CODEWITHSUNDEEP
keraskeras-layer
MadProgrammer
MadProgrammer

Reputation: 453

Keras Dimension Incompatibility

I would like to code a simple variational auto-encoder in Keras. Because its variational, I have to use the functional API for the encoder. However, its giving me a dimension mismatch error and I cant figure out why. Here is my code and error:

def sampling(args):
    z_mean, z_log_var = args
    epsilon = K.random_normal(shape=(100,2),
                              mean=0., stddev=1)
    return z_mean + K.exp(z_log_var) * epsilon

def testFcn():
    K.clear_session()

    # Create the input
    inPut = Input(shape=(3,))

    # Encoder Layers
    xEnc = Dense(128, input_shape=(3,), activation='relu')(inPut)
    xEnc = Dense(64, activation='relu')(xEnc)
    xEnc = Dense(32, activation='relu')(xEnc)

    # Distribution Embedding
    z_mean = Dense(2, activation='relu')(xEnc)
    z_log_var = Dense(2, activation='relu')(xEnc)
    z = Lambda(sampling, output_shape=(2,))([z_mean, z_log_var])

    # Tying together the model
    encoder = Model(inPut, z, name='encoder')
    print("\n Encoder Model")
    encoder.summary()

    return encoder

# Create some random data
X = np.random.multivariate_normal([0]*3,np.eye(3),size=(100))

# Create the model
encoder = testFcn()

# predict
encoder.predict(X)

This gives me the following error:

InvalidArgumentError: 2 root error(s) found.
  (0) Invalid argument: Incompatible shapes: [32,2] vs. [100,2]
     [[{{node lambda_1/mul}}]]
  (1) Invalid argument: Incompatible shapes: [32,2] vs. [100,2]
     [[{{node lambda_1/mul}}]]
     [[lambda_1/add/_23]]
0 successful operations.
0 derived errors ignored.

Any help here is really appreciated.

Upvotes: 0

Views: 72

Answers (1)

ad2004
ad2004

Reputation: 809

This might point you in the right direction - I was able to get success with this modification of adding the batch_size:

encoder.predict(X, batch_size=100)

The default is 32, which is conflicting with your design apparently. More details here: https://keras.io/models/model/#predict

I hope this helps.

Upvotes: 1

Related Questions