==== Skipping fold data generation
==== Training master network
# activation: relu
# batchsize: [256, 128]
# bproplen: 20
# datadir: data
# dev_data: fMLLR/fMLLR_dev_data.bin
# dev_ivectors: None
# dev_offsets: TIMIT_dev_offsets.bin
# dev_targets: ./targets_kaldiDBN_dev.bin
# dropout: [0.2]
# epoch: [20]
# fold_data_pattern: data/data_{0}.bin
# fold_network_pattern: Xfolds/fold_{0}.npz
# fold_offset_pattern: data/offsets_{0}.bin
# fold_output_pattern: data_out/data_{0}.bin
# fold_target_pattern: data/targets_{0}.bin
# frequency: -1
# ft: data/fMLLR/final.feature_transform
# gpu: 0
# layers: 4
# lr: [0.01, 0.001, 0.0001, 1e-05]
# network: lstm
# no_progress: True
# optimizer: ['adam', 'momentumsgd']
# out: models_master/lstm_4_1024_fmllr_0
# plot: False
# resume: 
# rpl_model: result_rpl/model
# shuffle_sequences: True
# splice: 0
# tdnn_ksize: [5]
# timedelay: 5
# train_data: fMLLR/fMLLR_train_data.bin
# train_fold: None
# train_ivectors: None
# train_offsets: TIMIT_train_offsets.bin
# train_rpl: False
# train_targets: ./targets_kaldiDBN.bin
# tri: True
# units: [1024]
# use_validation: True

=== Training stage 0: epoch = 20, batchsize = 256, optimizer = adam
epoch       main/loss   validation/main/loss  main/accuracy  validation/main/accuracy  elapsed_time
1           3.49149     2.18525               0.289411       0.43669                   85.609        
2           1.95632     1.79654               0.492047       0.515874                  170.974       
3           1.65348     1.64581               0.555966       0.550124                  254.731       
4           1.41282     1.59436               0.598127       0.567321                  342.106       
5           1.25841     1.55688               0.633758       0.572534                  428.855       
6           1.1723      1.51586               0.658492       0.582785                  514.567       
7           1.03772     1.55724               0.681431       0.580692                  602.493       
=== Training stage 1: epoch = 20, batchsize = 128, optimizer = momentumsgd, learning rate = 0.001
epoch       main/loss   validation/main/loss  main/accuracy  validation/main/accuracy  elapsed_time
1           0.819967    1.39609               0.748077       0.615734                  147.033       
2           0.757419    1.39113               0.764524       0.618737                  303.764       
3           0.744088    1.37088               0.771738       0.620688                  460.487       
4           0.713505    1.38315               0.778751       0.621644                  613.598       
=== Training stage 2: epoch = 20, batchsize = 128, optimizer = momentumsgd, learning rate = 0.0001
epoch       main/loss   validation/main/loss  main/accuracy  validation/main/accuracy  elapsed_time
1           0.692318    1.38076               0.78333        0.622174                  150.387       
2           0.684302    1.37933               0.785812       0.622904                  306.32        
3           0.697334    1.37083               0.78661        0.622615                  464.746       
4           0.684746    1.38471               0.786091       0.622174                  621.763       
=== Training stage 3: epoch = 20, batchsize = 128, optimizer = momentumsgd, learning rate = 1e-05
epoch       main/loss   validation/main/loss  main/accuracy  validation/main/accuracy  elapsed_time
1           0.691076    1.38437               0.785742       0.622631                  152.417       
2           0.684966    1.38445               0.784431       0.622527                  313.539       
3           0.699144    1.36757               0.783632       0.622848                  469.114       
4           0.677947    1.38436               0.788363       0.622415                  628.229       
==== Skipping training folds
==== Skipping predicting training and development data
==== Skipping training RPL layer
==== Evaluating -folds +master -rpl
................................................................................................................................................................................................
Writing results
Elapsed time: 416.602203 s
Loading master network
Calculating network outputs
Writing output files
PER: 15.09 %