for batch_index in xrange(valid_sets.cur_frame_num / batch_size): # loop over mini-batches
valid_error.append(valid_fn(index=batch_index))
valid_sets.initialize_read()
log('> epoch %d, lrate %f, validation error %f' % (lrate.epoch, lrate.get_rate(), numpy.mean(valid_error)))
lrate.get_next_rate(current_error = 100 * numpy.mean(valid_error))
# output both iVecNN and DNN
_nnet2file(dnn.ivec_layers, set_layer_num = len(ivec_nnet_layers)-1, filename=wdir + '/ivec.finetune.tmp', withfinal=True)
_nnet2file(dnn.sigmoid_layers, filename=wdir + '/nnet.finetune.tmp')
# determine whether it's BNF based on layer sizes
set_layer_num = -1
withfinal = True
bnf_layer_index = 1
while bnf_layer_index < len(hidden_layers_sizes):
if hidden_layers_sizes[bnf_layer_index] < hidden_layers_sizes[bnf_layer_index - 1]:
break
bnf_layer_index = bnf_layer_index + 1
if bnf_layer_index < len(hidden_layers_sizes): # is bottleneck
set_layer_num = bnf_layer_index+1
withfinal = False
# finally convert the nnet to kaldi or janus format
if do_maxout:
_nnet2kaldi_maxout(nnet_spec, pool_size = pool_size, set_layer_num = set_layer_num, filein = wdir + '/nnet.finetune.tmp', fileout = output_file, withfinal=withfinal)
else:
_nnet2kaldi(nnet_spec, set_layer_num = set_layer_num, filein = wdir + '/nnet.finetune.tmp', fileout = output_file, withfinal=withfinal)
_nnet2kaldi(ivec_nnet_spec, set_layer_num = len(ivec_nnet_layers)-1, filein = wdir + '/ivec.finetune.tmp', fileout = ivec_output_file, withfinal=False)
# determine whether it's BNF based on layer sizes
set_layer_num = -1
withfinal = True
bnf_layer_index = 1
while bnf_layer_index < len(hidden_layers_sizes):
if hidden_layers_sizes[bnf_layer_index] < hidden_layers_sizes[bnf_layer_index - 1]:
break
bnf_layer_index = bnf_layer_index + 1
if bnf_layer_index < len(hidden_layers_sizes): # is bottleneck
set_layer_num = bnf_layer_index+1
withfinal = False
# finally convert the nnet to kaldi or janus format
if output_format == 'kaldi':
if do_maxout:
_nnet2kaldi_maxout(nnet_spec, pool_size = pool_size, set_layer_num = set_layer_num, filein = wdir + '/nnet.finetune.tmp', fileout = output_file, withfinal=withfinal)
else:
_nnet2kaldi(nnet_spec, set_layer_num = set_layer_num, filein = wdir + '/nnet.finetune.tmp', fileout = output_file, withfinal=withfinal)
else: # janus format
if do_maxout:
_nnet2janus_maxout(nnet_spec, pool_size = pool_size, set_layer_num = set_layer_num, filein = wdir + '/nnet.finetune.tmp', fileout = output_file, withfinal=withfinal)
else:
_nnet2janus(nnet_spec, set_layer_num = set_layer_num, filein = wdir + '/nnet.finetune.tmp', fileout = output_file, withfinal=withfinal)
end_time = time.clock()
print >> sys.stderr, ('The code for file ' + os.path.split(__file__)[1] +
' ran for %.2fm' % ((end_time - start_time) / 60.))
valid_error = []
while (not valid_sets.is_finish()):
valid_sets.load_next_partition(valid_xy)
for batch_index in xrange(valid_sets.cur_frame_num / batch_size): # loop over mini-batches
valid_error.append(valid_fn(index=batch_index))
valid_sets.initialize_read()
log('> epoch %d, lrate %f, validation error %f' % (lrate.epoch, lrate.get_rate(), numpy.mean(valid_error)))
lrate.get_next_rate(current_error = 100 * numpy.mean(valid_error))
# output conv layer config
for i in xrange(len(conv_layer_configs)):
conv_layer_configs[i]['activation'] = activation_to_txt(conv_activation)
with open(wdir + '/conv.config.' + str(i), 'wb') as fp:
json.dump(conv_layer_configs[i], fp, indent=2, sort_keys = True)
fp.flush()
# output the conv part
_cnn2file(cnn.layers[0:len(conv_layer_configs)], filename=conv_output_file)
# output the full part
total_layer_number = len(cnn.layers)
_nnet2file(cnn.layers[len(conv_layer_configs):total_layer_number], filename=wdir + '/nnet.finetune.tmp')
_nnet2kaldi(str(cnn.conv_output_dim) + ':' + full_nnet_spec, filein = wdir + '/nnet.finetune.tmp', fileout = full_output_file)
end_time = time.clock()
print >> sys.stderr, ('The code for file ' +
os.path.split(__file__)[1] +
' ran for %.2fm' % ((end_time - start_time) / 60.))
