Keras: tf.matmul vs K.Dot

[lschaupp]

Quick question: (tensorflow 1.4/Keras 2.2)

Is there any difference between tf.matmul and keras dot function?
Seems to me that the dot function needs a specific axis, while the matmul function only
needs the two matrices. Is there any other difference? Any examples?

Thanks!

[ibealec]

They actually both work without specifying an axis. TensorFlow uses float64 by default whereas Keras uses float32. Using np.random.seed(42) and r = np.random.rand(4, 5),

x = tf.constant(r)  
cov = tf.matmul(x, tf.transpose(x))

returns Tensor("MatMul:0", shape=(4, 4), dtype=float64)

x = K.constant(x_val)
cov = K.dot(x, K.transpose(x))

returns Tensor("MatMul_1:0", shape=(4, 4), dtype=float32)

Note that there is a variance of about 1e-7 between the two, as pointed out in https://stackoverflow.com/a/44100246/6647539

[lschaupp]

thanks for the info! Why I'm asking: my model takes the following:

    input_T = Reshape((3, 3))(x)

    # forward net
    #K.batch_dot(x[0], BatchM, axes=[1,2])
    g = K.batch_dot(input, input_T,axes=[1,2])
    g = Convolution1D(64, 1, input_shape=(p.numPoints, 3), activation='relu')(g)

so I want to multiply input (28800,3) with input_T (3,3). But I'm receiving the error:
ValueError: Dimensions must be equal, but are 28800 and 3 for 'MatMul_3' (op: 'BatchMatMul') with input shapes: [?,28800,3], [?,3,3].

while

def cov_tf(x_val,y_val):
    x = tf.constant(x_val)
    y = tf.constant(y_val)
    cov = tf.matmul(x, y)
    return cov.eval(session=tf.Session())

def cov_keras(x_val, y_val):
    x = K.constant(x_val)
    y = K.constant(y_val)
    cov = K.dot(x, y)
    return cov.eval(session=tf.Session())

if __name__ == '__main__':
    x = np.random.rand(2048, 3)
    y = np.random.rand(3, 3)
    print (x.shape)
    print (cov_tf(x,y).shape)
    print (cov_keras(x,y).shape)
    delta = np.abs(cov_tf(x,y) - cov_keras(x,y)).max()

    print('Maximum absolute difference:', delta)

works:
(28800, 3)

[saravanabalagi]

Higher dimensions?

a.shape          # (125, 125, 1, 3)
b/shape          # (125, 125, 3, 1)
np.matmul(a,b).shape        # (125,125,1,1)
K.dot(ab).shape             # (125, 125, 1, 125, 125, 1)

What is Keras' equivalent?

[saravanabalagi]

I ended up using tf.matmul importing tf
It doesn't matter, we are doing tensor operations and it's completely fine

[hummosa]

I'm facing the same issue as @zekedran higher dimensions are a problem in Keras. Clarifying proper use would be appreciated.