BACKPROPAGATION AND OPTIMIZER¶

So far, we have learned how to use OneFlow to Dataset and DataLoader, Build Models,Autograd, and combine them so that we can train models by using backpropagation algorithms.

In oneflow.optim, there are various optimizers that simplify the code of back propagation.

This article will first introduce the basic concepts of back propagation and then show you how to use the oneflow.optim class.

Backpropagation by Numpy Code¶

In order to make it easier for readers to understand the relationship between backpropagation and autograd, a training process of a simple model implemented with numpy is provided here:

import numpy as np

ITER_COUNT = 500
LR = 0.01

# Forward propagation
def forward(x, w):
    return np.matmul(x, w)


# Loss function
def loss(y_pred, y):
    return ((y_pred - y) ** 2).sum()


# Calculate gradient
def gradient(x, y, y_pred):
    return np.matmul(x.T, 2 * (y_pred - y))


if __name__ == "__main__":
    # Train: Y = 2*X1 + 3*X2
    x = np.array([[1, 2], [2, 3], [4, 6], [3, 1]], dtype=np.float32)
    y = np.array([[8], [13], [26], [9]], dtype=np.float32)

    w = np.array([[2], [1]], dtype=np.float32)
    # Training cycle
    for i in range(0, ITER_COUNT):
        y_pred = forward(x, w)
        l = loss(y_pred, y)
        if (i + 1) % 50 == 0:
            print(f"{i+1}/{500} loss:{l}")

        grad = gradient(x, y, y_pred)
        w -= LR * grad

    print(f"w:{w}")

output：

50/500 loss:0.0034512376878410578
100/500 loss:1.965487399502308e-06
150/500 loss:1.05524122773204e-09
200/500 loss:3.865352482534945e-12
250/500 loss:3.865352482534945e-12
300/500 loss:3.865352482534945e-12
350/500 loss:3.865352482534945e-12
400/500 loss:3.865352482534945e-12
450/500 loss:3.865352482534945e-12
500/500 loss:3.865352482534945e-12
w:[[2.000001 ]
 [2.9999993]]

Note that the loss function expression we selected is \(\sum (y_{p} - y)^2\), so the code for gradient of loss to parameter w is：

def gradient(x, y, y_pred):
    return np.matmul(x.T, 2 * (y_pred - y))

SGD is used to update parameters：

grad = gradient(x, y, y_pred)
w -= LR*grad

In summary, a complete iteration in the training includes the following steps:

The model calculates the predicted value based on the input and parameters (y_pred)
Calculate loss, which is the error between the predicted value and the label
Calculate the gradient of loss to parameter
Update parameter(s)

1 and 2 are forward propagation process; 3 and 4 are back propagation process.

Hyperparameters¶

Hyperparameters are parameters related to model training settings, which can affect the efficiency and results of model training.As in the above code ITER_COUNT,LR are hyperparameters.

Using the optimizer class in `oneflow.optim`¶

Using the optimizer class in oneflow.optim for back propagation will be more concise.

First, prepare the data and model. The convenience of using Module is that you can place the hyperparameters in Module for management.

import oneflow as flow

x = flow.tensor([[1, 2], [2, 3], [4, 6], [3, 1]], dtype=flow.float32)
y = flow.tensor([[8], [13], [26], [9]], dtype=flow.float32)


class MyLrModule(flow.nn.Module):
    def __init__(self, lr, iter_count):
        super().__init__()
        self.w = flow.nn.Parameter(flow.tensor([[2], [1]], dtype=flow.float32))
        self.lr = lr
        self.iter_count = iter_count

    def forward(self, x):
        return flow.matmul(x, self.w)


model = MyLrModule(0.01, 500)

Loss function¶

Then, select the loss function. OneFlow comes with a variety of loss functions. We choose MSELoss here：

loss = flow.nn.MSELoss(reduction="sum")

Construct Optimizer¶

The logic of back propagation is wrapped in optimizer. We choose SGD here, You can choose other optimization algorithms as needed, such as Adam andAdamW .

optimizer = flow.optim.SGD(model.parameters(), model.lr)

When the optimizer is constructed, the model parameters and learning rate are given to SGD. Then the optimizer.step() is called, and it automatically completes the gradient of the model parameters and updates the model parameters according to the SGD algorithm.

Train¶

When the above preparations are completed, we can start training:

for i in range(0, model.iter_count):
    y_pred = model(x)
    l = loss(y_pred, y)
    if (i + 1) % 50 == 0:
        print(f"{i+1}/{model.iter_count} loss:{l.numpy()}")

    optimizer.zero_grad()
    l.backward()
    optimizer.step()

print(f"\nw: {model.w}")