8.5.rnn scratch
梯度裁剪
类似于球形投影,使梯度的 L2 范数不超过阈值。
训练
注意训练时候每个批量的隐状态是分别存储的,不会相互影响。
- 多久一次 backward?
- 要将一个批量内所有样本的所有时间步(在程序开始时设定,比如样本 size = 32,单次输入时间步 num_steps 为 35)forward 以后(每次 forward 仅进行一个时间步的预测),才进行 backward。因此权重参数矩阵会累乘,容易导致梯度爆炸,故进行梯度裁剪。d2l 代码中,梯度裁剪在 backward 后,updater 之前执行。
- 注意对每个批量而言,每次输入仅有一个 token,并非多个时间步 token 同时输入。
手绘
2:15:13
手动代码
import math
import torch
from torch import nn
from torch.nn import functional as F
from d2l import torch as d2l
batch_size, num_steps = 32, 35
train_iter, vocab = d2l.load_data_time_machine(batch_size, num_steps)
# 获取初始参数
def get_params(vocab_size, num_hiddens, device):
num_inputs = num_outputs = vocab_size
def normal(shape):
return torch.randn(size=shape, device=device) * 0.01
# 隐藏层参数
W_xh = normal((num_inputs, num_hiddens))
W_hh = normal((num_hiddens, num_hiddens))
b_h = torch.zeros(num_hiddens, device=device)
# 输出层参数
W_hq = normal((num_hiddens, num_outputs))
b_q = torch.zeros(num_outputs, device=device)
# 附加梯度
params = [W_xh, W_hh, b_h, W_hq, b_q]
for param in params:
param.requires_grad_(True)
return params
# 前向传播 forward
# 一次 forward 就将各个批量的所有时间步都生成完了
def rnn(inputs, state, params):
# inputs的形状:(时间步数量,批量大小,词表大小)
W_xh, W_hh, b_h, W_hq, b_q = params
H, = state
outputs = []
# 循环每个时间步
for X in inputs:
# X的形状:(批量大小,词表大小)
# 注意这里隐状态包含所有批量各自的隐状态
# H 形状: (批量大小,隐藏单元个数)
H = torch.tanh(torch.mm(X, W_xh) + torch.mm(H, W_hh) + b_h)
# 本步输出不会影响下一步输出,只有隐状态才会
# Y 形状: (批量大小,词表大小),因为前面已经指定 num_outputs = vocab_size
Y = torch.mm(H, W_hq) + b_q
outputs.append(Y)
# 返回一个元组,即所有时间步的输出 y 和最后一个隐状态
return torch.cat(outputs, dim=0), (H,)
# torch.cat 输出为 (ns * n, n_vocab)
def init_rnn_state(batch_size, num_hiddens, device):
return (torch.zeros((batch_size, num_hiddens), device=device), )
# 模型封装
class RNNModelScratch:
def __init__(self, vocab_size, num_hiddens, device,
get_params, init_state, forward_fn):
self.vocab_size, self.num_hiddens = vocab_size, num_hiddens
self.params = get_params(vocab_size, num_hiddens, device)
self.init_state, self.forward_fn = init_state, forward_fn
def __call__(self, X, state):
X = F.one_hot(X.T, self.vocab_size).type(torch.float32)
return self.forward_fn(X, state, self.params)
def begin_state(self, batch_size, device):
return self.init_state(batch_size, self.num_hiddens, device)
num_hiddens = 512
net = RNNModelScratch(len(vocab), num_hiddens, d2l.try_gpu(), get_params,
init_rnn_state, rnn)
state = net.begin_state(X.shape[0], d2l.try_gpu())
# 预测
def predict_ch8(prefix, num_preds, net, vocab, device):
"""在prefix后面生成新字符"""
state = net.begin_state(batch_size=1, device=device)
outputs = [vocab[prefix[0]]]
get_input = lambda: torch.tensor([outputs[-1]], device=device).reshape((1, 1))
for y in prefix[1:]: # 预热期
_, state = net(get_input(), state)
outputs.append(vocab[y])
for _ in range(num_preds): # 预测num_preds步
y, state = net(get_input(), state)
outputs.append(int(y.argmax(dim=1).reshape(1)))
return ''.join([vocab.idx_to_token[i] for i in outputs])
# 梯度裁剪
def grad_clipping(net, theta): #@save
"""裁剪梯度"""
if isinstance(net, nn.Module):
params = [p for p in net.parameters() if p.requires_grad]
else:
params = net.params
norm = torch.sqrt(sum(torch.sum((p.grad ** 2)) for p in params))
if norm > theta:
for param in params:
param.grad[:] *= theta / norm
# 训练一个 epoch
def train_epoch_ch8(net, train_iter, loss, updater, device, use_random_iter):
"""训练网络一个迭代周期(定义见第8章)"""
state, timer = None, d2l.Timer()
metric = d2l.Accumulator(2) # 训练损失之和,词元数量
for X, Y in train_iter:
if state is None or use_random_iter:
# 在第一次迭代或使用随机抽样时初始化state
state = net.begin_state(batch_size=X.shape[0], device=device)
else:
if isinstance(net, nn.Module) and not isinstance(state, tuple):
# state对于nn.GRU是个张量
state.detach_()
else:
# state对于nn.LSTM或对于我们从零开始实现的模型是个张量
for s in state:
s.detach_()
y = Y.T.reshape(-1)
# y.shape: (ns * n)
X, y = X.to(device), y.to(device)
y_hat, state = net(X, state)
l = loss(y_hat, y.long()).mean()
if isinstance(updater, torch.optim.Optimizer):
updater.zero_grad()
l.backward()
grad_clipping(net, 1)
updater.step()
else:
l.backward()
grad_clipping(net, 1)
# 因为已经调用了mean函数
updater(batch_size=1)
metric.add(l * y.numel(), y.numel())
return math.exp(metric[0] / metric[1]), metric[1] / timer.stop()
# 训练多个 epoch
def train_ch8(net, train_iter, vocab, lr, num_epochs, device,
use_random_iter=False):
"""训练模型(定义见第8章)"""
loss = nn.CrossEntropyLoss()
animator = d2l.Animator(xlabel='epoch', ylabel='perplexity',
legend=['train'], xlim=[10, num_epochs])
# 初始化
if isinstance(net, nn.Module):
updater = torch.optim.SGD(net.parameters(), lr)
else:
updater = lambda batch_size: d2l.sgd(net.params, lr, batch_size)
predict = lambda prefix: predict_ch8(prefix, 50, net, vocab, device)
# 训练和预测
for epoch in range(num_epochs):
ppl, speed = train_epoch_ch8(
net, train_iter, loss, updater, device, use_random_iter)
if (epoch + 1) % 10 == 0:
print(predict('time traveller'))
animator.add(epoch + 1, [ppl])
print(f'困惑度 {ppl:.1f}, {speed:.1f} 词元/秒 {str(device)}')
print(predict('time traveller'))
print(predict('traveller'))
num_epochs, lr = 500, 1
train_ch8(net, train_iter, vocab, lr, num_epochs, d2l.try_gpu())