Error with different dataset about semantic-segmentation-tensorflow-eager

[Shathe]

This repository is just an example (not a tutorial) that's why I did not commented anything about my Loader class.

The Loader class, loads the image labels but, they need to have values from 0 to the number of classes - 1. That is: if You have 2 classes the labels must be 0 and 1. If you have 10 classes, the labels must be 0,1,2,3,4,5,6,7,8,9.
Other label values will be ignore (new commit, pull new changes)

I think the values are 0 and 255 so -> Please change the 255 values to 1's.

Another thing to change is the number of classes. You have two classes not one. It's a binary problem. You have the class of "that's a thing I want to segment" and the class of "that is not the thing I want to segment"

[rafaelmarconiramos]

Thanks. I changed the code start, but now generate another error:

Traceback (most recent call last):
File "train_eager_seg.py", line 109, in
train(loader=loader, model=model, epochs=20, batch_size=8)
File "train_eager_seg.py", line 50, in train
loss = tf.losses.softmax_cross_entropy(y, y_, weights=mask)
File "/home/rafaelmr/.conda/envs/seg1/lib/python2.7/site-packages/tensorflow/python/ops/losses/losses_impl.py", line 793, in softmax_cross_entropy
logits.get_shape().assert_is_compatible_with(onehot_labels.get_shape())
File "/home/rafaelmr/.conda/envs/seg1/lib/python2.7/site-packages/tensorflow/python/framework/tensor_shape.py", line 847, in assert_is_compatible_with
raise ValueError("Shapes %s and %s are incompatible" % (self, other))
ValueError: Shapes (8, 160, 160, 2) and (8, 140, 140, 2) are incompatible

All images are 140x140, and I configured the parameters for that size.
Why the model generates a shape like (8, 160, 160, 2)?
Do I need to know something more about the data? Or the model?
I looked for in the project files, but I didn´t find.

Thanks for your attention.

[Shathe]

That is something concerning your code and yes, it's about sizes/shapes

Somewhere you must have specified the 160x160 (at code level) or the images are 140x140.

The image size is specified here https://github.com/Shathe/Semantic-Segmentation-Tensorflow-Eager/blob/master/train_eager.py#L100

The network input shape (input_shape parameter) here https://github.com/Shathe/Semantic-Segmentation-Tensorflow-Eager/blob/master/train_eager.py#L104

Please, take your time to code and to debug the code

[rafaelmarconiramos]

Sorry.

In this line generates the problem:

Semantic-Segmentation-Tensorflow-Eager/train_eager.py

Line 48 in 333b1a8

y_ = model(x, training=True)

x,y, mask are tf.Tensor(shape=(8, 140, 140, 2), dtype=uint8)
y_ is tf.Tensor(shape=(8, 160, 160, 2), dtype=float32)

I will try to understand what the line 48 do.

[Shathe]

y_ is the output, so the output has that size.
I cannot help more without seen your edited code.

[rafaelmarconiramos]

With the architecture of MnasnetFC, the code works just with the image sizes that are multiples of 32.
Look the shapes of your example (each execution of the out in MnasnetFC.call)
initial: (8, 112, 112, 32)
initial2: (8, 112, 112, 32)
initial3: (8, 112, 112, 32)
initial4: (8, 112, 112, 32)
initial5: (8, 112, 112, 16)
blocks: (8, 56, 56, 24)
blocks: (8, 56, 56, 24)
blocks: (8, 56, 56, 24)
blocks: (8, 28, 28, 40)
blocks: (8, 28, 28, 40)
blocks: (8, 28, 28, 40)
blocks: (8, 14, 14, 80)
blocks: (8, 14, 14, 80)
blocks: (8, 14, 14, 80)
blocks: (8, 14, 14, 96)
blocks: (8, 14, 14, 96)
blocks: (8, 7, 7, 192)
blocks: (8, 7, 7, 192)
blocks: (8, 7, 7, 192)
blocks: (8, 7, 7, 192)
blocks: (8, 7, 7, 320)
initial6: (8, 7, 7, 1152)
up: (8, 14, 14, 1152)
up: (8, 14, 14, 192)
up: (8, 14, 14, 192)
up: (8, 28, 28, 192)
up: (8, 28, 28, 80)
up: (8, 28, 28, 80)
up: (8, 56, 56, 80)
up: (8, 56, 56, 40)
up: (8, 112, 112, 40)
up: (8, 112, 112, 24)
up: (8, 224, 224, 24)
final: (8, 224, 224, 11)

Basic it decreases the shape dividing by two. After it increases the shape multiplying by two. It does this operation five times then it needs a multiply of 32.
In my case the division by two has rest then when it executes the multiplication don´t finish with correct number.

initial: (8, 70, 70, 32)
initial2: (8, 70, 70, 32)
initial3: (8, 70, 70, 32)
initial4: (8, 70, 70, 32)
initial5: (8, 70, 70, 16)
blocks: (8, 35, 35, 24)
blocks: (8, 35, 35, 24)
blocks: (8, 35, 35, 24)
blocks: (8, 18, 18, 40)
blocks: (8, 18, 18, 40)
blocks: (8, 18, 18, 40)
blocks: (8, 9, 9, 80)
blocks: (8, 9, 9, 80)
blocks: (8, 9, 9, 80)
blocks: (8, 9, 9, 96)
blocks: (8, 9, 9, 96)
blocks: (8, 5, 5, 192)
blocks: (8, 5, 5, 192)
blocks: (8, 5, 5, 192)
blocks: (8, 5, 5, 192)
blocks: (8, 5, 5, 320)
initial6: (8, 5, 5, 1152)
up: (8, 10, 10, 1152)
up: (8, 10, 10, 192)
up: (8, 10, 10, 192)
up: (8, 20, 20, 192)
up: (8, 20, 20, 80)
up: (8, 20, 20, 80)
up: (8, 40, 40, 80)
up: (8, 40, 40, 40)
up: (8, 80, 80, 40)
up: (8, 80, 80, 24)
up: (8, 160, 160, 24)
final: (8, 160, 160, 2)

Correct?
My code:
if name == "main":

    n_classes = 2
    dataset_path = 'seg'
    loader = Loader.Loader(dataFolderPath=dataset_path, n_classes=n_classes, problemType='segmentation', width=140,
                    height=140, ignore_label=n_classes)

    # build model and optimizer
    model = MnasnetEager.MnasnetFC(num_classes=n_classes)

    # optimizer
    optimizer = tf.train.AdamOptimizer(0.001)

    train(loader=loader, model=model, epochs=20, batch_size=8)
    get_params(model)

[Shathe]

Mmmm My Loader automatically resize the images to the specified size so although your images are 140, try to set the size to 224 or to 112

[Shathe]

Have you tried it?