【代码bug消除】TCN模型PyTorch版本代码时间步无法修改的小bug已解决
TCN模型PyTorch版本代码时间步无法修改的小bug已解决
- TCN模型PyTorch版本代码时间步无法修改的小bug已解决
-
- 修改版
-
- TCN.py文件处的修改
- TCN_torch的jupyter notebook主程序文件处的修改
TCN模型PyTorch版本代码时间步无法修改的小bug已解决
关于涡扇发动（CMPASS数据集）和滚动轴承RUL预测的pytorch版本代码的小问题——timestep（时间步）默认36，一旦修改，模型无法打印，后续也无法训练的问题，已经解决。
因为代码时我2020年的，当时一直在研究涡扇发动机的RUL预测，探究出的最佳时间步是36，所以我在TemporalBlock自定类中，默认写进去了，导致后续操作采用不同的序列长度也即时间步sequence_length无法修改
###时间步
sequence_length = 36
mask_value = 0
feats =features_colums
# feats = ['s' + str(i) for i in range(1, 1281)]##需要的特征数据维度，可以自己选择
with torch.no_grad():
fake_input = torch.randn(1,n_inputs,36)#根据自己设置的(miniBatch,features,timstep)修改
class TemporalBlock(nn.Module):
def __init__(self, n_inputs, n_outputs, kernel_size, stride, dilation, padding, dropout=0.2):
super(TemporalBlock, self).__init__()
self.conv1 =nn.Conv1d(n_inputs, n_outputs, kernel_size,
stride=stride, padding=padding, dilation=dilation)
self.chomp1 = Chomp1d(padding)
# self.relu1 = nn.ReLU()
self.leaky_relu1 = nn.LeakyReLU(negative_slope=0.1)
self.dropout1 = nn.Dropout2d(dropout)
self.conv2 = nn.Conv1d(n_outputs, n_outputs, kernel_size,
stride=stride, padding=padding, dilation=dilation)
self.chomp2 = Chomp1d(padding)
# self.relu2 = nn.ReLU()
self.leaky_relu2 = nn.LeakyReLU(negative_slope=0.1)
self.dropout2 = nn.Dropout2d(dropout)
# compute shape by doing a forward pass
with torch.no_grad():
fake_input = torch.randn(1,n_inputs,36)#根据自己设置的(miniBatch,features,timstep)修改
out = self.conv1(fake_input)
# out = self.chomp1(out)
bn1_size = out.size()[1:]
out = self.chomp1(out)
out = self.conv2(out)
# out = self.chomp2(out)
bn2_size = out.size()[1:]
self.layer_norm1 = LayerNorm(bn1_size)
self.layer_norm2 = LayerNorm(bn2_size)
self.net = nn.Sequential(self.conv1,self.layer_norm1,self.chomp1,self.leaky_relu1, self.dropout1,
self.conv2,self.layer_norm2,self.chomp2,self.leaky_relu2, self.dropout2)
self.downsample = nn.Conv1d(n_inputs, n_outputs, 1) if n_inputs != n_outputs else None
# self.relu = nn.ReLU()
self.leaky_relu = nn.LeakyReLU(negative_slope=0.1)
self.init_weights()
修改版
TCN.py文件处的修改
（1）tcn.py文件下的第 22行
class TemporalBlock(nn.Module):
def __init__(self, n_inputs,timestep, n_outputs, kernel_size, stride, dilation, padding, dropout=0.2):
super(TemporalBlock, self).__init__()
（2）tcn.py文件下的第 41行
with torch.no_grad():
fake_input = torch.randn(1,n_inputs,timestep)#根据自己设置的(miniBatch,features,timstep)修改
（3）tcn.py文件下的第 78行
class TemporalConvNet(nn.Module):
def __init__(self, num_inputs,timestep, num_channels, kernel_size=2, dropout=0.2, max_length=200, attention=False):
super(TemporalConvNet, self).__init__()
layers = []
（4）tcn.py文件下的第 86行
out_channels = num_channels[i]
layers += [TemporalBlock(in_channels, timestep, out_channels, kernel_size, stride=1, dilation=dilation_size,
padding=(kernel_size-1) * dilation_size, dropout=dropout)]
if attention == True:
TCN_torch的jupyter notebook主程序文件处的修改
（1）TCN模型构建处的修改
"""构建model"""
class TCN(nn.Module):
def __init__(self, input_size, timestep, output_size, num_channels, kernel_size, dropout):
super(TCN, self).__init__()
self.tcn = TemporalConvNet(input_size, timestep, num_channels, kernel_size=kernel_size, dropout=dropout)
self.linear = nn.Linear(num_channels[-1], output_size)
def forward(self, inputs):
"""Inputs have to have dimension (N, C_in, L_in)"""
y1 = self.tcn(inputs) # input should have dimension (N, C, L)
o = self.linear(y1[:, :, -1])
return F.relu(o)
（2）TCN模型初始化+打印模型处的修改
###设置随机数种子
np.random.seed(1234)
torch.manual_seed(1234)
batch_size=128
input_channels=x_train.shape[-1]
n_classes=1
"""此处就是时间步，需要修改一下"""
seq_length =sequence_length
epochs = 20
steps = 0
###TCN模型的每一层的卷积单元数目，卷积核大小设置
channel_sizes = [64, 32, 16,16, 8, 8]
kernel_size=3
model = TCN(input_channels, seq_length, n_classes, channel_sizes, kernel_size=kernel_size, dropout=0.25).to(device)
optimizer = torch.optim.Adam(model.parameters(), lr=0.001)##初始学习率可以自己自行设置
epoch_scheduler = torch.optim.lr_scheduler.ReduceLROnPlateau(optimizer,mode='min',factor=0.1,patience=50,
min_lr=0.0001)###设置学习率下降策略，连续50次loss没有下降，学习率为原来的0.1倍，最小值为0.0001
Losses = nn.MSELoss()
"""这一步是打印你的TCN模型，input_channels为特征维度，sequence_length为时间步，
轴承数据看你的特征列数"""
summary(model,(input_channels, sequence_length))
##下面diamante是当你已经保存好了训练的模型以及权重以后，离线读取模型权重
# model.load_state_dict(torch.load(os.path.join(save_dir, "model_onTestBest.pth")))
新的打印模型如下时间步取30：
大功告成，你们来去试试吧，目前还没有人提出这个问题，我先发出来了
----------------------------------------------------------------
Layer (type) Output Shape Param #
================================================================
Conv1d-1 [-1, 16, 32] 61,504
Conv1d-2 [-1, 16, 32] 61,504
LayerNorm-3 [-1, 16, 32] 1,024
LayerNorm-4 [-1, 16, 32] 1,024
Chomp1d-5 [-1, 16, 30] 0
Chomp1d-6 [-1, 16, 30] 0
LeakyReLU-7 [-1, 16, 30] 0
LeakyReLU-8 [-1, 16, 30] 0
Dropout2d-9 [-1, 16, 30] 0
Dropout2d-10 [-1, 16, 30] 0
Conv1d-11 [-1, 16, 32] 784
Conv1d-12 [-1, 16, 32] 784
LayerNorm-13 [-1, 16, 32] 1,024
LayerNorm-14 [-1, 16, 32] 1,024
Chomp1d-15 [-1, 16, 30] 0
Chomp1d-16 [-1, 16, 30] 0
LeakyReLU-17 [-1, 16, 30] 0
LeakyReLU-18 [-1, 16, 30] 0
Dropout2d-19 [-1, 16, 30] 0
Dropout2d-20 [-1, 16, 30] 0
Conv1d-21 [-1, 16, 30] 20,512
LeakyReLU-22 [-1, 16, 30] 0
TemporalBlock-23 [-1, 16, 30] 0
Conv1d-24 [-1, 16, 34] 784
Conv1d-25 [-1, 16, 34] 784
LayerNorm-26 [-1, 16, 34] 1,088
LayerNorm-27 [-1, 16, 34] 1,088
Chomp1d-28 [-1, 16, 30] 0
Chomp1d-29 [-1, 16, 30] 0
LeakyReLU-30 [-1, 16, 30] 0
LeakyReLU-31 [-1, 16, 30] 0
Dropout2d-32 [-1, 16, 30] 0
Dropout2d-33 [-1, 16, 30] 0
Conv1d-34 [-1, 16, 34] 784
Conv1d-35 [-1, 16, 34] 784
LayerNorm-36 [-1, 16, 34] 1,088
LayerNorm-37 [-1, 16, 34] 1,088
Chomp1d-38 [-1, 16, 30] 0
Chomp1d-39 [-1, 16, 30] 0
LeakyReLU-40 [-1, 16, 30] 0
LeakyReLU-41 [-1, 16, 30] 0
Dropout2d-42 [-1, 16, 30] 0
Dropout2d-43 [-1, 16, 30] 0
LeakyReLU-44 [-1, 16, 30] 0
TemporalBlock-45 [-1, 16, 30] 0
Conv1d-46 [-1, 8, 38] 392
Conv1d-47 [-1, 8, 38] 392
LayerNorm-48 [-1, 8, 38] 608
LayerNorm-49 [-1, 8, 38] 608
Chomp1d-50 [-1, 8, 30] 0
Chomp1d-51 [-1, 8, 30] 0
LeakyReLU-52 [-1, 8, 30] 0
LeakyReLU-53 [-1, 8, 30] 0
Dropout2d-54 [-1, 8, 30] 0
Dropout2d-55 [-1, 8, 30] 0
Conv1d-56 [-1, 8, 38] 200
Conv1d-57 [-1, 8, 38] 200
LayerNorm-58 [-1, 8, 38] 608
LayerNorm-59 [-1, 8, 38] 608
Chomp1d-60 [-1, 8, 30] 0
Chomp1d-61 [-1, 8, 30] 0
LeakyReLU-62 [-1, 8, 30] 0
LeakyReLU-63 [-1, 8, 30] 0
Dropout2d-64 [-1, 8, 30] 0
Dropout2d-65 [-1, 8, 30] 0
Conv1d-66 [-1, 8, 30] 136
LeakyReLU-67 [-1, 8, 30] 0
TemporalBlock-68 [-1, 8, 30] 0
Conv1d-69 [-1, 8, 46] 200
Conv1d-70 [-1, 8, 46] 200
LayerNorm-71 [-1, 8, 46] 736
LayerNorm-72 [-1, 8, 46] 736
Chomp1d-73 [-1, 8, 30] 0
Chomp1d-74 [-1, 8, 30] 0
LeakyReLU-75 [-1, 8, 30] 0
LeakyReLU-76 [-1, 8, 30] 0
Dropout2d-77 [-1, 8, 30] 0
Dropout2d-78 [-1, 8, 30] 0
Conv1d-79 [-1, 8, 46] 200
Conv1d-80 [-1, 8, 46] 200
LayerNorm-81 [-1, 8, 46] 736
LayerNorm-82 [-1, 8, 46] 736
Chomp1d-83 [-1, 8, 30] 0
Chomp1d-84 [-1, 8, 30] 0
LeakyReLU-85 [-1, 8, 30] 0
LeakyReLU-86 [-1, 8, 30] 0
Dropout2d-87 [-1, 8, 30] 0
Dropout2d-88 [-1, 8, 30] 0
LeakyReLU-89 [-1, 8, 30] 0
TemporalBlock-90 [-1, 8, 30] 0
TemporalConvNet-91 [-1, 8, 30] 0
Linear-92 [-1, 1] 9
================================================================
Total params: 164,177
Trainable params: 164,177
Non-trainable params: 0
----------------------------------------------------------------
Input size (MB): 0.15
Forward/backward pass size (MB): 0.27
Params size (MB): 0.63
Estimated Total Size (MB): 1.04
----------------------------------------------------------------