fferroni/dec-keras

Deep Embedding Clustering in Keras

DEC-Keras

Deep Embedding Clustering in Keras [https://arxiv.org/abs/1511.06335]

Similar to https://github.com/dmlc/mxnet/blob/master/example/dec/dec.py, but using Theano/Tensorflow backend rather than MXNET.

dec-keras's Issues

DEC in TensorFlow 1.3; AttributeError: can't set attribute

Thank you for posting the code! I attempted to implement your code in TF 1.3 making the following changes:

import tensorflow.contrib.keras.api.keras as K
from tensorflow.contrib.keras.api.keras.initializers import RandomNormal
from tensorflow.contrib.keras.api.keras.layers import Layer, InputSpec
from tensorflow.contrib.keras.api.keras.models import Model, Sequential
from tensorflow.contrib.keras.api.keras.layers import Dense, Dropout, Input
from tensorflow.contrib.keras.api.keras.optimizers import SGD
from tensorflow.contrib.keras.api.keras.callbacks import LearningRateScheduler
from tensorflow.contrib.keras.api.keras.backend import floatx

However, I encounter en error message pointing to line#71, which is the last line in the following:

def build(self, input_shape):
        assert len(input_shape) == 2
        input_dim = input_shape[1]
        self.input_spec = [InputSpec(dtype=K.backend.floatx(),
                                     shape=(None, input_dim))]
        
        self.W = K.backend.variable(self.initial_weights) 
        self.trainable_weights = [self.W]

layerwise pretrain and Finetuning autoencoder works fine on the test example (mnist), but then Initializing cluster centres with k-means. crushes with an error message:

self.trainable_weights = K.backend.variable(self.initial_weights)

AttributeError: can't set attribute

What could be the reason for that??? Thanks in advance.

Wrong ending condition

Hi, thanks for sharing your code. It is very helpful. However, I thought I found an error in keras_dec.py:

                delta_label = ((y_pred == self.y_pred).sum().astype(np.float32) / y_pred.shape[0])
                if y is not None:
                    acc = self.cluster_acc(y, y_pred)[0]
                    self.accuracy.append(acc)
                    print('Iteration '+str(iteration)+', Accuracy '+str(np.round(acc, 5)))
                else:
                    print(str(np.round(delta_label*100, 5))+'% change in label assignment')

                if delta_label < tol:
                    print('Reached tolerance threshold. Stopping training.')
                    train = False
                    np.save("cluster_loss", np.asarray(cluster_loss))
                    continue

It should be != instead of == when it comes to delta_label, isn't it?

from DEC import ClusteringLayer

Hi
I got this error when I try to run the script python 3.4.
ImportError: No module named 'DEC'
Can you help me?
thanks

The loss didn't converge to zero during fine-tunning in my own data.

DEC-Keras network is so wonderful for my data.

I changed the input as my data instead of mnist data.

The pre_trainning is so good.

But the loss value diverge to NaN during fine tunning.

I think that dynamic learning rate will be needed in fine tunning !

Could you suggest me the best method for applying dynamic learning rate.

Thank you in advance.

DeprecationWarning: linear_assignment_ module is deprecated

/usr/local/lib/python3.6/dist-packages/sklearn/utils/linear_assignment_.py:21: DeprecationWarning: The linear_assignment_ module is deprecated in 0.21 and will be removed from 0.23. Use scipy.optimize.linear_sum_assignment instead.
  DeprecationWarning)

new dataset

Hi,
Great work! Thanks a lot.
I'm just wondering how I can use my own dataset here. I have a folder full of images that I want to be clustered.

Thanks!

visualize

how can we visualize the clusters?

loss: nan

HI,

i have this value.
Any ideas?

Normalized Mutual Information (NMI) and Generalizability (G)

Are there any plans to implement NMI and G metrics as outlined in the paper?

Looking at the original Caffe version of DEC, NMI is defined using sklearn.metrics.normalized_mutual_info_score. https://github.com/piiswrong/dec/blob/master/dec/dec.py#L701

As for G, it should be a straightforward division of training loss and validation loss.

I plan to use use this to play around with some clustering. If/when I do I'm happy to take a stab at it. Or if you have time to work on it before then that would be great too.

Getting a 'initial_value must be specified' ValueError

Trying to load model for production predictions with the code:

from keras_dec import ClusteringLayer
from keras.models import load_model
model = load_model('./DEC.model', custom_objects={'ClusteringLayer': ClusteringLayer})
...

This is not working and producing:
ValueError: initial_value must be specified.
This error is referring to line #64 in keras_dec.py
What is the issue?

Multi GPU Support

This is a great implementation of DEC, would it be possible to add support for using multiple GPUs via Keras?

What is the accuracy obtained with this Keras implementation?

Hello @fferroni. Thank you for sharing this implementation.

With the MXNET model, I can get a 82% clustering accuracy on MNIST but get stuck around 60% with this Keras model out of the box?

Could you tell be if you manage to reach a higher accuracy? Thanks.

ValueError: too many values to unpack (expected 2)

Hi, thanks for your code. While runing the code it happend a error.

Do anyone know hoe to solve it ?
Thanks a lot

How to get the cluster label for new data X

After the cluster is formed, how to know which cluster does the new data belong to?
And also how to retrieve all images which rely on the same cluster?

Question: Unknown number of clusters?

Do you know if there have been any usage with an unknown number of clusters? Thanks for posting this btw! It makes a lot more sense looking at the code than interpreting math symbols for me.

How to insert the dataset?

the problem is : No such file or directory: './data/usps/usps_train.jf'