自定义编写层中的Tensorflow 2出错，该层在Tensorflow 1中正常工作（assert t.graph==cond_graph；AssertionError）
我已经编写了一个自定义编写层（在Keras中实现）。这一层在TF版本（1.13和1.15）中运行良好，我已经对自己进行了测试。另外，在TF2.0中，在导入Tensorflow模块后运行自定义编写层中的Tensorflow 2出错，该层在Tensorflow 1中正常工作（assert t.graph==cond_graph；AssertionError）,tensorflow,graph,keras-layer,tensorflow2.0,tf.keras,Tensorflow,Graph,Keras Layer,Tensorflow2.0,Tf.keras,我已经编写了一个自定义编写层（在Keras中实现）。这一层在TF版本（1.13和1.15）中运行良好，我已经对自己进行了测试。另外，在TF2.0中，在导入Tensorflow模块后运行TF.compat.v1.disable_v2_behavior（）line时，没有遇到错误。但是，如果我注释行tf.compat.v1.disable\u v2\u behavior（）以保持TF2 behavior处于启用状态，则将引发一个断言错误。 这是自定义写入层： from tensorflow.kera
TF.compat.v1.disable_v2_behavior（）tf.compat.v1.disable\u v2\u behavior（）这是自定义写入层：
from tensorflow.keras import backend as K
from tensorflow.keras.layers import Layer
import tensorflow as tf
import numpy as np
import math
class Conv1D_Sinc(tf.keras.layers.Layer):
@staticmethod
def to_mel(hz):
return 2595 * np.log10(1 + hz / 700)
@staticmethod
def to_hz(mel):
return 700 * (10 ** (mel / 2595) - 1)
def __init__(self,
filters,
kernel_size,
strides=1,
padding='same',
data_format='channels_last',
sample_rate=16000,
min_low_hz=50,
min_high_hz=50,
low_hz=30,
**kwargs):
super(Conv1D_Sinc, self).__init__(**kwargs)
self.filters = filters
self.kernel_size = kernel_size
self.strides = strides
self.padding = padding
self.data_format = data_format
self.sample_rate = sample_rate
self.min_low_hz = min_low_hz
self.min_high_hz = min_high_hz
self.low_hz = low_hz
self.weights_init_for_build = True
self.high_hz = self.sample_rate / 2 - (self.min_low_hz + self.min_high_hz)
if self.padding == 'causal':
if self.data_format != 'channels_last':
raise ValueError('When using causal padding in `Conv1D`, '
'`data_format` must be "channels_last" '
'(temporal data).')
try:
kernel_size_number = self.kernel_size[0]
if kernel_size_number % 2 == 0:
raise ValueError('''Speed Issue:
Length of filters ought to be odd (i.e, symmetric).''')
except:
if self.kernel_size % 2 == 0:
raise ValueError('''Speed Issue:
Length of filters ought to be odd (i.e, symmetric).''')
if self.data_format == 'channels_first':
self.channel_axis = 1
else:
self.channel_axis = -1
# mel : (81,) np array
mel = np.linspace(self.to_mel(self.low_hz), self.to_mel(self.high_hz),
self.filters + 1)
# hz : (81,) np array
hz = self.to_hz(mel)
# Hamming window
n_lin = np.linspace(0, self.kernel_size / 2 - 1, num=int(self.kernel_size / 2)) # computing
# only half of the window
self.window_ = np.array((0.54 - 0.46 * np.cos(2 * math.pi * n_lin / self.kernel_size)))
# (kernel_size, 1)
n = (self.kernel_size - 1) / 2.0
self.n_ = np.array(2 * math.pi *
np.reshape(np.arange(-n, 0), (1, -1)) /
self.sample_rate) # Due to symmetry, I only need half of the time axes
# filter lower frequency band (filters, 1) - TRAINABLE VECTORS
# hz[:,-1] : (80,)
self.low_vec = np.array(np.reshape(hz[:-1], (-1, 1))) # low_hz_ (80, 1)
# filter frequency band (filters, 1) - TRAINABLE VECTORS
# diff(hz) : (80,)
self.high_vec = np.array(np.reshape(np.diff(hz), (-1, 1))) # band_hz_ (80, 1)
def build(self, input_shape):
# Create a trainable weight variable for this layer.
if input_shape[self.channel_axis] is None:
raise ValueError('The channel dimension of the inputs '+\
'should be defined. Found `None`.')
input_dim = input_shape[self.channel_axis]
# kernel_size = 251
# kernel_shape = (251,) + (1, 80)
# kernel_shape = (251, 1, 80)
self.kernel_shape = (self.kernel_size,) + (input_dim, self.filters)
self.kernel_low = self.add_weight(shape=(self.filters, 1),
initializer='zeros', # should be zero.
# I will change it manually in call().
name='kernel_low',
regularizer=None,
constraint=None)
self.kernel_high = self.add_weight(shape=(self.filters, 1),
initializer='zeros', # should be zero.
# I will change it manually in call().
name='kernel_high',
regularizer=None,
constraint=None)
def call(self, inputs):
# input of the call function is : (None, 3200, 1) === (?, 3200, 1)
if self.weights_init_for_build:
self.weights_init_for_build = False
self.kernel_low = self.kernel_low + self.low_vec
self.kernel_high = self.kernel_high + self.high_vec
# low = self.min_low_hz + tf.math.abs(self.kernel_low)
low = self.min_low_hz + K.abs(self.kernel_low)
high = tf.clip_by_value(low + self.min_high_hz +
K.abs(self.kernel_high),
self.min_low_hz, self.sample_rate / 2)
band = (high - low)[:, 0]
f_times_t_low = K.dot(low, K.variable(self.n_))
f_times_t_high = K.dot(high, K.variable(self.n_))
band_pass_left = ((K.sin(f_times_t_high) - K.sin(f_times_t_low)) / (
self.n_ / 2)) * self.window_ # Equivalent of Eq.4 of the reference paper
# (SPEAKER RECOGNITION FROM RAW WAVEFORM WITH SINCNET). I just have expanded
# the sinc and simplified the terms. This way I avoid several useless computations.
band_pass_center = 2 * K.reshape(band, (-1, 1))
# band_pass_right = tf.reverse(band_pass_left, axis=[1])
band_pass_right = K.reverse(band_pass_left, axes=1)
band_pass = K.concatenate([band_pass_left, band_pass_center, band_pass_right], axis=1)
band_pass = band_pass / (2 * band[:, None])
self.kernel_all = K.reshape(band_pass, self.kernel_shape)
# if self.rank == 1:
outputs = K.conv1d(
inputs,
self.kernel_all,
strides=self.strides,
padding=self.padding,
data_format=self.data_format)
return outputs
def get_config(self):
config = super(tf.keras.layers.Conv1D, self).get_config()
if 'rank' in config.keys():
config.pop('rank')
return config
Traceback (most recent call last):
File "/media/deep/6BED7674319D2F8E/m-a-dastgheib/projects-code/keras/lightNet_v2/main_script_train.py", line 56, in <module>
history_of_batch = ml_model.train_on_batch(np.expand_dims(input_batch, -1), label_of_batch)
File "/home/deep/anaconda3/envs/tf2/lib/python3.7/site-packages/tensorflow_core/python/keras/engine/training.py", line 973, in train_on_batch
class_weight=class_weight, reset_metrics=reset_metrics)
File "/home/deep/anaconda3/envs/tf2/lib/python3.7/site-packages/tensorflow_core/python/keras/engine/training_v2_utils.py", line 264, in train_on_batch
output_loss_metrics=model._output_loss_metrics)
File "/home/deep/anaconda3/envs/tf2/lib/python3.7/site-packages/tensorflow_core/python/keras/engine/training_eager.py", line 311, in train_on_batch
output_loss_metrics=output_loss_metrics))
File "/home/deep/anaconda3/envs/tf2/lib/python3.7/site-packages/tensorflow_core/python/keras/engine/training_eager.py", line 268, in _process_single_batch
grads = tape.gradient(scaled_total_loss, trainable_weights)
File "/home/deep/anaconda3/envs/tf2/lib/python3.7/site-packages/tensorflow_core/python/eager/backprop.py", line 1014, in gradient
unconnected_gradients=unconnected_gradients)
File "/home/deep/anaconda3/envs/tf2/lib/python3.7/site-packages/tensorflow_core/python/eager/imperative_grad.py", line 76, in imperative_grad
compat.as_str(unconnected_gradients.value))
File "/home/deep/anaconda3/envs/tf2/lib/python3.7/site-packages/tensorflow_core/python/eager/backprop.py", line 138, in _gradient_function
return grad_fn(mock_op, *out_grads)
File "/home/deep/anaconda3/envs/tf2/lib/python3.7/site-packages/tensorflow_core/python/ops/cond_v2.py", line 166, in _IfGrad
true_grad_inputs = _resolve_grad_inputs(true_graph, true_grad_graph)
File "/home/deep/anaconda3/envs/tf2/lib/python3.7/site-packages/tensorflow_core/python/ops/cond_v2.py", line 424, in _resolve_grad_inputs
assert t.graph == cond_graph
AssertionError
回溯（最近一次呼叫最后一次）：
文件“/media/deep/6BED7674319D2F8E/m-a-dastgheib/projects code/keras/lightNet_v2/main_script_train.py”，第56行，in
批次历史记录=ml批次模型。在批次上训练（np.扩展dims（输入批次，-1），标记批次）
文件“/home/deep/anaconda3/envs/tf2/lib/python3.7/site packages/tensorflow\u core/python/keras/engine/training.py”，第973行，在批处理的列中
类别权重=类别权重，重置度量=重置度量）
文件“/home/deep/anaconda3/envs/tf2/lib/python3.7/site packages/tensorflow\u core/python/keras/engine/training\u v2\u utils.py”，第264行，在train\u on\u批处理中
输出\损失\度量=模型。\输出\损失\度量）
文件“/home/deep/anaconda3/envs/tf2/lib/python3.7/site-packages/tensorflow\u-core/python/keras/engine/training\u-eager.py”，第311行，在批量生产中
输出损失度量=输出损失度量）
文件“/home/deep/anaconda3/envs/tf2/lib/python3.7/site packages/tensorflow\u core/python/keras/engine/training\u eager.py”，第268行，在单批处理中
梯度=磁带梯度（标度总损耗、可训练重量）
文件“/home/deep/anaconda3/envs/tf2/lib/python3.7/site packages/tensorflow_core/python/eager/backprop.py”，第1014行，渐变格式
未连接的_渐变=未连接的_渐变）
文件“/home/deep/anaconda3/envs/tf2/lib/python3.7/site packages/tensorflow\u core/python/eager/祈使命令\u grad.py”，第76行，祈使命令\u grad
兼容as_str（未连接的梯度值））
文件“/home/deep/anaconda3/envs/tf2/lib/python3.7/site packages/tensorflow\u core/python/eager/backprop.py”，第138行，在函数中
返回级（模拟级，*非模拟级）
文件“/home/deep/anaconda3/envs/tf2/lib/python3.7/site packages/tensorflow_core/python/ops/cond_v2.py”，第166行，在IfGrad中
真梯度输入=\u解析梯度输入（真梯度图、真梯度图）
文件“/home/deep/anaconda3/envs/tf2/lib/python3.7/site packages/tensorflow\u core/python/ops/cond\u v2.py”，第424行，在“解析”梯度输入中
断言t.graph==cond_图
断言错误