Error when checking target: expected dense_3 to have shape (3,) but got array with shape (1,)
PythonPython 3.xTensorflowKerasPython Problem Overview
I am working on training a VGG16-like model in Keras, on a 3 classes subset from Places205, and encountered the following error:
ValueError: Error when checking target: expected dense_3 to have shape (3,) but got array with shape (1,)
I read multiple similar issues but none helped me so far. The error is on the last layer, where I've put 3 because this is the number of classes I'm trying right now.
The code is the following:
import keras from keras.datasets
import cifar10 from keras.preprocessing.image
import ImageDataGenerator from keras.models
import Sequential
from keras.layers import Dense, Dropout, Activation, Flatten from keras.layers import Conv2D, MaxPooling2D
from keras import backend as K import os
# Constants used
img_width, img_height = 224, 224
train_data_dir='places\\train'
validation_data_dir='places\\validation'
save_filename = 'vgg_trained_model.h5'
training_samples = 15
validation_samples = 5
batch_size = 5
epochs = 5
if K.image_data_format() == 'channels_first':
input_shape = (3, img_width, img_height) else:
input_shape = (img_width, img_height, 3)
model = Sequential([
# Block 1
Conv2D(64, (3, 3), activation='relu', input_shape=input_shape, padding='same'),
Conv2D(64, (3, 3), activation='relu', padding='same'),
MaxPooling2D(pool_size=(2, 2), strides=(2, 2)),
# Block 2
Conv2D(128, (3, 3), activation='relu', padding='same'),
Conv2D(128, (3, 3), activation='relu', padding='same'),
MaxPooling2D(pool_size=(2, 2), strides=(2, 2)),
# Block 3
Conv2D(256, (3, 3), activation='relu', padding='same'),
Conv2D(256, (3, 3), activation='relu', padding='same'),
Conv2D(256, (3, 3), activation='relu', padding='same'),
MaxPooling2D(pool_size=(2, 2), strides=(2, 2)),
# Block 4
Conv2D(512, (3, 3), activation='relu', padding='same'),
Conv2D(512, (3, 3), activation='relu', padding='same'),
Conv2D(512, (3, 3), activation='relu', padding='same'),
MaxPooling2D(pool_size=(2, 2), strides=(2, 2)),
# Block 5
Conv2D(512, (3, 3), activation='relu', padding='same',),
Conv2D(512, (3, 3), activation='relu', padding='same',),
Conv2D(512, (3, 3), activation='relu', padding='same',),
MaxPooling2D(pool_size=(2, 2), strides=(2, 2)),
# Top
Flatten(),
Dense(4096, activation='relu'),
Dense(4096, activation='relu'),
Dense(3, activation='softmax') ])
model.summary()
model.compile(loss='binary_crossentropy',
optimizer='rmsprop',
metrics=['accuracy'])
# no augmentation config train_datagen = ImageDataGenerator() validation_datagen = ImageDataGenerator()
train_generator = train_datagen.flow_from_directory(
train_data_dir,
target_size=(img_width, img_height),
batch_size=batch_size,
class_mode='binary')
validation_generator = validation_datagen.flow_from_directory(
validation_data_dir,
target_size=(img_width, img_height),
batch_size=batch_size,
class_mode='binary')
model.fit_generator(
train_generator,
steps_per_epoch=training_samples // batch_size,
epochs=epochs,
validation_data=validation_generator,
validation_steps=validation_samples // batch_size)
model.save_weights(save_filename)
Python Solutions
Solution 1 - Python
The problem is with your label-data shape. In a multiclass problem you are predicting the probabibility of every possible class, so must provide label data in (N, m) shape, where N is the number of training examples, and m is the number of possible classes (3 in your case).
Keras expects y-data in (N, 3) shape, not (N,) as you've problably provided, that's why it raises an error.
Use e.g. OneHotEncoder to convert your label data to one-hot encoded form.
Solution 2 - Python
As mentioned by others, Keras expects "one hot" encoding in multiclass problems.
Keras comes with a handy function to recode labels:
print(train_labels)
[1. 2. 2. ... 1. 0. 2.]
print(train_labels.shape)
(2000,)
Recode labels using to_categorical
to get the correct shape of inputs:
from keras.utils import to_categorical
train_labels = to_categorical(train_labels)
print(train_labels)
[[0. 1. 0.]
[0. 0. 1.]
[0. 0. 1.]
...
[0. 1. 0.]
[1. 0. 0.]
[0. 0. 1.]]
print(train_labels.shape)
(2000, 3) # viz. 2000 observations, 3 labels as 'one hot'
Other importent things to change/check in multiclass (compared to binary classification):
Set class_mode='categorical'
in the generator()
function(s).
Don't forget that the last dense layer must specify the number of labels (or classes):
model.add(layers.Dense(3, activation='softmax'))
Make sure that activation=
and loss=
is chosen so to suit multiclass problems, usually this means activation='softmax'
and loss='categorical_crossentropy'
.
Solution 3 - Python
Had the same issue. To solve the problem you can simply change in validation_generator and train_generator the class mode from 'binary' to 'categorical' - that's because you have 3 classes-which is not binary.
Solution 4 - Python
Problem : expected dense_3 to have shape (3,) but got array with shape (1,)
If you are using it for classification then the number of variables should be correct in the parameter for adding a dense layer.
variables_for_classification=5 #change it as per your number of categories
model.add(Dense(variables_for_classification, activation='softmax'))
model.fit(X_train, Y_train, epochs=epochs, batch_size=batch_size,validation_split=0.1,callbacks=[EarlyStopping(monitor='val_loss', patience=3, min_delta=0.0001)])
To make it more clear. As I was using the LSTM to predict the category of the news and the categories were 5- business, tech, politics, sports, entertainment
In that dense function when I put 5 it worked correctly.
Solution 5 - Python
The reason for this is you would have used 'binary' class_mode in the fit_generator() method for a multi class problem. Change that to 'categorical' and the error goes.
Solution 6 - Python
if take these errors u just need to specify the last with numbers of the classes > as an example u have 6 classes u have to make that :
model.add(Dense(6, activation='softmax'))
u can use this
num_classes=...
and the last layers will be
model.add(Dense(num_classes, activation='softmax'))
Solution 7 - Python
I also got the same error and solved it by setting class_mode
as categorical
instead of binary
Solution 8 - Python
The problem is with the shape of the labels of the data "Y".
The shape you have for the labels are (m,) and this will not work with the:
loss = "binary_crossentropy"
I believe if you don't want to play with the shape of the labels, then use:
loss = "sparse_categorical_crossentropy"
Solution 9 - Python
For me, this worked.
from keras.utils import to_categorical
num_labels=10 #for my case
train_labels=to_categorical(train_labels,10)
test_labels=to_categorical(test_labels,10)
Specifying the number of labels as an argument while categorically encoding my labels helped train effectively on my training set.