#!/usr/bin/env python
__author__ = "Pawel Swietojanski"
import sys, types, tables, numpy, theano
from pylearn2.scripts.pkl_to_pytables import ModelPyTables
from pylearn2.utils import serial
from theano import shared
def create_model(yaml_path):
from pylearn2.models.model import Model
model = serial.load_train_file(yaml_path)
assert isinstance(model, Model)
return model
if __name__=="__main__":
assert len(sys.argv)==3
#hardcoded_layer_names = {"softmax":"y"}
_, yaml_path, pytables_path_to = sys.argv
model = create_model(yaml_path)
params = ModelPyTables.pytables_to_params(pytables_path_to, name="Model")
model.set_params(params)
def main(args=None):
usage = "pool_adaptation.py [options] <si-model-dir> <sa-model-dir> <feats-scp> <targets-pdf> "
parser = OptionParser()
parser.add_option("--adapt-yaml", dest="adapt_yaml", default="",
help="Provide the adaptation yaml template to start with")
parser.add_option("--freeze-means", dest="freeze_means", default=False,
help="Skip means while updating pools")
parser.add_option("--freeze-betas", dest="freeze_betas", default=False,
help="Skip precisions while updating pools")
parser.add_option("--freeze-amp", dest="freeze_amp", default="true",
help="Skip activation function amplitudes while updating pools")
parser.add_option("--freeze-slopes", dest="freeze_slopes", default="true",
help="Skip activation function slopes while updating pools")
parser.add_option("--freeze-layer-ids", dest="freeze_layer_ids", default="",
help="update params only in this layers, i.e. --layer-ids 012")
parser.add_option("--job", dest="JOB", default=0,
help="JOB ID used to store model in")
parser.add_option("--debug", dest="debug", default=False,
help="Prints activations and shapes in text format rather than binary Kaldi archives")
(options,args) = parser.parse_args(args=args)
print options.adapt_yaml
print options.freeze_means
print options.freeze_betas
print options.freeze_amp
print options.freeze_slopes
print options.freeze_layer_ids
print options.JOB
print 'ARGS: ', args
#if options.adapt_yaml!='':
# NotImplementedError('Lodaing from pkl not yet supported due to GPU/CPU pickle issues.')
if len(args) != 5:
print usage
exit(1)
si_model_dir = args[1]
sa_model_dir = args[2]
feats_scp = args[3]
targets_pdf = args[4]
#print "si model dir is %s"%si_model_dir
model_yaml = "%s/adapt_final%s.yaml"%(si_model_dir, options.JOB)
model_params = "%s/cnn_best.h5"%si_model_dir
#print 'Yaml path', model_yaml
#print 'Params path', model_params
if not os.path.isfile(options.adapt_yaml):
raise Exception('File %s not found'%options.adapt_yaml)
if not os.path.isfile(model_params):
raise Exception('File %s not found'%model_params)
vars={}
vars['adapt_flist']=feats_scp
vars['adapt_pdfs']=targets_pdf
vars['adapt_lr']=0.05
vars['adapt_momentum']=0.5
vars['sa_dir'] = sa_model_dir
vars['JOB'] = options.JOB
#print vars
#print 'Locals: ',locals()
adapt_template = open(options.adapt_yaml, 'r').read()
adapt_template_str = adapt_template % vars
f = open(model_yaml, 'w')
f.write(adapt_template_str)
f.close()
print 'Building model %s'%model_yaml
train_obj = serial.load_train_file(model_yaml)
print 'Loading params from %s'%model_params
params = ModelPyTables.pytables_to_params(model_params, name='Model')
train_obj.model.set_params(params)
freeze_regex='softmax_[Wb]|h[0-9]_[Wb]|nlrf_[Wb]'
if options.freeze_layer_ids!='':
layers = options.freeze_layer_ids
freeze_regex = "%s|g[%s]p_u|g[%s]p_beta"%(freeze_regex, layers, layers)
if options.freeze_means == 'true':
freeze_regex = "%s|g[0-9]p_u"%(freeze_regex)
if options.freeze_betas == 'true':
freeze_regex = "%s|g[0-9]p_beta"%(freeze_regex)
if options.freeze_amp == 'true':
freeze_regex = "%s|g[0-9]p_amp"%(freeze_regex)
if options.freeze_slopes == 'true':
freeze_regex = "%s|g[0-9]p_arg"%(freeze_regex)
#print "Freeze regex is", freeze_regex
model_params = train_obj.model.get_params()
params_to_freeze = {}
for param in model_params:
if re.match(freeze_regex, str(param)) is not None:
if param not in params_to_freeze:
params_to_freeze[param] = param
#print params_to_freeze
if len(params_to_freeze)==len(model_params):
print 'None of the parameters were set to be updated. Freeze list is', params_to_freeze
exit(0)
train_obj.model.freeze(params_to_freeze.values())
print 'Will update those params only: ', train_obj.model.get_params()
train_obj.main_loop()
train_obj.model.freeze_set = set([]) #unfreeze so get_params will return all model params
print 'Unfreezed params are ', train_obj.model.get_params()