设置权重衰减的准则是什么(如l2罚则)--主要是,我如何做到 轨道 是否在整个训练中 "发挥作用"?即重量是否真的在衰减,以及 多少,相比之下,没有l2惩罚)。)
一个常见的方法是 "尝试一个数值范围,看看什么是有效的"--但它的缺陷是缺乏 正交性; l2=2e-4
在网络中可能效果最好 X但不是网络 Y. 一个变通的办法是引导重量衰减的方式。子网 方式。(1)分组层(如? Conv1D
栈&。LSTM
s分别),(2)设置目标权重规范,(3)跟踪。
(1): 看 此处同样的论点和建议的权重值将不适用于convs--因此需要进行各种分组
(2): 一个声音选项是: l2-常态 的权重矩阵进行正则化;然后就是哪个 轴心 来计算它与。面向特征提取的方法是选择 频道轴 (最后一个在Keras中),得到一个长度=通道特征数的向量,所以每个元素都是一个通道的l2-norm。
(3): l2-norm向量可以迭代地追加到一个列表中,也可以或许将其meanmax作为更简略的汇总统计--然后在训练结束时绘制。
下图所示的一个完整的例子;关键函数。weights_norm
,在底部给出,取自于 见RNN. 我还建议 Keras AdamW 以改善重量衰减。
解释:
wd=2e-3
衰减输出层强于 2e-4
,但不是输入,说明与瓶颈层存在平衡互动。wd=2e-3
产量 差异较小 相对于重量规范而言 2e-4
2e-3
,表明向产出的梯度更强BatchNormalization
附加代码& 解释;以下是做。
火车和轨道进度
n_batches
和 wd
(L2罚款)n_epochs
l2_stats
追踪进度weights_norm()
并附于 l2_stats
预处理进度数据,以便绘图
omit_names
l2_stats
便于附加到,但必须转换为 np.ndarray
适当的色调;拆开包装,以便 .shape == (n_epochs, n_layers, n_weights, n_batches) -> (n_rows, n_cols, hists_per_subplot)
. 请注意,这要求每层跟踪的权重矩阵数量相同。情节
xlims
和 ylim
互相比较 wd
价值观np.mean
(橙色),和 np.max
. 后者也是Keras的处理方式。maxnorm
权重正则化。import numpy as np
import tensorflow as tf
import random
np.random.seed(1)
random.seed(2)
tf.compat.v1.set_random_seed(3)
from keras.layers import Input, Conv1D
from keras.models import Model
from keras.regularizers import l2
from see_rnn import weights_norm, features_hist_v2
########### Model & data funcs ################################################
def make_model(batch_shape, layer_kw={}):
"""Conv1D autoencoder"""
dim = batch_shape[-1]
bdim = dim // 2
ipt = Input(batch_shape=batch_shape)
x = Conv1D(dim, 8, activation='relu', **layer_kw)(ipt)
x = Conv1D(bdim, 1, activation='relu', **layer_kw)(x) # bottleneck
out = Conv1D(dim, 8, activation='linear', **layer_kw)(x)
model = Model(ipt, out)
model.compile('adam', 'mse')
return model
def make_data(batch_shape, n_batches):
X = Y = np.random.randn(n_batches, *batch_shape)
return X, Y
########### Train setup #######################################################
batch_shape = (32, 100, 64)
n_epochs = 5
n_batches = 200
wd = 2e-3
layer_kw = dict(padding='same', kernel_regularizer=l2(wd))
model = make_model(batch_shape, layer_kw)
X, Y = make_data(batch_shape, n_batches)
## Train ####################
l2_stats = {}
for epoch in range(n_epochs):
l2_stats[epoch] = {}
for i, (x, y) in enumerate(zip(X, Y)):
model.train_on_batch(x, y)
print(end='.')
verbose = bool(i == len(X) - 1) # if last epoch iter, print last results
if verbose:
print()
l2_stats[epoch] = weights_norm(model, [1, 3], l2_stats[epoch],
omit_names='bias', verbose=verbose)
print("Epoch", epoch + 1, "finished")
print()
########### Preprocess funcs ##################################################
def _get_weight_names(model, layer_names, omit_names):
weight_names= []
for name in layer_names:
layer = model.get_layer(name=name)
for w in layer.weights:
if not any(to_omit in w.name for to_omit in omit_names):
weight_names.append(w.name)
return weight_names
def _merge_layers_and_weights(l2_stats):
stats_merged = []
for stats in l2_stats.values():
x = np.array(list(stats.values())) # (layers, weights, stats, batches)
x = x.reshape(-1, *x.shape[2:]) # (layers-weights, stats, batches)
stats_merged.append(x)
return stats_merged # (epochs, layer-weights, stats, batches)
########### Plot setup ########################################################
ylim = 5
xlims = (.4, 1.2)
omit_names = 'bias'
suptitle = "wd={:.0e}".format(wd).replace('0', '')
side_annot = "EP"
configs = {'side_annot': dict(xy=(.9, .9))}
layer_names = list(l2_stats[0].keys())
weight_names = _get_weight_names(model, layer_names, omit_names)
stats_merged = _merge_layers_and_weights(l2_stats)
## Plot ########
features_hist_v2(stats_merged, colnames=weight_names, title=suptitle,
xlims=xlims, ylim=ylim, side_annot=side_annot,
pad_xticks=True, configs=configs)