在脑电图分类的 CNN (Keras) 模型中，损失为“nan”，准确度为 0

Question

I am trying to classify an EEG dataset I found online. (BCI comp III, Dataset V)

I am just playing around with different models as a side project, and I though I'd start with a CNN.

This is the piece of code for extracting the data from the files:

def ext_data(subNum, rawNum):
file_name = r'C:\Users\me\OneDrive\Desktop\EEG\EEG datasets\BCI COMP III\Dataset V\train_subject{}_raw{:02d}.asc'.format(subNum, rawNum)
import numpy as np
data = np.loadtxt(file_name)
print("count of original data is {}".format(len(data)))
org_data = []
i = 0
while True:
    if i == len(data)-1:
        break
    meta = []
    cl = data[i,32]
    meta.append(data[i,0:32])
    while 1:
        i += 1
        if i == len(data)-1:
            meta.append(data[i, 0:32])
            org_data.append([np.array(meta).reshape(len(meta),32),cl])
            break
        if data[i,32] == cl:
            meta.append(data[i,0:32])
        else:
            org_data.append([np.array(meta).reshape(len(meta),32),cl])
            break
    #print(i)
print(len(org_data))
return np.array(org_data)

and then I make the training labels and data this way:

    # total number of inptu samples = 104
data = np.ones(shape=[104, 2], dtype=object)
index = 0
for i in range(1, 4):
    for j in range(1, 4):
        print("index is {}".format(index))
        dum = ext_data(i, j)
        data[index:len(dum) + index] = dum
        index += len(dum)
train_x = np.zeros(shape = [104,8000,32], dtype=np.float32)
for i in range(104):
    for j in range(32):
        train_x[i,:,j] = beta_filter(data[i,0][0:8000,j])
train_x = train_x/40
train_y = np.array(data[:,1])

beta_filter is just a band-pass filter and a notch filter:

def butter_bandpass(lowcut, highcut, fs, order=3):
    nyq = 0.5 * fs
    low = lowcut / nyq
    high = highcut / nyq
    sos = butter(order, [low, high], analog=False, btype='band', output='sos')
    return sos

def butter_bandpass_filter(data, lowcut, highcut, fs, order=3):
        sos = butter_bandpass(lowcut, highcut, fs, order=order)
        y = sosfiltfilt(sos, data)
        return y

def butter_notch_filter(data, lowcut, highcut, fs, order = 3):
    nyq = 0.5 * fs
    low = lowcut / nyq
    high = highcut / nyq
    sos = butter(order, [low, high], analog=False, btype='stop', output='sos')
    y = sosfiltfilt(sos, data)
    return y

def beta_filter(input):
    beta = butter_bandpass_filter(input, 12, 30, 512)
    beta = butter_notch_filter(beta,48,52,512)
    return beta

and finally, this is the model I made:

def generate_model():
model = tf.keras.Sequential([
    
    # first convolutional layer
    tf.keras.layers.Conv2D(32, (3, 3), strides=(1, 1), activation="relu"),
    tf.keras.layers.MaxPool2D(pool_size=2, strides=2),
    
    # second convolutional layer
    tf.keras.layers.Conv2D(16, (3, 3), strides=(1, 1), activation="relu"),
    tf.keras.layers.MaxPool2D(pool_size=2, strides=2),

    # fully connected classifier
    tf.keras.layers.Flatten(),
    tf.keras.layers.Dense(128, activation='relu'),
    tf.keras.layers.Dense(32, activation='relu'),
    tf.keras.layers.Dense(3, activation='softmax')  # 3 outputs
])
return model

and finally, compile the model like this:

model = generate_model()
model.compile(optimizer='adam',
            loss=tf.keras.losses.SparseCategoricalCrossentropy(from_logits=True),
            metrics=['accuracy'])

history = model.fit(train_x, train_y, epochs=15, batch_size=1)

but when training starts, I got nan for loss and 0 for accuracy! sometimes accuracy starts with a value of 0.0903, and then goes to 0 and stays there.

I generated matrices of random values ( np.random.rand(size) ) to check the model, and ofc got an accuracy of 0.33 (the chance accuracy, three classes).

I also checked for nan within the input, and got nothing.

what would be the problem? is it the model? or is the data fundamentally wrong?

What would be the problem?

PS : I only use up to 8000 data points for the input to avoid the variable length input, as inputs have different lengths.

PS : this is what the filter does to the data:

Unfiltered data:

filtered data:

This is one random input vector from the 32 columns of the input matrix.

Answer 1

change

loss=tf.keras.losses.SparseCategoricalCrossentropy(from_logits=True)

to

loss=tf.keras.losses.SparseCategoricalCrossentropy(from_logits=False)

You use from_logits=True when the last layer of your model doesn't have an activation like softmax or sigmoid (without that activation it will return logits and hence you can have it set to True else False ).