{ "id": 902, "title": "Benchmark minibatch k-means, step by step", "url": "/competitions/SemiSupervisedFeatureLearning/discussion/902", "author_name": "", "post_date": "2011-09-30T02:18:53.333Z", "votes": null, "comment_count": 2, "views": 4582, "content": "

I've just posted another benchmark, using the minibatch k-means from sofia-ml.  The cluster centers were learned on the combination of the unlabeled dataset and the training data set.  This benchmark used larger minibatch sizes and more iterations than the\r\n "example entry benchmark", and took about 25 minutes to train the cluster centers on a normal laptop rather than about 60 seconds.

\r\n

Since there had been some questions previously, I thought it might be helpful to give the full set of commands used to produce this benchmark:

\r\n
# Step 1.  Combine the unlabled data set with the training data.
\r\n
cat ../competition_data/unlabeled_data.svmlight.dat  ../competition_data/public_train_data.svmlight.dat > concatenated_data.dat
\r\n
 
\r\n
# Step 2.  Learn the cluster centers.
\r\n
$HOME/sofia-ml/sofia-kmeans \\
  --k 100 \\
  --opt_type mini_batch_kmeans \\
  --dimensionality 1000001 \\
  --training_file concatenated_data.dat \\
  --model_out full_kmeans_model.txt \\
  --iterations 10000 \\
  --mini_batch_size 1000 \\
  --objective_after_init \\
  --objective_after_training \\

\r\n
# Step 3.  Apply the learned cluster centers to the training data.
$HOME/sofia-ml/sofia-kmeans \\
  --model_in full_kmeans_model.txt \\
  --test_file public_train_data.svmlight.dat \\
  --objective_on_test \\
  --cluster_mapping_out full_kmeans.train.dat \\
  --cluster_mapping_type rbf_kernel \\
  --cluster_mapping_param 0.01 \\

# Step 4.  Apply the learned cluster centers to the test data.
$HOME/sofia-ml/sofia-kmeans \\
  --model_in full_kmeans_model.txt \\
  --test_filepublic_test_data.svmlight.dat \\
  --objective_on_test \\
  --cluster_mapping_out full_kmeans.test.dat \\
  --cluster_mapping_type rbf_kernel \\
  --cluster_mapping_param 0.01 \\
\r\n \r\n
# Step 5.  Create dense CSV format versions of the new data sets.
\r\n
./svmlightToDenseFormat.pl full_kmeans.train.dat > full_kmeans.train.dense.dat
\r\n
./svmlightToDenseFormat.pl full_kmeans.test.dat > full_kmeans.test.dense.dat
\r\n \r\n
# Step 6.  Execute the ./runLeaderboardEval.pl script.
\r\n
./runLeaderboardEval.pl full_kmeans.train.dense.dat ../competition_data/public_train.labels.dat full_kmeans.test.dense.dat /Users/dsculley/libsvm-3.1/ test.full_kmeans.out
", "messages": [ { "id": "5752", "postDate": "09/30/2011 02:18:53", "content": "

I've just posted another benchmark, using the minibatch k-means from sofia-ml.  The cluster centers were learned on the combination of the unlabeled dataset and the training data set.  This benchmark used larger minibatch sizes and more iterations than the\r\n "example entry benchmark", and took about 25 minutes to train the cluster centers on a normal laptop rather than about 60 seconds.

\r\n

Since there had been some questions previously, I thought it might be helpful to give the full set of commands used to produce this benchmark:

\r\n
# Step 1.  Combine the unlabled data set with the training data.
\r\n
cat ../competition_data/unlabeled_data.svmlight.dat  ../competition_data/public_train_data.svmlight.dat > concatenated_data.dat
\r\n
 
\r\n
# Step 2.  Learn the cluster centers.
\r\n
$HOME/sofia-ml/sofia-kmeans \\
  --k 100 \\
  --opt_type mini_batch_kmeans \\
  --dimensionality 1000001 \\
  --training_file concatenated_data.dat \\
  --model_out full_kmeans_model.txt \\
  --iterations 10000 \\
  --mini_batch_size 1000 \\
  --objective_after_init \\
  --objective_after_training \\

\r\n
# Step 3.  Apply the learned cluster centers to the training data.
$HOME/sofia-ml/sofia-kmeans \\
  --model_in full_kmeans_model.txt \\
  --test_file public_train_data.svmlight.dat \\
  --objective_on_test \\
  --cluster_mapping_out full_kmeans.train.dat \\
  --cluster_mapping_type rbf_kernel \\
  --cluster_mapping_param 0.01 \\

# Step 4.  Apply the learned cluster centers to the test data.
$HOME/sofia-ml/sofia-kmeans \\
  --model_in full_kmeans_model.txt \\
  --test_filepublic_test_data.svmlight.dat \\
  --objective_on_test \\
  --cluster_mapping_out full_kmeans.test.dat \\
  --cluster_mapping_type rbf_kernel \\
  --cluster_mapping_param 0.01 \\
\r\n \r\n
# Step 5.  Create dense CSV format versions of the new data sets.
\r\n
./svmlightToDenseFormat.pl full_kmeans.train.dat > full_kmeans.train.dense.dat
\r\n
./svmlightToDenseFormat.pl full_kmeans.test.dat > full_kmeans.test.dense.dat
\r\n \r\n
# Step 6.  Execute the ./runLeaderboardEval.pl script.
\r\n
./runLeaderboardEval.pl full_kmeans.train.dense.dat ../competition_data/public_train.labels.dat full_kmeans.test.dense.dat /Users/dsculley/libsvm-3.1/ test.full_kmeans.out
", "rawMarkdown": "", "votes": null }, { "id": "5960", "postDate": "10/08/2011 12:06:51", "content": "

For step 5 above, where is svmlightToDenseFormat.pl? (I only see denseFormatToSvmLight.pl in the competition download.)  I recreated it, but including it in the download or posting it might be helpful. Thanks!

", "rawMarkdown": "", "votes": null }, { "id": "5962", "postDate": "10/08/2011 12:25:40", "content": "

Oops, sorry not to have included that one.  Here it is, attached.

", "rawMarkdown": "", "votes": null } ], "comments": [ { "id": 5960, "author_name": "chefele", "author_url": "", "post_date": "10/08/2011 12:06:51", "content": "

For step 5 above, where is svmlightToDenseFormat.pl? (I only see denseFormatToSvmLight.pl in the competition download.)  I recreated it, but including it in the download or posting it might be helpful. Thanks!

", "votes": null, "replies": [] }, { "id": 5962, "author_name": "argv", "author_url": "", "post_date": "10/08/2011 12:25:40", "content": "

Oops, sorry not to have included that one.  Here it is, attached.

", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "5752": "", "5960": "", "5962": "" }, "source": "meta" }