A Deep Convolutional Encoder-Decoder Architecture for Image Segmentation

Hey, thanks for the example!

I found the code on your blog post first and there you're using the custom layer UnPooling2D but here you define it as well but then use the build-in UpSampling2D.

Do they perform exactly the same operation? Why did you change? I understand the code for UnPooling2D but I'm a little bit confused by the code of UpSampling2D..

Thanks in advance

Hi Pradyumna,

Thanks for sharing the code. Just curious, how difficult(or different) if migrating this implementation to Tensorflow-based? Or what are the major parts need to be modified? I am not familiar with Keras, and am learning Tensorflow.

I'm trying to run SegNet however, I'm seeing a lot of errors in the import statements. This is due to the fact that I'm using updated versions of most packages required.

Can you please provide the requirements.txt file containing the versions of packages that you guys used at the time of implementation. Also, can you please tell me what version of python did you guys use to implement this?

Thanks.

Hello,

I am just trying to run your code, I just tried to run the same example as it is mentioned is Segnet.ipynb but unfortunately I am not able to generate the same results as it is shown in it. So this is the image

what I tried to feed the network and this is the output I am getting.

Can you please help me what I am doing wrong.

Best, Muaz

On this line: https://github.com/pradyu1993/segnet/blob/master/segnet.py#L67

there is the number 367, it isn't the number of classes, perhaps it is the number of images?

Also what is the 7 in this line when going through the dataset? https://github.com/pradyu1993/segnet/blob/master/segnet.py#L61

sorry for the basic questions and thanks for your help!

Hi,

You are reshaping the output of last convolutional layer to ( 12 , 360 * 480 ) and permuting it to ( 360 * 480 , 12 ) so output shape after permutation will be ( None , 360 * 480 , 12 ) where None is the batch size as mentioned in ipython notebook.

On this ( None , 360 * 480 , 12 ) , a 3D tensor, you are applying softmax, which according to keras softmax defination will be consider as ( nb_samples , nb_timesteps , nb_channels ).

My question is what is this nb_timesteps?? which is 360*480 for here. As per my understanding in semantic segmentation each pixel is sample so output should be ( None * 360 * 480 , 12 ) , but here we don't know value of None in prior, so can you give me explanation of how to make sense of your implementation, as I am unable to understand this nb_timesteps part.

Hello.

After run this code:

output = autoencoder.predict_proba(train_data[8:9])
out_labels = np.argmax(output[0], axis=1)

pred = visualize(out_labels.reshape((height,width)), False)
plt.imshow(pred)
plt.figure(2)
plt.imshow(gt[8])
plt.show()

i have next result:

I change only Batch size and data size params. I work in PyCharm. It is incorrect prediction or visualisation? What i do wrong?

Hi there,

Thank you for your code.

I'm just wondering if it's a proper way to implement the decoder part using UpSampling2D since the paper mentioned that it uses max-pooling indices for upsampling.

Hi. I get the following error: AttributeError: 'Layer' object has no attribute 'is_placeholder' here: autoencoder.compile(loss="categorical_crossentropy", optimizer='adadelta') Does anyone know how to solve this?

Hi,

I have a question regarding the optimisation algo: why have chosen adadelta instead of SGD?

I trained my data on both and adadelta gave better results!

Thanks for sharing!!

Hello,

When I run this code, it gives me the following error:

MemoryError: Error allocating 442368000 bytes of device memory (out of memory). Apply node that caused the error: GpuElemwise{sub,no_inplace}(GpuElemwise{add,no_inplace}.0, GpuElemwise{Composite{(((i0 / i1) / i2) / i3)},no_inplace}.0) Toposort index: 366 Inputs types: [CudaNdarrayType(float32, 4D), CudaNdarrayType(float32, (True, True, True, False))] Inputs shapes: [(10, 64, 360, 480), (1, 1, 1, 480)] Inputs strides: [(11059200, 172800, 480, 1), (0, 0, 0, 1)] Inputs values: ['not shown', 'not shown'] Outputs clients: [[GpuElemwise{sqr,no_inplace}(GpuElemwise{sub,no_inplace}.0), GpuElemwise{mul,no_inplace}(CudaNdarrayConstant{[[[[ 2.]]]]}, GpuElemwise{Composite{(((i0 / i1) / i2) / i3)},no_inplace}.0, GpuElemwise{sub,no_inplace}.0), GpuElemwise{mul,no_inplace}(GpuElemwise{Composite{(((i0 / i1) / i2) / i3)},no_inplace}.0, GpuElemwise{sub,no_inplace}.0)]]

Tried using 660Ti - went upto 980 and 1080. Doesn't seem to go.

Any solutions???

Also, does this warning have anything to do with it ??

UserWarning: Model inputs must come from a Keras Input layer, they cannot be the output of a previous non-Input layer. Here, a tensor specified as input to "sequential_11_model" was not an Input tensor, it was generated by layer layer_10. Note that input tensors are instantiated via tensor = Input(shape). The tensor that caused the issue was: layer_input_10 str(x.name))