How to call TensorFlow model with linspace?

Question

I'm trying to call a TensorFlow model on a linspace but I can't seem to get even a very simple example (based on https://www.tensorflow.org/api_docs/python/tf/keras/Model) to work:

import tensorflow as tf

class FeedForward(tf.keras.Model):
    def __init__(self):
        super().__init__()
        self.dense1 = tf.keras.layers.Dense(4, activation=tf.nn.relu, name='lyr1')
        self.dense2 = tf.keras.layers.Dense(5, activation=tf.nn.softmax, name='lyr2')
    
    def call(self, inputs, training=False):
        x = self.dense1(inputs) # <--- error here
        return self.dense2(x)

model = FeedForward()
batchSize = 2048

preX = tf.linspace(0.0, 10.0, batchSize)
model(preX, training=True)

I get the following error (on the line indicated above):

ValueError: Exception encountered when calling layer "feed_forward_87" (type FeedForward).

Input 0 of layer "lyr1" is incompatible with the layer: expected min_ndim=2, found ndim=1. Full shape received: (2048,)

Call arguments received:
  • inputs=tf.Tensor(shape=(2048,), dtype=float32)
  • training=True

I have tried adding an input layer

import tensorflow as tf

class FeedForward(tf.keras.Model):
    def __init__(self):
        super().__init__()
        self.inLyr = tf.keras.layers.Input(shape=(2048,), name='inp')
        self.dense1 = tf.keras.layers.Dense(4, activation=tf.nn.relu, name='lyr1')
        self.dense2 = tf.keras.layers.Dense(5, activation=tf.nn.softmax, name='lyr2')
    
    def call(self, inputs, training=False):
        x = self.inLyr(inputs) # <---- error here
        x = self.dense1(x)
        return self.dense2(x)

model = FeedForward()
batchSize = 2048

preX = tf.linspace(0.0, 10.0, batchSize)
model(preX, training=True)

But then the error becomes

TypeError: Exception encountered when calling layer "feed_forward_88" (type FeedForward).

'KerasTensor' object is not callable

Call arguments received:
  • inputs=tf.Tensor(shape=(2048,), dtype=float32)
  • training=True

Any help is appreciated.

Answer 1

You do not need an explicit Input layer. You are just missing a dimension. Try using tf.expand_dims to get the required input shape (batch_size, features):

import tensorflow as tf

class FeedForward(tf.keras.Model):
    def __init__(self):
        super().__init__()
        self.dense1 = tf.keras.layers.Dense(4, activation=tf.nn.relu, name='lyr1')
        self.dense2 = tf.keras.layers.Dense(5, activation=tf.nn.softmax, name='lyr2')
    
    def call(self, inputs, training=False):
        x = self.dense1(inputs)
        return self.dense2(x)

model = FeedForward()
batchSize = 2048

preX = tf.expand_dims(tf.linspace(0.0, 10.0, batchSize), axis=-1)
model(preX, training=True)

<tf.Tensor: shape=(2048, 5), dtype=float32, numpy=
array([[2.0000000e-01, 2.0000000e-01, 2.0000000e-01, 2.0000000e-01,
        2.0000000e-01],
       [2.0061693e-01, 1.9939975e-01, 1.9954492e-01, 2.0115443e-01,
        1.9928394e-01],
       [2.0123295e-01, 1.9879854e-01, 1.9908811e-01, 2.0231272e-01,
        1.9856769e-01],
       ...,
       [4.1859145e-03, 1.6480753e-08, 7.3006660e-08, 9.9581391e-01,
        5.0233933e-09],
       [4.1747764e-03, 1.6337188e-08, 7.2423312e-08, 9.9582505e-01,
        4.9767293e-09],
       [4.1636685e-03, 1.6194873e-08, 7.1844603e-08, 9.9583614e-01,
        4.9305160e-09]], dtype=float32)>

Or with an Input layer:

class FeedForward(tf.keras.Model):
    def __init__(self):
        super().__init__()
        inputs = tf.keras.layers.Input((1,))
        dense1 = tf.keras.layers.Dense(4, activation=tf.nn.relu, name='lyr1')(inputs)
        dense2 = tf.keras.layers.Dense(5, activation=tf.nn.softmax, name='lyr2')(dense1)
        self.model = tf.keras.Model(inputs, dense2)
    def call(self, inputs, training=False):
        return self.model(inputs)

Answer 2

I study from your reply also and found their Fn may need to work as desns_1 and dense_2 then I reform it into the model of a sequence.

For the error this is to fixed !

preX = tf.linspace([0.0], [10.0], batchSize, axis=0)

[ Sample ]:

import tensorflow as tf
import numpy as np

import matplotlib.pyplot as plt

"""""""""""""""""""""""""""""""""""""""""""""""""""""""""
: Varibles
"""""""""""""""""""""""""""""""""""""""""""""""""""""""""
batchSize = 2048
preX = tf.linspace([0.0], [10.0], batchSize, axis=0)

"""""""""""""""""""""""""""""""""""""""""""""""""""""""""
: Class
"""""""""""""""""""""""""""""""""""""""""""""""""""""""""
class FeedForward(tf.keras.Model):
    def __init__(self):
        super().__init__()
        self.input1 = tf.keras.layers.InputLayer(input_shape=(2048, 1))
        self.dense1 = tf.keras.layers.Dense(4, activation=tf.nn.relu, name='lyr1')
        self.dense2 = tf.keras.layers.Dense(5, activation=tf.nn.softmax, name='lyr2')
        self.Q_Model = tf.keras.models.Sequential([ ])
        self.Q_Model.add(self.input1)
        self.Q_Model.add(self.dense1)
        self.Q_Model.add(self.dense2)
        
        self.optimizer = tf.keras.optimizers.Nadam(
            learning_rate=0.0001, beta_1=0.9, beta_2=0.999, epsilon=1e-07,
            name='Nadam'
        )
        self.lossfn = tf.keras.losses.MeanSquaredLogarithmicError(reduction=tf.keras.losses.Reduction.AUTO, name='mean_squared_logarithmic_error')
        self.Q_Model.compile(optimizer=self.optimizer, loss=self.lossfn, metrics=['accuracy'])
    
    def call(self, inputs, training=False):
        inputs = tf.constant( inputs, dtype=tf.float32, shape=(1, 2048, 1) )
        dataset = tf.data.Dataset.from_tensor_slices(([inputs], [inputs]))
        history = self.Q_Model.fit(dataset, epochs=10)
        
        return history


"""""""""""""""""""""""""""""""""""""""""""""""""""""""""
: Running
"""""""""""""""""""""""""""""""""""""""""""""""""""""""""
model = FeedForward()
history = model(preX, training=True)

"""""""""""""""""""""""""""""""""""""""""""""""""""""""""
: Summary
"""""""""""""""""""""""""""""""""""""""""""""""""""""""""
plt.plot(history.history['loss'])
plt.show()
plt.close()

[ Output ]:

Epoch 1/10
1/1 [==============================] - 1s 997ms/step - loss: 2.5892 - accuracy: 4.8828e-04
Epoch 2/10
1/1 [==============================] - 0s 7ms/step - loss: 2.5892 - accuracy: 4.8828e-04
Epoch 3/10
1/1 [==============================] - 0s 10ms/step - loss: 2.5891 - accuracy: 4.8828e-04
Epoch 4/10
1/1 [==============================] - 0s 6ms/step - loss: 2.5891 - accuracy: 4.8828e-04
Epoch 5/10
1/1 [==============================] - 0s 5ms/step - loss: 2.5890 - accuracy: 4.8828e-04
Epoch 6/10
1/1 [==============================] - 0s 6ms/step - loss: 2.5890 - accuracy: 4.8828e-04
Epoch 7/10
1/1 [==============================] - 0s 8ms/step - loss: 2.5889 - accuracy: 4.8828e-04
Epoch 8/10
1/1 [==============================] - 0s 5ms/step - loss: 2.5889 - accuracy: 4.8828e-04
Epoch 9/10
1/1 [==============================] - 0s 6ms/step - loss: 2.5888 - accuracy: 4.8828e-04
Epoch 10/10
1/1 [==============================] - 0s 7ms/step - loss: 2.5888 - accuracy: 0.0000e+00