Initializer

When using an initializer, you need to configure the name item to indicate which initializer to use, and then configure their respective parameters according to different initializer. The following are the names of each initializer.

Initializer	name
Zeros	zeros
Ones	ones
RandomUniform	random_uniform
RandomNormal	random_normal
TruncateNormal	truncate_normal

Zeros

0初始化 需要配置如下参数:

1. name: "zeros"

Ones

1初始化 需要配置如下参数:

1. name: "ones"

RandomUniform

均匀分布, 需要配置如下的一些参数:

1. name: "random_uniform"
2. min: 下限, 浮点数, default: -1.0
3. max: 上限, 浮点数, default: 1.0

RandomNormal

随机正态分布, 需要配置如下的一些参数:

1. name: "random_normal"
2. mean: 均值, 浮点数, default: 0.0
3. stddev: 标准差, 浮点数, default: 1.0

TruncateNormal

随机正态分布, 且2倍标准差外的数据丢弃重新生成 需要配置如下的一些参数:

1. name: "truncate_normal"
2. mean: 均值, 浮点数, default: 0.0
3. stddev: 标准差, 浮点数, default: 1.0

Example

import damo
import numpy as np

# zero
param = damo.Parameters()
param.insert("name", "zeros")
# must be float32
value = np.random.random(10).astype(np.float32)
obj = damo.PyInitializer(param)
obj.call(value)
print("zeros: ", value)

# ones
param = damo.Parameters()
param.insert("name", "ones")
# must be float32
value = np.random.random(10).astype(np.float32)
obj = damo.PyInitializer(param)
obj.call(value)
print("ones: ", value)

# random_uniform
param = damo.Parameters()
param.insert("name", "random_uniform")
param.insert("min", -1.0)
param.insert("max", 1.0)
# must be float32
value = np.random.random(10).astype(np.float32)
obj = damo.PyInitializer(param)
obj.call(value)
print("random_uniform: ", value)

# random_normal
param = damo.Parameters()
param.insert("name", "random_normal")
param.insert("mean", 0.0)
param.insert("stddev", 1.0)
# must be float32
value = np.random.random(10).astype(np.float32)
obj = damo.PyInitializer(param)
obj.call(value)
print("random_normal: ", value)

# truncate_normal
param = damo.Parameters()
param.insert("name", "truncate_normal")
param.insert("mean", 0.0)
param.insert("stddev", 1.0)
# must be float32
value = np.random.random(10).astype(np.float32)
obj = damo.PyInitializer(param)
obj.call(value)
print("truncate_normal: ", value)