def parse_config_common_sda(self, arguments):
if arguments.has_key('corruption_level'):
self.corruption_levels = [float(arguments['corruption_level']) for n in xrange(100)]
if arguments.has_key('learning_rate'):
self.learning_rates = [float(arguments['learning_rate']) for n in xrange(100)]
if arguments.has_key('batch_size'):
self.batch_size = int(arguments['batch_size'])
if arguments.has_key('epoch_number'):
self.epochs = int(arguments['epoch_number'])
if arguments.has_key('momentum'):
self.momentum = float(arguments['momentum'])
# parse DNN network structure
if arguments.has_key('nnet_spec'):
nnet_layers = arguments['nnet_spec'].split(':')
self.n_ins = int(nnet_layers[0])
self.hidden_layers_sizes = [int(nnet_layers[i]) for i in range(1, len(nnet_layers)-1)]
self.n_outs = int(nnet_layers[-1])
# parse pre-training layer number
self.ptr_layer_number = len(self.hidden_layers_sizes)
if arguments.has_key('ptr_layer_number'):
self.ptr_layer_number = int(arguments['ptr_layer_number'])
# parse activation function
if arguments.has_key('hidden_activation'):
self.hidden_activation = parse_activation(arguments['hidden_activation'])
if arguments['hidden_activation'].startswith('maxout'):
self.do_maxout = True; self.pool_size = int(arguments['hidden_activation'].replace('maxout:',''))
if arguments.has_key('1stlayer_reconstruct_activation'):
self.firstlayer_reconstruct_activation = parse_activation(arguments['1stlayer_reconstruct_activation'])
# parse sparsity setting
if arguments.has_key('sparsity'):
self.sparsity = float(arguments['sparsity'])
if arguments.has_key('sparsity_weight'):
self.sparsity_weight = float(arguments['sparsity_weight'])
# parse various paths for model saving
if arguments.has_key('cfg_output_file'):
self.cfg_output_file = arguments['cfg_output_file']
if arguments.has_key('param_output_file'):
self.param_output_file = arguments['param_output_file']
if arguments.has_key('kaldi_output_file'):
self.kaldi_output_file = arguments['kaldi_output_file']
def parse_config_cnn(self, arguments, nnet_spec, conv_nnet_spec):
self.parse_config_dnn(arguments, nnet_spec)
# parse convolutional layer structure
self.conv_layer_configs = parse_conv_spec(conv_nnet_spec, self.batch_size)
# parse convolutional layer activation
# parse activation function, including maxout
if arguments.has_key('conv_activation'):
self.conv_activation_text = arguments['conv_activation']
self.conv_activation = parse_activation(arguments['conv_activation'])
# maxout not supported yet
# whether we use the fast version of convolution
if arguments.has_key('use_fast'):
self.use_fast = string_2_bool(arguments['use_fast'])
def parse_config_common(self, arguments):
# parse batch_size, momentum and learning rate
self.parse_config_common_sda(arguments)
if arguments.has_key('batch_size'):
self.batch_size = int(arguments['batch_size'])
if arguments.has_key('momentum'):
self.momentum = float(arguments['momentum'])
if arguments.has_key('lrate'):
self.lrate = parse_lrate(arguments['lrate'])
if arguments.has_key('wdir'):
self.wdir = arguments['wdir']
if arguments.has_key('contraction_level'):
self.contraction_level = arguments['contraction_level']
# parse activation function, including maxout
if arguments.has_key('activation'):
self.activation_text = arguments['activation']
self.activation = parse_activation(arguments['activation'])
if arguments['activation'].startswith('maxout'):
self.do_maxout = True
self.pool_size = int(arguments['activation'].replace('maxout:',''))
self.activation_text = 'maxout'
# parse dropout. note that dropout can be applied to the input features only when dropout is also
# applied to hidden-layer outputs at the same time. that is, you cannot apply dropout only to the
# input features
if arguments.has_key('dropout_factor'):
self.do_dropout = True
self.dropout_factor = [float(arguments['dropout_factor']) for n in xrange(20)]
if arguments.has_key('input_dropout_factor'):
self.input_dropout_factor = float(arguments['input_dropout_factor'])
if arguments.has_key('l2_reg'):
self.l2_reg = arguments['l2_reg']
if arguments.has_key('l1_reg'):
self.l1_reg = arguments['l1_reg']
if arguments.has_key('cfg_output_file'):
self.cfg_output_file = arguments['cfg_output_file']
if arguments.has_key('param_output_file'):
self.param_output_file = arguments['param_output_file']
if arguments.has_key('kaldi_output_file'):
self.kaldi_output_file = arguments['kaldi_output_file']
if arguments.has_key('model_save_step'):
self.model_save_step = int(arguments['model_save_step'])
if arguments.has_key('non_updated_layers'):
layers = arguments['non_updated_layers'].split(",")
self.non_updated_layers = [int(layer) for layer in layers]