我目前正在開發一個 Autoencoder 類 - 其中一種方法如下：
@tf.function
def build_ae(self, inputs):
self.input_ae = inputs
self.encoded = self.encoder(self.input_ae)
self.decoded = self.decoder(self.encoded)
self.autoencoder = Model(self.input_ae,outputs=self.encoded)
如果我嘗試運行此功能：
inp = Input(shape=(Nx, Nx, Nu)) # Nx, Nu are both int
ae = Autoencoder(Nx, Nu, [32, 64, 128]) # Simply specifying layers input dimensions
ae.build_ae(inp)
我收到以下錯誤：
TypeError: Cannot convert a symbolic Keras input/output to a numpy array.
This error may indicate that you're trying to pass a symbolic value to a NumPy call,
which is not supported. Or, you may be trying to pass Keras symbolic inputs/outputs
to a TF API that does not register dispatching, preventing Keras from automatically
converting the API call to a lambda layer in the Functional Model.
但是，當我洗掉@tf.function裝飾器時，該功能按預期作業。
我試過寫一個簡單的測驗示例：
class Test:
@tf.function
def build_test(self, inputs):
self.inp = inputs
t = Test()
input_t = Input(shape=(3,3,3))
t.build_test(input_t)
再一次，這會導致相同的錯誤。
我試過禁用急切執行，但這沒有效果。
有誰知道為什么這可能不起作用？
更新：
這是完整的 Autoencoder 類：
import einops
import h5py
from pathlib import Path
from typing import List, Tuple
import numpy as np
import tensorflow as tf
from tensorflow.keras import Input
from tensorflow.keras.layers import (
Dense,
Conv2D,
MaxPool2D,
UpSampling2D,
concatenate,
BatchNormalization,
Conv2DTranspose,
Flatten,
PReLU,
Reshape,
Dropout,
AveragePooling2D,
add,
Lambda,
Layer,
TimeDistributed,
LSTM
)
from tensorflow.keras.regularizers import l2
from tensorflow.keras.optimizers import Adam
from tensorflow.keras.models import Model
from tensorflow.keras.callbacks import ModelCheckpoint, EarlyStopping
class Autoencoder:
def __init__(self,
nx: int,
nu: int,
features_layers: List[int],
latent_dim: int = 16,
filter_window: Tuple[int, int] = (3, 3),
act_fn: str = 'tanh',
batch_norm: bool = False,
dropout_rate: float = 0.0,
lmb: float = 0.0,
resize_method: str = 'bilinear') -> None:
self.nx = nx
self.nu = nu
self.features_layers = features_layers
self.latent_dim = latent_dim
self.filter_window = filter_window
self.act_fn = act_fn
self.batch_norm = batch_norm
self.dropout_rate = dropout_rate
self.lmb = lmb
self.resize_method = resize_method
self.train_history = None
self.val_history = None
self.encoder = Encoder(
self.nx,
self.nu,
self.features_layers,
self.latent_dim,
self.filter_window,
self.act_fn,
self.batch_norm,
self.dropout_rate,
self.lmb
)
self.decoder = Decoder(
self.nx,
self.nu,
self.features_layers,
self.latent_dim,
self.filter_window,
self.act_fn,
self.batch_norm,
self.dropout_rate,
self.lmb,
self.resize_method
)
@tf.function
def build_ae(self, inputs):
self.input_ae = inputs
self.encoded = self.encoder(self.input_ae)
self.decoded = self.decoder(self.encoded)
self.autoencoder = Model(self.input_ae, outputs=self.decoded)
@tf.function
def compile_ae(self, learning_rate, loss):
self.autoencoder.compile(optimizer=Adam(learning_rate=learning_rate), loss=loss)
def train(self,
inputs,
targets,
inputs_valid,
targets_valid,
n_epoch,
batch_size,
learning_rate,
patience,
filepath):
model_cb = ModelCheckpoint(filepath, monitor='val_loss', save_best_only=True, verbose=1, save_format="h5")
early_cb = EarlyStopping(monitor='val_loss', patience=patience, verbose=1)
cb = [model_cb, early_cb]
self.train_history = []
self.val_history = []
tf.keras.backend.set_value(self.autoencoder.optimizer.lr, learning_rate)
hist = self.autoencoder.fit(
inputs,
targets,
epochs=n_epoch,
batch_size=batch_size,
shuffle=True,
validation_data=(inputs_valid, targets_valid),
callbacks=cb
)
self.train_history.extend(hist.history['loss'])
self.train_history.extend(hist.history['val_loss'])
return self.train_history, self.val_history
uj5u.com熱心網友回復：
以下方法似乎都可以完成這項作業：
tf.config.run_functions_eagerly
class Test:
@tf.function
def build_test(self, inputs):
self.inp = inputs
t = Test()
input_t = tf.keras.layers.Input(shape=(3,3,3))
t.build_test(input_t)
和
class Test:
@tf.function
def build_test(self, inputs):
self.inp = inputs
t = Test()
t.build_test(tf.constant(1))
根據檔案，當您創建一個tf.keras.Model：
默認情況下，我們會嘗試將您的模型編譯為靜態圖以提供最佳執行性能。
您已經在您的build_ae方法中創建了一個模型，所以我認為省略@tf.function裝飾器不會影響您的性能。
轉載請註明出處，本文鏈接：https://www.uj5u.com/shujuku/327207.html