explain ground-thruth .mat file of an image for CNN

See also questions close to this topic

• Calling a Function in Matlab with Symbolic Functions as Arguments

  I have written the following Matlab function:

  function EulerMethod(t_min,t_max,h,f,Y,yzero)
  
  tlist = t_min:h:t_max;
  
  N = (t_max - t_min)/h;
  
  ylist = transpose(zeros(N+1,1));
  
  ylist(1) = yzero; 
  
  for i=1:N
      term = f(tlist(i),ylist(i))*h;
      ylist(i+1) = ylist(i) + term;
  end
  
  yrange = Y(tlist);
  
  % modified to generate a new figure window each time.
  figure;
  
  plot(tlist,yrange,'red','LineWidth', 2);
  hold;
  plot(tlist,ylist,'blue','LineWidth', 2);
  plot(tlist, abs(yrange - ylist),'magenta','LineWidth', 2)
  
  % modified to wrap the title.
  title({'Graphs of the True Solution, Euler Solution,', 'and the Absolute Value of the Global Error (GE)'})
  xlabel('t')
  ylabel('Y(t)')
  
  legend({'True Solution','Euler Solution', 'Absolute Value of the GE'},'Location','southwest')
  
  end 
  

  I now try and call this function in another script. The variables f and Y in the function are symbolic functions; so, in the other file, I first declare these functions before calling this function. Here's the code:

  clc
  
  syms f(t,y)
  syms Y(t)
  
  f(t,y) = -y + 2.0*cos(t); %This the derivative of the function whose solution we're trying to computed
  Y(t) = sin(t) + cos(t); %This is the true solution;
  
  % Calling function
  
  EulerMethod(0.0, 6.0, 0.2, f, Y, 1.0);
  

  I, however, get errors when I run the second script. Can anyone help me figure out what's going wrong? I suspect this may be because of the way I have both input and used the symbolic functions f and Y but I am not sure.
• How to plot a multidimensional array in matlab?

  I have a table as follows

  system:index 2017_06_18 2017_06_19 2017_06_20 2017_06_21 2 612.8099664 1174.656713 1282.083251 815.3828357 3 766.4103726 1345.135952 1322.726083 749.998993 4 765.0230453 1411.669136 1350.437586 610.9541838 5 553.5858458 1374.14789 1152.086957 566.7924468 6 466.9780908 1311.903756 1060.494001 559.1982264 7 257.1162602 1270.182385 988.5455285 562.9224932 8 230.6611542 1310.971988 1001.548768 502.3266959

  I want to plot a 2d-colormap representing system:index as y axis, dates as x axis and values under dates as colors. I tried with the following code but it did not give what I want.

    clear
    clc
    filename = 'TurbidityDailyMean.xlsx';
    data = xlsread(filename,'TurbidityDailyMean','A1:E8');
  
    figure;
    hold on
  
    for i = 2:5
       y = data(:,1);
       x = data(:,i);
     plot(x,y) 
    end
  

  I need to map a colormap as mentioned above. But from what I tried it gives something else. And another fact is that I can't insert system:index and dates row into matlab with relevant data.
• Regexp syntax MATLAB

  input1 = ' 8 BKN 15 BKN '  
  input2 = ' 2 X 3SM ' 
  
  regexp(input1, '\s{1}\d(1|2)\s{1}c{3}\s{1}')
  regexp(input2, ''\s{1}\d(1|2)\c{1}\s{1}c{1}\s{1}' )
  

  Have trouble getting regexp to work. I'm not at all great a debugging.

  The code needs to identify (one space, (one digit or two digits), one space, three characters [A-Z], and one space)

  The code needs (one space, one digit or two digits, one space, X the letter, one digit and two characters, and one space)
• Keras InvalidArgumentError: Incompatible shapes: [1,8,32] vs. [1,10,32]

  It stops training at the beginning of the first epoch.

  The error is:

  InvalidArgumentError: Incompatible shapes: [1,8,32] vs. [1,10,32]
       [[Node: training_8/RMSprop/gradients/loss_8/time_distributed_9_loss/mul_grad/BroadcastGradientArgs = BroadcastGradientArgs[T=DT_INT32, _class=["loc:@train.../Reshape_1"], _device="/job:localhost/replica:0/task:0/device:CPU:0"](training_8/RMSprop/gradients/loss_8/time_distributed_9_loss/mul_grad/Shape, training_8/RMSprop/gradients/loss_8/time_distributed_9_loss/mul_grad/Shape_1)]]
  

  I'm not sure if the error might be the format of the training data. This is a sample of train_x:

  [array([ 0,  1,  2,  3,  4,  5,  6,  3,  7,  8,  9, 10, 11, 12, 13, 14], dtype=int32), array([15, 16, 17, 18, 19], dtype=int32), array([20, 16, 17, 18, 21, 22, 23, 24, 25, 26], dtype=int32), array([27,  1, 17, 28, 18, 29, 30, 31, 24, 32], dtype=int32), array([33,  1, 17,  3, 34, 35, 36, 37, 18, 38], dtype=int32), array([39, 16, 40, 28, 41, 42], dtype=int32), array([39,  1, 40, 28, 41, 43], dtype=int32), array([44,  6,  3, 45], dtype=int32), array([15, 16, 40, 46, 47, 48,  3, 49, 50, 51, 52, 53], dtype=int32), array([ 0, 54, 28, 55, 56, 57, 58, 59], dtype=int32)]
  

  This is a sample of train_label

  [array([0, 1, 2, 1, 2, 3, 0, 1, 4, 2, 2, 4, 5, 6, 0, 7], dtype=int32), array([0, 5, 8, 9, 9], dtype=int32), array([10,  5,  8,  9,  7,  9,  7,  7], dtype=int32), array([10,  1,  8,  1,  9,  9,  3,  0,  7,  2,  7], dtype=int32), array([10,  1,  8,  1,  9,  9,  9,  7,  9,  7], dtype=int32), array([0, 5, 8, 1, 2, 2], dtype=int32), array([0, 1, 8, 1, 2, 2], dtype=int32), array([11,  0,  1,  9], dtype=int32), array([ 0,  5,  8,  7, 12, 13,  1,  2,  9,  9,  5,  6], dtype=int32), array([ 0, 14,  1,  2,  2, 12,  4, 15], dtype=int32)]
  

  This is the code:

  n_vocab = len(unique_words) #1517
  model = Sequential()
  model.add(Embedding(n_vocab,100))
  model.add(Convolution1D(128, 5, border_mode='same', activation='relu'))
  model.add(Dropout(0.25))
  model.add(GRU(100,return_sequences=True))
  model.add(TimeDistributed(Dense(n_classes, activation='softmax')))
  model.compile('rmsprop', 'categorical_crossentropy')
  
  n_epochs = 30
  train_x = [np.asarray(x, dtype=np.int32) for x in encoded_sentences]
  train_label = [np.asarray(x, dtype=np.int32) for x in encoded_POS]
  
  for i in range(n_epochs):
      print("Training epoch {}".format(i))
  
      bar = progressbar.ProgressBar(maxval=len(train_x))
      for n_batch, sent in bar(enumerate(train_x)):
          label = train_label[n_batch]
          # Make labels one hot
          label = np.eye(n_classes)[label][np.newaxis,:]
          # View each sentence as a batch
          sent = sent[np.newaxis,:]
  
          if sent.shape[1] > 1: #ignore 1 word sentences
              model.train_on_batch(sent, label)
  
• MNIST data denormalising does not give me back the same

  This is part of my learning. I Understood the normalization is really helping to improve accuracy, and hence divided by 255 the mnist values. This will divide all the pixels by 255, and hence all the pixels of 28*28 will have the values in range from 0.0 to 1.0 .

  Now i tired to multiply the same with 255,this essentially means we should get the original value back. but when i display the picture, both the original and de-normalised pictures are different.

  (trainX, trainY), (testX, testY) = mnist.load_data()
  
  
  plt.subplot(2,2,1)
  plt.imshow(trainX[143])
  
  trainX /= 255
  
  plt.subplot(2,2,2)
  plt.imshow(trainX[143])
  
  
  trainX *= 255
  
  plt.subplot(2,2,3)
  plt.imshow(trainX[143])
  plt.show()
  

  OUtput:

  

  What am i missing?. Anything related to float and int data type of the input data?
• Neural net not converging after enlargement of training dataset

  I used to train my network on a dataset of 3,500 images (5 classes, 700 from each class) and 21000 images (still the same 3,500 images, but apply random augmentation (rotate, flip, translate, stretch etc.) on each image to generate 5 more images, 3,500 + 3,500*5 = 21,000), and my network can always converge, but it will obviously overfit (train accuracy > 90% while test accuracy < 40%).

  Now I am using a significantly larger dataset, with around 35,000 original images (from the same dataset of the 3,500 mentioned above). This dataset is not balanced (25,000 from class 0, and class 1,2,3,4 share the remaining 10,000), therefore I apply random augmentation on class 1,2,3,4 so that each class have 25,000 images (original + augmentation).

  That is to say, now I have a dataset of 125,000 images in total.

  When I was training on the 3500 images, after 20 - 30 epochs, the loss will obviously reduce and the training accuracy will reach 30% (not an amazing number, but a good start that at least shows the network is learning something, because a random predictor for 5 classes will get 20% accuracy) .

  However, when it comes to 125,000 images, without any change to the structure of my network, it seems that my network won't learn anything. After 20 epochs, the loss won't reduce and the training accuracy is around 20%, which is no better than a random predictor.

  This is so weird and I just can't figure out what could be the reason for this.

  The following is the structure of the network I am training:

              units   filter stride size 
   1 Input                           112 
   2 Conv        32     5       2     56 
   3 Conv        32     3             56 
   4 MaxPool            3       2     27 
   5 Conv        64     5       2     14 
   6 Conv        64     3             14 
   7 Conv        64     3             14 
   8 MaxPool            3       2      6 
   9 Conv       128     3              6 
  10 Conv       128     3              6  
  11 Conv       128     3              6  
  12 RMSPool            3       3      2  
  13 Dropout(0.5)
  14 Dense     1024
  15 Maxout     512
  16 Dropout(0.5)
  17 Dense     1024
  18 Maxout     512
  19 Dense       1     (output prediction)
  

  This is a structure used in the solution of team o_O, the 2nd place in Kaggle Diabetic Retinopathy Detection competition.

  And it did learning something when I trained it on the dataset of 3500 images.

  What might be the reason that it can't learn anything when I use more training data?

  Edit:

  optimizer: I am using nesterov, optimizer = tf.train.MomentumOptimizer(learning_rate=my_learning_rate,momentum=0.9,use_nesterov=True) where my_learning_rate can be 0.003,0.0003 or 0.00003, which were planned to be the learning rate for the 1-150,151-200,200-220 epochs. And 0.003 worked well during my training on the 3500/21000 images.

  My codes to take in the training set for the 3500(or 21000 in fact) images:

  def my_train_estimator():
      dataset = tf.data.TextLineDataset(train_csv_3500).map(_parse_csv_train) 
      d1 = tf.data.TextLineDataset(train_csv_3500).map(_my_parse_Translation(myRandomAugsInOne))
      d2 = tf.data.TextLineDataset(train_csv_3500).map(_my_parse_Translation(myRandomAugsInOne))
      d3 = tf.data.TextLineDataset(train_csv_3500).map(_my_parse_Translation(myRandomAugsInOne))
      d4 = tf.data.TextLineDataset(train_csv_3500).map(_my_parse_Translation(myRandomAugsInOne))    
      d5 = tf.data.TextLineDataset(train_csv_3500).map(_my_parse_Translation(myRandomAugsInOne))
      dataset = dataset.concatenate(d1).concatenate(d2).concatenate(d3).concatenate(d4).concatenate(d5)
      dataset = dataset.shuffle(21000)
      dataset = dataset.batch(100)
      iterator = dataset.make_one_shot_iterator()  # create one shot iterator
      feature, label = iterator.get_next()
      return feature, label
  

  where myRandomAugsInOne do the random augmentations, and _my_parse_Translation is just parsing the csv and apply whatever transformation function to the images it got.

  def _my_parse_Translation(transform_func):
      def my_parse_func(line):
          img,label = _parse_csv_train(line)
          imgOut = transform_func(img)
          return imgOut, label
      return my_parse_func
  

  where _parse_csv_train is:

  def _parse_csv_train(line):
      parsed_line= tf.decode_csv(line, [[""], []])
      filename = my_file_folder_string + parsed_line[0] + ".jpeg"
      label = parsed_line[1]
      image_string = tf.read_file(filename)
      image_decoded = tf.image.decode_jpeg(image_string, channels=3)
      image_resized = tf.image.resize_images(image_decoded, image_resize)
      return image_resized, label
  

  My codes to take in the training set for the 125000(or 25000 in fact) images:

  def my_estimator_func_withAug():
      w = calResampleRate()
      wdr = np.round(w/w[0]).astype(np.int64)
      d0 = tf.data.TextLineDataset(train_csv_0).map(_my_parse_Translation(myRandomAllInOne))
      d1 = tf.data.TextLineDataset(train_csv_1).map(_my_parse_Translation(myRandomAllInOne))
      d2 = tf.data.TextLineDataset(train_csv_2).map(_my_parse_Translation(myRandomAllInOne))
      d3 = tf.data.TextLineDataset(train_csv_3).map(_my_parse_Translation(myRandomAllInOne))
      d4 = tf.data.TextLineDataset(train_csv_4).map(_my_parse_Translation(myRandomAllInOne))
      d1 = d1.repeat(wdr[1])
      d2 = d2.repeat(wdr[2])
      d3 = d3.repeat(wdr[3])
      d4 = d4.repeat(wdr[4])
      d0 = d0.shuffle(shuffle_buffer_size).take(take_size) 
      d1 = d1.shuffle(shuffle_buffer_size).take(take_size) 
      d2 = d2.shuffle(shuffle_buffer_size).take(take_size) 
      d3 = d3.shuffle(shuffle_buffer_size).take(take_size) 
      d4 = d4.shuffle(shuffle_buffer_size).take(take_size) 
      dataset = d0.concatenate(d1).concatenate(d2).concatenate(d3).concatenate(d4)    
      dataset = dataset.shuffle(shuffle_buffer_size)
      dataset = dataset.batch(128)
      iterator = dataset.make_one_shot_iterator()  
      feature, label = iterator.get_next()
      return feature, label
  

  where shuffle_buffer_size = 25000 and take_size = 5000, and calResampleRate() is just calculating the weights to balance each class, for example it could be something like [1.0,10.6,4.9,29.6,36.4] and then wdr = np.round(w/w[0]).astype(np.int64) will round it to [1,11,5,30,36], and because class 0 is the majority in the dataset, I just always set its weight to 1 and calculate other classes' weights according to it. That's why I only repeat() datasets for class 1,2,3,4.

  Please notice that the above is the only difference, I didn't change my network or my optimizer/ loss function

  The 5 classes approximately have 25000 images in class 0 , 2400 images in class 1, 5300 images in class 2, 870 images in class 3 and 710 images in class 4

  I have to admit that the ways I used my data in these two cases were different.

  In the first one I am training with (3500 original images (700 from each class) + 5x3500 duplicate(randomly augmented) images),

  while in the second one I am training with ( randomly picked 5000 images from "25000 randomly augmented class 0 images" + randomly picked 5000 images from "2400*11 randomly augmented class 1 images" + randomly picked 5000 images from "5300*5 randomly augmented class 2 images" + randomly picked 5000 images from "870*30 randomly augmented class 3 images" + randomly picked 5000 images from "710*36 randomly augmented class 4 images")

  In the second case, in every "epoch" my training data is just part of the "whole 25000 + 2400 + 5300 + 870 + 710 training dataset".

  Edit 2:

  Please notice that, it was not a binary classification, it was regression using MSE as loss function. The output will be clipped to values between 0 - 4, then apply thresholds at 0.5, 1.5, 2.5, 3.5 to class 0,1,2,3,4.

  Why did I (or why did team o_O) do this? Because the competition was using quadratic weighted kappa (qwk) instead of accuracy as evaluation criteria, and qwk punishes more on a mistake that deviates more from the true label.

  Then why did the competition use qwk ? Probably because 0-4 means different levels of diabetic retinopathy illness, where 0 represents healthy (and that's why we have class 0 as majority, most people are healthy !) and 4 represents the most severe level of diabetic retinopathy.
• Dataset containing spatial and temporal features (built on a CNN model)

  A dataset contains spatial and temporal features.

  It contains the time series data (2 min intervals) of the sections of a map.

  It is 320*480 (320 map sections and 480-time intervals). Each row belongs to a section and each column is a time interval. So each cell in the dataset is a value of data in a map section in a time interval.

  I would like to make a model to import all map sections in 40 min and predict the following 10 next min of all map sections.

  So the input shape equals 320*20 and the output shape will be 320*5.

  I split data to 320*20 as X and 320*5 as Y samples.

  The input and the output of the model are 2-dimensional matrixes.

  I can consider the input as a one channel image (320*20*1) and it is possible to use a CNN model.

  The codes of the final architecture of my model in Keras:

  model = models.Sequential()
  model.add(Conv2D(256,(3, 3),
               activation='relu',
               input_shape=(320,20,1), padding='same'))
  model.add(MaxPooling2D((2, 2)))
  
  model.add(Conv2D(128, (3, 3), activation='relu', padding='same'))
  model.add(MaxPooling2D((2, 2)))
  
  model.add(Conv2D(64, (3, 3), activation='relu', padding='same'))
  model.add(MaxPooling2D((2,2)))
  
  model.add(Flatten())
  model.add(Dense(1600))
  
  model.summary()
  

  The Loss value(MSE) is about 70. Is it lower than the random prediction?

  I reshaped the output to 1600 (320*5) which the final Dense layer expects, while the output is 2d. Does it make the model more prone to error?

  Thank you
• Share weights between keras layers without input shape

  Can I share weights between keras layers but have other parameters differ?

  The answer to the question above offered great insights in terms of sharing weights between two conv layers. While implementing it on sharing weights between conv2 and conv2', I'm having trouble to identify what the input shape should be for conv2 since it's not directly connect to the data.

  For example, in a model with the following structure:

  input (shape: 32,32,3) -> conv1 (connected to input [0][0],output shape (None,32,32,32) according to model.summary) -> conv2(connected to conv1 [29][0], output shape (None,30,30,32))

  conv1 = layers.Conv2D(32, 3, activation='relu', padding='same', input_shape=(32,32,3))
  conv2 = layers.Conv2D(32, 3, activation='relu',name = 'conv2')
  

  I'm trying to create conv1' that share weights from conv1, and conv2' that shares weight with conv2. I understand that the input shape for conv1 would be (32,32,3), but what would be the input shape for conv2? (I tried (32,32,32), doesn't work as conv2 and conv2' weights are different)

  Can I do it without specifying the input shape? also...why conv2 is connected to conv1[29][0]?
• Issues and failures in detecting multiple classes with YOLO v2 whereas single-class detection is nearly accurate

  For a course project, I am training a single-pass network to detect multiple instrument symbols in an image. A very small subset of these classes are shown below.

  Image_Class_1 Image_Class_2 Image_Class_3 Image_Class_4

  I am using YOLO v2 to train and detect multiple classes in an image. Since the dataset for such symbols doesn't exist I performed data augmentation in the following two ways on the four symbols shown above.

  • Rotation of each symbol type with a step of 0.5 degrees, which provides 720 images for each class.
  • Further, the 720 images obtained were passed through Gaussian blurring (5x5 filter).
    This augments the dataset to 1440 images per class.

  The configuration file for YOLO v2 was changed only on the following lines:

  • Line 3: batch = 64
  • Line 4: subdivision = 64 (subdivision size of 8 makes CUDA run out of memory).
  • Line 244: classes = 4
  • Line 237: filters=(classes + 5)*5 = 45.

  The image size used for training is 500x500 and the whole dataset is divided into 80% and 20% ratio for training and testing.

  The network training was run till 3100 iterations, and the average loss was 0.048.

  However, the training results have confused me a bit as the results are rather unexpected in case of multi-class detection. In the single-class classification on the test image, the network performs as expected (although not very good for two classes which perhaps can be improved) as shown below:
  Detection class_2_image (probability = 78%) Detection class_1_image (probability = 98%) Detection class_3_image (probability = 98%)

  However, When I tried to perform multi-class detection on the test images, the detection and classification fail completely as can be seen in the resulting images below :
  Multi-class Detection biased for class_3

  If I vary the spatial distance between the symbols in the test image, then the bounding boxes and class prediction becomes even more random. Hence, I have a few questions regarding this problem:

  1. Since YOLO also performs data augmentation implicitly, does my data augmentation process interfere with the training?
  2. Do I need to resize the image to exact 416X416 as the input required by YOLO v2 ?
  3. I understand that YOLO v2 is a very heavy CNN architecture which perhaps is not meant for such images but single-class classification results are too promising to ignore. So can someone figure out why it is failing entirely for multi-class detection?
  4. Should I use YOLO v2 as it is or modify it in order to detect these symbols in the images? I guess if I modify I risk losing more and deteriorating detection further.
• Sort by a depth dimension in OpenCV with C++

  I have a multidimensional matrix, how can I sort the cube by the third dimension? using a proper function of opencv?

  std::vector<int> sz = { 3,3,4 };
  cv::Mat M(3, sz.data(), CV_32FC1, cv::Scalar(0));
  

  In the documentation there is only an option of

  CV_SORT_EVERY_COLUMN //dimension-1
  CV_SORT_EVERY_ROW //dimension-2
  
• Make Angular table row expandable

  I want to make an entire row of my Angular material mat-table expandable:

  <mat-table [dataSource]="dataSource">
  
        <!-- ArticleNumber Column -->  
        <ng-container matColumnDef="articleNumber">
          <mat-header-cell *matHeaderCellDef>
            <b>Article Number</b>
          </mat-header-cell>
          <mat-cell *matCellDef="let item">{{item.articleNumber}}</mat-cell>
        </ng-container>
  
  
        <!-- Payment Column -->
        <ng-container matColumnDef="payment">
          <mat-header-cell *matHeaderCellDef>
            <b>Payment</b></mat-header-cell>
          <mat-cell *matCellDef="let item">{{item.payment}}</mat-cell>
        </ng-container>
  
        <!-- DeliveryTime Column -->
        <ng-container matColumnDef="deliveryTime">
          <mat-header-cell *matHeaderCellDef>
            <b>Delivery Time</b>
          </mat-header-cell>
          <mat-cell *matCellDef="let item">{{item.deliveryTime}}</mat-cell>
        </ng-container>
  
        <mat-header-row *matHeaderRowDef="displayedColumns"></mat-header-row>
  
        <mat-row *matRowDef="let basketItem; columns: displayedColumns;" (click)="onSelect(basketItem)"></mat-row>
  
  </mat-table>
  

  So I have 3 columns, ArticleNumber, Payment and DeliveryTime.

  When I click a row, an expansion panel should display some content.

  Here is the basic usage of such a panel:

  <mat-accordion>
    <mat-expansion-panel (opened)="panelOpenState = true"
                         (closed)="panelOpenState = false">
      <mat-expansion-panel-header>
        <mat-panel-title>
          ROW 1
        </mat-panel-title>
        <mat-panel-description>
          Currently I am {{panelOpenState ? 'open' : 'closed'}}
        </mat-panel-description>
      </mat-expansion-panel-header>
      <p>ROW 1 CONTENT visible because I am clicked</p>
    </mat-expansion-panel>
  
    <mat-expansion-panel (opened)="panelOpenState = true"
                         (closed)="panelOpenState = false">
      <mat-expansion-panel-header>
        <mat-panel-title>
          ROW 2
        </mat-panel-title>
        <mat-panel-description>
          Currently I am {{panelOpenState ? 'open' : 'closed'}}
        </mat-panel-description>
      </mat-expansion-panel-header>
      <p>ROW 2 CONTENT visible because I am clicked</p>
    </mat-expansion-panel>
  
  </mat-accordion>
  

  Is the table with expandable rows doable?
• Data retrieval from multiple DataSources

  I am working on a mat-table which is populated from itemsList using dataSource:

  component.ts:

  dataSource: MatTableDataSource<BasketItem>; 
  this.dataSource = new MatTableDataSource(this.itemsList);
  

  component.html:

  <mat-table [dataSource]="dataSource">
  </mat-table>
  

  Is it possible to populate with items from a second list itemsList2 simultaneously?
• How data pass for previous view controller using protocol and update collectionview cell swift

  First view controller is Collectionview cell and pass data to nextview controller. Nextviewcontroller Edit the text and pass the text in first view contoller-collectionview cell and update the collectionview cell.
• Could realtime EEG signal used directly in LSTM?

  I am doing an experiment that using raw EEG data for LSTM. I have some quenstion about the input shape of this networks. Some one says that "You need a feature dimension, a batch size dimension and a time dimension". I think that means the input shape should be (feature,batch,time)。 But now I want to use the raw data without feature extraction.I think it may be unfitted for the raw data(256Hz/s).

  For example,I have a long data for an hour(921600points). If I seperated in time-domian, each slice have 512 points. Can I use the 512 points as the directly input as the feature dimension?
• Meaning: improvement in RMSE during crossvalidation, although not on testset?

  In the code below I train a NN with crossvalidation on the first 20000 records in the dataset. The dataset contains 8 predictors.

  First I have split my data in 2 parts: the first 20.000 rows (trainset) and the last 4003 rows (out of sample test set)

  I have done 2 runs: run 1) a run with 3 predictors run 2) a run with all 8 predictors (see code below).

  Based on crossvalidation within the 20.000 rows from the trainset, the RMSE (for the optimal parametersetting) improves from 2.30 (run 1) to 2.11 (run 2).

  Although when I test both models on the 4003 rows from the out of sample test set, the RMSE improves only neglible from 2.64 (run 1) to 2.63 (run 2).

  What can be concluded from these contradiction in the results?

  Thanks!

  ### R code from Applied Predictive Modeling (2013) by Kuhn and Johnson.
  ### Chapter 7: Non-Linear Regression Models
  ### Required packages: AppliedPredictiveModeling, caret, doMC (optional), 
  ### earth, kernlab, lattice, nnet
  ################################################################################
  
  library(caret)    
  ### Load the data
  
  mydata <- read.csv(file="data.csv", header=TRUE, sep=",")
  
  validatiex <- mydata[20001:24003,c(1:8)]
  validatiey <- mydata[20001:24003,9]
  
  mydata <-  mydata[1:20000,]
  
  x <- mydata[,c(1:8)]
  y <- mydata[,9]
  
  parti <- createDataPartition(y, times = 1, p=0.8, list = FALSE)    
  x_train <- x[parti,]
  x_test <- x[-parti,]
  y_train <- y[parti]
  y_test <- y[-parti]
  
  
  set.seed(100)
  indx <- createFolds(y_train, returnTrain = TRUE)
  ctrl <- trainControl(method = "cv", index = indx)
  
  ## train neural net:
  
  nnetGrid <- expand.grid(decay = c(.1), 
                          size = c(5, 15, 30), 
                          bag = FALSE)
  
  set.seed(100)
  nnetTune <- train(x = x_train, y = y_train,
                    method = "avNNet",
                    tuneGrid = nnetGrid,
                    trControl = ctrl,
                    preProc = c("center", "scale"),
                    linout = TRUE,
                    trace = FALSE,
                    MaxNWts = 30 * (ncol(x_train) + 1) + 30 + 1,
                    maxit = 1000,
                    repeats = 25,
                    allowParallel = FALSE)
  nnetTune
  plot(nnetTune)
  
  
  
  predictions <- predict(nnetTune, validatiex, type="raw")
  mse <- mean((validatiey - predictions)^2)
  mse <- sqrt(mse)
  print (mse)