Functional and object-oriented models are not equivalent

TensorFlow General Discussion
,
1

I don’t quite understand why these two models defined with functional API:

inp = layers.Input((10,2))
x = layers.Flatten()(inp)
x = layers.Dense(5)(x)
m1 = keras.models.Model(inputs=inp, outputs=x)

and OO way:

class MyModel(tf.keras.Model):
    def __init__(self, inp_shape, out_size = 5):
        super(MyModel, self).__init__()
        self.inp = layers.InputLayer(input_shape=inp_shape)
        self.flatten = layers.Flatten()
        self.dense = layers.Dense(out_size)

    def call(self, a):
        x = self.inp(a)
        x = self.flatten(x)
        x = self.dense(x)
        return x
m2 = MyModel((10,2))
m2.build(input_shape = (10,2))

give different results:

> m1.summary()
Model: "functional_3"
_________________________________________________________________
Layer (type)                 Output Shape              Param #   
=================================================================
input_17 (InputLayer)        [(None, 10, 2)]           0         
_________________________________________________________________
flatten_19 (Flatten)         (None, 20)                0         
_________________________________________________________________
dense_18 (Dense)             (None, 5)                 105       
=================================================================
Total params: 105
Trainable params: 105
Non-trainable params: 0

> m2.summary()
Model: "my_model_9"
_________________________________________________________________
Layer (type)                 Output Shape              Param #   
=================================================================
input_18 (InputLayer)        [(None, 10, 2)]           0         
_________________________________________________________________
flatten_20 (Flatten)         multiple                  0         
_________________________________________________________________
dense_19 (Dense)             multiple                  15        
=================================================================
Total params: 15
Trainable params: 15
Non-trainable params: 0

When I test it with on some toy tensor, I get:

> tsta = np.random.randn(3,10,2)
> m1(tsta) # correct
> m2(tsta)
InvalidArgumentError: Matrix size-incompatible: In[0]: [3,20], In[1]: [2,5] [Op:MatMul]

What I want to achieve is to have exactly the same model as m1 but with subclassed API.

3

You could try my example:

import tensorflow as tf

tsta = tf.random.uniform([3,10,2])
tf.random.set_seed(111111)
inp = tf.keras.layers.Input((10,2))
x = tf.keras.layers.Flatten()(inp)
x = tf.keras.layers.Dense(5)(x)
m1 = tf.keras.Model(inputs=inp, outputs=x)
m1.build(input_shape = (3, 10, 2))
m1.summary()
print(m1(tsta))

tf.random.set_seed(111111)
class MyModel(tf.keras.Model):
   def __init__(self, out_size = 5):
       super(MyModel, self).__init__()
       self.flatten = tf.keras.layers.Flatten()
       self.dense = tf.keras.layers.Dense(out_size)

   def call(self, x):
       x = self.flatten(x)
       x = self.dense(x)
       return x
m2 = MyModel()
m2.build(input_shape = (3, 10, 2))
m2.summary()
print(m2(tsta))
4

Thanks, and why the summary of m2 gives:

Model: "my_model"
_________________________________________________________________
Layer (type)                 Output Shape              Param #   
=================================================================
flatten_4 (Flatten)          multiple                  0         
_________________________________________________________________
dense_4 (Dense)              multiple                  105       
=================================================================

so you see how there’s multiple in the output? The whole reason why I’m struggling with this is because I want to check the output sizes of my model objects (slightly more complicated than that), but what I’m getting so far is mulitple. Any remedy for this?

5

Subclass model doesn’t record any shape info about layers such as layer.output_shape.
So it prints multiple.