Function Differences with tf.compat.v1.layers.Dense

tf.compat.v1.layers.Dense

class tf.compat.v1.layers.Dense(
    units,
    activation=None,
    use_bias=True,
    kernel_initializer=None,
    bias_initializer=tf.compat.v1.zeros_initializer(),
    kernel_regularizer=None,
    bias_regularizer=None,
    activity_regularizer=None,
    kernel_constraint=None,
    bias_constraint=None,
    trainable=True,
    name=None,
    **kwargs
)(x) -> Tensor

For more information, see tf.compat.v1.layers.Dense.

mindspore.nn.Dense

class mindspore.nn.Dense(
    in_channels,
    out_channels,
    weight_init='normal',
    bias_init='zeros',
    has_bias=True,
    activation=None
)(x) -> Tensor

For more information, see mindspore.nn.Dense.

Differences

TensorFlow: Fully connected layer that implements the matrix multiplication operation.

MindSpore: MindSpore API basically implements the same function as TensorFlow.

Categories	Subcategories	TensorFlow	MindSpore	Differences
Parameters	Parameter 1	units	out_channels	Same function, different parameter names
	Parameter 2	activation	activation	-
	Parameter 3	use_bias	has_bias	Same function, different parameter names
	Parameter 4	kernel_initializer	weight_init	Same function, different parameter names
	Parameter 5	bias_initializer	bias_init	Same function, different parameter names
	Parameter 6	kernel_regularizer	-	The regular function of the weight matrix. MindSpore does not have this parameter.
	Parameter 7	bias_regularizer	-	The regularization function for the deviation. MindSpore does not have this parameter.
	Parameter 8	activity_regularizer	-	The regularization function for the output. MindSpore does not have this parameter.
	Parameter 9	kernel_constraint	-	Optional projection functions that will be applied to the kernel after the Optimizer program is updated (e.g., for implementing norm constraints or value constraints on layer weights). The function must take as input the unprojected variables and must return the projected variables (which must have the same shape). It is not safe to use constraints when doing asynchronous distributed training. MindSpore does not have this parameter
	Parameter 10	bias_constraint	-	Optional projection function to be applied to the deviation after being updated by Optimizer. MindSpore does not have this parameter
	Parameter 11	trainable	-	Boolean. If True, also adds the variable to the graph collection GraphKeys.TRAINABLE_VARIABLES. MindSpore does not have this parameter.
	Parameter 12	name	-	Not involved
	Parameter 13	**kwargs	-	Not involved
	Parameter 14	-	in_channels	The spatial dimension of the input. TensorFlow does not have this parameter
Input	Single input	x	x	-

Code Example

The two APIs achieve the same function and have the same usage.

# TensorFlow
import tensorflow as tf
from tensorflow.compat.v1 import layers
import numpy as np

model = layers.Dense(4)
x = tf.constant(np.array([[180, 234, 154], [244, 48, 247]]),dtype=tf.float32)
output = model(x)
print(output.shape)
# (2, 4)

# MindSpore
import mindspore
from mindspore import Tensor, nn
import numpy as np

x = Tensor(np.array([[180, 234, 154], [244, 48, 247]]), mindspore.float32)
net = nn.Dense(3, 4)
output = net(x)
print(output.shape)
# (2, 4)