what is the difference between s[:] and s if s is a torch.Tensor [duplicate] - indexing

Question

import numpy as np
import time
features, labels = d2l.get_data_ch7()
def init_adam_states():
v_w, v_b = torch.zeros((features.shape[1], 1),dtype=torch.float32), torch.zeros(1, dtype=torch.float32)
s_w, s_b = torch.zeros((features.shape[1], 1),dtype=torch.float32), torch.zeros(1, dtype=torch.float32)
return ((v_w, s_w), (v_b, s_b))
def adam(params, states, hyperparams):
beta1, beta2, eps = 0.9, 0.999, 1e-6
for p, (v, s) in zip(params, states):
v[:] = beta1 * v + (1 - beta1) * p.grad.data
s = beta2 * s + (1 - beta2) * p.grad.data**2
v_bias_corr = v / (1 - beta1 ** hyperparams['t'])
s_bias_corr = s / (1 - beta2 ** hyperparams['t'])
p.data -= hyperparams['lr'] * v_bias_corr / (torch.sqrt(s_bias_corr) + eps)
hyperparams['t'] += 1
def train_ch7(optimizer_fn, states, hyperparams, features, labels, batch_size=10, num_epochs=2):
# 初始化模型
net, loss = d2l.linreg, d2l.squared_loss
w = torch.nn.Parameter(torch.tensor(np.random.normal(0, 0.01, size=(features.shape[1], 1)), dtype=torch.float32),
requires_grad=True)
b = torch.nn.Parameter(torch.zeros(1, dtype=torch.float32), requires_grad=True)
def eval_loss():
return loss(net(features, w, b), labels).mean().item()
ls = [eval_loss()]
data_iter = torch.utils.data.DataLoader(torch.utils.data.TensorDataset(features, labels), batch_size, shuffle=True)
for _ in range(num_epochs):
start = time.time()
print(w)
print(b)
for batch_i, (X, y) in enumerate(data_iter):
l = loss(net(X, w, b), y).mean() # 使⽤平均损失
# 梯度清零
if w.grad is not None:
w.grad.data.zero_()
b.grad.data.zero_()
l.backward()
optimizer_fn([w, b], states, hyperparams) # 迭代模型参数
if (batch_i + 1) * batch_size % 100 == 0:
ls.append(eval_loss()) # 每100个样本记录下当前训练误差
# 打印结果和作图
print('loss: %f, %f sec per epoch' % (ls[-1], time.time() - start))
d2l.set_figsize()
d2l.plt.plot(np.linspace(0, num_epochs, len(ls)), ls)
d2l.plt.xlabel('epoch')
d2l.plt.ylabel('loss')
train_ch7(adam, init_adam_states(), {'lr': 0.01, 't': 1}, features, labels)
I want to implement the Adam algorithm in the follow code and I feel confused in the function named adam.
v = beta1 * v + (1 - beta1) * p.grad.data
s = beta2 * s + (1 - beta2) * p.grad.data**2
when I use the follow code, the loss function curve is figure 1.
figure 1
v[:] = beta1 * v + (1 - beta1) * p.grad.data
s = beta2 * s + (1 - beta2) * p.grad.data**2
or
v = beta1 * v + (1 - beta1) * p.grad.data
s[:] = beta2 * s + (1 - beta2) * p.grad.data**2
when I use the follow code, the loss function curve is figure 2.
figure 2
v[:] = beta1 * v + (1 - beta1) * p.grad.data
s[:] = beta2 * s + (1 - beta2) * p.grad.data**2
when I use the follow code, the loss function curve is figure 3.
figure 3
The loss function curve in case 3 has always been smoother than that in case 1.
The loss function curve in case 2 sometimes can't converge.
Why is different?

Answer 1

To answer the first question,
v = beta1 * v + (1 - beta1) * p.grad.data
is an out-of-place operation. Remember that python variables are references to objects. By assigning a new value to variable v, the underlying object which v referred to before this assignment will not be changed. Instead the expression beta1 * v + (1 - beta1) * p.grad.data results in a new tensor which is then referred to by v.
On the other hand
v[:] = beta1 * v + (1 - beta1) * p.grad.data
is an in-place operation. After this operation v still refers to the same underlying object, and the elements of that tensor are modified and replaced with the values of the new tensor beta1 * v + (1 - beta1) * p.grad.data.
Take a look at the following 3 lines to see why this matters
for p, (v, s) in zip(params, states):
v[:] = beta1 * v + (1 - beta1) * p.grad.data
s[:] = beta2 * s + (1 - beta2) * p.grad.data**2
v and s are actually referring to tensors which are stored in states. If we do in-place operations then the values in states are changed to reflect the value assigned to v[:] and s[:].
If out-of-place operations are used then the values in states remain unchanged.