_loader = DataLoader(cifar10_data, batch_size=n_image, shuffle=True)
dataiter = iter(cifar10_loader)
images, labels = next(dataiter)
plt.figure(figsize=(10,10)) # 画像の表示サイズ
for i in range(n_image):
ax = plt.subplot(5,5,i+1)
ax.imshow(images[i].permute(1, 2, 0))
label = cifar10_classes[labels[i]]
ax.set_title(label)
ax.get_xaxis().set_visible(False)
ax.get_yaxis().set_visible(False)
plt.show()
transform = transforms.Compose([transforms.ToTensor(),
transforms.RandomErasing(p=0.5)]) # 一部を消去
cifar10_data = CIFAR10(root="./data",
train=False,download=True,
transform=transform)
cifar10_loader = DataLoader(cifar10_data, batch_size=n_image, shuffle=True)
dataiter = iter(cifar10_loader)
images, labels = next(dataiter)
plt.figure(figsize=(10,10)) # 画像の表示サイズ
for i in range(n_image):
ax = plt.subplot(5,5,i+1)
ax.imshow(images[i].permute(1, 2, 0))
label = cifar10_classes[labels[i]]
ax.set_title(label)
ax.get_xaxis().set_visible(False)
ax.get_yaxis().set_visible(False)
plt.show()
from torchvision.datasets import CIFAR10
import torchvision.transforms as transforms
from torch.utils.data import DataLoader
cifar10_classes = ["airplane", "automobile", "bird", "cat", "deer",
"dog", "frog", "horse", "ship", "truck"]
affine = transforms.RandomAffine((-30, 30), scale=(0.8, 1.2)) # 回転とリサイズ
flip = transforms.RandomHorizontalFlip(p=0.5) # 左右反転
normalize = transforms.Normalize((0.0, 0.0, 0.0), (1.0, 1.0, 1.0)) # 平均値を0、標準偏差を1に
to_tensor = transforms.ToTensor()
transform_train = transforms.Compose([affine, flip, to_tensor, normalize])
transform_test = transforms.Compose([to_tensor, normalize])
cifar10_train = CIFAR10("./data", train=True, download=True, transform=transform_train)
cifar10_test = CIFAR10("./data", train=False, download=True, transform=transform_test)
# DataLoaderの設定
batch_size = 64
train_loader = DataLoader(cifar10_train, batch_size=batch_size, shuffle=True)
test_loader = DataLoader(cifar10_test, batch_size=batch_size, shuffle=False)
import torch.nn as nn
class Net(nn.Module):
def __init__(self):
super().__init__()
self.conv1 = nn.Conv2d(3, 8, 5) # 畳み込み層:(入力チャンネル数, フィルタ数、フィルタサイズ)
self.relu = nn.ReLU() # ReLU
self.pool = nn.MaxPool2d(2, 2) # プーリング層:（領域のサイズ, 領域の間隔）
self.conv2 = nn.Conv2d(8, 16, 5)
self.fc1 = nn.Linear(16*5*5, 256) # 全結合層
self.dropout = nn.Dropout(p=0.5) # ドロップアウト:(p=ドロップアウト率)
self.fc2 = nn.Linear(256, 10)
def forward(self, x):
x = self.relu(self.conv1(x))
x = self.pool(x)
x = self.relu(self.conv2(x))
x = self.pool(x)
x = x.view(-1, 16*5*5)
x = self.relu(self.fc1(x))
x = self.dropout(x)
x = self.fc2(x)
return x
net = Net()
net.cuda() # GPU対応
print(net)
from torch import optim
# 交差エントロピー誤差関数
loss_fnc = nn.CrossEntropyLoss()
# 最適化アルゴリズム
optimizer = optim.Adam(net.parameters())
# 損失のログ
record_loss_train = []
record_loss_test = []
# 学習
for i in range(20): # 20エポック学習
net.train() # 訓練モード
loss_train = 0
for j, (x, t) in enumerate(train_loader): # ミニバッチ（x, t）を取り出す
x, t = x.cuda(), t.cuda() # GPU対応
y = net(x)
loss = loss_fnc(y, t)
loss_train += loss.item()
optimizer.zero_grad()
loss.backward()
optimizer.step()
loss_train /= j+1
record_loss_train.append(loss_train)
net.eval() # 評価モード
loss_test = 0
for j, (x, t) in enumerate(test_loader): # ミニバッチ（x, t）を取り出す
x, t = x.cuda(), t.cuda()
y = net(x)
loss = loss_fnc(y, t)
loss_test += loss.item()
loss_test /= j+1
record_loss_test.append(loss_test)
if i%1 == 0:
print("Epoch:", i, "Loss_Train:", loss_train, "Loss_Test:", loss_test)
import matplotlib.pyplot as plt
plt.plot(range(len(record_loss_train)), record_loss_train, label="Train")
plt.plot(range(len(record_loss_test)), record_loss_test, label="Test")
plt.legend()
plt.xlabel("Epochs")
plt.ylabel("Error")
plt.show()
correct = 0
total = 0
net.eval() # 評価モード
for i, (x, t) in enumerate(test_loader):
x, t = x.cuda(), t.cuda() # GPU対応
y = net(x)
correct += (y.argmax(1) == t).sum().item()
total += len(x)
print("正解率:", str(correct/total*100) + "%")
cifar10_loader = DataLoader(cifar10_test, batch_size=1, shuffle=True)
dataiter = iter(cifar10_loader)
images, labels = next(dataiter) # サンプルを1つだけ取り出す
plt.imshow(images[0].permute(1, 2, 0)) # チャンネルを一番後ろに
plt.tick_params(labelbottom=False, labelleft=False, bottom=False, left=False) # ラベルとメモリを非表示に
plt.show()
net.eval() # 評価モード
x, t = images.cuda(), labels.cuda() # GPU対応
y = net(x)
print("正解:", cifar10_classes[labels[0]],
"予測結果:", cifar10_classes[y.argmax().item()])