Tensorflow model cannot add node after predict() called-CodePudding

import tensorflow as tf
from tensorflow.keras import Sequential
from tensorflow.keras.layers import LSTM, Conv1D, Dense
import numpy as np

#Create test data
X = np.array(list(range(100)))
X = np.reshape(X,(1,100,1))
print(X.shape)

#Model A
model_a = Sequential()
model_a.add(Conv1D(filters=1, kernel_size=10, strides=10, kernel_initializer = tf.keras.initializers.ones, use_bias = False, activation=None, padding='valid',input_shape=(100,1)))
model_a.add(Conv1D(filters=1, kernel_size=10, strides=10, kernel_initializer = tf.keras.initializers.ones, use_bias = False, activation=None, padding='valid',input_shape=(10,1)))
print(model_a.summary())
#Model A predict
predict_a = model_a.predict(X)
print(predict_a)
print(predict_a.shape)

#Model B
model_b = Sequential()
model_b.add(Conv1D(filters=1, kernel_size=10, strides=10, kernel_initializer = tf.keras.initializers.ones, use_bias = False, activation=None, padding='valid',input_shape=(100,1)))
#Make predict and add node after that
tmpPredict = model_b.predict(X)
#
model_b.add(Conv1D(filters=1, kernel_size=10, strides=10, kernel_initializer = tf.keras.initializers.ones, use_bias = False, activation=None, padding='valid',input_shape=(10,1)))
print(model_a.summary())
#Model B predict
predict_b = model_b.predict(X)
print(predict_b)
print(predict_b.shape)

The Model A seem fine, the final output shape is (1,1,1). But Model B seem cannot add the new node after the predict() function called. Why?

CodePudding user response：

Tensorflow automatically creates a static computation graph once the model is built (building happens when you first call the model). This is the default behavior according to the documentation of tf.keras.Sequential:

By default, we will attempt to compile your model to a static graph to deliver the best execution performance.

To instead run the model eagerly, you can disable this graph building by calling tf.config.run_functions_eagerly(True) before calling model.predict. Running the model eagerly should allow you to add layers after model.predict has been called. For more information on the static computation graphs, refer to https://www.tensorflow.org/guide/intro_to_graphs