High training accuracy, low validation accuracy CNN binary classification keras - python
I'm trying to create a binary classifier that can differentiate between MRIs of alzheimer's patients and healthy individuals.
These are the stats so far:
1032 training images
400 validation images
Running a simple model as shown below
I have both the raw 160x160 images as well as the images after edge detection
Model:
model = Sequential([
Conv2D(filters=16, kernel_size=(5, 5), activation='relu', padding = 'same', input_shape=(160,160,3)),
MaxPool2D(pool_size=(2, 2), strides=2),
Flatten(),
Dense(units=2, activation='softmax')
])
As you can see - it's very simple, something I've done purposefully to try and remedy the issue of overfitting.
Output:
11/11 [==============================] - 2s 194ms/step - loss: 0.7604 - accuracy: 0.5155 - val_loss: 0.7081 - val_accuracy: 0.5000
Epoch 2/20
11/11 [==============================] - 2s 185ms/step - loss: 0.6885 - accuracy: 0.5223 - val_loss: 0.6942 - val_accuracy: 0.4839
Epoch 3/20
11/11 [==============================] - 2s 185ms/step - loss: 0.6802 - accuracy: 0.5854 - val_loss: 0.6985 - val_accuracy: 0.4931
Epoch 4/20
11/11 [==============================] - 2s 185ms/step - loss: 0.6717 - accuracy: 0.5932 - val_loss: 0.6996 - val_accuracy: 0.4677
Epoch 5/20
11/11 [==============================] - 2s 195ms/step - loss: 0.6512 - accuracy: 0.6175 - val_loss: 0.7124 - val_accuracy: 0.5115
Epoch 6/20
11/11 [==============================] - 2s 185ms/step - loss: 0.6345 - accuracy: 0.6476 - val_loss: 0.7073 - val_accuracy: 0.5253
Epoch 7/20
11/11 [==============================] - 2s 185ms/step - loss: 0.6118 - accuracy: 0.6680 - val_loss: 0.6920 - val_accuracy: 0.5207
Epoch 8/20
11/11 [==============================] - 2s 185ms/step - loss: 0.5817 - accuracy: 0.7068 - val_loss: 0.6964 - val_accuracy: 0.5207
Epoch 9/20
11/11 [==============================] - 2s 184ms/step - loss: 0.5528 - accuracy: 0.7272 - val_loss: 0.7123 - val_accuracy: 0.5161
Epoch 10/20
11/11 [==============================] - 2s 193ms/step - loss: 0.5239 - accuracy: 0.7417 - val_loss: 0.7397 - val_accuracy: 0.5392
Epoch 11/20
11/11 [==============================] - 2s 186ms/step - loss: 0.5106 - accuracy: 0.7427 - val_loss: 0.7551 - val_accuracy: 0.5461
Epoch 12/20
11/11 [==============================] - 2s 197ms/step - loss: 0.4920 - accuracy: 0.7650 - val_loss: 0.7402 - val_accuracy: 0.5438
Epoch 13/20
11/11 [==============================] - 2s 190ms/step - loss: 0.4741 - accuracy: 0.7835 - val_loss: 0.7564 - val_accuracy: 0.5507
Epoch 14/20
11/11 [==============================] - 2s 188ms/step - loss: 0.4591 - accuracy: 0.7767 - val_loss: 0.7445 - val_accuracy: 0.5300
Epoch 15/20
11/11 [==============================] - 2s 185ms/step - loss: 0.4486 - accuracy: 0.7767 - val_loss: 0.7712 - val_accuracy: 0.5415
Epoch 16/20
11/11 [==============================] - 2s 185ms/step - loss: 0.4503 - accuracy: 0.7806 - val_loss: 0.7446 - val_accuracy: 0.5346
Epoch 17/20
11/11 [==============================] - 2s 188ms/step - loss: 0.4404 - accuracy: 0.7670 - val_loss: 0.7669 - val_accuracy: 0.5553
Epoch 18/20
11/11 [==============================] - 2s 184ms/step - loss: 0.4169 - accuracy: 0.8078 - val_loss: 0.7804 - val_accuracy: 0.5576
Epoch 19/20
11/11 [==============================] - 2s 184ms/step - loss: 0.3987 - accuracy: 0.7971 - val_loss: 0.7846 - val_accuracy: 0.5507
Epoch 20/20
11/11 [==============================] - 2s 192ms/step - loss: 0.3977 - accuracy: 0.7981 - val_loss: 0.8060 - val_accuracy: 0.5461
Things I've tried so far:
resizing the image to a smaller input
adding dropout layers
using preprocessed images where it's just the edges shown
ensuring both classes in both training and validation datasets are evenly distributed
changing learning rate
reducing number of parameters to be of the same magnitude of the number of training images i have
I am literally out of ideas, I'm not sure how to move forward with this so I would appreciate any tips or advice.
All my code:
# Use ImageDataGenerator to create 3 lots of batches
train_batches = ImageDataGenerator(
rescale=1/255).flow_from_directory(directory=train_path,
target_size=(80,80), classes=['cn', 'ad'], batch_size=100,
color_mode="rgb")
valid_batches = ImageDataGenerator(
rescale=1/255).flow_from_directory(directory=valid_path,
target_size=(80,80), classes=['cn', 'ad'], batch_size=100,
color_mode="rgb")
# test_batches = ImageDataGenerator(
# rescale=1/255).flow_from_directory(directory=test_path,
# target_size=(224,224), classes=['cn', 'ad'], batch_size=10,
# color_mode="rgb")
imgs, labels = next(train_batches)
# Test to see normalisation has occurred properly
print(imgs[1][8])
# Define method to plot MRIs
def plotImages(images_arr):
fig, axes = plt.subplots(1, 10, figsize=(20,20))
axes = axes.flatten()
for img, ax in zip( images_arr, axes):
ax.imshow(img)
ax.axis('off')
plt.tight_layout()
plt.show()
# Plot a sample of MRIs
plotImages(imgs)
# # Define the model
# # VGG16
# model = Sequential()
# model.add(Conv2D(input_shape=(160,160,3),filters=64,kernel_size=(3,3),padding="same", activation="relu"))
# model.add(Conv2D(filters=64,kernel_size=(3,3),padding="same", activation="relu"))
# model.add(MaxPool2D(pool_size=(2,2),strides=(2,2)))
# model.add(Conv2D(filters=128, kernel_size=(3,3), padding="same", activation="relu"))
# model.add(Conv2D(filters=128, kernel_size=(3,3), padding="same", activation="relu"))
# model.add(MaxPool2D(pool_size=(2,2),strides=(2,2)))
# model.add(Conv2D(filters=256, kernel_size=(3,3), padding="same", activation="relu"))
# model.add(Conv2D(filters=256, kernel_size=(3,3), padding="same", activation="relu"))
# model.add(Conv2D(filters=256, kernel_size=(3,3), padding="same", activation="relu"))
# model.add(MaxPool2D(pool_size=(2,2),strides=(2,2)))
# model.add(Conv2D(filters=512, kernel_size=(3,3), padding="same", activation="relu"))
# model.add(Conv2D(filters=512, kernel_size=(3,3), padding="same", activation="relu"))
# model.add(Conv2D(filters=512, kernel_size=(3,3), padding="same", activation="relu"))
# model.add(MaxPool2D(pool_size=(2,2),strides=(2,2)))
# model.add(Conv2D(filters=512, kernel_size=(3,3), padding="same", activation="relu"))
# model.add(Conv2D(filters=512, kernel_size=(3,3), padding="same", activation="relu"))
# model.add(Conv2D(filters=512, kernel_size=(3,3), padding="same", activation="relu"))
# model.add(MaxPool2D(pool_size=(2,2),strides=(2,2)))
# model.add(Flatten())
# model.add(Dense(units=1024,activation="relu"))
# model.add(Dense(units=128,activation="relu"))
# model.add(Dense(units=2, activation="softmax"))
# # Model from the paper
# model = Sequential([
# Conv2D(filters=32, kernel_size=(3, 3), activation='relu', padding = 'same', input_shape=(160,160,3)),
# Conv2D(filters=32, kernel_size=(3, 3), activation='relu', padding='same'),
# MaxPool2D(pool_size=(2, 2), strides=2),
# Flatten(),
# Dense(units=2, activation='softmax')
# ])
## Model from Dr Paul
# static_conv_layer=Conv2D(filters=16, kernel_size=(5, 5), activation='relu', padding = 'same')
#
# model = Sequential([
# Conv2D(filters=16, kernel_size=(5, 5), activation='relu', padding = 'same', input_shape=(32,32,3)),
# MaxPool2D(pool_size=(2, 2), strides=2),
# Dropout(0.1),
# static_conv_layer,
# MaxPool2D(pool_size=(2, 2), strides=2),
# Dropout(0.1),
# Flatten(),
# Dense(units=2, activation='softmax')
# ])
# This model hits around 75% train acc, 54% val acc
model = Sequential([
Conv2D(filters=16, kernel_size=(5, 5), activation='relu', padding = 'same', input_shape=(80,80,3)),
MaxPool2D(pool_size=(2, 2), strides=2),
# Dropout(0.1),
# Conv2D(filters=16, kernel_size=(3, 3), activation='relu', padding='same'),
# MaxPool2D(pool_size=(2, 2), strides=2),
# Conv2D(filters=16, kernel_size=(3, 3), activation='relu', padding='same'),
# MaxPool2D(pool_size=(2, 2), strides=2),
Flatten(),
Dense(units=2, activation='softmax')
])
# model = Sequential([
# Conv2D(filters=32, kernel_size=(3, 3), activation='relu', padding = 'same', input_shape=(160,160,3)),
# Conv2D(filters=32, kernel_size=(3, 3), activation='relu', padding='same'),
# MaxPool2D(pool_size=(2, 2), strides=2),
# Conv2D(filters=32, kernel_size=(3, 3), activation='relu', padding='same'),
# Flatten(),
# Dense(units=2, activation='softmax')
# ])
## Basic model with dropouts
# model = Sequential([
# Conv2D(filters=32, kernel_size=(3, 3), activation='relu', padding = 'same', input_shape=(224,224,3)),
# MaxPool2D(pool_size=(2, 2), strides=2),
# Dropout(0.1),
# Conv2D(filters=64, kernel_size=(3, 3), activation='relu', padding='same'),
# MaxPool2D(pool_size=(2, 2), strides=2),
# Dropout(0.2),
# Conv2D(filters=128, kernel_size=(3, 3), activation='relu', padding='same'),
# MaxPool2D(pool_size=(2, 2), strides=2),
# Dropout(0.3),
# Flatten(),
# Dense(units=1, activation='sigmoid')
# ])
# Summarise each layer of the model
print(model.summary())
# Compile and train the model
model.compile(optimizer=Adam(), loss='binary_crossentropy', metrics=['accuracy'])
model.fit(x=train_batches,
steps_per_epoch=len(train_batches),
validation_data=valid_batches,
validation_steps=len(valid_batches),
epochs=20,
verbose=1
)
EDIT:
This paper seems to be doing much better than me and completing a very similar task, it may be useful to look at the methodology for:
Things you can try.
It's a nice way to start with transfer learning. Using image net weights helps you to train just last layers and give much better accuracy.
Adding early stopping and learning rate reduction with validation accuracy as constraint.
Taking advantage of ImageDataGenerator and add much more data augmentation techniques.
Make your model much deeper and also try different optimizer(RMSprop), run for more epochs with early stopping.
Add callbacks and plot training validation accuracy graphs with respective to learning rate to see which lr proves best for the data.
It looks like your model is overfitting due to a lack of data. You can do some data augmentation to increase how many images you have. If you don't care about your aspect ratio you can warp the images, if you don't always need the full image you can crop it and you can rotate it if orientation is not important. These things can dramatically increase your dataset size and help mitigate overfitting.
Here is an example from the tensorflow documentation:
batch_size = 32
AUTOTUNE = tf.data.experimental.AUTOTUNE
def prepare(ds, shuffle=False, augment=False):
# Resize and rescale all datasets
ds = ds.map(lambda x, y: (resize_and_rescale(x), y),
num_parallel_calls=AUTOTUNE)
if shuffle:
ds = ds.shuffle(1000)
# Batch all datasets
ds = ds.batch(batch_size)
# Use data augmentation only on the training set
if augment:
ds = ds.map(lambda x, y: (data_augmentation(x, training=True), y),
num_parallel_calls=AUTOTUNE)
# Use buffered prefecting on all datasets
return ds.prefetch(buffer_size=AUTOTUNE)
Also, here is a great video to watch from the TensorFlow developers youtube channel which explains the idea of image augmentation and shows an example of how to implement it.
Related
Model accuracy and validation accuracy stuck at constant value for denoising autoencoder model
I am building a denoising autoencoder (DAE) to denoise respiratory signals. I pass through the model both noisy and clean versions of the signal (in frame sizes as multiples of 1024). I've set up my model up as follows: class NoiseReducer(tf.keras.Model): def __init__(self): super().__init__() self.encoder = tf.keras.Sequential([ # Input(shape=(window_size, 1)), Masking(mask_value=np.nan, input_shape=(window_size, 1)), Conv1D(filters=128, kernel_size=32, strides=1, kernel_constraint=max_norm(max_norm_value), padding='same', kernel_initializer='glorot_normal', activation='elu'), Dense(128, activation='elu'), Conv1D(filters=32, kernel_size=16, strides=1, kernel_constraint=max_norm(max_norm_value), padding='same', kernel_initializer='glorot_normal', activation='elu'), Conv1D(filters=16, kernel_size=8, strides=1, kernel_constraint=max_norm(max_norm_value), padding='same', kernel_initializer='glorot_normal', activation='elu') ]) self.decoder = tf.keras.Sequential([ Conv1DTranspose(filters=16, kernel_size=8, strides=1, kernel_constraint=max_norm(max_norm_value), padding='same', kernel_initializer='glorot_normal', activation='elu') Conv1DTranspose(filters=32, kernel_size=16, strides=1, kernel_constraint=max_norm(max_norm_value), padding='same', kernel_initializer='glorot_normal', activation='elu'), Dense(128, activation='elu'), Conv1DTranspose(filters=128, kernel_size=32, strides=1, kernel_constraint=max_norm(max_norm_value), padding='same', kernel_initializer='glorot_normal', activation='elu'), Conv1D(filters=1, kernel_size=2, strides=1, kernel_constraint=max_norm(max_norm_value), padding='same', activation='sigmoid') ]) def call(self, x): encoded = self.encoder(x) decoded = self.decoder(encoded) return decoded dae = NoiseReducer() adam_optimizer=tf.keras.optimizers.Adam( learning_rate=learning_rate, beta_1=0.9, beta_2=0.999, epsilon=None, decay=0.0, amsgrad=False) sgd_optimizer = tf.keras.optimizers.SGD(learning_rate=learning_rate) dae.compile(optimizer=sgd_optimizer, loss='mean_squared_error', metrics='accuracy') history = dae.fit(X_noisy_train, X_clean_train, epochs=epochs, batch_size=batch_size, shuffle=False, validation_split=0.3, callbacks=[tb_callback] ) RESULTS: 13/13 [==============================] - 16s 1s/step - loss: 0.2185 - accuracy: 0.8272 - val_loss: 0.2143 - val_accuracy: 0.8288 Epoch 2/100 13/13 [==============================] - 12s 898ms/step - loss: 0.2120 - accuracy: 0.8272 - val_loss: 0.2082 - val_accuracy: 0.8288 Epoch 3/100 13/13 [==============================] - 12s 908ms/step - loss: 0.2057 - accuracy: 0.8272 - val_loss: 0.2017 - val_accuracy: 0.8288 Epoch 4/100 13/13 [==============================] - 12s 906ms/step - loss: 0.1997 - accuracy: 0.8272 - val_loss: 0.1956 - val_accuracy: 0.8288 Epoch 5/100 13/13 [==============================] - 12s 907ms/step - loss: 0.1938 - accuracy: 0.8272 - val_loss: 0.1898 - val_accuracy: 0.8288 When running the model, the accuracy and validation accuracy is stuck at around 0.827 for both and does not change at all throughout the epochs (100 in total) suggesting that the model isn't learning anything. The MSE is however descreasing with epochs. For my datasets I have set any nan values to 0 In terms of solutions I have implemented the following changes to my model but to no success: Increased filter length of conv1D layers Tested different learning rates for both SGD and Adam Test with ELU (Exponential Linear Unit) activation instead of RELU Use more int. dense layers with more neurons. Using glorot (commonly known as Xavier) initializer Use SGD instead of Adam Change window size with different multiples of 1024 None of these seem to change the accuracy. After model completion and reconstructing the signal (from the noisy) I get a straight line cutting through 0.345 illustrating that the model has not learnt anything and can not reconstruct the signal. What other strategies/alleys should I explore around this?
Loss is decreasing and accuracy is not improving in siamese network with contrastive loss function
# We have 2 inputs, 1 for each picture
left_input = Input(img_size)
right_input = Input(img_size)
# We will use 2 instances of 1 network for this task
convnet = Sequential([
Conv2D(4,3, input_shape=img_size, padding='same'),
Activation('relu'),
MaxPooling2D(),
Conv2D(4,3, padding='same'),
Activation('relu'),
MaxPooling2D(),
Conv2D(4,3, padding='same'),
Activation('relu'),
MaxPooling2D(),
Conv2D(4,3, padding='same'),
Activation('relu'),
Flatten(),
Dropout(0.3),
Dense(18),
Activation('sigmoid')
])
# Connect each 'leg' of the network to each input
# Remember, they have the same weights
encoded_l = convnet(left_input)
encoded_r = convnet(right_input)
# Getting the L1 Distance between the 2 encodings
L1_layer = Lambda(lambda tensor:K.abs(tensor[0] - tensor[1]))
# Add the distance function to the network
L1_distance = L1_layer([encoded_l, encoded_r])
prediction = Dense(1,activation='sigmoid')(L1_distance)
siamese_net = Model(inputs=[left_input,right_input],outputs=prediction)
optimizer = Adam(lr, decay=2.5e-4)
#//TODO: get layerwise learning rates and momentum annealing scheme described in paperworking
siamese_net.compile(loss=tfa.losses.contrastive.contrastive_loss,optimizer=optimizer,metrics=['acc'])
siamese_net.summary()
my accuracy is decreasing eventhough loss value is decreasing?
I have got output like this
Epoch 1/10 6368/6368 [==============================] - 647s 102ms/step - loss: 0.1762 - acc: 0.2134 - val_loss: 0.1590 - val_acc: 0.1784
Epoch 2/10 6368/6368 [==============================] - 1094s 172ms/step - loss: 0.1649 - acc: 0.1881 - val_loss: 0.1548 - val_acc: 0.1784
Epoch 3/10 6368/6368 [==============================] - 1440s 226ms/step - loss: 0.1573 - acc: 0.1779 - val_loss: 0.1515 - val_acc: 0.1790
may i know what is the reason why my accuracy is reducing even though the loss function is reducing?
how to prevent overfitting
i'm doing for my final project and i'm new in ConVnets. i want to classifies which one is genuine image and spoof image. i have +-8000 data (combine). and i want to show you some of my training log.
Epoch 7/100
311/311 [==============================] - 20s 63ms/step - loss: 0.3274 - accuracy: 0.8675 - val_loss: 0.2481 - val_accuracy: 0.9002
Epoch 8/100
311/311 [==============================] - 20s 63ms/step - loss: 0.3189 - accuracy: 0.8691 - val_loss: 0.3015 - val_accuracy: 0.8684
Epoch 9/100
311/311 [==============================] - 19s 62ms/step - loss: 0.3201 - accuracy: 0.8667 - val_loss: 0.2460 - val_accuracy: 0.9036
Epoch 10/100
311/311 [==============================] - 19s 62ms/step - loss: 0.3063 - accuracy: 0.8723 - val_loss: 0.2752 - val_accuracy: 0.8901
Epoch 11/100
311/311 [==============================] - 19s 62ms/step - loss: 0.3086 - accuracy: 0.8749 - val_loss: 0.2717 - val_accuracy: 0.8988
[INFO] evaluating network...
model = Sequential()
inputShape = (height, width, depth)
chanDim = -1
if K.image_data_format() == "channels_first":
inputShape = (depth, height, width)
chanDim = 1
model.add(Conv2D(16, (3, 3), padding="same", input_shape=inputShape)) model.add(Activation("relu"))
model.add(BatchNormalization(axis=chanDim))
model.add(Conv2D(16, (3, 3), padding="same"))
model.add(Activation("relu"))
model.add(BatchNormalization(axis=chanDim))
model.add(MaxPooling2D(pool_size=(2, 2)))
model.add(Dropout(0.3))
model.add(Conv2D(32, (5, 5), padding="same"))
model.add(Activation("relu"))
model.add(BatchNormalization(axis=chanDim))
model.add(MaxPooling2D(pool_size=(2, 2)))
model.add(Dropout(0.3))
model.add(Flatten())
model.add(Dense(128))
model.add(Activation("relu"))
model.add(BatchNormalization())
model.add(Dropout(0.6))
model.add(Dense(classes))
model.add(Activation("softmax"))
the input is 32x32 and it has two classes. i used EarlyStopping in keras to prevent overfitting. and i always change the value of learning rate and try to change the number of neuron node but still the training always stop below 20 epoch. any advice to prevent overfitting ? since i'm beginner in convolutional neural network. Thanks in advance !
PS LR: 0.001 BS: 20 EPOCHS: 100
keras will only give accuracy of 0.5001
I'm running Keras with a tensorflow-backend. I try to predict images. My model looks like this: model = Sequential() model.add(Conv2D(32, (3, 3), activation='relu', input_shape=(50, 50, 1))) model.add(MaxPooling2D((2, 2))) model.add(Conv2D(64, (3, 3), activation='relu')) model.add(MaxPooling2D((2, 2))) model.add(Conv2D(64, (3, 3), activation='relu')) model.add(Flatten()) model.add(Dense(1, activation='softmax')) model.compile(optimizer='adam', loss='binary_crossentropy', metrics=['accuracy']) Running my code, it produces this output over 10 epochs: Epoch 1/10 24946/24946 [==============================] - 36s 1ms/sample - loss: 7.9693 - acc: 0.5001 Epoch 2/10 24946/24946 [==============================] - 35s 1ms/sample - loss: 7.9693 - acc: 0.5001 ... Epoch 9/10 24946/24946 [==============================] - 30s 1ms/sample - loss: 7.9693 - acc: 0.5001 Epoch 10/10 24946/24946 [==============================] - 30s 1ms/sample - loss: 7.9693 - acc: 0.5001 1/1 [==============================] - 0s 36ms/step [[1.]] Anyhow, I do not understand why the accuracy is always 0.5001 over all 10 epochs. My question is: Why does the accuracy not change within any epoch?
This part of your code makes no sense: model.add(Dense(1, activation='softmax')) Softmax with only one neuron with always produce output of constant 1.0, due to the normalization. If you want to do binary classification, you should use a sigmoid activation at the output.
Keras VGG model for MNIST: Disparity between training and validation accuracy
I have created the following model with Keras. The dataset is MNIST.
'''
conv - relu - conv- relu - pool -
conv - relu - conv- relu - pool -
conv - relu - conv- relu - pool -
affine - relu - dropout - affine - dropout - softmax
'''
model = Sequential()
model.add(Conv2D(16, kernel_size=(3, 3),
padding='same',
input_shape=input_shape))
model.add(Activation('relu'))
model.add(Conv2D(16, (3, 3), activation='relu'))
model.add(MaxPooling2D(pool_size=(2, 2)))
model.add(Conv2D(32, (3, 3), padding='same', activation='relu'))
model.add(Conv2D(32, (3, 3), activation='relu'))
model.add(MaxPooling2D(pool_size=(2, 2)))
model.add(Conv2D(64, (3, 3), padding='same', activation='relu'))
model.add(Conv2D(64, (3, 3), activation='relu'))
model.add(MaxPooling2D(pool_size=(2, 2)))
model.add(Flatten())
model.add(Dense(50, activation='relu'))
model.add(Dropout(0.5))
model.add(Dense(num_classes))
model.add(Dropout(0.5))
model.add(Activation('softmax'))
The following is the result:
60000/60000 [==============================] - 10s - loss: 1.2707 - acc: 0.5059 - val_loss: 0.0881 - val_acc: 0.9785
Epoch 2/20
60000/60000 [==============================] - 9s - loss: 0.9694 - acc: 0.5787 - val_loss: 0.0449 - val_acc: 0.9873
...
Epoch 19/20
60000/60000 [==============================] - 9s - loss: 0.8530 - acc: 0.6004 - val_loss: 0.0282 - val_acc: 0.9937
Epoch 20/20
60000/60000 [==============================] - 9s - loss: 0.8564 - acc: 0.5982 - val_loss: 0.0383 - val_acc: 0.9910
Test loss: 0.0382921607383
Test accuracy: 0.991
Why is the training accuracy so low, while the validation accururacy is so high?
The dropout on your last Dense layer removes half of your 10 neurons for your classes by random. Your last layer can only by accurate half of the times because in general half of the neurons are missing. Try to remove that and I assume you get even values.