10.8. RMSProp¶

In the experiment in Section 10.7, the learning rate of each element in the independent variable of the objective function declines (or remains unchanged) during iteration because the variable $$\mathbf{s}_t$$ in the denominator is increased by the square by element operation of the mini-batch stochastic gradient, adjusting the learning rate. Therefore, when the learning rate declines very fast during early iteration, yet the current solution is still not desirable, Adagrad might have difficulty finding a useful solution because the learning rate will be too small at later stages of iteration. To tackle this problem, the RMSProp algorithm [Tieleman.Hinton.2012] made a small modification to Adagrad.

10.8.1. The Algorithm¶

Unlike in Adagrad, the state variable $$\mathbf{s}_t$$ is the sum of the square by element all the mini-batch stochastic gradients $$\mathbf{g}_t$$ up to the time step $$t$$, RMSProp uses the exponentially weighted moving average on the square by element results of these gradients. Specifically, given the hyperparameter $$0 \leq \gamma < 1$$, RMSProp is computed at time step $$t>0$$.

(10.8.1)$\mathbf{s}_t \leftarrow \gamma \mathbf{s}_{t-1} + (1 - \gamma) \mathbf{g}_t \odot \mathbf{g}_t.$

Like Adagrad, RMSProp re-adjusts the learning rate of each element in the independent variable of the objective function with element operations and then updates the independent variable.

(10.8.2)$\mathbf{x}_t \leftarrow \mathbf{x}_{t-1} - \frac{\eta}{\sqrt{\mathbf{s}_t + \epsilon}} \odot \mathbf{g}_t,$

Here, $$\eta$$ is the learning rate while $$\epsilon$$ is a constant added to maintain numerical stability, such as $$10^{-6}$$.

10.8.1.1. Exponentially Weighted Moving Average (EWMA)¶

Now let expand the definition of $$\mathbf{s}_t$$, we can see that

(10.8.3)\begin{split}\begin{aligned} \mathbf{s}_t &= (1 - \gamma) \mathbf{g}_t \odot \mathbf{g}_t + \gamma \mathbf{s}_{t-1} \\ &= (1 - \gamma) \left(\mathbf{g}_t \odot \mathbf{g}_t + \gamma \mathbf{g}_{t-1} \odot \mathbf{g}_{t-1}\right) + \gamma^2 \mathbf{s}_{t-2} \\ &\cdots\\ &= (1 - \gamma)\left( \mathbf{g}_t \odot \mathbf{g}_t + \gamma \mathbf{g}_{t-1} \odot \mathbf{g}_{t-1} + \cdots + \gamma^{t-1}\mathbf{g}_{1} \odot \mathbf{g}_{1} \right). \end{aligned}\end{split}

In Section 10.6 we see that $$\frac{1}{1-\gamma} = 1 + \gamma + \gamma^2 + \cdots$$, so the sum of weights equals to 1. In addition, these weights decrease exponentially, it is called exponentially weighted moving average.

We visualize the weights in the past 40 time steps with various $$\gamma$$s.

%matplotlib inline
import d2l
import math
from mxnet import np, npx
npx.set_np()

gammas = [0.95, 0.9, 0.8, 0.7]
d2l.set_figsize((3.5, 2.5))
for gamma in gammas:
x = np.arange(40).asnumpy()
d2l.plt.plot(x, (1-gamma) * gamma ** x, label='gamma = %.2f'%gamma)
d2l.plt.xlabel('time'); 10.8.2. Implementation from Scratch¶

By convention, we will use the objective function $$f(\mathbf{x})=0.1x_1^2+2x_2^2$$ to observe the iterative trajectory of the independent variable in RMSProp. Recall that in Section 10.7, when we used Adagrad with a learning rate of 0.4, the independent variable moved less in later stages of iteration. However, at the same learning rate, RMSProp can approach the optimal solution faster.

def rmsprop_2d(x1, x2, s1, s2):
g1, g2, eps = 0.2 * x1, 4 * x2, 1e-6
s1 = gamma * s1 + (1 - gamma) * g1 ** 2
s2 = gamma * s2 + (1 - gamma) * g2 ** 2
x1 -= eta / math.sqrt(s1 + eps) * g1
x2 -= eta / math.sqrt(s2 + eps) * g2
return x1, x2, s1, s2

def f_2d(x1, x2):
return 0.1 * x1 ** 2 + 2 * x2 ** 2

eta, gamma = 0.4, 0.9
d2l.show_trace_2d(f_2d, d2l.train_2d(rmsprop_2d))
epoch 20, x1 -0.010599, x2 0.000000 Next, we implement RMSProp with the formula in the algorithm.

def init_rmsprop_states(feature_dim):
s_w = np.zeros((feature_dim, 1))
s_b = np.zeros(1)
return (s_w, s_b)

def rmsprop(params, states, hyperparams):
gamma, eps = hyperparams['gamma'], 1e-6
for p, s in zip(params, states):
s[:] = gamma * s + (1 - gamma) * np.square(p.grad)
p[:] -= hyperparams['lr'] * p.grad / np.sqrt(s + eps)

We set the initial learning rate to 0.01 and the hyperparameter $$\gamma$$ to 0.9. Now, the variable $$\boldsymbol{s}_t$$ can be treated as the weighted average of the square term $$\boldsymbol{g}_t \odot \boldsymbol{g}_t$$ from the last $$1/(1-0.9) = 10$$ time steps.

data_iter, feature_dim = d2l.get_data_ch10(batch_size=10)
d2l.train_ch10(rmsprop, init_rmsprop_states(feature_dim),
{'lr': 0.01, 'gamma': 0.9}, data_iter, feature_dim);
loss: 0.243, 0.058 sec/epoch 10.8.3. Concise Implementation¶

From the Trainer instance of the algorithm named “rmsprop”, we can implement the RMSProp algorithm with Gluon to train models. Note that the hyperparameter $$\gamma$$ is assigned by gamma1.

d2l.train_gluon_ch10('rmsprop', {'learning_rate': 0.01, 'gamma1': 0.9},
data_iter)
loss: 0.243, 0.033 sec/epoch 10.8.4. Summary¶

• The difference between RMSProp and Adagrad is that RMSProp uses an EWMA on the squares of elements in the mini-batch stochastic gradient to adjust the learning rate.

10.8.5. Exercises¶

• What happens to the experimental results if we set the value of $$\gamma$$ to 1? Why?
• Try using other combinations of initial learning rates and $$\gamma$$ hyperparameters and observe and analyze the experimental results.

10.8.6. Scan the QR Code to Discuss¶ 