딥러닝, 야 너도 할 수 있어(feat. PyTorch)
1 like591 views
The document discusses the development of machine learning models using PyTorch, including details on neural network architectures such as linear models, convolutional networks, and the use of activation functions like sigmoid and ReLU. It highlights the process of model training and evaluation on a food dataset sourced from an external file, along with the utilization of data loaders for handling batches of images. Additionally, it references the adaptation of pre-trained models for custom outputs tailored to the specific dataset.
![𝝈 𝒙 = 𝑠𝑖𝑔𝑚𝑜𝑖𝑑 𝑥 =
1
1 + e]
𝐇 𝒙 = 𝐚𝐜𝐭𝐢𝐯𝐚𝐭𝐢𝐨𝐧_𝐟𝐧(𝒘 𝟎 + 𝒘 𝟏 𝒙 𝟏 + ⋯ + 𝒘 𝒏 𝒙 𝒏)
𝑥f
bias(wl)
𝐻 𝑥
𝑥(
𝚺 𝝈
𝑥&
𝝈 𝒙 = 𝑟𝑒𝑙𝑢 𝑥 = r
𝑥 = 0 (𝑥 < 0)
𝑥 = 𝑥 (𝑥 ≥ 0)](https://image.slidesharecdn.com/nhnforward201937-200123071534/85/feat-PyTorch-23-320.jpg)
![𝐇 𝒙 = 𝐬𝐢𝐠𝐦𝐨𝐢𝐝 (𝒘 𝟎 + 𝒘 𝟏 𝒙 𝟏 + ⋯ + 𝒘 𝒏 𝒙 𝒏) =
𝟏
𝟏 + 𝐞(∑ 𝒘 𝒊 𝒙 𝒊)
𝑥f
w(
bias(wl)
H 𝑥𝚺 𝝈
𝝈 𝒙 = 𝑠𝑖𝑔𝑚𝑜𝑖𝑑 𝑥 =
1
1 + e]
𝑥(
𝑤f
𝑥( 𝑥&
𝑥f
𝑥&](https://image.slidesharecdn.com/nhnforward201937-200123071534/85/feat-PyTorch-24-320.jpg)
![class ConvModel(torch.nn.Module):
def __init__(self):
super().__init__()
self.conv1 = torch.nn.Conv2d(in_channels=1, out_channels=16,
kernel_size=3)
self.pool = torch.nn.MaxPool2d(kernel_size=2)
self.conv2 = torch.nn.Conv2d(in_channels=16, out_channels=32,
kernel_size=3)
self.fc1 = torch.nn.Linear(in_features=11*11*32, out_features=128)
self.fc2 = torch.nn.Linear(in_features=128, out_features=10)
self.relu = torch.nn.ReLU()
def forward(self, x):
x = self.conv1(x)
x = self.relu(x)
x = self.pool(x)
x = self.conv2(x)
x = self.relu(x)
x = x.view(x.shape[0], -1)
x = self.fc1(x)
x = self.relu(x)
x. = self.fc2(x)
return x](https://image.slidesharecdn.com/nhnforward201937-200123071534/85/feat-PyTorch-40-320.jpg)
![class FoodDataset(torch.utils.data.Dataset):
def __init__(self, datafile):
self.samples = []
lines = open(datafile).read().splitlines()
for line in lines:
path, label = line.split(‘:’)
self.samples.append((path, label))
def __getitem__(self, idx):
path, label = self.samples[idx]
image = cv2.imread(path)
...
return image, float(label)
def __len__(self):
return len(self.samples)](https://image.slidesharecdn.com/nhnforward201937-200123071534/85/feat-PyTorch-50-320.jpg)
![train_dataset = FoodDataset('dataset/data.list’)
train_loader = torch.utils.data.DataLoader(dataset=train_dataset, batch_size=batch_size, shuffle=True)
image_converter = transforms.ToPILImage()
for images, labels in train_loader:
label = str(labels[0].numpy())
image = image_converter(images[0])
print (label, id_to_names[label])
plt.imshow(image)
plt.show()](https://image.slidesharecdn.com/nhnforward201937-200123071534/85/feat-PyTorch-51-320.jpg)
딥러닝, 야 너도 할 수 있어(feat. PyTorch)
- 18. Hours Score 10 90 9 80 3 50 2 30
- 19. (“ ” – ” ”)& Model 1 𝐻( 𝑥 = 6.6𝑥 + 22.9 Model 2 𝐻& 𝑥 = 5.5𝑥 + 34.5 Model 3 𝐻3 𝑥 = 8.25𝑥 + 23 𝐱 𝐲 𝐇 𝟏 𝐱 (𝐇 𝟏 𝐱 − 𝐲) 𝟐 𝐇 𝟐 𝐱 (𝐇 𝟐 𝐱 − 𝐲) 𝟐 𝐇 𝟑 𝐱 (𝐇 𝟑 𝐱 − 𝐲) 𝟐
- 20. 0 20 40 60 80 100 0 5 10 15 𝐇 𝐱 = 𝐚𝐱 + 𝐛 𝐇 𝐱 = 𝐰 𝟏×𝐱 + 𝐰 𝟎×𝟏 𝐖 = 𝐰 𝟏 𝐰 𝟎 𝐗 = 𝐱 𝟏 𝐇 𝐱 = 𝐖𝐗 𝐖 𝐖
- 21. 𝛁𝐋𝐨𝐬𝐬 𝐋𝐨𝐬𝐬 𝐰 𝛂 𝐰 𝐖 𝟏 𝐦 ∑𝐢K𝟏 𝐦 𝐖×𝐗 𝐢 − 𝐲 𝐢 𝟐 (예측 값 – 실제 값)& 𝐇 𝐱 = 𝐖𝐗 𝐦𝐢𝐧𝐢𝐦𝐢𝐳𝐞 𝑳𝒐𝒔𝒔 𝐰 𝐖
- 22. !"#$$ "#$$ % & % 𝛁𝐋𝐨𝐬𝐬 𝐖 = 𝐖 − 𝛂𝛁𝐋𝐨𝐬𝐬
- 23. 𝝈 𝒙 = 𝑠𝑖𝑔𝑚𝑜𝑖𝑑 𝑥 = 1 1 + e] 𝐇 𝒙 = 𝐚𝐜𝐭𝐢𝐯𝐚𝐭𝐢𝐨𝐧_𝐟𝐧(𝒘 𝟎 + 𝒘 𝟏 𝒙 𝟏 + ⋯ + 𝒘 𝒏 𝒙 𝒏) 𝑥f bias(wl) 𝐻 𝑥 𝑥( 𝚺 𝝈 𝑥& 𝝈 𝒙 = 𝑟𝑒𝑙𝑢 𝑥 = r 𝑥 = 0 (𝑥 < 0) 𝑥 = 𝑥 (𝑥 ≥ 0)
- 24. 𝐇 𝒙 = 𝐬𝐢𝐠𝐦𝐨𝐢𝐝 (𝒘 𝟎 + 𝒘 𝟏 𝒙 𝟏 + ⋯ + 𝒘 𝒏 𝒙 𝒏) = 𝟏 𝟏 + 𝐞(∑ 𝒘 𝒊 𝒙 𝒊) 𝑥f w( bias(wl) H 𝑥𝚺 𝝈 𝝈 𝒙 = 𝑠𝑖𝑔𝑚𝑜𝑖𝑑 𝑥 = 1 1 + e] 𝑥( 𝑤f 𝑥( 𝑥& 𝑥f 𝑥&
- 25. 𝑥f w( bias(wl) H 𝑥𝚺 𝝈 𝑥( 𝑤f class MyModel(torch.nn.Module): def __init__(self): super(MyScoreModel, self).__init__() self.weights = torch.nn.Parameter(torch.randn(n)) self.bias = torch.nn.Parameter(torch.randn(1)) self.activation_fn = torch.nn.Sigmoid() def forward(self, x): output = self.weights * x + self.bias output = self.activation_fn(output) return output
- 26. 𝑥f bias 𝐻( 𝑥 𝑥( class MyModel(torch.nn.Module): def __init__(self, input_num=n, output_num=1): super(MyScoreModel, self).__init__() self.linear = torch.nn.Linear(in_features=n, out_features=1, bias = True) self.activation_fn = torch.nn.Sigmoid() def forward(self, x): output = self.linear(x) output = self.activation_fn(output) return output
- 29. 𝑥f 𝑥( Weights 𝑦& 𝑦( Weights Weights class MyModel(torch.nn.Module): def __init__(self): super(MyScoreModel, self).__init__() self.linear1 = torch.nn.Linear(in_features=n, out_features =m) self.linear2 = torch.nn.Linear(in_features=m, out_features =p) self.linear3 = torch.nn.Linear(in_features=p, out_features =2) self.activation_fn = torch.nn.Sigmoid() def forward(self, x): output = self.activation_fn(self.linear1(x)) output = self.activation_fn(self.linear2(output)) output = self.linear3(output) return output 𝑜(( 𝑜(• 𝑜&( 𝑜&‚
- 37. 𝑤( 𝑤& 𝑤3 𝑤{ 𝝈
- 40. class ConvModel(torch.nn.Module): def __init__(self): super().__init__() self.conv1 = torch.nn.Conv2d(in_channels=1, out_channels=16, kernel_size=3) self.pool = torch.nn.MaxPool2d(kernel_size=2) self.conv2 = torch.nn.Conv2d(in_channels=16, out_channels=32, kernel_size=3) self.fc1 = torch.nn.Linear(in_features=11*11*32, out_features=128) self.fc2 = torch.nn.Linear(in_features=128, out_features=10) self.relu = torch.nn.ReLU() def forward(self, x): x = self.conv1(x) x = self.relu(x) x = self.pool(x) x = self.conv2(x) x = self.relu(x) x = x.view(x.shape[0], -1) x = self.fc1(x) x = self.relu(x) x. = self.fc2(x) return x
- 48. 0 0 0
- 50. class FoodDataset(torch.utils.data.Dataset): def __init__(self, datafile): self.samples = [] lines = open(datafile).read().splitlines() for line in lines: path, label = line.split(‘:’) self.samples.append((path, label)) def __getitem__(self, idx): path, label = self.samples[idx] image = cv2.imread(path) ... return image, float(label) def __len__(self): return len(self.samples)
- 51. train_dataset = FoodDataset('dataset/data.list’) train_loader = torch.utils.data.DataLoader(dataset=train_dataset, batch_size=batch_size, shuffle=True) image_converter = transforms.ToPILImage() for images, labels in train_loader: label = str(labels[0].numpy()) image = image_converter(images[0]) print (label, id_to_names[label]) plt.imshow(image) plt.show()
- 52. fc1000 import torchvision.models as models class FoodModel(torch.nn.Module): def __init__(self): super().__init__() self.model = models.resnet18(pretrained=True) self.model.fc = torch.nn.Linear(512, 20) def forward(self, x): return self.model(x)