R：為時間序列中的異常檢測擬合深度學習模型時出錯

Question

我找到了 DeepAnT（時間序列中無監督異常檢測的深度學習方法）的示例，並想嘗試一下。

所以我嘗試將 Python 腳本（ https://github.com/swlee23/deep-learning-time-series-anomaly-detection/blob/master/deep-ant-main.ipynb ）轉換為 R 並被卡住，因為我擬合模型時出錯。

代碼：

library(ggplot2)
library(rlist)
library(keras)
library(tensorflow)
library(readr)

### Anomaly Detection ###
anomaly_detector <- function(prediction_seq, ground_truth_seq){
  
  # calculate Euclidean between actual seq and predicted seq
  dist <- norm(ground_truth_seq - prediction_seq, type = "F")  
  
  if (dist > anm_det_thr){
    return (TRUE)  # anomaly
  }
  
  else{
    return (FALSE) # normal
  }
  
}

### Hyperparameters ###

w = 2000                 # History window (number of time stamps taken into account) 
# i.e., filter(kernel) size       
p_w = 300                # Prediction window (number of time stampes required to be 
# predicted)
n_features = 1           # Univariate time series

kernel_size = 2          # Size of filter in conv layers
num_filt_1 = 32          # Number of filters in first conv layer
num_filt_2 = 32          # Number of filters in second conv layer
num_nrn_dl = 40          # Number of neurons in dense layer
num_nrn_ol = p_w         # Number of neurons in output layer

conv_strides = 1
pool_size_1 = 2          # Length of window of pooling layer 1
pool_size_2 = 2          # Length of window of pooling layer 2
pool_strides_1 = 2       # Stride of window of pooling layer 1
pool_strides_2 = 2       # Stride of window of pooling layer 2

epochs = 30
dropout_rate = 0.5       # Dropout rate in the fully connected layer
learning_rate = 2e-5  
anm_det_thr = 0.8        # Threshold for classifying anomaly (0.5~0.8)

# Loading data
df_sine <- read_csv('https://raw.githubusercontent.com/swlee23/Deep-Learning-Time-Series-Anomaly-Detection/master/data/sinewave.csv')


### Data preprocessing ###

# split a univariate sequence into samples
split_sequence <- function(sequence){
  
  X <- list() 
  y <- list()
  
  
  for(i in 1:length(sequence)){
    # find the end of this pattern
    end_ix <- i + w
    out_end_ix <- end_ix + p_w
    # check if we are beyond the sequence
    if (out_end_ix > length(sequence)){
      # print("if-break")
      break
    }
    
    # gather input and output parts of the pattern
    seq_x <- list(sequence[i:end_ix])
    seq_y <- list(sequence[end_ix:out_end_ix])

    X[length(X)+1] <- seq_x
    y[length(y)+1] <- seq_y

  }
  
  result <- list("x" = X, "y" = y)
  return(result)
}


# define input sequence
raw_seq = df_sine['sinewave'][[1]]


# split into samples
batch_ <- split_sequence(raw_seq)
batch_sample <- batch_$x
batch_label <-batch_$y

# tried to convert to matrix (did not change anything)
batch_sample_2 <- matrix(unlist(batch_sample), ncol = 2001, byrow = TRUE)
batch_label_2 <- matrix(unlist(batch_label), ncol = 301, byrow = TRUE)


### Generate model for predictor ###
model <- keras_model_sequential() %>% 
  layer_conv_1d(filters = num_filt_1,
                kernel_size = kernel_size,
                strides = conv_strides,
                padding = 'valid',
                activation = 'relu',
                input_shape = c(w,n_features)
  ) %>% layer_max_pooling_1d(
    pool_size = pool_size_1
  )%>% layer_conv_1d(
    filters = num_filt_2,
    kernel_size = kernel_size,
    strides = conv_strides,
    padding = 'valid',
    activation = 'relu'
  ) %>% layer_max_pooling_1d(
    pool_size = pool_size_2
  )

# Flatten tensor into a batch of vectors
# Input Tensor Shape: [batch_size, 0.25 * w, num_filt_1 * num_filt_2]
# Output Tensor Shape: [batch_size, 0.25 * w * num_filt_1 * num_filt_2]
model <- model %>% layer_flatten()

# Dense Layer (Output layer)
# Densely connected layer with 1024 neurons
# Input Tensor Shape: [batch_size, 0.25 * w * num_filt_1 * num_filt_2]
# Output Tensor Shape: [batch_size, 1024]
model <- model %>% layer_dense(units = num_nrn_dl, activation = 'relu') 

# Dropout
# Prevents overfitting in deep neural networks
model <- model %>% layer_dropout(rate = dropout_rate)


# Output layer
# Input Tensor Shape: [batch_size, 1024]
# Output Tensor Shape: [batch_size, p_w]
model <- model %>% layer_dense(units = num_nrn_ol)

# Summarize model structure
summary(model)


### Configure model ###

model <- model %>% compile(optimizer = 'adam',
                           loss = 'mean_absolute_error')

### Training ###

model_fit <- model %>% fit(batch_sample_2,
                           batch_label_2,
                           epochs = epochs,
                           verbose = 1)

我得到的錯誤：

Error in py_call_impl(callable, dots$args, dots$keywords) : 
ValueError: Error when checking input: expected conv1d_6_input to have 3 dimensions, but got array with shape (2701, 2001) 

在 Python 中， batch_sample如下所示：

[[0.8737364 ]
 [0.90255357]
 [0.92780878]
 ...
 [0.7671179 ]
 [0.80588467]
 [0.84147098]]

batch_label像這樣：

[[0.84147098 0.8737364  0.90255357 ... 0.72532366 0.7671179  0.80588467]
 [0.8737364  0.90255357 0.92780878 ... 0.7671179  0.80588467 0.84147098]
 [0.90255357 0.92780878 0.94940235 ... 0.80588467 0.84147098 0.8737364 ]
 ...
 [0.80588467 0.84147098 0.8737364  ... 0.68066691 0.72532366 0.7671179 ]
 [0.84147098 0.8737364  0.90255357 ... 0.72532366 0.7671179  0.80588467]
 [0.8737364  0.90255357 0.92780878 ... 0.7671179  0.80588467 0.84147098]]

所以我需要為擬合過程重塑數據，但我嘗試的一切都不起作用。 嘗試過列表列表（就像在 Python 中一樣）以及矩陣和向量，但每次都會出現維度錯誤。 我需要如何對數據進行整形以使其符合擬合函數的要求？

Answer 1

您的輸入形狀不正確。 模型需要 3 個暗淡的形狀輸入：(n_samples, window_size, n_features)。
在您的代碼中，n_features 等於 1，因此您需要做的就是將當前輸入形狀重塑為模型期望的形狀（基於輸入配置），因此：(2701, 2001) -> (2701, 2001, 1)

為此，您可以使用listarrays 中的 expand_dims 之類的函數

expand_dims(x, -1)