Input_shape (None, None, 3)

[pluniak]

Referring to closed issue #65:

I still can't load the model with input_shape=(None, None, 3) using the latest (TF2-based) version of this repo:

Deeplabv3(weights='pascal_voc', input_shape=(None, None, 3), classes=1, backbone='xception', OS=16)

gives me

ValueError: Cannot convert a partially known TensorShape to a Tensor: (None, None)

caused by lines 417-419 in model.py. I can load the model when replacing x.shape with tf.shape(x) in line 418, but then I'm getting another shape error during training (model.fit). Am I doing something wrong? Do others get the same error message? I tried to solve the issue on my own, but it seems to be beyond my skills.

Additional info: using input_shape of (248,248,3) causes

ValueError: A Concatenate layer requires inputs with matching shapes except for the concat axis. Got inputs shapes: [(None, 64, 64, 256), (None, 62, 62, 48)]

in Line 425 in model.py. The TF1.13-based version of this repo accepts 248*248 input images.

[yingshaoxo]

With tensorflow 2.0.0-beta1

    b4 = Lambda(lambda x: tf.compat.v1.image.resize(x, size_before[1:3],
                                                    method='bilinear', align_corners=True))(b4)

tf.image.resize_images(
    images,
    size,
    method=ResizeMethodV1.BILINEAR,
    align_corners=False,
    preserve_aspect_ratio=False,
    name=None
)

The size parameter must be a 1-D int32 Tensor

https://www.tensorflow.org/api_docs/python/tf/image/resize_images#args

[yingshaoxo]

If I use deeplab_model = Deeplabv3(input_shape=(None, None, 3), classes=4), I'll get :

size_before:  (None, None, None, 320)
size_before[1:3]:  (None, None)

which cause error:

Traceback (most recent call last):
  File "main.py", line 3, in <module>
    deeplab_model = Deeplabv3(input_shape=(None, None, 3), classes=4)
  File "/home/yingshaoxo/Codes/keras-deeplab-v3-plus/model.py", line 385, in Deeplabv3
    method='bilinear', align_corners=True))(b4)
  File "/usr/lib/python3.7/site-packages/tensorflow/python/keras/engine/base_layer.py", line 662, in __call__
    outputs = call_fn(inputs, *args, **kwargs)
  File "/usr/lib/python3.7/site-packages/tensorflow/python/keras/layers/core.py", line 785, in call
    return self.function(inputs, **arguments)
  File "/home/yingshaoxo/Codes/keras-deeplab-v3-plus/model.py", line 385, in <lambda>
    method='bilinear', align_corners=True))(b4)
  File "/usr/lib/python3.7/site-packages/tensorflow/python/ops/image_ops_impl.py", line 1180, in resize_images
    skip_resize_if_same=True)
  File "/usr/lib/python3.7/site-packages/tensorflow/python/ops/image_ops_impl.py", line 1041, in _resize_images_common
    raise ValueError('\'size\' must be a 1-D int32 Tensor')
ValueError: 'size' must be a 1-D int32 Tensor

[yingshaoxo]

In the end, I have no choice but reload model every time before taking a new image as input:

import numpy as np
from PIL import Image
from model import Deeplabv3

image = np.array(Image.open('imgs/image1.jpg'))
image_shape = image.shape

deeplab_model = Deeplabv3(input_shape=(image_shape[0], image_shape[1], 3), classes=4)

image = np.expand_dims(image, axis=0)
y = deeplab_model.predict(image)

print(y)
print(y.shape)

[yingshaoxo]

Wait for a minute.

Maybe we don't need this feature at all.

Because normally, you'll deal same resolution images at one specific task.

For example, process a whole length of movie or video.

(Or you can still use a fixed resolution image as input, then scale the resulting mask by yourself.

[pluniak]

That's a workaround indeed. Resizing might affect segmentation performance though, as object dimensions change. Moreover, it's computationally expensive when doing on the fly or you need a second copy of your images on hard disk if you wanna do it up front. That's why flexible model inputs are desirable I think.
Still I wann give great props to @bonlime for the Keras implementation of this model. Awesome job and many thanks!