CODEWITHSUNDEEP
tensorflowkeras
Derk
Derk

Reputation: 1395

Keras and tensorflow eager execution

I have a model from https://github.com/bonlime/keras-deeplab-v3-plus that I try to customize a bit.

I want to run it with Tensorflow Eager mode

from model import Deeplabv3
import tensorflow as tf
tf.enable_eager_execution()

model = Deeplabv3(weights='pascal_voc', input_shape=(200,200,3), backbone='mobilenetv2', classes=64)
batch = tf.zeros((1,200,200,3))
f = model(batch)

However this gives the error:

model.py", line 236, in _inverted_res_block in_channels = inputs._keras_shape[-1] AttributeError: 'DeferredTensor' object has no attribute '_keras_shape'

It is about this part of the code

def _inverted_res_block(inputs, expansion, stride, alpha, filters, block_id, skip_connection, rate=1):
    in_channels = inputs._keras_shape[-1]
    #in_channels = inputs.get_shape()[-1].value
    #in_channels = tf.shape(inputs)[-1]

    import pdb;pdb.set_trace()

    pointwise_conv_filters = int(filters * alpha)
    pointwise_filters = _make_divisible(pointwise_conv_filters, 8)
    x = inputs

    prefix = 'expanded_conv_{}_'.format(block_id)
    if block_id:
        # Expand

        x = Conv2D(expansion * in_channels, kernel_size=1, padding='same',
                   use_bias=False, activation=None,
                   name=prefix + 'expand')(x)

How to solve this?

Upvotes: 2

Views: 1533

Answers (2)

mmonlab
mmonlab

Reputation: 51

I have made this change:

Instead this line: in_channels = inputs.shape[-1].value or this other line: inputs._keras_shape[-1]

I have used this other: in_channels = inputs.shape.as_list()[-1]

and it works for me.

Upvotes: 0

ash
ash

Reputation: 6751

As P-gn pointed out:

  • tf.keras (included with TensorFlow) supports eager execution, the keras module does not.
  • tf.keras implements the keras API spec, so it should be a drop-in replacement for any program using keras (e.g., change references to keras.Model to tf.keras.Model). Plus it additionally supports eager execution in TensorFlow.

Upvotes: 1

Related Questions