从零开始深度学习0614——pytorch入门之RNN实现图像分类和回归预测_transforms 输出hidden state

RNN实现图像分类

用RNN处理图像

如何将图像的处理理解为时间序列

可以理解为时间序顺序为从上到下

Mnist图像的处理  一个图像为28*28 pixel

时间顺序就是从上往下，从第一行到第28行

# Hyper Parameters
EPOCH = 1     
 
BATCH_SIZE = 64
TIME_STEP = 28          # rnn time step / image height   一共输入time_step次。 时序步长数  seq_len
INPUT_SIZE = 28         # rnn input size / image width   每次输入多少   输入维度
LR = 0.01               # learning rate
DOWNLOAD_MNIST = True   # set to True if haven't download the data

 

self.rnn = nn.LSTM(         # if use nn.RNN(), it hardly learns
    input_size=INPUT_SIZE,
    hidden_size=64,         # rnn hidden unit  隐藏层神经元的个数
    num_layers=1,           # number of rnn layer  多少层
    batch_first=True,
 
)

 

https://www.jianshu.com/p/41c15d301542 

理解为什么RNN输入默认不是batch first=True？这是为了便于并行计算。

r_out, (h_n, h_c) = self.rnn(x, None)

 

# x shape (batch, time_step, input_size)
# r_out shape (batch, time_step, output_size)
# h_n 的形状是 (n_layers, batch, hidden_size) t=time_step 时刻的隐层状态  分线剧情  hidden_state = (h_n, h_c)
# h_c 的形状是 (n_layers, batch, hidden_size) t=time_step 时刻的细胞状态  主线剧情
# 每一次处理完后会输出hidden_state 产生output , 结合下一次读取图像处理完输出的hidden_state 又产生output 这样循环
# 意思就是每一时刻的输入 会包括上一时刻的输入
# 其中hidden_state，又分为 h_n, h_c == h_state, c_state。

 

# h_n和output的关系: output包括了time_step中每一个时间点的隐层状态,
#                   而h_n是第time_step时刻的隐层状态, 所以output中最后一个元素就是h_n, 即output[-1] == h_n.

 

 

b_x = b_x.view(-1, 28, 28)     # 变一下维度 在pytorch中是.view()的形式表示reshape

 

完整代码：

LSTM实现是写数字识别

 
import torch
from torch import nn
import torchvision.datasets as dsets
import torchvision.transforms as transforms
import matplotlib.pyplot as plt
 
 
# torch.manual_seed(1)    # reproducible
 
# Hyper Parameters
EPOCH = 1               # train the training data n times, to save time, we just train 1 epoch
BATCH_SIZE = 64
TIME_STEP = 28          # rnn time step / image height   一共输入time_step次。 时序步长数  seq_len
INPUT_SIZE = 28         # rnn input size / image width   每次输入多少   输入维度
LR = 0.01               # learning rate
DOWNLOAD_MNIST = True   # set to True if haven't download the data
 
 
# Mnist digital dataset
train_data = dsets.MNIST(
    root='./mnist/',
    train=True,                         # this is training data
    transform=transforms.ToTensor(),    # Converts a PIL.Image or numpy.ndarray to
                                        # torch.FloatTensor of shape (C x H x W) and normalize in the range [0.0, 1.0]
    download=DOWNLOAD_MNIST,            # download it if you don't have it
)
 
# plot one example
print(train_data.train_data.size())     # (60000, 28, 28)
print(train_data.train_labels.size())   # (60000)
plt.imshow(train_data.train_data[0].numpy(), cmap='gray')
plt.title('%i' % train_data.train_labels[0])
plt.show()
 
# Data Loader for easy mini-batch return in training
train_loader = torch.utils.data.DataLoader(dataset=train_data, batch_size=BATCH_SIZE, shuffle=True)
 
# convert test data into Variable, pick 2000 samples to speed up testing
test_data = dsets.MNIST(root='./mnist/', train=False, transform=transforms.ToTensor())
test_x = test_data.test_data.type(torch.FloatTensor)[:2000]/255.   # shape (2000, 28, 28) value in range(0,1)
test_y = test_data.test_labels.numpy()[:2000]    # covert to numpy array
 
 
class RNN(nn.Module):
    def __init__(self):
        super(RNN, self).__init__()
 
        # LSTM 函数的参数和RNN都是一致的, 区别在于输入输出不同,LSTM 多了一个细胞的状态, 所以每一个循环层都增加了一个细胞状态h_c的输出.
 
        self.rnn = nn.LSTM(         # if use nn.RNN(), it hardly learns
            input_size=INPUT_SIZE,
            hidden_size=64,         # rnn hidden unit  隐藏层神经元的个数
            num_layers=1,           # number of rnn layer  多少层
            batch_first=True,
            # https://www.jianshu.com/p/41c15d301542  理解为什么RNN输入默认不是batch first=True？这是为了便于并行计算。
            # input & output will has batch size as 1s dimension. e.g. (batch, time_step, input_size)
            # 默认是(time_step, batch, input_size) 如果batch_first=True, 则 (batch, time_step, input_size)
        )
 
        self.out = nn.Linear(64, 10)
 
    def forward(self, x):
        # x shape (batch, time_step, input_size)
        # r_out shape (batch, time_step, output_size)
        # h_n 的形状是 (n_layers, batch, hidden_size) t=time_step 时刻的隐层状态  分线剧情  hidden_state = (h_n, h_c)
        # h_c 的形状是 (n_layers, batch, hidden_size) t=time_step 时刻的细胞状态  主线剧情
        # 每一次处理完后会输出hidden_state 产生output , 结合下一次读取图像处理完输出的hidden_state 又产生output 这样循环
        # 意思就是每一时刻的输入 会包括上一时刻的输入
        # 其中hidden_state，又分为 h_n, h_c == h_state, c_state。
 
        r_out, (h_n, h_c) = self.rnn(x, None)   # None represents zero initial hidden state
 
        # h_n和output的关系: output包括了time_step中每一个时间点的隐层状态,
        #                   而h_n是第time_step时刻的隐层状态, 所以output中最后一个元素就是h_n, 即output[-1] == h_n.
 
        # choose r_out at the last time step
        out = self.out(r_out[:, -1, :])  # r_out shape (batch, time_step, output_size)  在time_step位置插上-1就表示最后一个时刻
        return out
 
 
rnn = RNN()
print(rnn)
 
optimizer = torch.optim.Adam(rnn.parameters(), lr=LR)   # optimize all cnn parameters
loss_func = nn.CrossEntropyLoss()                       # the target label is not one-hotted
 
# training and testing
for epoch in range(EPOCH):
    for step, (b_x, b_y) in enumerate(train_loader):        # gives batch data
        b_x = b_x.view(-1, 28, 28)                      # 变一下维度 在pytorch中是.view()的形式表示reshape    reshape x to (batch, time_step, input_size)
 
        output = rnn(b_x)                               # rnn output
        loss = loss_func(output, b_y)                   # cross entropy loss
        optimizer.zero_grad()                           # clear gradients for this training step
        loss.backward()                                 # backpropagation, compute gradients
        optimizer.step()                                # apply gradients
 
        if step % 50 == 0:
            test_output = rnn(test_x)                   # (samples, time_step, input_size)
            pred_y = torch.max(test_output, 1)[1].data.numpy()
            accuracy = float((pred_y == test_y).astype(int).sum()) / float(test_y.size)
            print('Epoch: ', epoch, '| train loss: %.4f' % loss.data.numpy(), '| test accuracy: %.2f' % accuracy)
 
# print 10 predictions from test data
test_output = rnn(test_x[:10].view(-1, 28, 28))
pred_y = torch.max(test_output, 1)[1].data.numpy()
print(pred_y, 'prediction number')
print(test_y[:10], 'real number')

运行结果：

 

-------------------------------------------------RNN regressor----------------------------------------------------------------------

##################################################################################

############################################################################

目的

通过sin曲线

去生成cos曲线

 

其中

def forward(self, x, h_state):
    # x (batch, time_step, input_size)
    # h_state (n_layers, batch, hidden_size)
    # r_out (batch, time_step, hidden_size)
    # r_out 保存所有time_step的hidden_state
    r_out, h_state = self.rnn(x, h_state)
 
    outs = []  # save all predictions
    for time_step in range(r_out.size(1)):  # calculate output for each time step
        outs.append(self.out(r_out[:, time_step, :]))
    return torch.stack(outs, dim=1), h_state

其中有一个类似递归的思想

不断的产生h_state 然后再作为输入

所以在后面调用的时候需要第一次传入一个h_state

 

其次self.rnn()  会生成r_out , h_state

区别于 self.lstm()   会生成r_out , (h_n , h_c)

 

将每一次time_step 的r_out 作为输入到out中

将结果存入outs[ ]

因为r_out shape (batch, time_step, hidden_size)

所以

outs = []  # save all predictions
for time_step in range(r_out.size(1)):  # calculate output for each time step
    outs.append(self.out(r_out[:, time_step, :]))

 

最后的返回值将outs[ ] 是一个list 将其变为Tensor的形式，将里面的东西压在一起

return torch.stack(outs, dim=1), h_state

 

 

在训练阶段

step 是训练的步数

 

start, end = step * np.pi, (step + 1) * np.pi

截取一小段距离

 

steps = np.linspace(start, end, TIME_STEP, dtype=np.float32,endpoint=False)  # float32 for converting torch FloatTensor
x_np = np.sin(steps)
y_np = np.cos(steps)

在每段距离上撒点  生成训练sin曲线 x_np   预测曲线cos曲线  y_np

 

x = torch.from_numpy(
    x_np[np.newaxis, :, np.newaxis])  # shape (batch, time_step, input_size)  增加了2个维度 batch 和 input_size 为1
y = torch.from_numpy(y_np[np.newaxis, :, np.newaxis])

增加两个维度  变为pytorch接收的维度

 

# !! next step is important !!
h_state = h_state.data  # repack the hidden state, break the connection from last iteration

这步非常重要

将每次训练的h_state结果 变为h_state.data 形式赋值给h_state

以下是完整的代码结果

 
import torch
 
from torch import nn
 
import numpy as np
 
import matplotlib.pyplot as plt
 
 
 
# torch.manual_seed(1)    # reproducible
 
 
 
# Hyper Parameters
 
TIME_STEP = 10      # rnn time step
 
INPUT_SIZE = 1      # rnn input size
 
LR = 0.02           # learning rate
 
 
 
# show data
 
steps = np.linspace(0, np.pi*2, 100, dtype=np.float32)  # float32 for converting torch FloatTensor
 
x_np = np.sin(steps)
 
y_np = np.cos(steps)
 
plt.plot(steps, y_np, 'r-', label='target (cos)')
 
plt.plot(steps, x_np, 'b-', label='input (sin)')
 
plt.legend(loc='best')
 
plt.show()
 
 
 
 
 
class RNN(nn.Module):
 
    def __init__(self):
 
        super(RNN, self).__init__()
 
 
 
        self.rnn = nn.RNN(
 
            input_size=INPUT_SIZE,
 
            hidden_size=32,     # rnn hidden unit
 
            num_layers=1,       # number of rnn layer
 
            batch_first=True,   # input & output will has batch size as 1s dimension. e.g. (batch, time_step, input_size)
 
        )
 
        self.out = nn.Linear(32, 1)
 
 
 
    def forward(self, x, h_state):
 
        # x (batch, time_step, input_size)
 
        # h_state (n_layers, batch, hidden_size)
 
        # r_out (batch, time_step, hidden_size)
 
        r_out, h_state = self.rnn(x, h_state)
 
 
 
        outs = []    # save all predictions
 
        for time_step in range(r_out.size(1)):    # calculate output for each time step
 
            outs.append(self.out(r_out[:, time_step, :]))
 
        return torch.stack(outs, dim=1), h_state
 
 
 
        # instead, for simplicity, you can replace above codes by follows
 
        # r_out = r_out.view(-1, 32)
 
        # outs = self.out(r_out)
 
        # outs = outs.view(-1, TIME_STEP, 1)
 
        # return outs, h_state
 
        
 
        # or even simpler, since nn.Linear can accept inputs of any dimension 
 
        # and returns outputs with same dimension except for the last
 
        # outs = self.out(r_out)
 
        # return outs
 
 
 
rnn = RNN()
 
print(rnn)
 
 
 
optimizer = torch.optim.Adam(rnn.parameters(), lr=LR)   # optimize all cnn parameters
 
loss_func = nn.MSELoss()
 
 
 
h_state = None      # for initial hidden state
 
 
 
plt.figure(1, figsize=(12, 5))
 
plt.ion()           # continuously plot
 
 
 
for step in range(100):
 
    start, end = step * np.pi, (step+1)*np.pi   # time range  截取一小段的距离
 
    # use sin predicts cos
 
    steps = np.linspace(start, end, TIME_STEP, dtype=np.float32, endpoint=False)  # float32 for converting torch FloatTensor
 
    x_np = np.sin(steps)
 
    y_np = np.cos(steps)
 
 
 
    x = torch.from_numpy(x_np[np.newaxis, :, np.newaxis])    # shape (batch, time_step, input_size)  增加了2个维度 batch 和 input_size 为1
 
    y = torch.from_numpy(y_np[np.newaxis, :, np.newaxis])
 
 
 
    prediction, h_state = rnn(x, h_state)   # rnn output
 
    # !! next step is important !!
 
    h_state = h_state.data        # repack the hidden state, break the connection from last iteration
 
 
 
    loss = loss_func(prediction, y)         # calculate loss
 
    optimizer.zero_grad()                   # clear gradients for this training step
 
    loss.backward()                         # backpropagation, compute gradients
 
    optimizer.step()                        # apply gradients
 
 
 
    # plotting
 
    plt.plot(steps, y_np.flatten(), 'r-')
 
    plt.plot(steps, prediction.data.numpy().flatten(), 'b-')
 
    plt.draw(); plt.pause(0.05)
 
 
 
plt.ioff()
 
plt.show()

运行结果：