PyTorch一小時掌握之autograd機制篇
概述
PyTorch 干的最厲害的一件事情就是幫我們把反向傳播全部計算好了.
代碼實現(xiàn)
手動定義求導
import torch # 方法一 x = torch.randn(3, 4, requires_grad=True) # 方法二 x = torch.randn(3,4) x.requires_grad = True
b = torch.randn(3, 4, requires_grad=True) t = x + b y = t.sum() print(y) print(y.backward()) print(b.grad) print(x.requires_grad) print(b.requires_grad) print(t.requires_grad)
輸出結果:
tensor(1.1532, grad_fn=<SumBackward0>)
None
tensor([[1., 1., 1., 1.],
[1., 1., 1., 1.],
[1., 1., 1., 1.]])
True
True
True
計算流量
# 計算流量 x = torch.rand(1) w = torch.rand(1, requires_grad=True) b = torch.rand(1, requires_grad=True) y = w * x z = y + b print(x.requires_grad, w.requires_grad,b.requires_grad, z.requires_grad) print(x.is_leaf, w.is_leaf, b.is_leaf, y.is_leaf,z.is_leaf)
輸出結果:
False True True True
True True True False False
反向傳播計算
# 反向傳播 z.backward(retain_graph= True) # 如果不清空會累加起來 print(w.grad) print(b.grad)
輸出結果:
tensor([0.1485])
tensor([1.])
線性回歸
導包
import numpy as np import torch import torch.nn as nn
構造 x, y
# 構造數(shù)據(jù) X_values = [i for i in range(11)] X_train = np.array(X_values, dtype=np.float32) X_train = X_train.reshape(-1, 1) print(X_train.shape) # (11, 1) y_values = [2 * i + 1 for i in X_values] y_train = np.array(y_values, dtype=np.float32) y_train = y_train.reshape(-1,1) print(y_train.shape) # (11, 1)
輸出結果:
(11, 1)
(11, 1)
構造模型
# 構造模型 class LinerRegressionModel(nn.Module): def __init__(self, input_dim, output_dim): super(LinerRegressionModel, self).__init__() self.liner = nn.Linear(input_dim, output_dim) def forward(self, x): out = self.liner(x) return out input_dim = 1 output_dim = 1 model = LinerRegressionModel(input_dim, output_dim) print(model)
輸出結果:
LinerRegressionModel(
(liner): Linear(in_features=1, out_features=1, bias=True)
)
參數(shù) & 損失函數(shù)
# 超參數(shù) enpochs = 1000 learning_rate = 0.01 # 損失函數(shù) optimizer = torch.optim.SGD(model.parameters(), lr=learning_rate) criterion = nn.MSELoss()
訓練模型
# 訓練模型
for epoch in range(enpochs):
# 轉成tensor
inputs = torch.from_numpy(X_train)
labels = torch.from_numpy(y_train)
# 梯度每次迭代清零
optimizer.zero_grad()
# 前向傳播
outputs = model(inputs)
# 計算損失
loss = criterion(outputs, labels)
# 反向傳播
loss.backward()
# 更新參數(shù)
optimizer.step()
if epoch % 50 == 0:
print("epoch {}, loss {}".format(epoch, loss.item()))
輸出結果:
epoch 0, loss 114.47456359863281
epoch 50, loss 0.00021522105089388788
epoch 100, loss 0.00012275540211703628
epoch 150, loss 7.001651829341426e-05
epoch 200, loss 3.9934264350449666e-05
epoch 250, loss 2.2777328922529705e-05
epoch 300, loss 1.2990592040296178e-05
epoch 350, loss 7.409254521917319e-06
epoch 400, loss 4.227155841363128e-06
epoch 450, loss 2.410347860859474e-06
epoch 500, loss 1.3751249525739695e-06
epoch 550, loss 7.844975016269018e-07
epoch 600, loss 4.4756839656656666e-07
epoch 650, loss 2.5517596213830984e-07
epoch 700, loss 1.4577410922811396e-07
epoch 750, loss 8.30393886985803e-08
epoch 800, loss 4.747753479250605e-08
epoch 850, loss 2.709844615367274e-08
epoch 900, loss 1.5436164346738224e-08
epoch 950, loss 8.783858973515635e-09
完整代碼
import numpy as np
import torch
import torch.nn as nn
# 構造數(shù)據(jù)
X_values = [i for i in range(11)]
X_train = np.array(X_values, dtype=np.float32)
X_train = X_train.reshape(-1, 1)
print(X_train.shape) # (11, 1)
y_values = [2 * i + 1 for i in X_values]
y_train = np.array(y_values, dtype=np.float32)
y_train = y_train.reshape(-1,1)
print(y_train.shape) # (11, 1)
# 構造模型
class LinerRegressionModel(nn.Module):
def __init__(self, input_dim, output_dim):
super(LinerRegressionModel, self).__init__()
self.liner = nn.Linear(input_dim, output_dim)
def forward(self, x):
out = self.liner(x)
return out
input_dim = 1
output_dim = 1
model = LinerRegressionModel(input_dim, output_dim)
print(model)
# 超參數(shù)
enpochs = 1000
learning_rate = 0.01
# 損失函數(shù)
optimizer = torch.optim.SGD(model.parameters(), lr=learning_rate)
criterion = nn.MSELoss()
# 訓練模型
for epoch in range(enpochs):
# 轉成tensor
inputs = torch.from_numpy(X_train)
labels = torch.from_numpy(y_train)
# 梯度每次迭代清零
optimizer.zero_grad()
# 前向傳播
outputs = model(inputs)
# 計算損失
loss = criterion(outputs, labels)
# 反向傳播
loss.backward()
# 更新參數(shù)
optimizer.step()
if epoch % 50 == 0:
print("epoch {}, loss {}".format(epoch, loss.item()))
到此這篇關于PyTorch一小時掌握之autograd機制篇的文章就介紹到這了,更多相關PyTorch autograd內容請搜索本站以前的文章或繼續(xù)瀏覽下面的相關文章希望大家以后多多支持本站！
版權聲明：本站文章來源標注為YINGSOO的內容版權均為本站所有，歡迎引用、轉載，請保持原文完整并注明來源及原文鏈接。禁止復制或仿造本網站，禁止在非maisonbaluchon.cn所屬的服務器上建立鏡像，否則將依法追究法律責任。本站部分內容來源于網友推薦、互聯(lián)網收集整理而來，僅供學習參考，不代表本站立場，如有內容涉嫌侵權，請聯(lián)系alex-e#qq.com處理。
關注官方微信