How to get the intermediate tensor output of pretrained model in TF2

Question

I want to use attention model to extract attention score.But I can't find any TF2 API to use. simple code :

import tensorflow as tf

model = train_model()
func = tf.function(model)
tensor_specs1 = tf.TensorSpec.from_tensor(model.input)

call = func.get_concrete_function(tensor_specs1)
graph = call.graph

tensor_names = [n for n in graph.as_graph_def().node]
for name in tensor_names:
    print(name)

outputs = graph.get_tensor_by_name('model_1/word_encoder_time/word_attention/Softmax:0')

pred_model = tf.keras.models.Model(model.input,outputs)

results = pred_model(tensor_specs1)

print(results)

but raise an exception:

    raise ValueError("Tensor %s is not an element of this graph." % obj)
ValueError: Tensor Tensor("model_1/word_encoder_time/model/word_attention/BiasAdd:0", shape=(?, 10), dtype=float32) is not an element of this graph

It's working ,but it's not what I want:

outputs = [model.get_layer(name=output).get_output_at(0) for output in output_layers]
pred_model = tf.keras.models.Model(model.input,outputs)

I want to get intermediate tensor,not layer's output.

Answer 1

In order to evaluate the output of a arbitary layer in a Keras model, you can use of Keras functions to avoid working with sessions and graphs.

import tensorflow as tf
print('TensorFlow: ', tf.__version__, end='\n\n')

input_layer = tf.keras.Input(shape=[100])
x = tf.keras.layers.Dense(16, activation='relu')(input_layer)
x = tf.keras.layers.Dense(64, activation='relu')(x)
x = tf.keras.layers.Dense(32, activation='relu')(x)
x = tf.keras.layers.Dense(10, activation='relu')(x)
output_layer = tf.keras.layers.Dense(5, activation='softmax')(x)

model = tf.keras.Model(inputs=[input_layer], outputs=[output_layer])

a = model.layers[3].output
print(a)

fn = tf.keras.backend.function(input_layer, a)  # create a Keras Function
random_input = tf.random.normal([1, 100])  # random noise

a_eval = fn(random_input)
print('\nLayer Output:\n', a_eval)

Output:

TensorFlow:  2.3.0-dev20200611

Tensor("dense_73/Identity:0", shape=(None, 32), dtype=float32)

Layer Output:
 [[0.         0.         0.46475422 0.0961322  0.         0.
  0.23016977 0.         0.         0.05861767 0.03298267 0.11953808
  0.         0.         0.97043467 0.         0.         0.6384926
  0.         0.         0.         0.2346505  0.1822727  0.0145395
  0.08411474 0.         0.         0.37601566 0.         0.
  0.29435986 0.44069782]]