PyTorch学习笔记（2）

内容来自Datacamp相关课程：https://app.datacamp.com/learn/courses/introduction-to-deep-learning-with-pytorch

Forward Pass

What is a forward pass?

• input data is passed forwarf or propagated through a network
• computations performed at each layer
• outputs of each layer passed to each subsequent layer
• output of final layer: "prediction"
• used for both training and prediction

Some possible outputs:

• Binary classification
  • dingle probability between 0 and 1
• Multiclass classification
  • distribution of probabilities summing to 1
• Regression values
  • continuous numerical predictions

Backward Pass

• Backward pass, or backpropagation is used to update weights and biases during training
• In the "training loop", we:
  1. Propagate data forward
  2. Compare outputs to true values (ground-truth)
  3. Backpropagate to update model weights and biases
  4. Repeat until weigths and biases are tuned to produce useful outputs

代码示例

Binary classification: forward pass

input_data = torch.trnsor(...)

# create binary classification model
model = nn.Sequential(
    nn.Linear(6, 4),
    nn.Linear(4, 1),
    nn.Sigmoid()
)

# pass input data through model
output = model(input_data)

Multi-class classification: forward pass

n_classes = 3

# create multi-class classification model
model = nn.Sequential(
    nn.Linear(6, 4),
    nn.Linear(4, n_classes),
    nn.Softmax(dim=-1) # indicate the samples have the same last dimension as the last linear layer's output
)

# pass input data through model
output = model(input_data)

Loss Functions

Why do we need a loss function?

• gives feedback to model during training
• takes in model prediction \(\hat{y}\) and ground truth \(y\)
• outputs a float