readme

Author: piiswrong

README.md

@@ -1 +1,12 @@
-# dec
+# Deep Embedded Clustering
+
+This package implements the algorithm described in paper "Unsupervised Deep Embedding for Clustering Analysis". It depends on opencv, numpy, scipy and Caffe.
+
+This implementation is intended for reproducing the results in the paper. If you only want to try the algorithm and find caffe too difficault to install, there is an experimental implementation in MXNet: https://github.com/dmlc/mxnet/blob/master/example/dec/dec.py. MXNet is a flexible deep learning library with fewer dependencies. You are welcome to try it. Installation guide can be found here: https://mxnet.readthedocs.org/en/latest/build.html. Once you install MXNet, simple go into directory examples/dec and run ```python dec.py'''.
+
+## Usage
+To run, please first build our custom version of Caffe included in this package following the official guide: http://caffe.berkeleyvision.org/installation.html. 
+
+Then download the data set you want to experiment on. We provide scripts for downloading the datasets used in the paper. For example you can download MNIST by ```cd mnist; ./get_data.sh'''. Once download completes, run ```cd dec; python make_mnist_data.py''' to prepare data for Caffe.
+
+After data is ready, run ```python dec.py DB''' to run experiment on with DB. DB can be one of mnist, stl, reutersidf10k, reutersidf. We provide pretrained autoencoder weights with this package. You can use dec/pretrain.py to train your own autoencoder. Please read source for usage info.

dec/dec.py

@@ -327,7 +327,6 @@ def make_net(fnet, layers):
 
     fnet.write('name: "net"\n')
     for k,v in layers:
-        print k,v
         fnet.write(layer_dict[k].format(*v))
     fnet.close()
 
@@ -748,14 +747,16 @@ def DisKmeans(db, update_interval = None):
 
 if __name__ == '__main__':
     db = sys.argv[1]
+    lam = 160
     """acc_list = []
     nmi_list = []
     for i in xrange(0,9):
-      acc, nmi = DisKmeans(db, 10*(2**i))
+      lam = 10*(2**i)
+      acc, nmi = DisKmeans(db, lam)
       acc_list.append(acc)
       nmi_list.append(nmi)
     print acc_list
     print nmi_list"""
-    DisKmeans(db, 20)
+    DisKmeans(db, lam)
 
 

dec/pretrain.py

@@ -72,7 +72,7 @@ def main(db, params):
 snapshot_prefix: "exp/{4}/save"
 snapshot_after_train:true
 solver_mode: GPU
-debug_info: true
+debug_info: false 
 device_id: 0""".format(rate, params['step'][0], niter, params['decay'][0],db ))
 
 
@@ -138,7 +138,7 @@ def pretrain_main(db, params):
 snapshot_prefix: "exp/{4}/save"
 snapshot_after_train:true
 solver_mode: GPU
-debug_info: true
+debug_info: false 
 device_id: 0""".format(rate, params['step'][0], params['pt_iter'][0], params['decay'][0], db))
 
         if i > 0:
@@ -160,7 +160,6 @@ def pretrain_main(db, params):
         del net
 
     net = caffe.Net('pt_net.prototxt', 'exp/'+db+'/save_iter_%d.caffemodel'%params['pt_iter'][0])
-    print net.params.keys()
     for i in xrange(n_layer):
         if i == 0:
             k = 'd_data'
@@ -176,10 +175,11 @@ def pretrain_main(db, params):
 
 if __name__ == '__main__':
     db = 'mnist'
+    input_dim = 784
     #dec.make_mnist_data()
-    print main(db, {'n_layer':[4], 'dim': [784, 500, 500, 2000, 10],
+    print main(db, {'n_layer':[4], 'dim': [input_dim, 500, 500, 2000, 10],
                'drop': [0.0], 'rate': [0.1], 'step': [20000], 'iter':[100000], 'decay': [0.0000]})
-    print pretrain_main(db, {'dim': [784, 500, 500, 2000, 10], 'pt_iter': [50000],
+    print pretrain_main(db, {'dim': [input_dim, 500, 500, 2000, 10], 'pt_iter': [50000],
               'drop': [0.2], 'rate': [0.1], 'step': [20000], 'iter':[100000], 'decay': [0.0000]})
     os.system("caffe train --solver=ft_solver.prototxt --weights=stack_init_final.caffemodel")