def CheckArgs(args):
if not os.path.exists(args.config_dir):
os.makedirs(args.config_dir)
## Check arguments.
if args.feat_dir is not None:
args.feat_dim = common_lib.get_feat_dim(args.feat_dir)
if args.ali_dir is not None:
args.num_targets = common_lib.get_number_of_leaves_from_tree(args.ali_dir)
elif args.tree_dir is not None:
args.num_targets = common_lib.get_number_of_leaves_from_tree(args.tree_dir)
if args.ivector_dir is not None:
args.ivector_dim = common_lib.get_ivector_dim(args.ivector_dir)
if not args.feat_dim > 0:
raise Exception("feat-dim has to be postive")
if not args.num_targets > 0:
print(args.num_targets)
raise Exception("num_targets has to be positive")
if not args.ivector_dim >= 0:
raise Exception("ivector-dim has to be non-negative")
if (
not args.max_change_per_component >= 0
or not args.max_change_per_component_final >= 0
):
raise Exception(
"max-change-per-component and max_change-per-component-final should be non-negative"
)
if args.num_lstm_layers < 1:
sys.exit("--num-lstm-layers has to be a positive integer")
if args.clipping_threshold < 0 or args.zeroing_threshold < 0:
sys.exit("--clipping-threshold and --zeroing-threshold have to be non-negative")
if not args.zeroing_interval > 0:
raise Exception("--zeroing-interval has to be positive")
if args.lstm_delay is None:
args.lstm_delay = [[-1]] * args.num_lstm_layers
else:
try:
args.lstm_delay = ParseLstmDelayString(args.lstm_delay.strip())
except ValueError:
sys.exit(
"--lstm-delay has incorrect format value. Provided value is '{0}'".format(
args.lstm_delay
)
)
if len(args.lstm_delay) != args.num_lstm_layers:
sys.exit(
"--lstm-delay: Number of delays provided has to match --num-lstm-layers"
)
return args
def CheckArgs(args):
if not os.path.exists(args.config_dir):
os.makedirs(args.config_dir)
## Check arguments.
if args.feat_dir is not None:
args.feat_dim = common_lib.get_feat_dim(args.feat_dir)
if args.ali_dir is not None:
args.num_targets = common_lib.get_number_of_leaves_from_tree(args.ali_dir)
elif args.tree_dir is not None:
args.num_targets = common_lib.get_number_of_leaves_from_tree(args.tree_dir)
if args.ivector_dir is not None:
args.ivector_dim = common_lib.get_ivector_dim(args.ivector_dir)
if not args.feat_dim > 0:
raise Exception("feat-dim has to be postive")
if not args.num_targets > 0:
print(args.num_targets)
raise Exception("num_targets has to be positive")
if not args.ivector_dim >= 0:
raise Exception("ivector-dim has to be non-negative")
if not args.max_change_per_component >= 0 or not args.max_change_per_component_final >= 0:
raise Exception("max-change-per-component and max_change-per-component-final should be non-negative")
if (args.num_lstm_layers < 1):
sys.exit("--num-lstm-layers has to be a positive integer")
if (args.clipping_threshold < 0 or args.zeroing_threshold < 0):
sys.exit("--clipping-threshold and --zeroing-threshold have to be non-negative")
if not args.zeroing_interval > 0:
raise Exception("--zeroing-interval has to be positive")
if args.lstm_delay is None:
args.lstm_delay = [[-1]] * args.num_lstm_layers
else:
try:
args.lstm_delay = ParseLstmDelayString(args.lstm_delay.strip())
except ValueError:
sys.exit("--lstm-delay has incorrect format value. Provided value is '{0}'".format(args.lstm_delay))
if len(args.lstm_delay) != args.num_lstm_layers:
sys.exit("--lstm-delay: Number of delays provided has to match --num-lstm-layers")
return args
def train(args, run_opts):
""" The main function for training.
Args:
args: a Namespace object with the required parameters
obtained from the function process_args()
run_opts: RunOpts object obtained from the process_args()
"""
arg_string = pprint.pformat(vars(args))
logger.info("Arguments for the experiment\n{0}".format(arg_string))
# Check files
chain_lib.check_for_required_files(args.feat_dir, args.tree_dir,
args.lat_dir)
# Copy phones.txt from tree-dir to dir. Later, steps/nnet3/decode.sh will
# use it to check compatibility between training and decoding phone-sets.
shutil.copy('{0}/phones.txt'.format(args.tree_dir), args.dir)
# Set some variables.
num_jobs = common_lib.get_number_of_jobs(args.tree_dir)
feat_dim = common_lib.get_feat_dim(args.feat_dir)
ivector_dim = common_lib.get_ivector_dim(args.online_ivector_dir)
ivector_id = common_lib.get_ivector_extractor_id(args.online_ivector_dir)
# split the training data into parts for individual jobs
# we will use the same number of jobs as that used for alignment
common_lib.execute_command("utils/split_data.sh {0} {1}"
"".format(args.feat_dir, num_jobs))
with open('{0}/num_jobs'.format(args.dir), 'w') as f:
f.write(str(num_jobs))
if args.input_model is None:
config_dir = '{0}/configs'.format(args.dir)
var_file = '{0}/vars'.format(config_dir)
variables = common_train_lib.parse_generic_config_vars_file(var_file)
else:
# If args.input_model is specified, the model left and right contexts
# are computed using input_model.
variables = common_train_lib.get_input_model_info(args.input_model)
# Set some variables.
try:
model_left_context = variables['model_left_context']
model_right_context = variables['model_right_context']
except KeyError as e:
raise Exception("KeyError {0}: Variables need to be defined in "
"{1}".format(str(e), '{0}/configs'.format(args.dir)))
left_context = args.chunk_left_context + model_left_context
right_context = args.chunk_right_context + model_right_context
left_context_initial = (args.chunk_left_context_initial + model_left_context if
args.chunk_left_context_initial >= 0 else -1)
right_context_final = (args.chunk_right_context_final + model_right_context if
args.chunk_right_context_final >= 0 else -1)
# Initialize as "raw" nnet, prior to training the LDA-like preconditioning
# matrix. This first config just does any initial splicing that we do;
# we do this as it's a convenient way to get the stats for the 'lda-like'
# transform.
if (args.stage <= -6):
logger.info("Creating phone language-model")
chain_lib.create_phone_lm(args.dir, args.tree_dir, run_opts,
lm_opts=args.lm_opts)
if (args.stage <= -5):
logger.info("Creating denominator FST")
shutil.copy('{0}/tree'.format(args.tree_dir), args.dir)
chain_lib.create_denominator_fst(args.dir, args.tree_dir, run_opts)
if ((args.stage <= -4) and
os.path.exists("{0}/configs/init.config".format(args.dir))
and (args.input_model is None)):
logger.info("Initializing a basic network for estimating "
"preconditioning matrix")
common_lib.execute_command(
"""{command} {dir}/log/nnet_init.log \
nnet3-init --srand=-2 {dir}/configs/init.config \
{dir}/init.raw""".format(command=run_opts.command,
dir=args.dir))
egs_left_context = left_context + args.frame_subsampling_factor / 2
egs_right_context = right_context + args.frame_subsampling_factor / 2
# note: the '+ args.frame_subsampling_factor / 2' is to allow for the
# fact that we'll be shifting the data slightly during training to give
# variety to the training data.
egs_left_context_initial = (left_context_initial +
args.frame_subsampling_factor / 2 if
left_context_initial >= 0 else -1)
egs_right_context_final = (right_context_final +
args.frame_subsampling_factor / 2 if
right_context_final >= 0 else -1)
default_egs_dir = '{0}/egs'.format(args.dir)
if ((args.stage <= -3) and args.egs_dir is None):
logger.info("Generating egs")
if (not os.path.exists("{0}/den.fst".format(args.dir)) or
not os.path.exists("{0}/normalization.fst".format(args.dir)) or
not os.path.exists("{0}/tree".format(args.dir))):
raise Exception("Chain egs generation expects {0}/den.fst, "
"{0}/normalization.fst and {0}/tree "
"to exist.".format(args.dir))
# this is where get_egs.sh is called.
chain_lib.generate_chain_egs(
dir=args.dir, data=args.feat_dir,
lat_dir=args.lat_dir, egs_dir=default_egs_dir,
left_context=egs_left_context,
right_context=egs_right_context,
left_context_initial=egs_left_context_initial,
right_context_final=egs_right_context_final,
run_opts=run_opts,
left_tolerance=args.left_tolerance,
right_tolerance=args.right_tolerance,
frame_subsampling_factor=args.frame_subsampling_factor,
alignment_subsampling_factor=args.alignment_subsampling_factor,
frames_per_eg_str=args.chunk_width,
srand=args.srand,
egs_opts=args.egs_opts,
cmvn_opts=args.cmvn_opts,
online_ivector_dir=args.online_ivector_dir,
frames_per_iter=args.frames_per_iter,
transform_dir=args.transform_dir,
stage=args.egs_stage)
if args.egs_dir is None:
egs_dir = default_egs_dir
else:
egs_dir = args.egs_dir
[egs_left_context, egs_right_context,
frames_per_eg_str, num_archives] = (
common_train_lib.verify_egs_dir(egs_dir, feat_dim,
ivector_dim, ivector_id,
egs_left_context, egs_right_context,
egs_left_context_initial,
egs_right_context_final))
assert(args.chunk_width == frames_per_eg_str)
num_archives_expanded = num_archives * args.frame_subsampling_factor
if (args.num_jobs_final > num_archives_expanded):
raise Exception('num_jobs_final cannot exceed the '
'expanded number of archives')
# copy the properties of the egs to dir for
# use during decoding
logger.info("Copying the properties from {0} to {1}".format(egs_dir, args.dir))
common_train_lib.copy_egs_properties_to_exp_dir(egs_dir, args.dir)
if ((args.stage <= -2) and (os.path.exists(args.dir+"/configs/init.config"))
and (args.input_model is None)):
logger.info('Computing the preconditioning matrix for input features')
chain_lib.compute_preconditioning_matrix(
args.dir, egs_dir, num_archives, run_opts,
max_lda_jobs=args.max_lda_jobs,
rand_prune=args.rand_prune)
if (args.stage <= -1):
logger.info("Preparing the initial acoustic model.")
chain_lib.prepare_initial_acoustic_model(args.dir, run_opts,
input_model=args.input_model)
with open("{0}/frame_subsampling_factor".format(args.dir), "w") as f:
f.write(str(args.frame_subsampling_factor))
# set num_iters so that as close as possible, we process the data
# $num_epochs times, i.e. $num_iters*$avg_num_jobs) ==
# $num_epochs*$num_archives, where
# avg_num_jobs=(num_jobs_initial+num_jobs_final)/2.
num_archives_to_process = int(args.num_epochs * num_archives_expanded)
num_archives_processed = 0
num_iters = ((num_archives_to_process * 2)
/ (args.num_jobs_initial + args.num_jobs_final))
# If do_final_combination is True, compute the set of models_to_combine.
# Otherwise, models_to_combine will be none.
if args.do_final_combination:
models_to_combine = common_train_lib.get_model_combine_iters(
num_iters, args.num_epochs,
num_archives_expanded, args.max_models_combine,
args.num_jobs_final)
else:
models_to_combine = None
min_deriv_time = None
max_deriv_time_relative = None
if args.deriv_truncate_margin is not None:
min_deriv_time = -args.deriv_truncate_margin - model_left_context
max_deriv_time_relative = \
args.deriv_truncate_margin + model_right_context
logger.info("Training will run for {0} epochs = "
"{1} iterations".format(args.num_epochs, num_iters))
for iter in range(num_iters):
if (args.exit_stage is not None) and (iter == args.exit_stage):
logger.info("Exiting early due to --exit-stage {0}".format(iter))
return
current_num_jobs = int(0.5 + args.num_jobs_initial
+ (args.num_jobs_final - args.num_jobs_initial)
* float(iter) / num_iters)
if args.stage <= iter:
model_file = "{dir}/{iter}.mdl".format(dir=args.dir, iter=iter)
lrate = common_train_lib.get_learning_rate(iter, current_num_jobs,
num_iters,
num_archives_processed,
num_archives_to_process,
args.initial_effective_lrate,
args.final_effective_lrate)
shrinkage_value = 1.0 - (args.proportional_shrink * lrate)
if shrinkage_value <= 0.5:
raise Exception("proportional-shrink={0} is too large, it gives "
"shrink-value={1}".format(args.proportional_shrink,
shrinkage_value))
if args.shrink_value < shrinkage_value:
shrinkage_value = (args.shrink_value
if common_train_lib.should_do_shrinkage(
iter, model_file,
args.shrink_saturation_threshold)
else shrinkage_value)
percent = num_archives_processed * 100.0 / num_archives_to_process
epoch = (num_archives_processed * args.num_epochs
/ num_archives_to_process)
shrink_info_str = ''
if shrinkage_value != 1.0:
shrink_info_str = 'shrink: {0:0.5f}'.format(shrinkage_value)
logger.info("Iter: {0}/{1} "
"Epoch: {2:0.2f}/{3:0.1f} ({4:0.1f}% complete) "
"lr: {5:0.6f} {6}".format(iter, num_iters - 1,
epoch, args.num_epochs,
percent,
lrate, shrink_info_str))
chain_lib.train_one_iteration(
dir=args.dir,
iter=iter,
srand=args.srand,
egs_dir=egs_dir,
num_jobs=current_num_jobs,
num_archives_processed=num_archives_processed,
num_archives=num_archives,
learning_rate=lrate,
dropout_edit_string=common_train_lib.get_dropout_edit_string(
args.dropout_schedule,
float(num_archives_processed) / num_archives_to_process,
iter),
train_opts=' '.join(args.train_opts),
shrinkage_value=shrinkage_value,
num_chunk_per_minibatch_str=args.num_chunk_per_minibatch,
apply_deriv_weights=args.apply_deriv_weights,
min_deriv_time=min_deriv_time,
max_deriv_time_relative=max_deriv_time_relative,
l2_regularize=args.l2_regularize,
xent_regularize=args.xent_regularize,
leaky_hmm_coefficient=args.leaky_hmm_coefficient,
momentum=args.momentum,
max_param_change=args.max_param_change,
shuffle_buffer_size=args.shuffle_buffer_size,
frame_subsampling_factor=args.frame_subsampling_factor,
run_opts=run_opts,
backstitch_training_scale=args.backstitch_training_scale,
backstitch_training_interval=args.backstitch_training_interval)
if args.cleanup:
# do a clean up everything but the last 2 models, under certain
# conditions
common_train_lib.remove_model(
args.dir, iter-2, num_iters, models_to_combine,
args.preserve_model_interval)
if args.email is not None:
reporting_iter_interval = num_iters * args.reporting_interval
if iter % reporting_iter_interval == 0:
# lets do some reporting
[report, times, data] = (
nnet3_log_parse.generate_acc_logprob_report(
args.dir, "log-probability"))
message = report
subject = ("Update : Expt {dir} : "
"Iter {iter}".format(dir=args.dir, iter=iter))
common_lib.send_mail(message, subject, args.email)
num_archives_processed = num_archives_processed + current_num_jobs
if args.stage <= num_iters:
if args.do_final_combination:
logger.info("Doing final combination to produce final.mdl")
chain_lib.combine_models(
dir=args.dir, num_iters=num_iters,
models_to_combine=models_to_combine,
num_chunk_per_minibatch_str=args.num_chunk_per_minibatch,
egs_dir=egs_dir,
leaky_hmm_coefficient=args.leaky_hmm_coefficient,
l2_regularize=args.l2_regularize,
xent_regularize=args.xent_regularize,
run_opts=run_opts,
max_objective_evaluations=args.max_objective_evaluations)
else:
logger.info("Copying the last-numbered model to final.mdl")
common_lib.force_symlink("{0}.mdl".format(num_iters),
"{0}/final.mdl".format(args.dir))
common_lib.force_symlink("compute_prob_valid.{iter}.log"
"".format(iter=num_iters-1),
"{dir}/log/compute_prob_valid.final.log".format(
dir=args.dir))
if args.cleanup:
logger.info("Cleaning up the experiment directory "
"{0}".format(args.dir))
remove_egs = args.remove_egs
if args.egs_dir is not None:
# this egs_dir was not created by this experiment so we will not
# delete it
remove_egs = False
# leave the last-two-numbered models, for diagnostic reasons.
common_train_lib.clean_nnet_dir(
args.dir, num_iters - 1, egs_dir,
preserve_model_interval=args.preserve_model_interval,
remove_egs=remove_egs)
# do some reporting
[report, times, data] = nnet3_log_parse.generate_acc_logprob_report(
args.dir, "log-probability")
if args.email is not None:
common_lib.send_mail(report, "Update : Expt {0} : "
"complete".format(args.dir), args.email)
with open("{dir}/accuracy.report".format(dir=args.dir), "w") as f:
f.write(report)
common_lib.execute_command("steps/info/chain_dir_info.pl "
"{0}".format(args.dir))
def train(args, run_opts, background_process_handler):
""" The main function for training.
Args:
args: a Namespace object with the required parameters
obtained from the function process_args()
run_opts: RunOpts object obtained from the process_args()
"""
arg_string = pprint.pformat(vars(args))
logger.info("Arguments for the experiment\n{0}".format(arg_string))
# Set some variables.
feat_dim = common_lib.get_feat_dim(args.feat_dir)
ivector_dim = common_lib.get_ivector_dim(args.online_ivector_dir)
ivector_id = common_lib.get_ivector_extractor_id(args.online_ivector_dir)
config_dir = '{0}/configs'.format(args.dir)
var_file = '{0}/vars'.format(config_dir)
variables = common_train_lib.parse_generic_config_vars_file(var_file)
# Set some variables.
try:
model_left_context = variables['model_left_context']
model_right_context = variables['model_right_context']
# this is really the number of times we add layers to the network for
# discriminative pretraining
num_hidden_layers = variables['num_hidden_layers']
add_lda = common_lib.str_to_bool(variables['add_lda'])
include_log_softmax = common_lib.str_to_bool(
variables['include_log_softmax'])
except KeyError as e:
raise Exception("KeyError {0}: Variables need to be defined in "
"{1}".format(str(e), '{0}/configs'.format(args.dir)))
left_context = args.chunk_left_context + model_left_context
right_context = args.chunk_right_context + model_right_context
left_context_initial = (args.chunk_left_context_initial + model_left_context if
args.chunk_left_context_initial >= 0 else -1)
right_context_final = (args.chunk_right_context_final + model_right_context if
args.chunk_right_context_final >= 0 else -1)
# Initialize as "raw" nnet, prior to training the LDA-like preconditioning
# matrix. This first config just does any initial splicing that we do;
# we do this as it's a convenient way to get the stats for the 'lda-like'
# transform.
if (args.stage <= -4):
logger.info("Initializing a basic network")
common_lib.run_job(
"""{command} {dir}/log/nnet_init.log \
nnet3-init --srand=-2 {dir}/configs/init.config \
{dir}/init.raw""".format(command=run_opts.command,
dir=args.dir))
default_egs_dir = '{0}/egs'.format(args.dir)
if (args.stage <= -3) and args.egs_dir is None:
logger.info("Generating egs")
if args.use_dense_targets:
target_type = "dense"
try:
num_targets = int(variables['num_targets'])
if (common_lib.get_feat_dim_from_scp(args.targets_scp)
!= num_targets):
raise Exception("Mismatch between num-targets provided to "
"script vs configs")
except KeyError as e:
num_targets = -1
else:
target_type = "sparse"
try:
num_targets = int(variables['num_targets'])
except KeyError as e:
raise Exception("KeyError {0}: Variables need to be defined "
"in {1}".format(
str(e), '{0}/configs'.format(args.dir)))
train_lib.raw_model.generate_egs_using_targets(
data=args.feat_dir, targets_scp=args.targets_scp,
egs_dir=default_egs_dir,
left_context=left_context,
right_context=right_context,
left_context_initial=left_context_initial,
right_context_final=right_context_final,
run_opts=run_opts,
frames_per_eg_str=args.chunk_width,
srand=args.srand,
egs_opts=args.egs_opts,
cmvn_opts=args.cmvn_opts,
online_ivector_dir=args.online_ivector_dir,
samples_per_iter=args.samples_per_iter,
transform_dir=args.transform_dir,
stage=args.egs_stage,
target_type=target_type,
num_targets=num_targets)
if args.egs_dir is None:
egs_dir = default_egs_dir
else:
egs_dir = args.egs_dir
[egs_left_context, egs_right_context,
frames_per_eg_str, num_archives] = (
common_train_lib.verify_egs_dir(egs_dir, feat_dim,
ivector_dim, ivector_id,
left_context, right_context))
if args.chunk_width != frames_per_eg_str:
raise Exception("mismatch between --egs.chunk-width and the frames_per_eg "
"in the egs dir {0} vs {1}".format(args.chunk_width,
frames_per_eg_str))
if (args.num_jobs_final > num_archives):
raise Exception('num_jobs_final cannot exceed the number of archives '
'in the egs directory')
# copy the properties of the egs to dir for
# use during decoding
common_train_lib.copy_egs_properties_to_exp_dir(egs_dir, args.dir)
if (add_lda and args.stage <= -2):
logger.info('Computing the preconditioning matrix for input features')
train_lib.common.compute_preconditioning_matrix(
args.dir, egs_dir, num_archives, run_opts,
max_lda_jobs=args.max_lda_jobs,
rand_prune=args.rand_prune)
if (args.stage <= -1):
logger.info("Preparing the initial network.")
common_train_lib.prepare_initial_network(args.dir, run_opts)
# set num_iters so that as close as possible, we process the data
# $num_epochs times, i.e. $num_iters*$avg_num_jobs) ==
# $num_epochs*$num_archives, where
# avg_num_jobs=(num_jobs_initial+num_jobs_final)/2.
num_archives_to_process = int(args.num_epochs * num_archives)
num_archives_processed = 0
num_iters = ((num_archives_to_process * 2)
/ (args.num_jobs_initial + args.num_jobs_final))
models_to_combine = common_train_lib.verify_iterations(
num_iters, args.num_epochs,
num_hidden_layers, num_archives,
args.max_models_combine, args.add_layers_period,
args.num_jobs_final)
def learning_rate(iter, current_num_jobs, num_archives_processed):
return common_train_lib.get_learning_rate(iter, current_num_jobs,
num_iters,
num_archives_processed,
num_archives_to_process,
args.initial_effective_lrate,
args.final_effective_lrate)
min_deriv_time = None
max_deriv_time_relative = None
if args.deriv_truncate_margin is not None:
min_deriv_time = -args.deriv_truncate_margin - model_left_context
max_deriv_time_relative = \
args.deriv_truncate_margin + model_right_context
logger.info("Training will run for {0} epochs = "
"{1} iterations".format(args.num_epochs, num_iters))
for iter in range(num_iters):
if (args.exit_stage is not None) and (iter == args.exit_stage):
logger.info("Exiting early due to --exit-stage {0}".format(iter))
return
current_num_jobs = int(0.5 + args.num_jobs_initial
+ (args.num_jobs_final - args.num_jobs_initial)
* float(iter) / num_iters)
if args.stage <= iter:
model_file = "{dir}/{iter}.raw".format(dir=args.dir, iter=iter)
shrinkage_value = 1.0
if args.shrink_value != 1.0:
shrinkage_value = (args.shrink_value
if common_train_lib.do_shrinkage(
iter, model_file,
args.shrink_saturation_threshold,
get_raw_nnet_from_am=False)
else 1
)
train_lib.common.train_one_iteration(
dir=args.dir,
iter=iter,
srand=args.srand,
egs_dir=egs_dir,
num_jobs=current_num_jobs,
num_archives_processed=num_archives_processed,
num_archives=num_archives,
learning_rate=learning_rate(iter, current_num_jobs,
num_archives_processed),
dropout_edit_string=common_train_lib.get_dropout_edit_string(
args.dropout_schedule,
float(num_archives_processed) / num_archives_to_process,
iter),
shrinkage_value=shrinkage_value,
minibatch_size_str=args.num_chunk_per_minibatch,
num_hidden_layers=num_hidden_layers,
add_layers_period=args.add_layers_period,
left_context=left_context,
right_context=right_context,
min_deriv_time=min_deriv_time,
max_deriv_time_relative=max_deriv_time_relative,
momentum=args.momentum,
max_param_change=args.max_param_change,
shuffle_buffer_size=args.shuffle_buffer_size,
run_opts=run_opts,
get_raw_nnet_from_am=False,
background_process_handler=background_process_handler)
if args.cleanup:
# do a clean up everythin but the last 2 models, under certain
# conditions
common_train_lib.remove_model(
args.dir, iter-2, num_iters, models_to_combine,
args.preserve_model_interval,
get_raw_nnet_from_am=False)
if args.email is not None:
reporting_iter_interval = num_iters * args.reporting_interval
if iter % reporting_iter_interval == 0:
# lets do some reporting
[report, times, data] = (
nnet3_log_parse.generate_acc_logprob_report(args.dir))
message = report
subject = ("Update : Expt {dir} : "
"Iter {iter}".format(dir=args.dir, iter=iter))
common_lib.send_mail(message, subject, args.email)
num_archives_processed = num_archives_processed + current_num_jobs
if args.stage <= num_iters:
logger.info("Doing final combination to produce final.raw")
train_lib.common.combine_models(
dir=args.dir, num_iters=num_iters,
models_to_combine=models_to_combine, egs_dir=egs_dir,
left_context=left_context, right_context=right_context,
minibatch_size_str=args.num_chunk_per_minibatch,
run_opts=run_opts, chunk_width=args.chunk_width,
background_process_handler=background_process_handler,
get_raw_nnet_from_am=False,
sum_to_one_penalty=args.combine_sum_to_one_penalty)
if include_log_softmax and args.stage <= num_iters + 1:
logger.info("Getting average posterior for purposes of "
"adjusting the priors.")
train_lib.common.compute_average_posterior(
dir=args.dir, iter='final', egs_dir=egs_dir,
num_archives=num_archives,
left_context=left_context, right_context=right_context,
prior_subset_size=args.prior_subset_size, run_opts=run_opts,
get_raw_nnet_from_am=False)
if args.cleanup:
logger.info("Cleaning up the experiment directory "
"{0}".format(args.dir))
remove_egs = args.remove_egs
if args.egs_dir is not None:
# this egs_dir was not created by this experiment so we will not
# delete it
remove_egs = False
common_train_lib.clean_nnet_dir(
nnet_dir=args.dir, num_iters=num_iters, egs_dir=egs_dir,
preserve_model_interval=args.preserve_model_interval,
remove_egs=remove_egs,
get_raw_nnet_from_am=False)
# do some reporting
[report, times, data] = nnet3_log_parse.generate_acc_logprob_report(args.dir)
if args.email is not None:
common_lib.send_mail(report, "Update : Expt {0} : "
"complete".format(args.dir), args.email)
with open("{dir}/accuracy.report".format(dir=args.dir), "w") as f:
f.write(report)
common_lib.run_job("steps/info/nnet3_dir_info.pl "
"{0}".format(args.dir))
def train(args, run_opts):
""" The main function for training.
Args:
args: a Namespace object with the required parameters
obtained from the function process_args()
run_opts: RunOpts object obtained from the process_args()
"""
arg_string = pprint.pformat(vars(args))
logger.info("Arguments for the experiment\n{0}".format(arg_string))
# Check files
files = ['{0}/feats.scp'.format(args.feat_dir), '{0}/fst.1.scp'.format(args.tree_dir),
'{0}/final.mdl'.format(args.tree_dir), '{0}/tree'.format(args.tree_dir),
'{0}/phone_lm.fst'.format(args.tree_dir),
'{0}/num_jobs'.format(args.tree_dir)]
for file in files:
if not os.path.isfile(file):
raise Exception('Expected {0} to exist.'.format(file))
# Set some variables.
num_jobs = common_lib.get_number_of_jobs(args.tree_dir)
feat_dim = common_lib.get_feat_dim(args.feat_dir)
ivector_dim = common_lib.get_ivector_dim(args.online_ivector_dir)
ivector_id = common_lib.get_ivector_extractor_id(args.online_ivector_dir)
logger.info("feat-dim: {}, ivector-dim: {}".format(feat_dim, ivector_dim))
# split the training data into parts for individual jobs
# we will use the same number of jobs as that used for compiling FSTs
common_lib.execute_command("utils/split_data.sh {0} {1}".format(
args.feat_dir, num_jobs))
shutil.copy('{0}/tree'.format(args.tree_dir), args.dir)
shutil.copy('{0}/phones.txt'.format(args.tree_dir), args.dir)
shutil.copy('{0}/phone_lm.fst'.format(args.tree_dir), args.dir)
shutil.copy('{0}/0.trans_mdl'.format(args.tree_dir), args.dir)
with open('{0}/num_jobs'.format(args.dir), 'w') as f:
f.write(str(num_jobs))
config_dir = '{0}/configs'.format(args.dir)
var_file = '{0}/vars'.format(config_dir)
variables = common_train_lib.parse_generic_config_vars_file(var_file)
# Set some variables.
try:
model_left_context = variables['model_left_context']
model_right_context = variables['model_right_context']
except KeyError as e:
raise Exception("KeyError {0}: Variables need to be defined in "
"{1}".format(str(e), '{0}/configs'.format(args.dir)))
left_context = args.chunk_left_context + model_left_context
right_context = args.chunk_right_context + model_right_context
left_context_initial = (args.chunk_left_context_initial + model_left_context if
args.chunk_left_context_initial >= 0 else -1)
right_context_final = (args.chunk_right_context_final + model_right_context if
args.chunk_right_context_final >= 0 else -1)
# Initialize as "raw" nnet, prior to training the LDA-like preconditioning
# matrix. This first config just does any initial splicing that we do;
# we do this as it's a convenient way to get the stats for the 'lda-like'
# transform.
if (args.stage <= -5):
logger.info("Creating denominator FST")
chain_lib.create_denominator_fst(args.dir, args.tree_dir, run_opts)
if (args.stage <= -4):
logger.info("Initializing a basic network...")
common_lib.execute_command(
"""{command} {dir}/log/nnet_init.log \
nnet3-init --srand=-2 {dir}/configs/final.config \
{dir}/init.raw""".format(command=run_opts.command,
dir=args.dir))
egs_left_context = left_context + args.frame_subsampling_factor / 2
egs_right_context = right_context + args.frame_subsampling_factor / 2
egs_left_context_initial = (left_context_initial + args.frame_subsampling_factor / 2 if
left_context_initial >= 0 else -1)
egs_right_context_final = (right_context_final + args.frame_subsampling_factor / 2 if
right_context_final >= 0 else -1)
default_egs_dir = '{0}/egs'.format(args.dir)
if (args.stage <= -3) and args.egs_dir is None:
logger.info("Generating end-to-end egs...")
common_lib.execute_command(
"""steps/nnet3/chain/e2e/get_egs_e2e.sh {egs_opts} \
--cmd "{command}" \
--cmvn-opts "{cmvn_opts}" \
--online-ivector-dir "{ivector_dir}" \
--left-context {left_context} \
--right-context {right_context} \
--left-context-initial {left_context_initial} \
--right-context-final {right_context_final} \
--frame-subsampling-factor {frame_subsampling_factor} \
--stage {stage} \
--frames-per-iter {frames_per_iter} \
--srand {srand} \
{data} {dir} {fst_dir} {egs_dir}""".format(
command=run_opts.command,
cmvn_opts=args.cmvn_opts if args.cmvn_opts is not None else '',
ivector_dir=(args.online_ivector_dir
if args.online_ivector_dir is not None
else ''),
left_context=egs_left_context,
right_context=egs_right_context,
left_context_initial=egs_left_context_initial,
right_context_final=egs_right_context_final,
frame_subsampling_factor=args.frame_subsampling_factor,
stage=args.egs_stage, frames_per_iter=args.frames_per_iter,
srand=args.srand,
data=args.feat_dir, dir=args.dir, fst_dir=args.tree_dir,
egs_dir=default_egs_dir,
egs_opts=args.egs_opts if args.egs_opts is not None else ''))
if args.egs_dir is None:
egs_dir = default_egs_dir
else:
egs_dir = args.egs_dir
[egs_left_context, egs_right_context,
frames_per_eg_str, num_archives] = (
common_train_lib.verify_egs_dir(egs_dir, feat_dim,
ivector_dim, ivector_id,
egs_left_context, egs_right_context,
egs_left_context_initial,
egs_right_context_final))
num_archives_expanded = num_archives * args.frame_subsampling_factor
if (args.num_jobs_final > num_archives_expanded):
raise Exception('num_jobs_final cannot exceed the '
'expanded number of archives')
# copy the properties of the egs to dir for
# use during decoding
logger.info("Copying the properties from {0} to {1}".format(egs_dir, args.dir))
common_train_lib.copy_egs_properties_to_exp_dir(egs_dir, args.dir)
if (args.stage <= -1):
logger.info("Preparing the initial acoustic model.")
chain_lib.prepare_initial_acoustic_model(args.dir, run_opts)
with open("{0}/frame_subsampling_factor".format(args.dir), "w") as f:
f.write(str(args.frame_subsampling_factor))
# set num_iters so that as close as possible, we process the data
# $num_epochs times, i.e. $num_iters*$avg_num_jobs) ==
# $num_epochs*$num_archives, where
# avg_num_jobs=(num_jobs_initial+num_jobs_final)/2.
num_archives_to_process = int(args.num_epochs * num_archives_expanded)
num_archives_processed = 0
num_iters = ((num_archives_to_process * 2)
/ (args.num_jobs_initial + args.num_jobs_final))
models_to_combine = common_train_lib.get_model_combine_iters(
num_iters, args.num_epochs,
num_archives_expanded, args.max_models_combine,
args.num_jobs_final)
min_deriv_time = None
max_deriv_time_relative = None
if args.deriv_truncate_margin is not None:
min_deriv_time = -args.deriv_truncate_margin - model_left_context
max_deriv_time_relative = \
args.deriv_truncate_margin + model_right_context
logger.info("Training will run for {0} epochs = "
"{1} iterations".format(args.num_epochs, num_iters))
for iter in range(num_iters):
percent = num_archives_processed * 100.0 / num_archives_to_process
epoch = (num_archives_processed * args.num_epochs
/ num_archives_to_process)
if (args.exit_stage is not None) and (iter == args.exit_stage):
logger.info("Exiting early due to --exit-stage {0}".format(iter))
return
current_num_jobs = int(0.5 + args.num_jobs_initial
+ (args.num_jobs_final - args.num_jobs_initial)
* float(iter) / num_iters)
if args.stage <= iter:
model_file = "{dir}/{iter}.mdl".format(dir=args.dir, iter=iter)
lrate = common_train_lib.get_learning_rate(iter, current_num_jobs,
num_iters,
num_archives_processed,
num_archives_to_process,
args.initial_effective_lrate,
args.final_effective_lrate)
shrinkage_value = 1.0 - (args.proportional_shrink * lrate)
if shrinkage_value <= 0.5:
raise Exception("proportional-shrink={0} is too large, it gives "
"shrink-value={1}".format(args.proportional_shrink,
shrinkage_value))
if args.shrink_value < shrinkage_value:
shrinkage_value = (args.shrink_value
if common_train_lib.should_do_shrinkage(
iter, model_file,
args.shrink_saturation_threshold)
else shrinkage_value)
shrink_info_str = ''
if shrinkage_value != 1.0:
shrink_info_str = 'shrink: {0:0.5f}'.format(shrinkage_value)
logger.info("Iter: {0}/{1} "
"Epoch: {2:0.2f}/{3:0.1f} ({4:0.1f}% complete) "
"lr: {5:0.6f} {6}".format(iter, num_iters - 1,
epoch, args.num_epochs,
percent,
lrate, shrink_info_str))
chain_lib.train_one_iteration(
dir=args.dir,
iter=iter,
srand=args.srand,
egs_dir=egs_dir,
num_jobs=current_num_jobs,
num_archives_processed=num_archives_processed,
num_archives=num_archives,
learning_rate=lrate,
dropout_edit_string=common_train_lib.get_dropout_edit_string(
args.dropout_schedule,
float(num_archives_processed) / num_archives_to_process,
iter),
shrinkage_value=shrinkage_value,
num_chunk_per_minibatch_str=args.num_chunk_per_minibatch,
apply_deriv_weights=args.apply_deriv_weights,
min_deriv_time=min_deriv_time,
max_deriv_time_relative=max_deriv_time_relative,
l2_regularize=args.l2_regularize,
xent_regularize=args.xent_regularize,
leaky_hmm_coefficient=args.leaky_hmm_coefficient,
momentum=args.momentum,
max_param_change=args.max_param_change,
shuffle_buffer_size=args.shuffle_buffer_size,
frame_subsampling_factor=args.frame_subsampling_factor,
run_opts=run_opts)
if args.cleanup:
# do a clean up everything but the last 2 models, under certain
# conditions
common_train_lib.remove_model(
args.dir, iter-2, num_iters, models_to_combine,
args.preserve_model_interval)
if args.email is not None:
reporting_iter_interval = num_iters * args.reporting_interval
if iter % reporting_iter_interval == 0:
# lets do some reporting
[report, times, data] = (
nnet3_log_parse.generate_acc_logprob_report(
args.dir, "log-probability"))
message = report
subject = ("Update : Expt {dir} : "
"Iter {iter}".format(dir=args.dir, iter=iter))
common_lib.send_mail(message, subject, args.email)
num_archives_processed = num_archives_processed + current_num_jobs
if args.stage <= num_iters:
logger.info("Doing final combination to produce final.mdl")
chain_lib.combine_models(
dir=args.dir, num_iters=num_iters,
models_to_combine=models_to_combine,
num_chunk_per_minibatch_str=args.num_chunk_per_minibatch,
egs_dir=egs_dir,
leaky_hmm_coefficient=args.leaky_hmm_coefficient,
l2_regularize=args.l2_regularize,
xent_regularize=args.xent_regularize,
run_opts=run_opts)
if args.cleanup:
logger.info("Cleaning up the experiment directory "
"{0}".format(args.dir))
remove_egs = args.remove_egs
if args.egs_dir is not None:
# this egs_dir was not created by this experiment so we will not
# delete it
remove_egs = False
common_train_lib.clean_nnet_dir(
args.dir, num_iters, egs_dir,
preserve_model_interval=args.preserve_model_interval,
remove_egs=remove_egs)
# do some reporting
[report, times, data] = nnet3_log_parse.generate_acc_logprob_report(
args.dir, "log-probability")
if args.email is not None:
common_lib.send_mail(report, "Update : Expt {0} : "
"complete".format(args.dir), args.email)
with open("{dir}/accuracy.report".format(dir=args.dir), "w") as f:
f.write(report)
common_lib.execute_command("steps/info/chain_dir_info.pl "
"{0}".format(args.dir))
def train(args, run_opts):
""" The main function for training.
Args:
args: a Namespace object with the required parameters
obtained from the function process_args()
run_opts: RunOpts object obtained from the process_args()
"""
arg_string = pprint.pformat(vars(args))
logger.info("Arguments for the experiment\n{0}".format(arg_string))
# Copy phones.txt from ali-dir to dir. Later, steps/nnet3/decode.sh will
# use it to check compatibility between training and decoding phone-sets.
shutil.copy('{0}/phones.txt'.format(args.ali_dir), args.dir)
# Set some variables.
num_jobs = common_lib.get_number_of_jobs(args.ali_dir)
feat_dim = common_lib.get_feat_dim(args.feat_dir)
ivector_dim = common_lib.get_ivector_dim(args.online_ivector_dir)
ivector_id = common_lib.get_ivector_extractor_id(args.online_ivector_dir)
# split the training data into parts for individual jobs
# we will use the same number of jobs as that used for alignment
common_lib.execute_command("utils/split_data.sh {0} {1}".format(
args.feat_dir, num_jobs))
shutil.copy('{0}/tree'.format(args.ali_dir), args.dir)
with open('{0}/num_jobs'.format(args.dir), 'w') as f:
f.write('{}'.format(num_jobs))
config_dir = '{0}/configs'.format(args.dir)
var_file = '{0}/vars'.format(config_dir)
if args.input_model is None:
config_dir = '{0}/configs'.format(args.dir)
var_file = '{0}/vars'.format(config_dir)
variables = common_train_lib.parse_generic_config_vars_file(var_file)
else:
# If args.input_model is specified, the model left and right contexts
# are computed using input_model.
variables = common_train_lib.get_input_model_info(args.input_model)
# Set some variables.
try:
model_left_context = variables['model_left_context']
model_right_context = variables['model_right_context']
except KeyError as e:
raise Exception("KeyError {0}: Variables need to be defined in "
"{1}".format(str(e), '{0}/configs'.format(args.dir)))
left_context = args.chunk_left_context + model_left_context
right_context = args.chunk_right_context + model_right_context
left_context_initial = (args.chunk_left_context_initial + model_left_context if
args.chunk_left_context_initial >= 0 else -1)
right_context_final = (args.chunk_right_context_final + model_right_context if
args.chunk_right_context_final >= 0 else -1)
# Initialize as "raw" nnet, prior to training the LDA-like preconditioning
# matrix. This first config just does any initial splicing that we do;
# we do this as it's a convenient way to get the stats for the 'lda-like'
# transform.
if (args.stage <= -5) and (args.input_model is None):
logger.info("Initializing a basic network for estimating "
"preconditioning matrix")
common_lib.execute_command(
"""{command} {dir}/log/nnet_init.log \
nnet3-init --srand=-2 {dir}/configs/init.config \
{dir}/init.raw""".format(command=run_opts.command,
dir=args.dir))
default_egs_dir = '{0}/egs'.format(args.dir)
if args.stage <= -4 and args.egs_dir is None:
logger.info("Generating egs")
if args.feat_dir is None:
raise Exception("--feat-dir option is required if you don't supply --egs-dir")
train_lib.acoustic_model.generate_egs(
data=args.feat_dir, alidir=args.ali_dir,
egs_dir=default_egs_dir,
left_context=left_context,
right_context=right_context,
left_context_initial=left_context_initial,
right_context_final=right_context_final,
run_opts=run_opts,
frames_per_eg_str=args.chunk_width,
srand=args.srand,
egs_opts=args.egs_opts,
cmvn_opts=args.cmvn_opts,
online_ivector_dir=args.online_ivector_dir,
samples_per_iter=args.samples_per_iter,
stage=args.egs_stage)
if args.egs_dir is None:
egs_dir = default_egs_dir
else:
egs_dir = args.egs_dir
[egs_left_context, egs_right_context,
frames_per_eg_str, num_archives] = (
common_train_lib.verify_egs_dir(egs_dir, feat_dim,
ivector_dim, ivector_id,
left_context, right_context,
left_context_initial, right_context_final))
if args.chunk_width != frames_per_eg_str:
raise Exception("mismatch between --egs.chunk-width and the frames_per_eg "
"in the egs dir {0} vs {1}".format(args.chunk_width,
frames_per_eg_str))
if args.num_jobs_final > num_archives:
raise Exception('num_jobs_final cannot exceed the number of archives '
'in the egs directory')
# copy the properties of the egs to dir for
# use during decoding
common_train_lib.copy_egs_properties_to_exp_dir(egs_dir, args.dir)
if args.stage <= -3 and (args.input_model is None):
logger.info('Computing the preconditioning matrix for input features')
train_lib.common.compute_preconditioning_matrix(
args.dir, egs_dir, num_archives, run_opts,
max_lda_jobs=args.max_lda_jobs,
rand_prune=args.rand_prune)
if args.stage <= -2 and (args.input_model is None):
logger.info("Computing initial vector for FixedScaleComponent before"
" softmax, using priors^{prior_scale} and rescaling to"
" average 1".format(
prior_scale=args.presoftmax_prior_scale_power))
common_train_lib.compute_presoftmax_prior_scale(
args.dir, args.ali_dir, num_jobs, run_opts,
presoftmax_prior_scale_power=args.presoftmax_prior_scale_power)
if args.stage <= -1:
logger.info("Preparing the initial acoustic model.")
train_lib.acoustic_model.prepare_initial_acoustic_model(
args.dir, args.ali_dir, run_opts,
input_model=args.input_model)
# set num_iters so that as close as possible, we process the data
# $num_epochs times, i.e. $num_iters*$avg_num_jobs) ==
# $num_epochs*$num_archives, where
# avg_num_jobs=(num_jobs_initial+num_jobs_final)/2.
num_archives_to_process = int(args.num_epochs * num_archives)
num_archives_processed = 0
num_iters = int((num_archives_to_process * 2) / (args.num_jobs_initial + args.num_jobs_final))
# If do_final_combination is True, compute the set of models_to_combine.
# Otherwise, models_to_combine will be none.
if args.do_final_combination:
models_to_combine = common_train_lib.get_model_combine_iters(
num_iters, args.num_epochs,
num_archives, args.max_models_combine,
args.num_jobs_final)
else:
models_to_combine = None
min_deriv_time = None
max_deriv_time_relative = None
if args.deriv_truncate_margin is not None:
min_deriv_time = -args.deriv_truncate_margin - model_left_context
max_deriv_time_relative = \
args.deriv_truncate_margin + model_right_context
logger.info("Training will run for {0} epochs = "
"{1} iterations".format(args.num_epochs, num_iters))
for iter in range(num_iters):
if (args.exit_stage is not None) and (iter == args.exit_stage):
logger.info("Exiting early due to --exit-stage {0}".format(iter))
return
current_num_jobs = int(0.5 + args.num_jobs_initial
+ (args.num_jobs_final - args.num_jobs_initial)
* float(iter) / num_iters)
if args.stage <= iter:
model_file = "{dir}/{iter}.mdl".format(dir=args.dir, iter=iter)
lrate = common_train_lib.get_learning_rate(iter, current_num_jobs,
num_iters,
num_archives_processed,
num_archives_to_process,
args.initial_effective_lrate,
args.final_effective_lrate)
shrinkage_value = 1.0 - (args.proportional_shrink * lrate)
if shrinkage_value <= 0.5:
raise Exception("proportional-shrink={0} is too large, it gives "
"shrink-value={1}".format(args.proportional_shrink,
shrinkage_value))
if args.shrink_value < shrinkage_value:
shrinkage_value = (args.shrink_value
if common_train_lib.should_do_shrinkage(
iter, model_file,
args.shrink_saturation_threshold) else 1.0)
percent = num_archives_processed * 100.0 / num_archives_to_process
epoch = (num_archives_processed * args.num_epochs
/ num_archives_to_process)
shrink_info_str = ''
if shrinkage_value != 1.0:
shrink_info_str = 'shrink: {0:0.5f}'.format(shrinkage_value)
logger.info("Iter: {0}/{1} "
"Epoch: {2:0.2f}/{3:0.1f} ({4:0.1f}% complete) "
"lr: {5:0.6f} {6}".format(iter, num_iters - 1,
epoch, args.num_epochs,
percent,
lrate, shrink_info_str))
train_lib.common.train_one_iteration(
dir=args.dir,
iter=iter,
srand=args.srand,
egs_dir=egs_dir,
num_jobs=current_num_jobs,
num_archives_processed=num_archives_processed,
num_archives=num_archives,
learning_rate=lrate,
dropout_edit_string=common_train_lib.get_dropout_edit_string(
args.dropout_schedule,
float(num_archives_processed) / num_archives_to_process,
iter),
train_opts=' '.join(args.train_opts),
shrinkage_value=shrinkage_value,
minibatch_size_str=args.num_chunk_per_minibatch,
min_deriv_time=min_deriv_time,
max_deriv_time_relative=max_deriv_time_relative,
momentum=args.momentum,
max_param_change=args.max_param_change,
shuffle_buffer_size=args.shuffle_buffer_size,
run_opts=run_opts,
backstitch_training_scale=args.backstitch_training_scale,
backstitch_training_interval=args.backstitch_training_interval,
compute_per_dim_accuracy=args.compute_per_dim_accuracy)
if args.cleanup:
# do a clean up everythin but the last 2 models, under certain
# conditions
common_train_lib.remove_model(
args.dir, iter-2, num_iters, models_to_combine,
args.preserve_model_interval)
if args.email is not None:
reporting_iter_interval = num_iters * args.reporting_interval
if iter % reporting_iter_interval == 0:
# lets do some reporting
[report, times, data] = (
nnet3_log_parse.generate_acc_logprob_report(args.dir))
message = report
subject = ("Update : Expt {dir} : "
"Iter {iter}".format(dir=args.dir, iter=iter))
common_lib.send_mail(message, subject, args.email)
num_archives_processed = num_archives_processed + current_num_jobs
if args.stage <= num_iters:
if args.do_final_combination:
logger.info("Doing final combination to produce final.mdl")
train_lib.common.combine_models(
dir=args.dir, num_iters=num_iters,
models_to_combine=models_to_combine, egs_dir=egs_dir,
run_opts=run_opts,
minibatch_size_str=args.num_chunk_per_minibatch,
chunk_width=args.chunk_width,
max_objective_evaluations=args.max_objective_evaluations,
compute_per_dim_accuracy=args.compute_per_dim_accuracy)
if args.stage <= num_iters + 1:
logger.info("Getting average posterior for purposes of "
"adjusting the priors.")
# If args.do_final_combination is true, we will use the combined model.
# Otherwise, we will use the last_numbered model.
real_iter = 'combined' if args.do_final_combination else num_iters
avg_post_vec_file = train_lib.common.compute_average_posterior(
dir=args.dir, iter=real_iter, egs_dir=egs_dir,
num_archives=num_archives,
prior_subset_size=args.prior_subset_size, run_opts=run_opts)
logger.info("Re-adjusting priors based on computed posteriors")
combined_or_last_numbered_model = "{dir}/{iter}.mdl".format(dir=args.dir,
iter=real_iter)
final_model = "{dir}/final.mdl".format(dir=args.dir)
train_lib.common.adjust_am_priors(args.dir, combined_or_last_numbered_model,
avg_post_vec_file, final_model,
run_opts)
if args.cleanup:
logger.info("Cleaning up the experiment directory "
"{0}".format(args.dir))
remove_egs = args.remove_egs
if args.egs_dir is not None:
# this egs_dir was not created by this experiment so we will not
# delete it
remove_egs = False
common_train_lib.clean_nnet_dir(
nnet_dir=args.dir, num_iters=num_iters, egs_dir=egs_dir,
preserve_model_interval=args.preserve_model_interval,
remove_egs=remove_egs)
# do some reporting
[report, times, data] = nnet3_log_parse.generate_acc_logprob_report(args.dir)
if args.email is not None:
common_lib.send_mail(report, "Update : Expt {0} : "
"complete".format(args.dir), args.email)
with open("{dir}/accuracy.report".format(dir=args.dir), "w") as f:
f.write(report)
common_lib.execute_command("steps/info/nnet3_dir_info.pl "
"{0}".format(args.dir))
def egs_prepare(args, run_opts):
""" The main function for egs_generate.
Args:
args: a Namespace object with the required parameters
obtained from the function process_args()
run_opts: RunOpts object obtained from the process_args()
"""
arg_string = pprint.pformat(vars(args))
logger.info("Arguments for the experiment\n{0}".format(arg_string))
# Check files
chain_lib.check_for_required_files(args.feat_dir, args.tree_dir,
args.lat_dir)
# Set some variables.
num_jobs = common_lib.get_number_of_jobs(args.tree_dir)
feat_dim = common_lib.get_feat_dim(args.feat_dir)
ivector_dim = common_lib.get_ivector_dim(args.online_ivector_dir)
ivector_id = common_lib.get_ivector_extractor_id(args.online_ivector_dir)
# split the training data into parts for individual jobs
# we will use the same number of jobs as that used for alignment
common_lib.execute_command("utils/split_data.sh {0} {1}"
"".format(args.feat_dir, num_jobs))
with open('{0}/num_jobs'.format(args.dir), 'w') as f:
f.write(str(num_jobs))
if args.input_model is None:
config_dir = '{0}/configs'.format(args.dir)
var_file = '{0}/vars'.format(config_dir)
variables = common_train_lib.parse_generic_config_vars_file(var_file)
else:
# If args.input_model is specified, the model left and right contexts
# are computed using input_model.
variables = common_train_lib.get_input_model_info(args.input_model)
# Set some variables.
try:
model_left_context = variables['model_left_context']
model_right_context = variables['model_right_context']
except KeyError as e:
raise Exception("KeyError {0}: Variables need to be defined in "
"{1}".format(str(e), '{0}/configs'.format(args.dir)))
left_context = args.chunk_left_context + model_left_context
right_context = args.chunk_right_context + model_right_context
left_context_initial = (args.chunk_left_context_initial + model_left_context if
args.chunk_left_context_initial >= 0 else -1)
right_context_final = (args.chunk_right_context_final + model_right_context if
args.chunk_right_context_final >= 0 else -1)
# Initialize as "raw" nnet, prior to training the LDA-like preconditioning
# matrix. This first config just does any initial splicing that we do;
# we do this as it's a convenient way to get the stats for the 'lda-like'
# transform.
if (args.stage <= -6):
logger.info("Creating phone language-model")
chain_lib.create_phone_lm(args.dir, args.tree_dir, run_opts,
lm_opts=args.lm_opts)
if (args.stage <= -5):
logger.info("Creating denominator FST")
shutil.copy('{0}/tree'.format(args.tree_dir), args.dir)
chain_lib.create_denominator_fst(args.dir, args.tree_dir, run_opts)
if ((args.stage <= -4) and
os.path.exists("{0}/configs/init.config".format(args.dir))
and (args.input_model is None)):
logger.info("Initializing a basic network for estimating "
"preconditioning matrix")
common_lib.execute_command(
"""{command} {dir}/log/nnet_init.log \
nnet3-init --srand=-2 {dir}/configs/init.config \
{dir}/init.raw""".format(command=run_opts.command,
dir=args.dir))
egs_left_context = left_context + args.frame_subsampling_factor / 2
egs_right_context = right_context + args.frame_subsampling_factor / 2
# note: the '+ args.frame_subsampling_factor / 2' is to allow for the
# fact that we'll be shifting the data slightly during training to give
# variety to the training data.
egs_left_context_initial = (left_context_initial +
args.frame_subsampling_factor / 2 if
left_context_initial >= 0 else -1)
egs_right_context_final = (right_context_final +
args.frame_subsampling_factor / 2 if
right_context_final >= 0 else -1)
def train(args, run_opts):
""" The main function for training.
Args:
args: a Namespace object with the required parameters
obtained from the function process_args()
run_opts: RunOpts object obtained from the process_args()
"""
arg_string = pprint.pformat(vars(args))
logger.info("Arguments for the experiment\n{0}".format(arg_string))
# Set some variables.
feat_dim = common_lib.get_feat_dim(args.feat_dir)
ivector_dim = common_lib.get_ivector_dim(args.online_ivector_dir)
ivector_id = common_lib.get_ivector_extractor_id(args.online_ivector_dir)
if args.input_model is None:
config_dir = '{0}/configs'.format(args.dir)
var_file = '{0}/vars'.format(config_dir)
variables = common_train_lib.parse_generic_config_vars_file(var_file)
else:
# If args.input_model is specified, the model left and right contexts
# are computed using input_model.
variables = common_train_lib.get_input_model_info(args.input_model)
# Set some variables.
try:
model_left_context = variables['model_left_context']
model_right_context = variables['model_right_context']
except KeyError as e:
raise Exception("KeyError {0}: Variables need to be defined in "
"{1}".format(str(e), '{0}/configs'.format(args.dir)))
left_context = args.chunk_left_context + model_left_context
right_context = args.chunk_right_context + model_right_context
left_context_initial = (args.chunk_left_context_initial + model_left_context if
args.chunk_left_context_initial >= 0 else -1)
right_context_final = (args.chunk_right_context_final + model_right_context if
args.chunk_right_context_final >= 0 else -1)
# Initialize as "raw" nnet, prior to training the LDA-like preconditioning
# matrix. This first config just does any initial splicing that we do;
# we do this as it's a convenient way to get the stats for the 'lda-like'
# transform.
if (args.stage <= -4) and os.path.exists(args.dir+"/configs/init.config") and \
(args.input_model is None):
logger.info("Initializing the network for computing the LDA stats")
common_lib.execute_command(
"""{command} {dir}/log/nnet_init.log \
nnet3-init --srand=-2 {dir}/configs/init.config \
{dir}/init.raw""".format(command=run_opts.command,
dir=args.dir))
default_egs_dir = '{0}/egs'.format(args.dir)
if (args.stage <= -3) and args.egs_dir is None:
logger.info("Generating egs")
if args.use_dense_targets:
target_type = "dense"
try:
num_targets = int(variables['num_targets'])
if (common_lib.get_feat_dim_from_scp(args.targets_scp)
!= num_targets):
raise Exception("Mismatch between num-targets provided to "
"script vs configs")
except KeyError as e:
num_targets = -1
else:
target_type = "sparse"
try:
num_targets = int(variables['num_targets'])
except KeyError as e:
raise Exception("KeyError {0}: Variables need to be defined "
"in {1}".format(
str(e), '{0}/configs'.format(args.dir)))
train_lib.raw_model.generate_egs_using_targets(
data=args.feat_dir, targets_scp=args.targets_scp,
egs_dir=default_egs_dir,
left_context=left_context,
right_context=right_context,
left_context_initial=left_context_initial,
right_context_final=right_context_final,
run_opts=run_opts,
frames_per_eg_str=args.chunk_width,
srand=args.srand,
egs_opts=args.egs_opts,
cmvn_opts=args.cmvn_opts,
online_ivector_dir=args.online_ivector_dir,
samples_per_iter=args.samples_per_iter,
stage=args.egs_stage,
target_type=target_type,
num_targets=num_targets)
if args.egs_dir is None:
egs_dir = default_egs_dir
else:
egs_dir = args.egs_dir
[egs_left_context, egs_right_context,
frames_per_eg_str, num_archives] = (
common_train_lib.verify_egs_dir(egs_dir, feat_dim,
ivector_dim, ivector_id,
left_context, right_context,
left_context_initial,
right_context_final))
if args.chunk_width != frames_per_eg_str:
raise Exception("mismatch between --egs.chunk-width and the frames_per_eg "
"in the egs dir {0} vs {1}".format(args.chunk_width,
frames_per_eg_str))
if args.num_jobs_final > num_archives:
raise Exception('num_jobs_final cannot exceed the number of archives '
'in the egs directory')
# copy the properties of the egs to dir for
# use during decoding
common_train_lib.copy_egs_properties_to_exp_dir(egs_dir, args.dir)
if args.stage <= -2 and os.path.exists(args.dir+"/configs/init.config") and \
(args.input_model is None):
logger.info('Computing the preconditioning matrix for input features')
train_lib.common.compute_preconditioning_matrix(
args.dir, egs_dir, num_archives, run_opts,
max_lda_jobs=args.max_lda_jobs,
rand_prune=args.rand_prune)
if args.stage <= -1:
logger.info("Preparing the initial network.")
common_train_lib.prepare_initial_network(args.dir, run_opts, args.srand, args.input_model)
# set num_iters so that as close as possible, we process the data
# $num_epochs times, i.e. $num_iters*$avg_num_jobs) ==
# $num_epochs*$num_archives, where
# avg_num_jobs=(num_jobs_initial+num_jobs_final)/2.
num_archives_to_process = int(args.num_epochs * num_archives)
num_archives_processed = 0
num_iters = int((num_archives_to_process * 2) / (args.num_jobs_initial + args.num_jobs_final))
# If do_final_combination is True, compute the set of models_to_combine.
# Otherwise, models_to_combine will be none.
if args.do_final_combination:
models_to_combine = common_train_lib.get_model_combine_iters(
num_iters, args.num_epochs,
num_archives, args.max_models_combine,
args.num_jobs_final)
else:
models_to_combine = None
if (os.path.exists('{0}/valid_diagnostic.scp'.format(egs_dir))):
if (os.path.exists('{0}/valid_diagnostic.egs'.format(egs_dir))):
raise Exception('both {0}/valid_diagnostic.egs and '
'{0}/valid_diagnostic.scp exist.'
'This script expects only one of them to exist.'
''.format(egs_dir))
use_multitask_egs = True
else:
if (not os.path.exists('{0}/valid_diagnostic.egs'
''.format(egs_dir))):
raise Exception('neither {0}/valid_diagnostic.egs nor '
'{0}/valid_diagnostic.scp exist.'
'This script expects one of them.'
''.format(egs_dir))
use_multitask_egs = False
min_deriv_time = None
max_deriv_time_relative = None
if args.deriv_truncate_margin is not None:
min_deriv_time = -args.deriv_truncate_margin - model_left_context
max_deriv_time_relative = \
args.deriv_truncate_margin + model_right_context
logger.info("Training will run for {0} epochs = "
"{1} iterations".format(args.num_epochs, num_iters))
for iter in range(num_iters):
if (args.exit_stage is not None) and (iter == args.exit_stage):
logger.info("Exiting early due to --exit-stage {0}".format(iter))
return
current_num_jobs = common_train_lib.get_current_num_jobs(
iter, num_iters,
args.num_jobs_initial, args.num_jobs_step, args.num_jobs_final)
if args.stage <= iter:
model_file = "{dir}/{iter}.raw".format(dir=args.dir, iter=iter)
lrate = common_train_lib.get_learning_rate(iter, current_num_jobs,
num_iters,
num_archives_processed,
num_archives_to_process,
args.initial_effective_lrate,
args.final_effective_lrate)
# shrinkage_value is a scale on the parameters.
shrinkage_value = 1.0 - (args.proportional_shrink * lrate)
if shrinkage_value <= 0.5:
raise Exception("proportional-shrink={0} is too large, it gives "
"shrink-value={1}".format(args.proportional_shrink,
shrinkage_value))
if args.shrink_value < shrinkage_value:
shrinkage_value = (args.shrink_value
if common_train_lib.should_do_shrinkage(
iter, model_file,
args.shrink_saturation_threshold,
get_raw_nnet_from_am=False)
else shrinkage_value)
percent = num_archives_processed * 100.0 / num_archives_to_process
epoch = (num_archives_processed * args.num_epochs
/ num_archives_to_process)
shrink_info_str = ''
if shrinkage_value != 1.0:
shrink_info_str = 'shrink: {0:0.5f}'.format(shrinkage_value)
logger.info("Iter: {0}/{1} Jobs: {2} "
"Epoch: {3:0.2f}/{4:0.1f} ({5:0.1f}% complete) "
"lr: {6:0.6f} {7}".format(iter, num_iters - 1,
current_num_jobs,
epoch, args.num_epochs,
percent,
lrate, shrink_info_str))
train_lib.common.train_one_iteration(
dir=args.dir,
iter=iter,
srand=args.srand,
egs_dir=egs_dir,
num_jobs=current_num_jobs,
num_archives_processed=num_archives_processed,
num_archives=num_archives,
learning_rate=lrate,
dropout_edit_string=common_train_lib.get_dropout_edit_string(
args.dropout_schedule,
float(num_archives_processed) / num_archives_to_process,
iter),
train_opts=' '.join(args.train_opts),
shrinkage_value=shrinkage_value,
minibatch_size_str=args.num_chunk_per_minibatch,
min_deriv_time=min_deriv_time,
max_deriv_time_relative=max_deriv_time_relative,
momentum=args.momentum,
max_param_change=args.max_param_change,
shuffle_buffer_size=args.shuffle_buffer_size,
run_opts=run_opts,
get_raw_nnet_from_am=False,
use_multitask_egs=use_multitask_egs,
compute_per_dim_accuracy=args.compute_per_dim_accuracy)
if args.cleanup:
# do a clean up everythin but the last 2 models, under certain
# conditions
common_train_lib.remove_model(
args.dir, iter-2, num_iters, models_to_combine,
args.preserve_model_interval,
get_raw_nnet_from_am=False)
if args.email is not None:
reporting_iter_interval = num_iters * args.reporting_interval
if iter % reporting_iter_interval == 0:
# lets do some reporting
[report, times, data] = (
nnet3_log_parse.generate_acc_logprob_report(args.dir))
message = report
subject = ("Update : Expt {dir} : "
"Iter {iter}".format(dir=args.dir, iter=iter))
common_lib.send_mail(message, subject, args.email)
num_archives_processed = num_archives_processed + current_num_jobs
if args.stage <= num_iters:
if args.do_final_combination:
logger.info("Doing final combination to produce final.raw")
train_lib.common.combine_models(
dir=args.dir, num_iters=num_iters,
models_to_combine=models_to_combine, egs_dir=egs_dir,
minibatch_size_str=args.num_chunk_per_minibatch,
run_opts=run_opts, chunk_width=args.chunk_width,
get_raw_nnet_from_am=False,
compute_per_dim_accuracy=args.compute_per_dim_accuracy,
max_objective_evaluations=args.max_objective_evaluations,
use_multitask_egs=use_multitask_egs)
else:
common_lib.force_symlink("{0}.raw".format(num_iters),
"{0}/final.raw".format(args.dir))
if args.compute_average_posteriors and args.stage <= num_iters + 1:
logger.info("Getting average posterior for purposes of "
"adjusting the priors.")
train_lib.common.compute_average_posterior(
dir=args.dir, iter='final', egs_dir=egs_dir,
num_archives=num_archives,
prior_subset_size=args.prior_subset_size, run_opts=run_opts,
get_raw_nnet_from_am=False)
if args.cleanup:
logger.info("Cleaning up the experiment directory "
"{0}".format(args.dir))
remove_egs = args.remove_egs
if args.egs_dir is not None:
# this egs_dir was not created by this experiment so we will not
# delete it
remove_egs = False
common_train_lib.clean_nnet_dir(
nnet_dir=args.dir, num_iters=num_iters, egs_dir=egs_dir,
preserve_model_interval=args.preserve_model_interval,
remove_egs=remove_egs,
get_raw_nnet_from_am=False)
# do some reporting
[report, times, data] = nnet3_log_parse.generate_acc_logprob_report(args.dir)
if args.email is not None:
common_lib.send_mail(report, "Update : Expt {0} : "
"complete".format(args.dir), args.email)
with open("{dir}/accuracy.report".format(dir=args.dir), "w") as f:
f.write(report)
common_lib.execute_command("subtools/kaldi/steps/info/nnet3_dir_info.pl "
"{0}".format(args.dir))
def train(args, run_opts):
""" The main function for training.
Args:
args: a Namespace object with the required parameters
obtained from the function process_args()
run_opts: RunOpts object obtained from the process_args()
"""
arg_string = pprint.pformat(vars(args))
logger.info("Arguments for the experiment\n{0}".format(arg_string))
# Copy phones.txt from ali-dir to dir. Later, steps/nnet3/decode.sh will
# use it to check compatibility between training and decoding phone-sets.
shutil.copy('{0}/phones.txt'.format(args.ali_dir), args.dir)
# Set some variables.
# num_leaves = common_lib.get_number_of_leaves_from_tree(args.ali_dir)
num_jobs = common_lib.get_number_of_jobs(args.ali_dir)
feat_dim = common_lib.get_feat_dim(args.feat_dir)
ivector_dim = common_lib.get_ivector_dim(args.online_ivector_dir)
ivector_id = common_lib.get_ivector_extractor_id(args.online_ivector_dir)
# split the training data into parts for individual jobs
# we will use the same number of jobs as that used for alignment
common_lib.execute_command("utils/split_data.sh {0} {1}".format(
args.feat_dir, num_jobs))
shutil.copy('{0}/tree'.format(args.ali_dir), args.dir)
with open('{0}/num_jobs'.format(args.dir), 'w') as f:
f.write(str(num_jobs))
if args.input_model is None:
config_dir = '{0}/configs'.format(args.dir)
var_file = '{0}/vars'.format(config_dir)
variables = common_train_lib.parse_generic_config_vars_file(var_file)
else:
# If args.input_model is specified, the model left and right contexts
# are computed using input_model.
variables = common_train_lib.get_input_model_info(args.input_model)
# Set some variables.
try:
model_left_context = variables['model_left_context']
model_right_context = variables['model_right_context']
except KeyError as e:
raise Exception("KeyError {0}: Variables need to be defined in "
"{1}".format(str(e), '{0}/configs'.format(args.dir)))
left_context = model_left_context
right_context = model_right_context
# Initialize as "raw" nnet, prior to training the LDA-like preconditioning
# matrix. This first config just does any initial splicing that we do;
# we do this as it's a convenient way to get the stats for the 'lda-like'
# transform.
if (args.stage <= -5) and os.path.exists(args.dir+"/configs/init.config") and \
(args.input_model is None):
logger.info("Initializing a basic network for estimating "
"preconditioning matrix")
common_lib.execute_command("""{command} {dir}/log/nnet_init.log \
nnet3-init --srand=-2 {dir}/configs/init.config \
{dir}/init.raw""".format(command=run_opts.command,
dir=args.dir))
default_egs_dir = '{0}/egs'.format(args.dir)
if (args.stage <= -4) and args.egs_dir is None:
logger.info("Generating egs")
if args.feat_dir is None:
raise Exception(
"--feat-dir option is required if you don't supply --egs-dir")
train_lib.acoustic_model.generate_egs(
data=args.feat_dir,
alidir=args.ali_dir,
egs_dir=default_egs_dir,
left_context=left_context,
right_context=right_context,
run_opts=run_opts,
frames_per_eg_str=str(args.frames_per_eg),
srand=args.srand,
egs_opts=args.egs_opts,
cmvn_opts=args.cmvn_opts,
online_ivector_dir=args.online_ivector_dir,
samples_per_iter=args.samples_per_iter,
stage=args.egs_stage)
if args.egs_dir is None:
egs_dir = default_egs_dir
else:
egs_dir = args.egs_dir
[egs_left_context, egs_right_context, frames_per_eg_str,
num_archives] = (common_train_lib.verify_egs_dir(egs_dir, feat_dim,
ivector_dim, ivector_id,
left_context,
right_context))
assert str(args.frames_per_eg) == frames_per_eg_str
if args.num_jobs_final > num_archives:
raise Exception('num_jobs_final cannot exceed the number of archives '
'in the egs directory')
# copy the properties of the egs to dir for
# use during decoding
common_train_lib.copy_egs_properties_to_exp_dir(egs_dir, args.dir)
if args.stage <= -3 and os.path.exists(args.dir +
"/configs/init.config") and (
args.input_model is None):
logger.info('Computing the preconditioning matrix for input features')
train_lib.common.compute_preconditioning_matrix(
args.dir,
egs_dir,
num_archives,
run_opts,
max_lda_jobs=args.max_lda_jobs,
rand_prune=args.rand_prune)
if args.stage <= -2 and (args.input_model is None):
logger.info(
"Computing initial vector for FixedScaleComponent before"
" softmax, using priors^{prior_scale} and rescaling to"
" average 1".format(prior_scale=args.presoftmax_prior_scale_power))
common_train_lib.compute_presoftmax_prior_scale(
args.dir,
args.ali_dir,
num_jobs,
run_opts,
presoftmax_prior_scale_power=args.presoftmax_prior_scale_power)
if args.stage <= -1:
logger.info("Preparing the initial acoustic model.")
train_lib.acoustic_model.prepare_initial_acoustic_model(
args.dir, args.ali_dir, run_opts, input_model=args.input_model)
# set num_iters so that as close as possible, we process the data
# $num_epochs times, i.e. $num_iters*$avg_num_jobs) ==
# $num_epochs*$num_archives, where
# avg_num_jobs=(num_jobs_initial+num_jobs_final)/2.
num_archives_expanded = num_archives * args.frames_per_eg
num_archives_to_process = int(args.num_epochs * num_archives_expanded)
num_archives_processed = 0
num_iters = ((num_archives_to_process * 2) /
(args.num_jobs_initial + args.num_jobs_final))
# If do_final_combination is True, compute the set of models_to_combine.
# Otherwise, models_to_combine will be none.
if args.do_final_combination:
models_to_combine = common_train_lib.get_model_combine_iters(
num_iters, args.num_epochs, num_archives_expanded,
args.max_models_combine, args.num_jobs_final)
else:
models_to_combine = None
logger.info("Training will run for {0} epochs = "
"{1} iterations".format(args.num_epochs, num_iters))
for iter in range(num_iters):
if (args.exit_stage is not None) and (iter == args.exit_stage):
logger.info("Exiting early due to --exit-stage {0}".format(iter))
return
current_num_jobs = int(0.5 + args.num_jobs_initial +
(args.num_jobs_final - args.num_jobs_initial) *
float(iter) / num_iters)
if args.stage <= iter:
lrate = common_train_lib.get_learning_rate(
iter, current_num_jobs, num_iters, num_archives_processed,
num_archives_to_process, args.initial_effective_lrate,
args.final_effective_lrate)
shrinkage_value = 1.0 - (args.proportional_shrink * lrate)
if shrinkage_value <= 0.5:
raise Exception(
"proportional-shrink={0} is too large, it gives "
"shrink-value={1}".format(args.proportional_shrink,
shrinkage_value))
percent = num_archives_processed * 100.0 / num_archives_to_process
epoch = (num_archives_processed * args.num_epochs /
num_archives_to_process)
shrink_info_str = ''
if shrinkage_value != 1.0:
shrink_info_str = 'shrink: {0:0.5f}'.format(shrinkage_value)
logger.info("Iter: {0}/{1} "
"Epoch: {2:0.2f}/{3:0.1f} ({4:0.1f}% complete) "
"lr: {5:0.6f} {6}".format(iter, num_iters - 1,
epoch, args.num_epochs,
percent, lrate,
shrink_info_str))
train_lib.common.train_one_iteration(
dir=args.dir,
iter=iter,
srand=args.srand,
egs_dir=egs_dir,
num_jobs=current_num_jobs,
num_archives_processed=num_archives_processed,
num_archives=num_archives,
learning_rate=lrate,
dropout_edit_string=common_train_lib.get_dropout_edit_string(
args.dropout_schedule,
float(num_archives_processed) / num_archives_to_process,
iter),
train_opts=' '.join(args.train_opts),
minibatch_size_str=args.minibatch_size,
frames_per_eg=args.frames_per_eg,
momentum=args.momentum,
max_param_change=args.max_param_change,
shrinkage_value=shrinkage_value,
shuffle_buffer_size=args.shuffle_buffer_size,
run_opts=run_opts)
if args.cleanup:
# do a clean up everythin but the last 2 models, under certain
# conditions
common_train_lib.remove_model(args.dir, iter - 2, num_iters,
models_to_combine,
args.preserve_model_interval)
if args.email is not None:
reporting_iter_interval = num_iters * args.reporting_interval
if iter % reporting_iter_interval == 0:
# lets do some reporting
[report, times,
data] = (nnet3_log_parse.generate_acc_logprob_report(
args.dir))
message = report
subject = ("Update : Expt {dir} : "
"Iter {iter}".format(dir=args.dir, iter=iter))
common_lib.send_mail(message, subject, args.email)
num_archives_processed = num_archives_processed + current_num_jobs
if args.stage <= num_iters:
if args.do_final_combination:
logger.info("Doing final combination to produce final.mdl")
train_lib.common.combine_models(
dir=args.dir,
num_iters=num_iters,
models_to_combine=models_to_combine,
egs_dir=egs_dir,
minibatch_size_str=args.minibatch_size,
run_opts=run_opts,
max_objective_evaluations=args.max_objective_evaluations)
if args.stage <= num_iters + 1:
logger.info("Getting average posterior for purposes of "
"adjusting the priors.")
# If args.do_final_combination is true, we will use the combined model.
# Otherwise, we will use the last_numbered model.
real_iter = 'combined' if args.do_final_combination else num_iters
avg_post_vec_file = train_lib.common.compute_average_posterior(
dir=args.dir,
iter=real_iter,
egs_dir=egs_dir,
num_archives=num_archives,
prior_subset_size=args.prior_subset_size,
run_opts=run_opts)
logger.info("Re-adjusting priors based on computed posteriors")
combined_or_last_numbered_model = "{dir}/{iter}.mdl".format(
dir=args.dir, iter=real_iter)
final_model = "{dir}/final.mdl".format(dir=args.dir)
train_lib.common.adjust_am_priors(args.dir,
combined_or_last_numbered_model,
avg_post_vec_file, final_model,
run_opts)
if args.cleanup:
logger.info("Cleaning up the experiment directory "
"{0}".format(args.dir))
remove_egs = args.remove_egs
if args.egs_dir is not None:
# this egs_dir was not created by this experiment so we will not
# delete it
remove_egs = False
common_train_lib.clean_nnet_dir(
nnet_dir=args.dir,
num_iters=num_iters,
egs_dir=egs_dir,
preserve_model_interval=args.preserve_model_interval,
remove_egs=remove_egs)
# do some reporting
[report, times,
data] = nnet3_log_parse.generate_acc_logprob_report(args.dir)
if args.email is not None:
common_lib.send_mail(
report, "Update : Expt {0} : "
"complete".format(args.dir), args.email)
with open("{dir}/accuracy.report".format(dir=args.dir), "w") as f:
f.write(report)
common_lib.execute_command("steps/info/nnet3_dir_info.pl "
"{0}".format(args.dir))
def one_iter(args, run_opts):
""" The main function for training.
Args:
args: a Namespace object with the required parameters
obtained from the function process_args()
run_opts: RunOpts object obtained from the process_args()
"""
arg_string = pprint.pformat(vars(args))
logger.info("Arguments for the experiment\n{0}".format(arg_string))
# Check files
chain_lib.check_for_required_files(args.feat_dir, args.tree_dir,
args.lat_dir)
# Set some variables.
num_jobs = common_lib.get_number_of_jobs(args.tree_dir)
feat_dim = common_lib.get_feat_dim(args.feat_dir)
ivector_dim = common_lib.get_ivector_dim(args.online_ivector_dir)
ivector_id = common_lib.get_ivector_extractor_id(args.online_ivector_dir)
# split the training data into parts for individual jobs
# we will use the same number of jobs as that used for alignment
common_lib.execute_command("utils/split_data.sh {0} {1}"
"".format(args.feat_dir, num_jobs))
with open('{0}/num_jobs'.format(args.dir), 'w') as f:
f.write(str(num_jobs))
if args.input_model is None:
config_dir = '{0}/configs'.format(args.dir)
var_file = '{0}/vars'.format(config_dir)
variables = common_train_lib.parse_generic_config_vars_file(var_file)
else:
# If args.input_model is specified, the model left and right contexts
# are computed using input_model.
variables = common_train_lib.get_input_model_info(args.input_model)
# Set some variables.
try:
model_left_context = variables['model_left_context']
model_right_context = variables['model_right_context']
except KeyError as e:
raise Exception("KeyError {0}: Variables need to be defined in "
"{1}".format(str(e), '{0}/configs'.format(args.dir)))
left_context = args.chunk_left_context + model_left_context
right_context = args.chunk_right_context + model_right_context
left_context_initial = (args.chunk_left_context_initial +
model_left_context
if args.chunk_left_context_initial >= 0 else -1)
right_context_final = (args.chunk_right_context_final + model_right_context
if args.chunk_right_context_final >= 0 else -1)
egs_left_context = left_context + args.frame_subsampling_factor / 2
egs_right_context = right_context + args.frame_subsampling_factor / 2
# note: the '+ args.frame_subsampling_factor / 2' is to allow for the
# fact that we'll be shifting the data slightly during training to give
# variety to the training data.
egs_left_context_initial = (left_context_initial +
args.frame_subsampling_factor / 2
if left_context_initial >= 0 else -1)
egs_right_context_final = (right_context_final +
args.frame_subsampling_factor / 2
if right_context_final >= 0 else -1)
default_egs_dir = '{0}/egs'.format(args.dir)
if args.egs_dir is None:
egs_dir = default_egs_dir
else:
egs_dir = args.egs_dir
[egs_left_context, egs_right_context, frames_per_eg_str,
num_archives] = (common_train_lib.verify_egs_dir(
egs_dir, feat_dim, ivector_dim, ivector_id, egs_left_context,
egs_right_context, egs_left_context_initial, egs_right_context_final))
assert (args.chunk_width == frames_per_eg_str)
num_archives_expanded = num_archives * args.frame_subsampling_factor
if (args.num_jobs_final > num_archives_expanded):
raise Exception('num_jobs_final cannot exceed the '
'expanded number of archives')
# copy the properties of the egs to dir for
# use during decoding
logger.info("Copying the properties from {0} to {1}".format(
egs_dir, args.dir))
common_train_lib.copy_egs_properties_to_exp_dir(egs_dir, args.dir)
if ((args.stage <= -2)
and (os.path.exists(args.dir + "/configs/init.config"))
and (args.input_model is None)):
logger.info('Computing the preconditioning matrix for input features')
chain_lib.compute_preconditioning_matrix(
args.dir,
egs_dir,
num_archives,
run_opts,
max_lda_jobs=args.max_lda_jobs,
rand_prune=args.rand_prune)
if (args.stage <= -1):
logger.info("Preparing the initial acoustic model.")
chain_lib.prepare_initial_acoustic_model(args.dir,
run_opts,
input_model=args.input_model)
with open("{0}/frame_subsampling_factor".format(args.dir), "w") as f:
f.write(str(args.frame_subsampling_factor))
# set num_iters so that as close as possible, we process the data
# $num_epochs times, i.e. $num_iters*$avg_num_jobs) ==
# $num_epochs*$num_archives, where
# avg_num_jobs=(num_jobs_initial+num_jobs_final)/2.
num_archives_to_process = int(args.num_epochs * num_archives_expanded)
num_archives_processed = 0
num_iters = ((num_archives_to_process * 2) /
(args.num_jobs_initial + args.num_jobs_final))
# If do_final_combination is True, compute the set of models_to_combine.
# Otherwise, models_to_combine will be none.
if args.do_final_combination:
models_to_combine = common_train_lib.get_model_combine_iters(
num_iters, args.num_epochs, num_archives_expanded,
args.max_models_combine, args.num_jobs_final)
else:
models_to_combine = None
min_deriv_time = None
max_deriv_time_relative = None
if args.deriv_truncate_margin is not None:
min_deriv_time = -args.deriv_truncate_margin - model_left_context
max_deriv_time_relative = \
args.deriv_truncate_margin + model_right_context
logger.info("Training will run for {0} epochs = "
"{1} iterations".format(args.num_epochs, num_iters))
if (args.exit_stage is not None) and (args.iter == args.exit_stage):
logger.info("Exiting early due to --exit-stage {0}".format(args.iter))
return
current_num_jobs = int(0.5 + args.num_jobs_initial +
(args.num_jobs_final - args.num_jobs_initial) *
float(args.iter) / num_iters)
model_file = "{dir}/{iter}.mdl".format(dir=args.dir, iter=args.iter)
lrate = common_train_lib.get_learning_rate(args.iter, current_num_jobs,
num_iters,
num_archives_processed,
num_archives_to_process,
args.initial_effective_lrate,
args.final_effective_lrate)
shrinkage_value = 1.0 - (args.proportional_shrink * lrate)
if shrinkage_value <= 0.5:
raise Exception("proportional-shrink={0} is too large, it gives "
"shrink-value={1}".format(args.proportional_shrink,
shrinkage_value))
if args.shrink_value < shrinkage_value:
shrinkage_value = (
args.shrink_value if common_train_lib.should_do_shrinkage(
args.iter, model_file,
args.shrink_saturation_threshold) else shrinkage_value)
percent = num_archives_processed * 100.0 / num_archives_to_process
epoch = (num_archives_processed * args.num_epochs /
num_archives_to_process)
shrink_info_str = ''
if shrinkage_value != 1.0:
shrink_info_str = 'shrink: {0:0.5f}'.format(shrinkage_value)
logger.info("Iter: {0}/{1} "
"Epoch: {2:0.2f}/{3:0.1f} ({4:0.1f}% complete) "
"lr: {5:0.6f} {6}".format(args.iter, num_iters - 1, epoch,
args.num_epochs, percent, lrate,
shrink_info_str))
chain_lib.train_one_iteration(
dir=args.dir,
iter=args.iter,
srand=args.srand,
egs_dir=egs_dir,
num_jobs=current_num_jobs,
num_archives_processed=num_archives_processed,
num_archives=num_archives,
learning_rate=lrate,
dropout_edit_string=common_train_lib.get_dropout_edit_string(
args.dropout_schedule,
float(num_archives_processed) / num_archives_to_process,
args.iter),
shrinkage_value=shrinkage_value,
num_chunk_per_minibatch_str=args.num_chunk_per_minibatch,
apply_deriv_weights=args.apply_deriv_weights,
min_deriv_time=min_deriv_time,
max_deriv_time_relative=max_deriv_time_relative,
l2_regularize=args.l2_regularize,
xent_regularize=args.xent_regularize,
leaky_hmm_coefficient=args.leaky_hmm_coefficient,
momentum=args.momentum,
max_param_change=args.max_param_change,
shuffle_buffer_size=args.shuffle_buffer_size,
frame_subsampling_factor=args.frame_subsampling_factor,
run_opts=run_opts,
backstitch_training_scale=args.backstitch_training_scale,
backstitch_training_interval=args.backstitch_training_interval)
num_archives_processed = num_archives_processed + current_num_jobs
# do some reporting
[report, times, data
] = nnet3_log_parse.generate_acc_logprob_report(args.dir,
"log-probability")
if args.email is not None:
common_lib.send_mail(
report, "Update : Expt {0} : "
"complete".format(args.dir), args.email)
with open("{dir}/accuracy.report".format(dir=args.dir), "w") as f:
f.write(report)
common_lib.execute_command("steps/info/nnet3_dir_info.pl "
"{0}".format(args.dir))
def CheckArgs(args):
if not os.path.exists(args.config_dir):
os.makedirs(args.config_dir)
## Check arguments.
if args.feat_dir is not None:
args.feat_dim = common_lib.get_feat_dim(args.feat_dir)
if args.ali_dir is not None:
args.num_targets = common_lib.get_number_of_leaves_from_tree(args.ali_dir)
elif args.tree_dir is not None:
args.num_targets = common_lib.get_number_of_leaves_from_tree(args.tree_dir)
if args.ivector_dir is not None:
args.ivector_dim = common_lib.get_ivector_dim(args.ivector_dir)
if not args.feat_dim > 0:
raise Exception("feat-dim has to be postive")
if not args.num_targets > 0:
print(args.num_targets)
raise Exception("num_targets has to be positive")
if not args.ivector_dim >= 0:
raise Exception("ivector-dim has to be non-negative")
if (args.subset_dim < 0):
raise Exception("--subset-dim has to be non-negative")
if not args.relu_dim is None:
if not args.pnorm_input_dim is None or not args.pnorm_output_dim is None or not args.relu_dim_init is None:
raise Exception("--relu-dim argument not compatible with "
"--pnorm-input-dim or --pnorm-output-dim or --relu-dim-init options");
args.nonlin_input_dim = args.relu_dim
args.nonlin_output_dim = args.relu_dim
args.nonlin_output_dim_final = None
args.nonlin_output_dim_init = None
args.nonlin_type = 'relu'
elif not args.relu_dim_final is None:
if not args.pnorm_input_dim is None or not args.pnorm_output_dim is None:
raise Exception("--relu-dim-final argument not compatible with "
"--pnorm-input-dim or --pnorm-output-dim options")
if args.relu_dim_init is None:
raise Exception("--relu-dim-init argument should also be provided with --relu-dim-final")
if args.relu_dim_init > args.relu_dim_final:
raise Exception("--relu-dim-init has to be no larger than --relu-dim-final")
args.nonlin_input_dim = None
args.nonlin_output_dim = None
args.nonlin_output_dim_final = args.relu_dim_final
args.nonlin_output_dim_init = args.relu_dim_init
args.nonlin_type = 'relu'
else:
if not args.relu_dim_init is None:
raise Exception("--relu-dim-final argument not compatible with "
"--pnorm-input-dim or --pnorm-output-dim options")
if not args.pnorm_input_dim > 0 or not args.pnorm_output_dim > 0:
raise Exception("--relu-dim not set, so expected --pnorm-input-dim and "
"--pnorm-output-dim to be provided.");
args.nonlin_input_dim = args.pnorm_input_dim
args.nonlin_output_dim = args.pnorm_output_dim
if (args.nonlin_input_dim < args.nonlin_output_dim) or (args.nonlin_input_dim % args.nonlin_output_dim != 0):
raise Exception("Invalid --pnorm-input-dim {0} and --pnorm-output-dim {1}".format(args.nonlin_input_dim, args.nonlin_output_dim))
args.nonlin_output_dim_final = None
args.nonlin_output_dim_init = None
args.nonlin_type = 'pnorm'
if args.add_final_sigmoid and args.include_log_softmax:
raise Exception("--include-log-softmax and --add-final-sigmoid cannot both be true.")
if args.xent_separate_forward_affine and args.add_final_sigmoid:
raise Exception("It does not make sense to have --add-final-sigmoid=true when xent-separate-forward-affine is true")
if args.add_lda and args.cnn_layer is not None:
args.add_lda = False
warnings.warn("--add-lda is set to false as CNN layers are used.")
if not args.max_change_per_component >= 0 or not args.max_change_per_component_final >= 0:
raise Exception("max-change-per-component and max_change-per-component-final should be non-negative")
return args
def model_init(args, run_opts):
""" The main function for training.
Args:
args: a Namespace object with the required parameters
obtained from the function process_args()
run_opts: RunOpts object obtained from the process_args()
"""
arg_string = pprint.pformat(vars(args))
logger.info("Arguments for the experiment\n{0}".format(arg_string))
# Check files
chain_lib.check_for_required_files(args.feat_dir, args.tree_dir,
args.lat_dir)
# Set some variables.
num_jobs = common_lib.get_number_of_jobs(args.tree_dir)
feat_dim = common_lib.get_feat_dim(args.feat_dir)
ivector_dim = common_lib.get_ivector_dim(args.online_ivector_dir)
ivector_id = common_lib.get_ivector_extractor_id(args.online_ivector_dir)
# split the training data into parts for individual jobs
# we will use the same number of jobs as that used for alignment
common_lib.execute_command("utils/split_data.sh {0} {1}"
"".format(args.feat_dir, num_jobs))
with open('{0}/num_jobs'.format(args.dir), 'w') as f:
f.write(str(num_jobs))
if args.input_model is None:
config_dir = '{0}/configs'.format(args.dir)
var_file = '{0}/vars'.format(config_dir)
variables = common_train_lib.parse_generic_config_vars_file(var_file)
else:
# If args.input_model is specified, the model left and right contexts
# are computed using input_model.
variables = common_train_lib.get_input_model_info(args.input_model)
# Set some variables.
try:
model_left_context = variables['model_left_context']
model_right_context = variables['model_right_context']
except KeyError as e:
raise Exception("KeyError {0}: Variables need to be defined in "
"{1}".format(str(e), '{0}/configs'.format(args.dir)))
left_context = args.chunk_left_context + model_left_context
right_context = args.chunk_right_context + model_right_context
left_context_initial = (args.chunk_left_context_initial + model_left_context if
args.chunk_left_context_initial >= 0 else -1)
right_context_final = (args.chunk_right_context_final + model_right_context if
args.chunk_right_context_final >= 0 else -1)
egs_left_context = left_context + args.frame_subsampling_factor / 2
egs_right_context = right_context + args.frame_subsampling_factor / 2
# note: the '+ args.frame_subsampling_factor / 2' is to allow for the
# fact that we'll be shifting the data slightly during training to give
# variety to the training data.
egs_left_context_initial = (left_context_initial +
args.frame_subsampling_factor / 2 if
left_context_initial >= 0 else -1)
egs_right_context_final = (right_context_final +
args.frame_subsampling_factor / 2 if
right_context_final >= 0 else -1)
default_egs_dir = '{0}/egs'.format(args.dir)
num_jobs = common_lib.get_number_of_jobs(args.tree_dir)
feat_dim = common_lib.get_feat_dim(args.feat_dir)
ivector_dim = common_lib.get_ivector_dim(args.online_ivector_dir)
ivector_id = common_lib.get_ivector_extractor_id(args.online_ivector_dir)
# split the training data into parts for individual jobs
# we will use the same number of jobs as that used for alignment
common_lib.execute_command("utils/split_data.sh {0} {1}"
"".format(args.feat_dir, num_jobs))
with open('{0}/num_jobs'.format(args.dir), 'w') as f:
f.write(str(num_jobs))
if args.input_model is None:
config_dir = '{0}/configs'.format(args.dir)
var_file = '{0}/vars'.format(config_dir)
variables = common_train_lib.parse_generic_config_vars_file(var_file)
else:
# If args.input_model is specified, the model left and right contexts
# are computed using input_model.
variables = common_train_lib.get_input_model_info(args.input_model)
# Set some variables.
try:
model_left_context = variables['model_left_context']
model_right_context = variables['model_right_context']
except KeyError as e:
raise Exception("KeyError {0}: Variables need to be defined in "
"{1}".format(str(e), '{0}/configs'.format(args.dir)))
left_context = args.chunk_left_context + model_left_context
right_context = args.chunk_right_context + model_right_context
left_context_initial = (args.chunk_left_context_initial + model_left_context if
args.chunk_left_context_initial >= 0 else -1)
right_context_final = (args.chunk_right_context_final + model_right_context if
args.chunk_right_context_final >= 0 else -1)
if args.egs_dir is None:
egs_dir = default_egs_dir
else:
egs_dir = args.egs_dir
[egs_left_context, egs_right_context,
frames_per_eg_str, num_archives] = (
common_train_lib.verify_egs_dir(egs_dir, feat_dim,
ivector_dim, ivector_id,
egs_left_context, egs_right_context,
egs_left_context_initial,
egs_right_context_final))
assert(args.chunk_width == frames_per_eg_str)
num_archives_expanded = num_archives * args.frame_subsampling_factor
if (args.num_jobs_final > num_archives_expanded):
raise Exception('num_jobs_final cannot exceed the '
'expanded number of archives')
# copy the properties of the egs to dir for
# use during decoding
logger.info("Copying the properties from {0} to {1}".format(egs_dir, args.dir))
common_train_lib.copy_egs_properties_to_exp_dir(egs_dir, args.dir)
if ((args.stage <= -2) and (os.path.exists(args.dir+"/configs/init.config"))
and (args.input_model is None)):
logger.info('Computing the preconditioning matrix for input features')
chain_lib.compute_preconditioning_matrix(
args.dir, egs_dir, num_archives, run_opts,
max_lda_jobs=args.max_lda_jobs,
rand_prune=args.rand_prune)
if (args.stage <= -1):
logger.info("Preparing the initial acoustic model.")
chain_lib.prepare_initial_acoustic_model(args.dir, run_opts,
input_model=args.input_model)
with open("{0}/frame_subsampling_factor".format(args.dir), "w") as f:
f.write(str(args.frame_subsampling_factor))
# set num_iters so that as close as possible, we process the data
# $num_epochs times, i.e. $num_iters*$avg_num_jobs) ==
# $num_epochs*$num_archives, where
# avg_num_jobs=(num_jobs_initial+num_jobs_final)/2.
num_archives_to_process = int(args.num_epochs * num_archives_expanded)
num_archives_processed = 0
num_iters = ((num_archives_to_process * 2)
/ (args.num_jobs_initial + args.num_jobs_final))
# If do_final_combination is True, compute the set of models_to_combine.
# Otherwise, models_to_combine will be none.
if args.do_final_combination:
models_to_combine = common_train_lib.get_model_combine_iters(
num_iters, args.num_epochs,
num_archives_expanded, args.max_models_combine,
args.num_jobs_final)
else:
models_to_combine = None
min_deriv_time = None
max_deriv_time_relative = None
if args.deriv_truncate_margin is not None:
min_deriv_time = -args.deriv_truncate_margin - model_left_context
max_deriv_time_relative = \
args.deriv_truncate_margin + model_right_context
logger.info("Training will run for {0} epochs = "
"{1} iterations".format(args.num_epochs, num_iters))
def train(args, run_opts):
""" The main function for training.
Args:
args: a Namespace object with the required parameters
obtained from the function process_args()
run_opts: RunOpts object obtained from the process_args()
"""
arg_string = pprint.pformat(vars(args))
logger.info("Arguments for the experiment\n{0}".format(arg_string))
# Set some variables.
# num_leaves = common_lib.get_number_of_leaves_from_tree(args.ali_dir)
num_jobs = common_lib.get_number_of_jobs(args.ali_dir)
feat_dim = common_lib.get_feat_dim(args.feat_dir)
ivector_dim = common_lib.get_ivector_dim(args.online_ivector_dir)
ivector_id = common_lib.get_ivector_extractor_id(args.online_ivector_dir)
# split the training data into parts for individual jobs
# we will use the same number of jobs as that used for alignment
# 将特征数据 按照job数量进行划分, 会使用相同数量的job 进行对齐.
# feat_dir -- data/train_sp_hires 划分为jos 个子集.
common_lib.execute_command("utils/split_data.sh {0} {1}".format(
args.feat_dir, num_jobs))
# copy tri5a_sp_ali/tree(最后训练好的模型tree) ---> exp/nnet3/tdnn_sp
shutil.copy('{0}/tree'.format(args.ali_dir), args.dir)
# 将ali_dir(exp/tri5a_sp_ali/num_jobs) 读取写入 目标目录 exp/nnet3/tdnn_sp/num_jos
with open('{0}/num_jobs'.format(args.dir), 'w') as f:
f.write(str(num_jobs))
# 目标目录中 存入 nnet3结构/ 以及nnet3的 left right context配置信息
config_dir = '{0}/configs'.format(args.dir)
var_file = '{0}/vars'.format(config_dir)
variables = common_train_lib.parse_generic_config_vars_file(var_file)
# liujunnan@innovem:configs$ cat vars
# model_left_context=16
# model_right_context=12
# liujunnan@innovem:configs$
# Set some variables.
model_left_context = variables['model_left_context']
model_right_context = variables['model_right_context']
left_context = model_left_context
right_context = model_right_context
# 初始化 原始nnet结构, 在训练 LDA预处理矩阵之前
# 第一个配置 只做了一些初始化拼接操作
# 这样做,是因为这样很方变能够获得LDA变换矩阵的统计信息.??????
if (args.stage <= -5) and os.path.exists(args.dir+"/configs/init.config"):
logger.info("Initializing a basic network for estimating preconditioning matrix")
# nnet3-init 利用 exp/nnet3/tdnn_sp/configs/init.config ===> exp/nnet3/tdnn_sp/init.raw.
common_lib.execute_command(
"""{command} {dir}/log/nnet_init.log \
nnet3-init --srand=-2 {dir}/configs/init.config \
{dir}/init.raw""".format(command=run_opts.command,
dir=args.dir))
def train(args, run_opts, background_process_handler):
""" The main function for training.
Args:
args: a Namespace object with the required parameters
obtained from the function process_args()
run_opts: RunOpts object obtained from the process_args()
"""
arg_string = pprint.pformat(vars(args))
logger.info("Arguments for the experiment\n{0}".format(arg_string))
# Check files
chain_lib.check_for_required_files(args.feat_dir, args.tree_dir,
args.lat_dir)
# Set some variables.
num_jobs = common_lib.get_number_of_jobs(args.tree_dir)
feat_dim = common_lib.get_feat_dim(args.feat_dir)
ivector_dim = common_lib.get_ivector_dim(args.online_ivector_dir)
# split the training data into parts for individual jobs
# we will use the same number of jobs as that used for alignment
common_lib.split_data(args.feat_dir, num_jobs)
shutil.copy('{0}/tree'.format(args.tree_dir), args.dir)
with open('{0}/num_jobs'.format(args.dir), 'w') as f:
f.write(str(num_jobs))
config_dir = '{0}/configs'.format(args.dir)
var_file = '{0}/vars'.format(config_dir)
variables = common_train_lib.parse_generic_config_vars_file(var_file)
# Set some variables.
try:
model_left_context = variables['model_left_context']
model_right_context = variables['model_right_context']
# this is really the number of times we add layers to the network for
# discriminative pretraining
num_hidden_layers = variables['num_hidden_layers']
except KeyError as e:
raise Exception("KeyError {0}: Variables need to be defined in "
"{1}".format(str(e), '{0}/configs'.format(args.dir)))
left_context = args.chunk_left_context + model_left_context
right_context = args.chunk_right_context + model_right_context
# Initialize as "raw" nnet, prior to training the LDA-like preconditioning
# matrix. This first config just does any initial splicing that we do;
# we do this as it's a convenient way to get the stats for the 'lda-like'
# transform.
if (args.stage <= -6):
logger.info("Creating phone language-model")
chain_lib.create_phone_lm(args.dir, args.tree_dir, run_opts,
lm_opts=args.lm_opts)
if (args.stage <= -5):
logger.info("Creating denominator FST")
chain_lib.create_denominator_fst(args.dir, args.tree_dir, run_opts)
if (args.stage <= -4):
logger.info("Initializing a basic network for estimating "
"preconditioning matrix")
common_lib.run_kaldi_command(
"""{command} {dir}/log/nnet_init.log \
nnet3-init --srand=-2 {dir}/configs/init.config \
{dir}/init.raw""".format(command=run_opts.command,
dir=args.dir))
egs_left_context = left_context + args.frame_subsampling_factor/2
egs_right_context = right_context + args.frame_subsampling_factor/2
default_egs_dir = '{0}/egs'.format(args.dir)
if (args.stage <= -3) and args.egs_dir is None:
logger.info("Generating egs")
# this is where get_egs.sh is called.
chain_lib.generate_chain_egs(
dir=args.dir, data=args.feat_dir,
lat_dir=args.lat_dir, egs_dir=default_egs_dir,
left_context=egs_left_context,
right_context=egs_right_context,
run_opts=run_opts,
left_tolerance=args.left_tolerance,
right_tolerance=args.right_tolerance,
frame_subsampling_factor=args.frame_subsampling_factor,
alignment_subsampling_factor=args.alignment_subsampling_factor,
frames_per_eg=args.chunk_width,
srand=args.srand,
egs_opts=args.egs_opts,
cmvn_opts=args.cmvn_opts,
online_ivector_dir=args.online_ivector_dir,
frames_per_iter=args.frames_per_iter,
transform_dir=args.transform_dir,
stage=args.egs_stage)
if args.egs_dir is None:
egs_dir = default_egs_dir
else:
egs_dir = args.egs_dir
[egs_left_context, egs_right_context,
frames_per_eg, num_archives] = (
common_train_lib.verify_egs_dir(egs_dir, feat_dim, ivector_dim,
egs_left_context, egs_right_context))
assert(args.chunk_width == frames_per_eg)
num_archives_expanded = num_archives * args.frame_subsampling_factor
if (args.num_jobs_final > num_archives_expanded):
raise Exception('num_jobs_final cannot exceed the '
'expanded number of archives')
# copy the properties of the egs to dir for
# use during decoding
common_train_lib.copy_egs_properties_to_exp_dir(egs_dir, args.dir)
if (args.stage <= -2):
logger.info('Computing the preconditioning matrix for input features')
chain_lib.compute_preconditioning_matrix(
args.dir, egs_dir, num_archives, run_opts,
max_lda_jobs=args.max_lda_jobs,
rand_prune=args.rand_prune)
if (args.stage <= -1):
logger.info("Preparing the initial acoustic model.")
chain_lib.prepare_initial_acoustic_model(args.dir, run_opts)
with open("{0}/frame_subsampling_factor".format(args.dir), "w") as f:
f.write(str(args.frame_subsampling_factor))
# set num_iters so that as close as possible, we process the data
# $num_epochs times, i.e. $num_iters*$avg_num_jobs) ==
# $num_epochs*$num_archives, where
# avg_num_jobs=(num_jobs_initial+num_jobs_final)/2.
num_archives_to_process = args.num_epochs * num_archives_expanded
num_archives_processed = 0
num_iters = ((num_archives_to_process * 2)
/ (args.num_jobs_initial + args.num_jobs_final))
models_to_combine = common_train_lib.verify_iterations(
num_iters, args.num_epochs,
num_hidden_layers, num_archives_expanded,
args.max_models_combine, args.add_layers_period,
args.num_jobs_final)
def learning_rate(iter, current_num_jobs, num_archives_processed):
return common_train_lib.get_learning_rate(iter, current_num_jobs,
num_iters,
num_archives_processed,
num_archives_to_process,
args.initial_effective_lrate,
args.final_effective_lrate)
min_deriv_time = None
max_deriv_time = None
if args.deriv_truncate_margin is not None:
min_deriv_time = -args.deriv_truncate_margin - model_left_context
max_deriv_time = (args.chunk_width - 1 + args.deriv_truncate_margin
+ model_right_context)
logger.info("Training will run for {0} epochs = "
"{1} iterations".format(args.num_epochs, num_iters))
for iter in range(num_iters):
if (args.exit_stage is not None) and (iter == args.exit_stage):
logger.info("Exiting early due to --exit-stage {0}".format(iter))
return
current_num_jobs = int(0.5 + args.num_jobs_initial
+ (args.num_jobs_final - args.num_jobs_initial)
* float(iter) / num_iters)
if args.stage <= iter:
model_file = "{dir}/{iter}.mdl".format(dir=args.dir, iter=iter)
shrinkage_value = 1.0
if args.shrink_value != 1.0:
shrinkage_value = (args.shrink_value
if common_train_lib.do_shrinkage(
iter, model_file,
args.shrink_saturation_threshold)
else 1
)
chain_lib.train_one_iteration(
dir=args.dir,
iter=iter,
srand=args.srand,
egs_dir=egs_dir,
num_jobs=current_num_jobs,
num_archives_processed=num_archives_processed,
num_archives=num_archives,
learning_rate=learning_rate(iter, current_num_jobs,
num_archives_processed),
dropout_edit_string=common_lib.get_dropout_edit_string(
args.dropout_schedule,
float(num_archives_processed) / num_archives_to_process,
iter),
shrinkage_value=shrinkage_value,
num_chunk_per_minibatch=args.num_chunk_per_minibatch,
num_hidden_layers=num_hidden_layers,
add_layers_period=args.add_layers_period,
left_context=left_context,
right_context=right_context,
apply_deriv_weights=args.apply_deriv_weights,
min_deriv_time=min_deriv_time,
max_deriv_time=max_deriv_time,
l2_regularize=args.l2_regularize,
xent_regularize=args.xent_regularize,
leaky_hmm_coefficient=args.leaky_hmm_coefficient,
momentum=args.momentum,
max_param_change=args.max_param_change,
shuffle_buffer_size=args.shuffle_buffer_size,
frame_subsampling_factor=args.frame_subsampling_factor,
truncate_deriv_weights=args.truncate_deriv_weights,
run_opts=run_opts,
background_process_handler=background_process_handler)
if args.cleanup:
# do a clean up everythin but the last 2 models, under certain
# conditions
common_train_lib.remove_model(
args.dir, iter-2, num_iters, models_to_combine,
args.preserve_model_interval)
if args.email is not None:
reporting_iter_interval = num_iters * args.reporting_interval
if iter % reporting_iter_interval == 0:
# lets do some reporting
[report, times, data] = (
nnet3_log_parse.generate_accuracy_report(
args.dir, "log-probability"))
message = report
subject = ("Update : Expt {dir} : "
"Iter {iter}".format(dir=args.dir, iter=iter))
common_lib.send_mail(message, subject, args.email)
num_archives_processed = num_archives_processed + current_num_jobs
if args.stage <= num_iters:
logger.info("Doing final combination to produce final.mdl")
chain_lib.combine_models(
dir=args.dir, num_iters=num_iters,
models_to_combine=models_to_combine,
num_chunk_per_minibatch=args.num_chunk_per_minibatch,
egs_dir=egs_dir,
left_context=left_context, right_context=right_context,
leaky_hmm_coefficient=args.leaky_hmm_coefficient,
l2_regularize=args.l2_regularize,
xent_regularize=args.xent_regularize,
run_opts=run_opts,
background_process_handler=background_process_handler)
if args.cleanup:
logger.info("Cleaning up the experiment directory "
"{0}".format(args.dir))
remove_egs = args.remove_egs
if args.egs_dir is not None:
# this egs_dir was not created by this experiment so we will not
# delete it
remove_egs = False
common_train_lib.clean_nnet_dir(
args.dir, num_iters, egs_dir,
preserve_model_interval=args.preserve_model_interval,
remove_egs=remove_egs)
# do some reporting
[report, times, data] = nnet3_log_parse.generate_accuracy_report(
args.dir, "log-probability")
if args.email is not None:
common_lib.send_mail(report, "Update : Expt {0} : "
"complete".format(args.dir), args.email)
with open("{dir}/accuracy.report".format(dir=args.dir), "w") as f:
f.write(report)
common_lib.run_kaldi_command("steps/info/nnet3_dir_info.pl "
"{0}".format(args.dir))
def train(args, run_opts):
""" The main function for training.
Args:
args: a Namespace object with the required parameters
obtained from the function process_args()
run_opts: RunOpts object obtained from the process_args()
"""
arg_string = pprint.pformat(vars(args))
logger.info("Arguments for the experiment\n{0}".format(arg_string))
# Check files
chain_lib.check_for_required_files(
args.feat_dir, args.tree_dir,
args.lat_dir if args.egs_dir is None else None)
# Copy phones.txt from tree-dir to dir. Later, steps/nnet3/decode.sh will
# use it to check compatibility between training and decoding phone-sets.
shutil.copy('{0}/phones.txt'.format(args.tree_dir), args.dir)
# Set some variables.
num_jobs = common_lib.get_number_of_jobs(args.tree_dir)
feat_dim = common_lib.get_feat_dim(args.feat_dir)
ivector_dim = common_lib.get_ivector_dim(args.online_ivector_dir)
ivector_id = common_lib.get_ivector_extractor_id(args.online_ivector_dir)
# split the training data into parts for individual jobs
# we will use the same number of jobs as that used for alignment
common_lib.execute_command("utils/split_data.sh {0} {1}"
"".format(args.feat_dir, num_jobs))
with open('{0}/num_jobs'.format(args.dir), 'w') as f:
f.write(str(num_jobs))
if args.input_model is None:
config_dir = '{0}/configs'.format(args.dir)
var_file = '{0}/vars'.format(config_dir)
variables = common_train_lib.parse_generic_config_vars_file(var_file)
else:
# If args.input_model is specified, the model left and right contexts
# are computed using input_model.
variables = common_train_lib.get_input_model_info(args.input_model)
# Set some variables.
try:
model_left_context = variables['model_left_context']
model_right_context = variables['model_right_context']
except KeyError as e:
raise Exception("KeyError {0}: Variables need to be defined in "
"{1}".format(str(e), '{0}/configs'.format(args.dir)))
left_context = args.chunk_left_context + model_left_context
right_context = args.chunk_right_context + model_right_context
left_context_initial = (args.chunk_left_context_initial +
model_left_context
if args.chunk_left_context_initial >= 0 else -1)
right_context_final = (args.chunk_right_context_final + model_right_context
if args.chunk_right_context_final >= 0 else -1)
# Initialize as "raw" nnet, prior to training the LDA-like preconditioning
# matrix. This first config just does any initial splicing that we do;
# we do this as it's a convenient way to get the stats for the 'lda-like'
# transform.
if (args.stage <= -6):
logger.info("Creating phone language-model")
chain_lib.create_phone_lm(args.dir,
args.tree_dir,
run_opts,
lm_opts=args.lm_opts)
if (args.stage <= -5):
logger.info("Creating denominator FST")
shutil.copy('{0}/tree'.format(args.tree_dir), args.dir)
chain_lib.create_denominator_fst(args.dir, args.tree_dir, run_opts)
if ((args.stage <= -4)
and os.path.exists("{0}/configs/init.config".format(args.dir))
and (args.input_model is None)):
logger.info("Initializing a basic network for estimating "
"preconditioning matrix")
common_lib.execute_command("""{command} {dir}/log/nnet_init.log \
nnet3-init --srand=-2 {dir}/configs/init.config \
{dir}/init.raw""".format(command=run_opts.command, dir=args.dir))
egs_left_context = left_context + args.frame_subsampling_factor // 2
egs_right_context = right_context + args.frame_subsampling_factor // 2
# note: the '+ args.frame_subsampling_factor / 2' is to allow for the
# fact that we'll be shifting the data slightly during training to give
# variety to the training data.
egs_left_context_initial = (left_context_initial +
args.frame_subsampling_factor // 2
if left_context_initial >= 0 else -1)
egs_right_context_final = (right_context_final +
args.frame_subsampling_factor // 2
if right_context_final >= 0 else -1)
default_egs_dir = '{0}/egs'.format(args.dir)
if ((args.stage <= -3) and args.egs_dir is None):
logger.info("Generating egs")
if (not os.path.exists("{0}/den.fst".format(args.dir))
or not os.path.exists("{0}/normalization.fst".format(args.dir))
or not os.path.exists("{0}/tree".format(args.dir))):
raise Exception("Chain egs generation expects {0}/den.fst, "
"{0}/normalization.fst and {0}/tree "
"to exist.".format(args.dir))
# this is where get_egs.sh is called.
chain_lib.generate_chain_egs(
dir=args.dir,
data=args.feat_dir,
lat_dir=args.lat_dir,
egs_dir=default_egs_dir,
left_context=egs_left_context,
right_context=egs_right_context,
left_context_initial=egs_left_context_initial,
right_context_final=egs_right_context_final,
run_opts=run_opts,
left_tolerance=args.left_tolerance,
right_tolerance=args.right_tolerance,
frame_subsampling_factor=args.frame_subsampling_factor,
alignment_subsampling_factor=args.alignment_subsampling_factor,
frames_per_eg_str=args.chunk_width,
srand=args.srand,
egs_opts=args.egs_opts,
cmvn_opts=args.cmvn_opts,
online_ivector_dir=args.online_ivector_dir,
frames_per_iter=args.frames_per_iter,
stage=args.egs_stage)
if args.egs_dir is None:
egs_dir = default_egs_dir
else:
egs_dir = args.egs_dir
[egs_left_context, egs_right_context, frames_per_eg_str,
num_archives] = (common_train_lib.verify_egs_dir(
egs_dir, feat_dim, ivector_dim, ivector_id, egs_left_context,
egs_right_context, egs_left_context_initial, egs_right_context_final))
assert (args.chunk_width == frames_per_eg_str)
num_archives_expanded = num_archives * args.frame_subsampling_factor
if (args.num_jobs_final > num_archives_expanded):
raise Exception('num_jobs_final cannot exceed the '
'expanded number of archives')
# copy the properties of the egs to dir for
# use during decoding
logger.info("Copying the properties from {0} to {1}".format(
egs_dir, args.dir))
common_train_lib.copy_egs_properties_to_exp_dir(egs_dir, args.dir)
if not os.path.exists('{0}/valid_diagnostic.cegs'.format(egs_dir)):
if (not os.path.exists('{0}/valid_diagnostic.scp'.format(egs_dir))):
raise Exception('Neither {0}/valid_diagnostic.cegs nor '
'{0}/valid_diagnostic.scp exist.'
'This script expects one of them.'.format(egs_dir))
use_multitask_egs = True
else:
use_multitask_egs = False
if ((args.stage <= -2)
and (os.path.exists(args.dir + "/configs/init.config"))
and (args.input_model is None)):
logger.info('Computing the preconditioning matrix for input features')
chain_lib.compute_preconditioning_matrix(
args.dir,
egs_dir,
num_archives,
run_opts,
max_lda_jobs=args.max_lda_jobs,
rand_prune=args.rand_prune,
use_multitask_egs=use_multitask_egs)
if (args.stage <= -1):
logger.info("Preparing the initial acoustic model.")
chain_lib.prepare_initial_acoustic_model(args.dir,
run_opts,
input_model=args.input_model)
with open("{0}/frame_subsampling_factor".format(args.dir), "w") as f:
f.write(str(args.frame_subsampling_factor))
# set num_iters so that as close as possible, we process the data
# $num_epochs times, i.e. $num_iters*$avg_num_jobs) ==
# $num_epochs*$num_archives, where
# avg_num_jobs=(num_jobs_initial+num_jobs_final)/2.
num_archives_to_process = int(args.num_epochs * num_archives_expanded)
num_archives_processed = 0
num_iters = ((num_archives_to_process * 2) //
(args.num_jobs_initial + args.num_jobs_final))
# If do_final_combination is True, compute the set of models_to_combine.
# Otherwise, models_to_combine will be none.
if args.do_final_combination:
models_to_combine = common_train_lib.get_model_combine_iters(
num_iters, args.num_epochs, num_archives_expanded,
args.max_models_combine, args.num_jobs_final)
else:
models_to_combine = None
min_deriv_time = None
max_deriv_time_relative = None
if args.deriv_truncate_margin is not None:
min_deriv_time = -args.deriv_truncate_margin - model_left_context
max_deriv_time_relative = \
args.deriv_truncate_margin + model_right_context
logger.info("Training will run for {0} epochs = "
"{1} iterations".format(args.num_epochs, num_iters))
for iter in range(num_iters):
if (args.exit_stage is not None) and (iter == args.exit_stage):
logger.info("Exiting early due to --exit-stage {0}".format(iter))
return
current_num_jobs = int(0.5 + args.num_jobs_initial +
(args.num_jobs_final - args.num_jobs_initial) *
float(iter) / num_iters)
if args.stage <= iter:
model_file = "{dir}/{iter}.mdl".format(dir=args.dir, iter=iter)
lrate = common_train_lib.get_learning_rate(
iter, current_num_jobs, num_iters, num_archives_processed,
num_archives_to_process, args.initial_effective_lrate,
args.final_effective_lrate)
shrinkage_value = 1.0 - (args.proportional_shrink * lrate)
if shrinkage_value <= 0.5:
raise Exception(
"proportional-shrink={0} is too large, it gives "
"shrink-value={1}".format(args.proportional_shrink,
shrinkage_value))
if args.shrink_value < shrinkage_value:
shrinkage_value = (
args.shrink_value if common_train_lib.should_do_shrinkage(
iter, model_file,
args.shrink_saturation_threshold) else shrinkage_value)
percent = num_archives_processed * 100.0 / num_archives_to_process
epoch = (num_archives_processed * args.num_epochs /
num_archives_to_process)
shrink_info_str = ''
if shrinkage_value != 1.0:
shrink_info_str = 'shrink: {0:0.5f}'.format(shrinkage_value)
logger.info("Iter: {0}/{1} "
"Epoch: {2:0.2f}/{3:0.1f} ({4:0.1f}% complete) "
"lr: {5:0.6f} {6}".format(iter, num_iters - 1,
epoch, args.num_epochs,
percent, lrate,
shrink_info_str))
chain_lib.train_one_iteration(
dir=args.dir,
iter=iter,
srand=args.srand,
egs_dir=egs_dir,
num_jobs=current_num_jobs,
num_archives_processed=num_archives_processed,
num_archives=num_archives,
learning_rate=lrate,
dropout_edit_string=common_train_lib.get_dropout_edit_string(
args.dropout_schedule,
float(num_archives_processed) / num_archives_to_process,
iter),
train_opts=' '.join(args.train_opts),
shrinkage_value=shrinkage_value,
num_chunk_per_minibatch_str=args.num_chunk_per_minibatch,
apply_deriv_weights=args.apply_deriv_weights,
min_deriv_time=min_deriv_time,
max_deriv_time_relative=max_deriv_time_relative,
l2_regularize=args.l2_regularize,
xent_regularize=args.xent_regularize,
leaky_hmm_coefficient=args.leaky_hmm_coefficient,
momentum=args.momentum,
max_param_change=args.max_param_change,
shuffle_buffer_size=args.shuffle_buffer_size,
frame_subsampling_factor=args.frame_subsampling_factor,
run_opts=run_opts,
backstitch_training_scale=args.backstitch_training_scale,
backstitch_training_interval=args.backstitch_training_interval,
use_multitask_egs=use_multitask_egs)
if args.cleanup:
# do a clean up everything but the last 2 models, under certain
# conditions
common_train_lib.remove_model(args.dir, iter - 2, num_iters,
models_to_combine,
args.preserve_model_interval)
if args.email is not None:
reporting_iter_interval = num_iters * args.reporting_interval
if iter % reporting_iter_interval == 0:
# lets do some reporting
[report, times,
data] = (nnet3_log_parse.generate_acc_logprob_report(
args.dir, "log-probability"))
message = report
subject = ("Update : Expt {dir} : "
"Iter {iter}".format(dir=args.dir, iter=iter))
common_lib.send_mail(message, subject, args.email)
num_archives_processed = num_archives_processed + current_num_jobs
if args.stage <= num_iters:
if args.do_final_combination:
logger.info("Doing final combination to produce final.mdl")
chain_lib.combine_models(
dir=args.dir,
num_iters=num_iters,
models_to_combine=models_to_combine,
num_chunk_per_minibatch_str=args.num_chunk_per_minibatch,
egs_dir=egs_dir,
leaky_hmm_coefficient=args.leaky_hmm_coefficient,
l2_regularize=args.l2_regularize,
xent_regularize=args.xent_regularize,
run_opts=run_opts,
max_objective_evaluations=args.max_objective_evaluations,
use_multitask_egs=use_multitask_egs)
else:
logger.info("Copying the last-numbered model to final.mdl")
common_lib.force_symlink("{0}.mdl".format(num_iters),
"{0}/final.mdl".format(args.dir))
chain_lib.compute_train_cv_probabilities(
dir=args.dir,
iter=num_iters,
egs_dir=egs_dir,
l2_regularize=args.l2_regularize,
xent_regularize=args.xent_regularize,
leaky_hmm_coefficient=args.leaky_hmm_coefficient,
run_opts=run_opts,
use_multitask_egs=use_multitask_egs)
common_lib.force_symlink(
"compute_prob_valid.{iter}.log"
"".format(iter=num_iters),
"{dir}/log/compute_prob_valid.final.log".format(dir=args.dir))
if args.cleanup:
logger.info("Cleaning up the experiment directory "
"{0}".format(args.dir))
remove_egs = args.remove_egs
if args.egs_dir is not None:
# this egs_dir was not created by this experiment so we will not
# delete it
remove_egs = False
# leave the last-two-numbered models, for diagnostic reasons.
common_train_lib.clean_nnet_dir(
args.dir,
num_iters - 1,
egs_dir,
preserve_model_interval=args.preserve_model_interval,
remove_egs=remove_egs)
# do some reporting
[report, times, data
] = nnet3_log_parse.generate_acc_logprob_report(args.dir,
"log-probability")
if args.email is not None:
common_lib.send_mail(
report, "Update : Expt {0} : "
"complete".format(args.dir), args.email)
with open("{dir}/accuracy.report".format(dir=args.dir), "w") as f:
f.write(report)
common_lib.execute_command("steps/info/chain_dir_info.pl "
"{0}".format(args.dir))
def train(args, run_opts, background_process_handler):
""" The main function for training.
Args:
args: a Namespace object with the required parameters
obtained from the function process_args()
run_opts: RunOpts object obtained from the process_args()
"""
arg_string = pprint.pformat(vars(args))
logger.info("Arguments for the experiment\n{0}".format(arg_string))
# Check files
chain_lib.check_for_required_files(args.feat_dir, args.tree_dir,
args.lat_dir)
# Set some variables.
num_jobs = common_lib.get_number_of_jobs(args.tree_dir)
feat_dim = common_lib.get_feat_dim(args.feat_dir)
ivector_dim = common_lib.get_ivector_dim(args.online_ivector_dir)
# split the training data into parts for individual jobs
# we will use the same number of jobs as that used for alignment
common_lib.split_data(args.feat_dir, num_jobs)
shutil.copy('{0}/tree'.format(args.tree_dir), args.dir)
with open('{0}/num_jobs'.format(args.dir), 'w') as f:
f.write(str(num_jobs))
config_dir = '{0}/configs'.format(args.dir)
var_file = '{0}/vars'.format(config_dir)
variables = common_train_lib.parse_generic_config_vars_file(var_file)
# Set some variables.
try:
model_left_context = variables['model_left_context']
model_right_context = variables['model_right_context']
# this is really the number of times we add layers to the network for
# discriminative pretraining
num_hidden_layers = variables['num_hidden_layers']
except KeyError as e:
raise Exception("KeyError {0}: Variables need to be defined in "
"{1}".format(str(e), '{0}/configs'.format(args.dir)))
left_context = args.chunk_left_context + model_left_context
right_context = args.chunk_right_context + model_right_context
# Initialize as "raw" nnet, prior to training the LDA-like preconditioning
# matrix. This first config just does any initial splicing that we do;
# we do this as it's a convenient way to get the stats for the 'lda-like'
# transform.
if (args.stage <= -6):
logger.info("Creating phone language-model")
chain_lib.create_phone_lm(args.dir, args.tree_dir, run_opts,
lm_opts=args.lm_opts)
if (args.stage <= -5):
logger.info("Creating denominator FST")
chain_lib.create_denominator_fst(args.dir, args.tree_dir, run_opts)
if (args.stage <= -4):
logger.info("Initializing a basic network for estimating "
"preconditioning matrix")
common_lib.run_kaldi_command(
"""{command} {dir}/log/nnet_init.log \
nnet3-init --srand=-2 {dir}/configs/init.config \
{dir}/init.raw""".format(command=run_opts.command,
dir=args.dir))
egs_left_context = left_context + args.frame_subsampling_factor/2
egs_right_context = right_context + args.frame_subsampling_factor/2
default_egs_dir = '{0}/egs'.format(args.dir)
if (args.stage <= -3) and args.egs_dir is None:
logger.info("Generating egs")
# this is where get_egs.sh is called.
chain_lib.generate_chain_egs(
dir=args.dir, data=args.feat_dir,
lat_dir=args.lat_dir, egs_dir=default_egs_dir,
left_context=egs_left_context,
right_context=egs_right_context,
run_opts=run_opts,
left_tolerance=args.left_tolerance,
right_tolerance=args.right_tolerance,
frame_subsampling_factor=args.frame_subsampling_factor,
alignment_subsampling_factor=args.alignment_subsampling_factor,
frames_per_eg=args.chunk_width,
srand=args.srand,
egs_opts=args.egs_opts,
cmvn_opts=args.cmvn_opts,
online_ivector_dir=args.online_ivector_dir,
frames_per_iter=args.frames_per_iter,
transform_dir=args.transform_dir,
stage=args.egs_stage)
if args.egs_dir is None:
egs_dir = default_egs_dir
else:
egs_dir = args.egs_dir
[egs_left_context, egs_right_context,
frames_per_eg, num_archives] = (
common_train_lib.verify_egs_dir(egs_dir, feat_dim, ivector_dim,
egs_left_context, egs_right_context))
assert(args.chunk_width == frames_per_eg)
num_archives_expanded = num_archives * args.frame_subsampling_factor
if (args.num_jobs_final > num_archives_expanded):
raise Exception('num_jobs_final cannot exceed the '
'expanded number of archives')
# copy the properties of the egs to dir for
# use during decoding
common_train_lib.copy_egs_properties_to_exp_dir(egs_dir, args.dir)
if (args.stage <= -2):
logger.info('Computing the preconditioning matrix for input features')
chain_lib.compute_preconditioning_matrix(
args.dir, egs_dir, num_archives, run_opts,
max_lda_jobs=args.max_lda_jobs,
rand_prune=args.rand_prune)
if (args.stage <= -1):
logger.info("Preparing the initial acoustic model.")
chain_lib.prepare_initial_acoustic_model(args.dir, run_opts)
with open("{0}/frame_subsampling_factor".format(args.dir), "w") as f:
f.write(str(args.frame_subsampling_factor))
# set num_iters so that as close as possible, we process the data
# $num_epochs times, i.e. $num_iters*$avg_num_jobs) ==
# $num_epochs*$num_archives, where
# avg_num_jobs=(num_jobs_initial+num_jobs_final)/2.
num_archives_to_process = args.num_epochs * num_archives_expanded
num_archives_processed = 0
num_iters = ((num_archives_to_process * 2)
/ (args.num_jobs_initial + args.num_jobs_final))
models_to_combine = common_train_lib.verify_iterations(
num_iters, args.num_epochs,
num_hidden_layers, num_archives_expanded,
args.max_models_combine, args.add_layers_period,
args.num_jobs_final)
def learning_rate(iter, current_num_jobs, num_archives_processed):
return common_train_lib.get_learning_rate(iter, current_num_jobs,
num_iters,
num_archives_processed,
num_archives_to_process,
args.initial_effective_lrate,
args.final_effective_lrate)
min_deriv_time = None
max_deriv_time = None
if args.deriv_truncate_margin is not None:
min_deriv_time = -args.deriv_truncate_margin - model_left_context
max_deriv_time = (args.chunk_width - 1 + args.deriv_truncate_margin
+ model_right_context)
logger.info("Training will run for {0} epochs = "
"{1} iterations".format(args.num_epochs, num_iters))
for iter in range(num_iters):
if (args.exit_stage is not None) and (iter == args.exit_stage):
logger.info("Exiting early due to --exit-stage {0}".format(iter))
return
current_num_jobs = int(0.5 + args.num_jobs_initial
+ (args.num_jobs_final - args.num_jobs_initial)
* float(iter) / num_iters)
if args.stage <= iter:
model_file = "{dir}/{iter}.mdl".format(dir=args.dir, iter=iter)
shrinkage_value = 1.0
if args.shrink_value != 1.0:
shrinkage_value = (args.shrink_value
if common_train_lib.do_shrinkage(
iter, model_file,
args.shrink_saturation_threshold)
else 1
)
logger.info("On iteration {0}, learning rate is {1} and "
"shrink value is {2}.".format(
iter, learning_rate(iter, current_num_jobs,
num_archives_processed),
shrinkage_value))
chain_lib.train_one_iteration(
dir=args.dir,
iter=iter,
srand=args.srand,
egs_dir=egs_dir,
num_jobs=current_num_jobs,
num_archives_processed=num_archives_processed,
num_archives=num_archives,
learning_rate=learning_rate(iter, current_num_jobs,
num_archives_processed),
shrinkage_value=shrinkage_value,
num_chunk_per_minibatch=args.num_chunk_per_minibatch,
num_hidden_layers=num_hidden_layers,
add_layers_period=args.add_layers_period,
left_context=left_context,
right_context=right_context,
apply_deriv_weights=args.apply_deriv_weights,
min_deriv_time=min_deriv_time,
max_deriv_time=max_deriv_time,
l2_regularize=args.l2_regularize,
xent_regularize=args.xent_regularize,
leaky_hmm_coefficient=args.leaky_hmm_coefficient,
momentum=args.momentum,
max_param_change=args.max_param_change,
shuffle_buffer_size=args.shuffle_buffer_size,
frame_subsampling_factor=args.frame_subsampling_factor,
truncate_deriv_weights=args.truncate_deriv_weights,
run_opts=run_opts,
background_process_handler=background_process_handler)
if args.cleanup:
# do a clean up everythin but the last 2 models, under certain
# conditions
common_train_lib.remove_model(
args.dir, iter-2, num_iters, models_to_combine,
args.preserve_model_interval)
if args.email is not None:
reporting_iter_interval = num_iters * args.reporting_interval
if iter % reporting_iter_interval == 0:
# lets do some reporting
[report, times, data] = (
nnet3_log_parse.generate_accuracy_report(
args.dir, "log-probability"))
message = report
subject = ("Update : Expt {dir} : "
"Iter {iter}".format(dir=args.dir, iter=iter))
common_lib.send_mail(message, subject, args.email)
num_archives_processed = num_archives_processed + current_num_jobs
if args.stage <= num_iters:
logger.info("Doing final combination to produce final.mdl")
chain_lib.combine_models(
dir=args.dir, num_iters=num_iters,
models_to_combine=models_to_combine,
num_chunk_per_minibatch=args.num_chunk_per_minibatch,
egs_dir=egs_dir,
left_context=left_context, right_context=right_context,
leaky_hmm_coefficient=args.leaky_hmm_coefficient,
l2_regularize=args.l2_regularize,
xent_regularize=args.xent_regularize,
run_opts=run_opts,
background_process_handler=background_process_handler)
if args.cleanup:
logger.info("Cleaning up the experiment directory "
"{0}".format(args.dir))
remove_egs = args.remove_egs
if args.egs_dir is not None:
# this egs_dir was not created by this experiment so we will not
# delete it
remove_egs = False
common_train_lib.clean_nnet_dir(
args.dir, num_iters, egs_dir,
preserve_model_interval=args.preserve_model_interval,
remove_egs=remove_egs)
# do some reporting
[report, times, data] = nnet3_log_parse.generate_accuracy_report(
args.dir, "log-probability")
if args.email is not None:
common_lib.send_mail(report, "Update : Expt {0} : "
"complete".format(args.dir), args.email)
with open("{dir}/accuracy.report".format(dir=args.dir), "w") as f:
f.write(report)
common_lib.run_kaldi_command("steps/info/nnet3_dir_info.pl "
"{0}".format(args.dir))
def train(args, run_opts):
""" The main function for training.
Args:
args: a Namespace object with the required parameters
obtained from the function process_args()
run_opts: RunOpts object obtained from the process_args()
"""
arg_string = pprint.pformat(vars(args))
logger.info("Arguments for the experiment\n{0}".format(arg_string))
# Set some variables.
feat_dim = common_lib.get_feat_dim(args.feat_dir)
ivector_dim = common_lib.get_ivector_dim(args.online_ivector_dir)
ivector_id = common_lib.get_ivector_extractor_id(args.online_ivector_dir)
config_dir = '{0}/configs'.format(args.dir)
var_file = '{0}/vars'.format(config_dir)
variables = common_train_lib.parse_generic_config_vars_file(var_file)
# Set some variables.
try:
model_left_context = variables['model_left_context']
model_right_context = variables['model_right_context']
except KeyError as e:
raise Exception("KeyError {0}: Variables need to be defined in "
"{1}".format(str(e), '{0}/configs'.format(args.dir)))
left_context = args.chunk_left_context + model_left_context
right_context = args.chunk_right_context + model_right_context
left_context_initial = (args.chunk_left_context_initial + model_left_context if
args.chunk_left_context_initial >= 0 else -1)
right_context_final = (args.chunk_right_context_final + model_right_context if
args.chunk_right_context_final >= 0 else -1)
# Initialize as "raw" nnet, prior to training the LDA-like preconditioning
# matrix. This first config just does any initial splicing that we do;
# we do this as it's a convenient way to get the stats for the 'lda-like'
# transform.
if (args.stage <= -4) and os.path.exists(args.dir+"/configs/init.config"):
logger.info("Initializing the network for computing the LDA stats")
common_lib.execute_command(
"""{command} {dir}/log/nnet_init.log \
nnet3-init --srand=-2 {dir}/configs/init.config \
{dir}/init.raw""".format(command=run_opts.command,
dir=args.dir))
default_egs_dir = '{0}/egs'.format(args.dir)
if (args.stage <= -3) and args.egs_dir is None:
logger.info("Generating egs")
if args.use_dense_targets:
target_type = "dense"
try:
num_targets = int(variables['num_targets'])
if (common_lib.get_feat_dim_from_scp(args.targets_scp)
!= num_targets):
raise Exception("Mismatch between num-targets provided to "
"script vs configs")
except KeyError as e:
num_targets = -1
else:
target_type = "sparse"
try:
num_targets = int(variables['num_targets'])
except KeyError as e:
raise Exception("KeyError {0}: Variables need to be defined "
"in {1}".format(
str(e), '{0}/configs'.format(args.dir)))
train_lib.raw_model.generate_egs_using_targets(
data=args.feat_dir, targets_scp=args.targets_scp,
egs_dir=default_egs_dir,
left_context=left_context,
right_context=right_context,
left_context_initial=left_context_initial,
right_context_final=right_context_final,
run_opts=run_opts,
frames_per_eg_str=args.chunk_width,
srand=args.srand,
egs_opts=args.egs_opts,
cmvn_opts=args.cmvn_opts,
online_ivector_dir=args.online_ivector_dir,
samples_per_iter=args.samples_per_iter,
stage=args.egs_stage,
target_type=target_type,
num_targets=num_targets)
if args.egs_dir is None:
egs_dir = default_egs_dir
else:
egs_dir = args.egs_dir
[egs_left_context, egs_right_context,
frames_per_eg_str, num_archives] = (
common_train_lib.verify_egs_dir(egs_dir, feat_dim,
ivector_dim, ivector_id,
left_context, right_context,
left_context_initial,
right_context_final))
if args.chunk_width != frames_per_eg_str:
raise Exception("mismatch between --egs.chunk-width and the frames_per_eg "
"in the egs dir {0} vs {1}".format(args.chunk_width,
frames_per_eg_str))
if args.num_jobs_final > num_archives:
raise Exception('num_jobs_final cannot exceed the number of archives '
'in the egs directory')
# copy the properties of the egs to dir for
# use during decoding
common_train_lib.copy_egs_properties_to_exp_dir(egs_dir, args.dir)
if args.stage <= -2 and os.path.exists(args.dir+"/configs/init.config"):
logger.info('Computing the preconditioning matrix for input features')
train_lib.common.compute_preconditioning_matrix(
args.dir, egs_dir, num_archives, run_opts,
max_lda_jobs=args.max_lda_jobs,
rand_prune=args.rand_prune)
if args.stage <= -1:
logger.info("Preparing the initial network.")
common_train_lib.prepare_initial_network(args.dir, run_opts)
# set num_iters so that as close as possible, we process the data
# $num_epochs times, i.e. $num_iters*$avg_num_jobs) ==
# $num_epochs*$num_archives, where
# avg_num_jobs=(num_jobs_initial+num_jobs_final)/2.
num_archives_to_process = int(args.num_epochs * num_archives)
num_archives_processed = 0
num_iters = ((num_archives_to_process * 2)
/ (args.num_jobs_initial + args.num_jobs_final))
# If do_final_combination is True, compute the set of models_to_combine.
# Otherwise, models_to_combine will be none.
if args.do_final_combination:
models_to_combine = common_train_lib.get_model_combine_iters(
num_iters, args.num_epochs,
num_archives, args.max_models_combine,
args.num_jobs_final)
else:
models_to_combine = None
if (os.path.exists('{0}/valid_diagnostic.scp'.format(egs_dir))):
if (os.path.exists('{0}/valid_diagnostic.egs'.format(egs_dir))):
raise Exception('both {0}/valid_diagnostic.egs and '
'{0}/valid_diagnostic.scp exist.'
'This script expects only one of them to exist.'
''.format(egs_dir))
use_multitask_egs = True
else:
if (not os.path.exists('{0}/valid_diagnostic.egs'
''.format(egs_dir))):
raise Exception('neither {0}/valid_diagnostic.egs nor '
'{0}/valid_diagnostic.scp exist.'
'This script expects one of them.'
''.format(egs_dir))
use_multitask_egs = False
min_deriv_time = None
max_deriv_time_relative = None
if args.deriv_truncate_margin is not None:
min_deriv_time = -args.deriv_truncate_margin - model_left_context
max_deriv_time_relative = \
args.deriv_truncate_margin + model_right_context
logger.info("Training will run for {0} epochs = "
"{1} iterations".format(args.num_epochs, num_iters))
for iter in range(num_iters):
if (args.exit_stage is not None) and (iter == args.exit_stage):
logger.info("Exiting early due to --exit-stage {0}".format(iter))
return
current_num_jobs = int(0.5 + args.num_jobs_initial
+ (args.num_jobs_final - args.num_jobs_initial)
* float(iter) / num_iters)
if args.stage <= iter:
model_file = "{dir}/{iter}.raw".format(dir=args.dir, iter=iter)
lrate = common_train_lib.get_learning_rate(iter, current_num_jobs,
num_iters,
num_archives_processed,
num_archives_to_process,
args.initial_effective_lrate,
args.final_effective_lrate)
# shrinkage_value is a scale on the parameters.
shrinkage_value = 1.0 - (args.proportional_shrink * lrate)
if shrinkage_value <= 0.5:
raise Exception("proportional-shrink={0} is too large, it gives "
"shrink-value={1}".format(args.proportional_shrink,
shrinkage_value))
if args.shrink_value < shrinkage_value:
shrinkage_value = (args.shrink_value
if common_train_lib.should_do_shrinkage(
iter, model_file,
args.shrink_saturation_threshold,
get_raw_nnet_from_am=False)
else shrinkage_value)
percent = num_archives_processed * 100.0 / num_archives_to_process
epoch = (num_archives_processed * args.num_epochs
/ num_archives_to_process)
shrink_info_str = ''
if shrinkage_value != 1.0:
shrink_info_str = 'shrink: {0:0.5f}'.format(shrinkage_value)
logger.info("Iter: {0}/{1} "
"Epoch: {2:0.2f}/{3:0.1f} ({4:0.1f}% complete) "
"lr: {5:0.6f} {6}".format(iter, num_iters - 1,
epoch, args.num_epochs,
percent,
lrate, shrink_info_str))
train_lib.common.train_one_iteration(
dir=args.dir,
iter=iter,
srand=args.srand,
egs_dir=egs_dir,
num_jobs=current_num_jobs,
num_archives_processed=num_archives_processed,
num_archives=num_archives,
learning_rate=lrate,
dropout_edit_string=common_train_lib.get_dropout_edit_string(
args.dropout_schedule,
float(num_archives_processed) / num_archives_to_process,
iter),
train_opts=' '.join(args.train_opts),
shrinkage_value=shrinkage_value,
minibatch_size_str=args.num_chunk_per_minibatch,
min_deriv_time=min_deriv_time,
max_deriv_time_relative=max_deriv_time_relative,
momentum=args.momentum,
max_param_change=args.max_param_change,
shuffle_buffer_size=args.shuffle_buffer_size,
run_opts=run_opts,
get_raw_nnet_from_am=False,
use_multitask_egs=use_multitask_egs,
compute_per_dim_accuracy=args.compute_per_dim_accuracy)
if args.cleanup:
# do a clean up everythin but the last 2 models, under certain
# conditions
common_train_lib.remove_model(
args.dir, iter-2, num_iters, models_to_combine,
args.preserve_model_interval,
get_raw_nnet_from_am=False)
if args.email is not None:
reporting_iter_interval = num_iters * args.reporting_interval
if iter % reporting_iter_interval == 0:
# lets do some reporting
[report, times, data] = (
nnet3_log_parse.generate_acc_logprob_report(args.dir))
message = report
subject = ("Update : Expt {dir} : "
"Iter {iter}".format(dir=args.dir, iter=iter))
common_lib.send_mail(message, subject, args.email)
num_archives_processed = num_archives_processed + current_num_jobs
if args.stage <= num_iters:
if args.do_final_combination:
logger.info("Doing final combination to produce final.raw")
train_lib.common.combine_models(
dir=args.dir, num_iters=num_iters,
models_to_combine=models_to_combine, egs_dir=egs_dir,
minibatch_size_str=args.num_chunk_per_minibatch,
run_opts=run_opts, chunk_width=args.chunk_width,
get_raw_nnet_from_am=False,
compute_per_dim_accuracy=args.compute_per_dim_accuracy,
max_objective_evaluations=args.max_objective_evaluations)
else:
common_lib.force_symlink("{0}.raw".format(num_iters),
"{0}/final.raw".format(args.dir))
if args.compute_average_posteriors and args.stage <= num_iters + 1:
logger.info("Getting average posterior for purposes of "
"adjusting the priors.")
train_lib.common.compute_average_posterior(
dir=args.dir, iter='final', egs_dir=egs_dir,
num_archives=num_archives,
prior_subset_size=args.prior_subset_size, run_opts=run_opts,
get_raw_nnet_from_am=False)
if args.cleanup:
logger.info("Cleaning up the experiment directory "
"{0}".format(args.dir))
remove_egs = args.remove_egs
if args.egs_dir is not None:
# this egs_dir was not created by this experiment so we will not
# delete it
remove_egs = False
common_train_lib.clean_nnet_dir(
nnet_dir=args.dir, num_iters=num_iters, egs_dir=egs_dir,
preserve_model_interval=args.preserve_model_interval,
remove_egs=remove_egs,
get_raw_nnet_from_am=False)
# do some reporting
[report, times, data] = nnet3_log_parse.generate_acc_logprob_report(args.dir)
if args.email is not None:
common_lib.send_mail(report, "Update : Expt {0} : "
"complete".format(args.dir), args.email)
with open("{dir}/accuracy.report".format(dir=args.dir), "w") as f:
f.write(report)
common_lib.execute_command("steps/info/nnet3_dir_info.pl "
"{0}".format(args.dir))
def CheckArgs(args):
if not os.path.exists(args.config_dir):
os.makedirs(args.config_dir)
## Check arguments.
if args.feat_dir is not None:
args.feat_dim = common_lib.get_feat_dim(args.feat_dir)
if args.ali_dir is not None:
args.num_targets = common_lib.get_number_of_leaves_from_tree(
args.ali_dir)
elif args.tree_dir is not None:
args.num_targets = common_lib.get_number_of_leaves_from_tree(
args.tree_dir)
if args.ivector_dir is not None:
args.ivector_dim = common_lib.get_ivector_dim(args.ivector_dir)
if not args.feat_dim > 0:
raise Exception("feat-dim has to be postive")
if not args.num_targets > 0:
print(args.num_targets)
raise Exception("num_targets has to be positive")
if not args.ivector_dim >= 0:
raise Exception("ivector-dim has to be non-negative")
if (args.subset_dim < 0):
raise Exception("--subset-dim has to be non-negative")
if not args.relu_dim is None:
if not args.pnorm_input_dim is None or not args.pnorm_output_dim is None or not args.relu_dim_init is None:
raise Exception(
"--relu-dim argument not compatible with "
"--pnorm-input-dim or --pnorm-output-dim or --relu-dim-init options"
)
args.nonlin_input_dim = args.relu_dim
args.nonlin_output_dim = args.relu_dim
args.nonlin_output_dim_final = None
args.nonlin_output_dim_init = None
args.nonlin_type = 'relu'
elif not args.relu_dim_final is None:
if not args.pnorm_input_dim is None or not args.pnorm_output_dim is None:
raise Exception("--relu-dim-final argument not compatible with "
"--pnorm-input-dim or --pnorm-output-dim options")
if args.relu_dim_init is None:
raise Exception(
"--relu-dim-init argument should also be provided with --relu-dim-final"
)
if args.relu_dim_init > args.relu_dim_final:
raise Exception(
"--relu-dim-init has to be no larger than --relu-dim-final")
args.nonlin_input_dim = None
args.nonlin_output_dim = None
args.nonlin_output_dim_final = args.relu_dim_final
args.nonlin_output_dim_init = args.relu_dim_init
args.nonlin_type = 'relu'
else:
if not args.relu_dim_init is None:
raise Exception("--relu-dim-final argument not compatible with "
"--pnorm-input-dim or --pnorm-output-dim options")
if not args.pnorm_input_dim > 0 or not args.pnorm_output_dim > 0:
raise Exception(
"--relu-dim not set, so expected --pnorm-input-dim and "
"--pnorm-output-dim to be provided.")
args.nonlin_input_dim = args.pnorm_input_dim
args.nonlin_output_dim = args.pnorm_output_dim
if (args.nonlin_input_dim < args.nonlin_output_dim) or (
args.nonlin_input_dim % args.nonlin_output_dim != 0):
raise Exception(
"Invalid --pnorm-input-dim {0} and --pnorm-output-dim {1}".
format(args.nonlin_input_dim, args.nonlin_output_dim))
args.nonlin_output_dim_final = None
args.nonlin_output_dim_init = None
args.nonlin_type = 'pnorm'
if args.add_final_sigmoid and args.include_log_softmax:
raise Exception(
"--include-log-softmax and --add-final-sigmoid cannot both be true."
)
if args.xent_separate_forward_affine and args.add_final_sigmoid:
raise Exception(
"It does not make sense to have --add-final-sigmoid=true when xent-separate-forward-affine is true"
)
if args.add_lda and args.cnn_layer is not None:
args.add_lda = False
warnings.warn("--add-lda is set to false as CNN layers are used.")
if not args.max_change_per_component >= 0 or not args.max_change_per_component_final >= 0:
raise Exception(
"max-change-per-component and max_change-per-component-final should be non-negative"
)
return args
def train(args, run_opts, background_process_handler):
""" The main function for training.
Args:
args: a Namespace object with the required parameters
obtained from the function process_args()
run_opts: RunOpts object obtained from the process_args()
"""
arg_string = pprint.pformat(vars(args))
logger.info("Arguments for the experiment\n{0}".format(arg_string))
# Set some variables.
# num_leaves = common_lib.get_number_of_leaves_from_tree(args.ali_dir)
num_jobs = common_lib.get_number_of_jobs(args.ali_dir)
feat_dim = common_lib.get_feat_dim(args.feat_dir)
ivector_dim = common_lib.get_ivector_dim(args.online_ivector_dir)
# split the training data into parts for individual jobs
# we will use the same number of jobs as that used for alignment
common_lib.split_data(args.feat_dir, num_jobs)
shutil.copy('{0}/tree'.format(args.ali_dir), args.dir)
with open('{0}/num_jobs'.format(args.dir), 'w') as f:
f.write(str(num_jobs))
config_dir = '{0}/configs'.format(args.dir)
var_file = '{0}/vars'.format(config_dir)
variables = common_train_lib.parse_generic_config_vars_file(var_file)
# Set some variables.
try:
model_left_context = variables['model_left_context']
model_right_context = variables['model_right_context']
# this is really the number of times we add layers to the network for
# discriminative pretraining
num_hidden_layers = variables['num_hidden_layers']
except KeyError as e:
raise Exception("KeyError {0}: Variables need to be defined in "
"{1}".format(str(e), '{0}/configs'.format(args.dir)))
left_context = args.chunk_left_context + model_left_context
right_context = args.chunk_right_context + model_right_context
# Initialize as "raw" nnet, prior to training the LDA-like preconditioning
# matrix. This first config just does any initial splicing that we do;
# we do this as it's a convenient way to get the stats for the 'lda-like'
# transform.
if (args.stage <= -5):
logger.info("Initializing a basic network for estimating "
"preconditioning matrix")
common_lib.run_job(
"""{command} {dir}/log/nnet_init.log \
nnet3-init --srand=-2 {dir}/configs/init.config \
{dir}/init.raw""".format(command=run_opts.command,
dir=args.dir))
default_egs_dir = '{0}/egs'.format(args.dir)
if (args.stage <= -4) and args.egs_dir is None:
logger.info("Generating egs")
train_lib.acoustic_model.generate_egs(
data=args.feat_dir, alidir=args.ali_dir, egs_dir=default_egs_dir,
left_context=left_context, right_context=right_context,
run_opts=run_opts,
frames_per_eg=args.frames_per_eg,
srand=args.srand,
egs_opts=args.egs_opts,
cmvn_opts=args.cmvn_opts,
online_ivector_dir=args.online_ivector_dir,
samples_per_iter=args.samples_per_iter,
transform_dir=args.transform_dir,
stage=args.egs_stage)
if args.egs_dir is None:
egs_dir = default_egs_dir
else:
egs_dir = args.egs_dir
[egs_left_context, egs_right_context,
frames_per_eg, num_archives] = (
common_train_lib.verify_egs_dir(egs_dir, feat_dim, ivector_dim,
left_context, right_context))
assert(args.frames_per_eg == frames_per_eg)
if (args.num_jobs_final > num_archives):
raise Exception('num_jobs_final cannot exceed the number of archives '
'in the egs directory')
# copy the properties of the egs to dir for
# use during decoding
common_train_lib.copy_egs_properties_to_exp_dir(egs_dir, args.dir)
if (args.stage <= -3):
logger.info('Computing the preconditioning matrix for input features')
train_lib.common.compute_preconditioning_matrix(
args.dir, egs_dir, num_archives, run_opts,
max_lda_jobs=args.max_lda_jobs,
rand_prune=args.rand_prune)
if (args.stage <= -2):
logger.info("Computing initial vector for FixedScaleComponent before"
" softmax, using priors^{prior_scale} and rescaling to"
" average 1".format(
prior_scale=args.presoftmax_prior_scale_power))
common_train_lib.compute_presoftmax_prior_scale(
args.dir, args.ali_dir, num_jobs, run_opts,
presoftmax_prior_scale_power=args.presoftmax_prior_scale_power)
if (args.stage <= -1):
logger.info("Preparing the initial acoustic model.")
train_lib.acoustic_model.prepare_initial_acoustic_model(
args.dir, args.ali_dir, run_opts)
# set num_iters so that as close as possible, we process the data
# $num_epochs times, i.e. $num_iters*$avg_num_jobs) ==
# $num_epochs*$num_archives, where
# avg_num_jobs=(num_jobs_initial+num_jobs_final)/2.
num_archives_expanded = num_archives * args.frames_per_eg
num_archives_to_process = args.num_epochs * num_archives_expanded
num_archives_processed = 0
num_iters = ((num_archives_to_process * 2)
/ (args.num_jobs_initial + args.num_jobs_final))
models_to_combine = common_train_lib.verify_iterations(
num_iters, args.num_epochs,
num_hidden_layers, num_archives_expanded,
args.max_models_combine, args.add_layers_period,
args.num_jobs_final)
def learning_rate(iter, current_num_jobs, num_archives_processed):
return common_train_lib.get_learning_rate(iter, current_num_jobs,
num_iters,
num_archives_processed,
num_archives_to_process,
args.initial_effective_lrate,
args.final_effective_lrate)
logger.info("Training will run for {0} epochs = "
"{1} iterations".format(args.num_epochs, num_iters))
for iter in range(num_iters):
if (args.exit_stage is not None) and (iter == args.exit_stage):
logger.info("Exiting early due to --exit-stage {0}".format(iter))
return
current_num_jobs = int(0.5 + args.num_jobs_initial
+ (args.num_jobs_final - args.num_jobs_initial)
* float(iter) / num_iters)
if args.stage <= iter:
logger.info("On iteration {0}, learning rate is {1}.".format(
iter, learning_rate(iter, current_num_jobs,
num_archives_processed)))
train_lib.common.train_one_iteration(
dir=args.dir,
iter=iter,
srand=args.srand,
egs_dir=egs_dir,
num_jobs=current_num_jobs,
num_archives_processed=num_archives_processed,
num_archives=num_archives,
learning_rate=learning_rate(iter, current_num_jobs,
num_archives_processed),
minibatch_size=args.minibatch_size,
frames_per_eg=args.frames_per_eg,
num_hidden_layers=num_hidden_layers,
add_layers_period=args.add_layers_period,
left_context=left_context,
right_context=right_context,
momentum=args.momentum,
max_param_change=args.max_param_change,
shuffle_buffer_size=args.shuffle_buffer_size,
run_opts=run_opts,
background_process_handler=background_process_handler)
if args.cleanup:
# do a clean up everythin but the last 2 models, under certain
# conditions
common_train_lib.remove_model(
args.dir, iter-2, num_iters, models_to_combine,
args.preserve_model_interval)
if args.email is not None:
reporting_iter_interval = num_iters * args.reporting_interval
if iter % reporting_iter_interval == 0:
# lets do some reporting
[report, times, data] = (
nnet3_log_parse.generate_accuracy_report(args.dir))
message = report
subject = ("Update : Expt {dir} : "
"Iter {iter}".format(dir=args.dir, iter=iter))
common_lib.send_mail(message, subject, args.email)
num_archives_processed = num_archives_processed + current_num_jobs
if args.stage <= num_iters:
logger.info("Doing final combination to produce final.mdl")
train_lib.common.combine_models(
dir=args.dir, num_iters=num_iters,
models_to_combine=models_to_combine,
egs_dir=egs_dir,
left_context=left_context, right_context=right_context,
run_opts=run_opts,
background_process_handler=background_process_handler)
if args.stage <= num_iters + 1:
logger.info("Getting average posterior for purposes of "
"adjusting the priors.")
avg_post_vec_file = train_lib.common.compute_average_posterior(
dir=args.dir, iter='combined', egs_dir=egs_dir,
num_archives=num_archives,
left_context=left_context, right_context=right_context,
prior_subset_size=args.prior_subset_size, run_opts=run_opts)
logger.info("Re-adjusting priors based on computed posteriors")
combined_model = "{dir}/combined.mdl".format(dir=args.dir)
final_model = "{dir}/final.mdl".format(dir=args.dir)
train_lib.common.adjust_am_priors(args.dir, combined_model,
avg_post_vec_file, final_model,
run_opts)
if args.cleanup:
logger.info("Cleaning up the experiment directory "
"{0}".format(args.dir))
remove_egs = args.remove_egs
if args.egs_dir is not None:
# this egs_dir was not created by this experiment so we will not
# delete it
remove_egs = False
common_train_lib.clean_nnet_dir(
nnet_dir=args.dir, num_iters=num_iters, egs_dir=egs_dir,
preserve_model_interval=args.preserve_model_interval,
remove_egs=remove_egs)
# do some reporting
[report, times, data] = nnet3_log_parse.generate_accuracy_report(args.dir)
if args.email is not None:
common_lib.send_mail(report, "Update : Expt {0} : "
"complete".format(args.dir), args.email)
with open("{dir}/accuracy.report".format(dir=args.dir), "w") as f:
f.write(report)
common_lib.run_job("steps/info/nnet3_dir_info.pl "
"{0}".format(args.dir))
def train(args, run_opts, background_process_handler):
""" The main function for training.
Args:
args: a Namespace object with the required parameters
obtained from the function process_args()
run_opts: RunOpts object obtained from the process_args()
"""
arg_string = pprint.pformat(vars(args))
logger.info("Arguments for the experiment\n{0}".format(arg_string))
# Set some variables.
feat_dim = common_lib.get_feat_dim(args.feat_dir)
ivector_dim = common_lib.get_ivector_dim(args.online_ivector_dir)
config_dir = '{0}/configs'.format(args.dir)
var_file = '{0}/vars'.format(config_dir)
variables = common_train_lib.parse_generic_config_vars_file(var_file)
# Set some variables.
try:
model_left_context = variables['model_left_context']
model_right_context = variables['model_right_context']
# this is really the number of times we add layers to the network for
# discriminative pretraining
num_hidden_layers = variables['num_hidden_layers']
add_lda = common_lib.str_to_bool(variables['add_lda'])
include_log_softmax = common_lib.str_to_bool(
variables['include_log_softmax'])
except KeyError as e:
raise Exception("KeyError {0}: Variables need to be defined in "
"{1}".format(str(e), '{0}/configs'.format(args.dir)))
left_context = args.chunk_left_context + model_left_context
right_context = args.chunk_right_context + model_right_context
# Initialize as "raw" nnet, prior to training the LDA-like preconditioning
# matrix. This first config just does any initial splicing that we do;
# we do this as it's a convenient way to get the stats for the 'lda-like'
# transform.
if (args.stage <= -5):
logger.info("Initializing a basic network")
common_lib.run_job(
"""{command} {dir}/log/nnet_init.log \
nnet3-init --srand=-2 {dir}/configs/init.config \
{dir}/init.raw""".format(command=run_opts.command,
dir=args.dir))
default_egs_dir = '{0}/egs'.format(args.dir)
if (args.stage <= -4) and args.egs_dir is None:
logger.info("Generating egs")
if args.use_dense_targets:
target_type = "dense"
try:
num_targets = int(variables['num_targets'])
if (common_lib.get_feat_dim_from_scp(args.targets_scp)
!= num_targets):
raise Exception("Mismatch between num-targets provided to "
"script vs configs")
except KeyError as e:
num_targets = -1
else:
target_type = "sparse"
try:
num_targets = int(variables['num_targets'])
except KeyError as e:
raise Exception("KeyError {0}: Variables need to be defined "
"in {1}".format(
str(e), '{0}/configs'.format(args.dir)))
train_lib.raw_model.generate_egs_using_targets(
data=args.feat_dir, targets_scp=args.targets_scp,
egs_dir=default_egs_dir,
left_context=left_context, right_context=right_context,
valid_left_context=left_context, valid_right_context=right_context,
run_opts=run_opts,
frames_per_eg=args.frames_per_eg,
srand=args.srand,
egs_opts=args.egs_opts,
cmvn_opts=args.cmvn_opts,
online_ivector_dir=args.online_ivector_dir,
samples_per_iter=args.samples_per_iter,
transform_dir=args.transform_dir,
stage=args.egs_stage,
target_type=target_type,
num_targets=num_targets)
if args.egs_dir is None:
egs_dir = default_egs_dir
else:
egs_dir = args.egs_dir
[egs_left_context, egs_right_context,
frames_per_eg, num_archives] = (
common_train_lib.verify_egs_dir(egs_dir, feat_dim, ivector_dim,
left_context, right_context))
assert(args.frames_per_eg == frames_per_eg)
if (args.num_jobs_final > num_archives):
raise Exception('num_jobs_final cannot exceed the number of archives '
'in the egs directory')
# copy the properties of the egs to dir for
# use during decoding
common_train_lib.copy_egs_properties_to_exp_dir(egs_dir, args.dir)
if (add_lda and args.stage <= -3):
logger.info('Computing the preconditioning matrix for input features')
train_lib.common.compute_preconditioning_matrix(
args.dir, egs_dir, num_archives, run_opts,
max_lda_jobs=args.max_lda_jobs,
rand_prune=args.rand_prune)
if (args.stage <= -1):
logger.info("Preparing the initial network.")
common_train_lib.prepare_initial_network(args.dir, run_opts)
# set num_iters so that as close as possible, we process the data
# $num_epochs times, i.e. $num_iters*$avg_num_jobs) ==
# $num_epochs*$num_archives, where
# avg_num_jobs=(num_jobs_initial+num_jobs_final)/2.
num_archives_expanded = num_archives * args.frames_per_eg
num_archives_to_process = args.num_epochs * num_archives_expanded
num_archives_processed = 0
num_iters = ((num_archives_to_process * 2)
/ (args.num_jobs_initial + args.num_jobs_final))
models_to_combine = common_train_lib.verify_iterations(
num_iters, args.num_epochs,
num_hidden_layers, num_archives_expanded,
args.max_models_combine, args.add_layers_period,
args.num_jobs_final)
def learning_rate(iter, current_num_jobs, num_archives_processed):
return common_train_lib.get_learning_rate(iter, current_num_jobs,
num_iters,
num_archives_processed,
num_archives_to_process,
args.initial_effective_lrate,
args.final_effective_lrate)
logger.info("Training will run for {0} epochs = "
"{1} iterations".format(args.num_epochs, num_iters))
for iter in range(num_iters):
if (args.exit_stage is not None) and (iter == args.exit_stage):
logger.info("Exiting early due to --exit-stage {0}".format(iter))
return
current_num_jobs = int(0.5 + args.num_jobs_initial
+ (args.num_jobs_final - args.num_jobs_initial)
* float(iter) / num_iters)
if args.stage <= iter:
train_lib.common.train_one_iteration(
dir=args.dir,
iter=iter,
srand=args.srand,
egs_dir=egs_dir,
num_jobs=current_num_jobs,
num_archives_processed=num_archives_processed,
num_archives=num_archives,
learning_rate=learning_rate(iter, current_num_jobs,
num_archives_processed),
dropout_edit_string=common_lib.get_dropout_edit_string(
args.dropout_schedule,
float(num_archives_processed) / num_archives_to_process,
iter),
minibatch_size=args.minibatch_size,
frames_per_eg=args.frames_per_eg,
num_hidden_layers=num_hidden_layers,
add_layers_period=args.add_layers_period,
left_context=left_context,
right_context=right_context,
momentum=args.momentum,
max_param_change=args.max_param_change,
shuffle_buffer_size=args.shuffle_buffer_size,
run_opts=run_opts,
get_raw_nnet_from_am=False,
background_process_handler=background_process_handler)
if args.cleanup:
# do a clean up everythin but the last 2 models, under certain
# conditions
common_train_lib.remove_model(
args.dir, iter-2, num_iters, models_to_combine,
args.preserve_model_interval,
get_raw_nnet_from_am=False)
if args.email is not None:
reporting_iter_interval = num_iters * args.reporting_interval
if iter % reporting_iter_interval == 0:
# lets do some reporting
[report, times, data] = (
nnet3_log_parse.generate_accuracy_report(args.dir))
message = report
subject = ("Update : Expt {dir} : "
"Iter {iter}".format(dir=args.dir, iter=iter))
common_lib.send_mail(message, subject, args.email)
num_archives_processed = num_archives_processed + current_num_jobs
if args.stage <= num_iters:
logger.info("Doing final combination to produce final.raw")
train_lib.common.combine_models(
dir=args.dir, num_iters=num_iters,
models_to_combine=models_to_combine, egs_dir=egs_dir,
left_context=left_context, right_context=right_context,
run_opts=run_opts,
background_process_handler=background_process_handler,
get_raw_nnet_from_am=False)
if include_log_softmax and args.stage <= num_iters + 1:
logger.info("Getting average posterior for purposes of "
"adjusting the priors.")
train_lib.common.compute_average_posterior(
dir=args.dir, iter='final', egs_dir=egs_dir,
num_archives=num_archives,
left_context=left_context, right_context=right_context,
prior_subset_size=args.prior_subset_size, run_opts=run_opts,
get_raw_nnet_from_am=False)
if args.cleanup:
logger.info("Cleaning up the experiment directory "
"{0}".format(args.dir))
remove_egs = args.remove_egs
if args.egs_dir is not None:
# this egs_dir was not created by this experiment so we will not
# delete it
remove_egs = False
common_train_lib.clean_nnet_dir(
nnet_dir=args.dir, num_iters=num_iters, egs_dir=egs_dir,
preserve_model_interval=args.preserve_model_interval,
remove_egs=remove_egs,
get_raw_nnet_from_am=False)
# do some reporting
[report, times, data] = nnet3_log_parse.generate_accuracy_report(args.dir)
if args.email is not None:
common_lib.send_mail(report, "Update : Expt {0} : "
"complete".format(args.dir), args.email)
with open("{dir}/accuracy.report".format(dir=args.dir), "w") as f:
f.write(report)
common_lib.run_job("steps/info/nnet3_dir_info.pl "
"{0}".format(args.dir))
def train(args, run_opts):
""" The main function for training.
Args:
args: a Namespace object with the required parameters
obtained from the function process_args()
run_opts: RunOpts object obtained from the process_args()
"""
arg_string = pprint.pformat(vars(args))
logger.info("Arguments for the experiment\n{0}".format(arg_string))
# Set some variables.
# note, feat_dim gets set to 0 if args.feat_dir is unset (None).
feat_dim = common_lib.get_feat_dim(args.feat_dir)
ivector_dim = common_lib.get_ivector_dim(args.online_ivector_dir)
ivector_id = common_lib.get_ivector_extractor_id(args.online_ivector_dir)
config_dir = '{0}/configs'.format(args.dir)
var_file = '{0}/vars'.format(config_dir)
variables = common_train_lib.parse_generic_config_vars_file(var_file)
# Set some variables.
try:
model_left_context = variables['model_left_context']
model_right_context = variables['model_right_context']
except KeyError as e:
raise Exception("KeyError {0}: Variables need to be defined in "
"{1}".format(str(e), '{0}/configs'.format(args.dir)))
left_context = model_left_context
right_context = model_right_context
# Initialize as "raw" nnet, prior to training the LDA-like preconditioning
# matrix. This first config just does any initial splicing that we do;
# we do this as it's a convenient way to get the stats for the 'lda-like'
# transform.
if (args.stage <= -5) and os.path.exists(args.dir +
"/configs/init.config"):
logger.info("Initializing the network for computing the LDA stats")
common_lib.execute_command("""{command} {dir}/log/nnet_init.log \
nnet3-init --srand=-2 {dir}/configs/init.config \
{dir}/init.raw""".format(command=run_opts.command,
dir=args.dir))
default_egs_dir = '{0}/egs'.format(args.dir)
if (args.stage <= -4) and args.egs_dir is None:
if args.targets_scp is None or args.feat_dir is None:
raise Exception(
"If you don't supply the --egs-dir option, the "
"--targets-scp and --feat-dir options are required.")
logger.info("Generating egs")
if args.use_dense_targets:
target_type = "dense"
try:
num_targets = int(variables['num_targets'])
if (common_lib.get_feat_dim_from_scp(args.targets_scp) !=
num_targets):
raise Exception("Mismatch between num-targets provided to "
"script vs configs")
except KeyError as e:
num_targets = -1
else:
target_type = "sparse"
try:
num_targets = int(variables['num_targets'])
except KeyError as e:
raise Exception("KeyError {0}: Variables need to be defined "
"in {1}".format(str(e), '{0}/configs'.format(
args.dir)))
train_lib.raw_model.generate_egs_using_targets(
data=args.feat_dir,
targets_scp=args.targets_scp,
egs_dir=default_egs_dir,
left_context=left_context,
right_context=right_context,
run_opts=run_opts,
frames_per_eg_str=str(args.frames_per_eg),
srand=args.srand,
egs_opts=args.egs_opts,
cmvn_opts=args.cmvn_opts,
online_ivector_dir=args.online_ivector_dir,
samples_per_iter=args.samples_per_iter,
transform_dir=args.transform_dir,
stage=args.egs_stage,
target_type=target_type,
num_targets=num_targets)
if args.egs_dir is None:
egs_dir = default_egs_dir
else:
egs_dir = args.egs_dir
[egs_left_context, egs_right_context, frames_per_eg_str,
num_archives] = (common_train_lib.verify_egs_dir(egs_dir, feat_dim,
ivector_dim, ivector_id,
left_context,
right_context))
assert str(args.frames_per_eg) == frames_per_eg_str
if args.num_jobs_final > num_archives:
raise Exception('num_jobs_final cannot exceed the number of archives '
'in the egs directory')
# copy the properties of the egs to dir for
# use during decoding
common_train_lib.copy_egs_properties_to_exp_dir(egs_dir, args.dir)
if args.stage <= -3 and os.path.exists(args.dir + "/configs/init.config"):
logger.info('Computing the preconditioning matrix for input features')
train_lib.common.compute_preconditioning_matrix(
args.dir,
egs_dir,
num_archives,
run_opts,
max_lda_jobs=args.max_lda_jobs,
rand_prune=args.rand_prune)
if args.stage <= -1:
logger.info("Preparing the initial network.")
common_train_lib.prepare_initial_network(args.dir, run_opts)
# set num_iters so that as close as possible, we process the data
# $num_epochs times, i.e. $num_iters*$avg_num_jobs) ==
# $num_epochs*$num_archives, where
# avg_num_jobs=(num_jobs_initial+num_jobs_final)/2.
num_archives_expanded = num_archives * args.frames_per_eg
num_archives_to_process = int(args.num_epochs * num_archives_expanded)
num_archives_processed = 0
num_iters = ((num_archives_to_process * 2) /
(args.num_jobs_initial + args.num_jobs_final))
# If do_final_combination is True, compute the set of models_to_combine.
# Otherwise, models_to_combine will be none.
if args.do_final_combination:
models_to_combine = common_train_lib.get_model_combine_iters(
num_iters, args.num_epochs, num_archives_expanded,
args.max_models_combine, args.num_jobs_final)
else:
models_to_combine = None
if os.path.exists('{0}/valid_diagnostic.scp'.format(egs_dir)):
if os.path.exists('{0}/valid_diagnostic.egs'.format(egs_dir)):
raise Exception('both {0}/valid_diagnostic.egs and '
'{0}/valid_diagnostic.scp exist.'
'This script expects only one of them to exist.'
''.format(egs_dir))
use_multitask_egs = True
else:
if not os.path.exists('{0}/valid_diagnostic.egs'.format(egs_dir)):
raise Exception('neither {0}/valid_diagnostic.egs nor '
'{0}/valid_diagnostic.scp exist.'
'This script expects one of them.'
''.format(egs_dir))
use_multitask_egs = False
logger.info("Training will run for {0} epochs = "
"{1} iterations".format(args.num_epochs, num_iters))
for iter in range(num_iters):
if (args.exit_stage is not None) and (iter == args.exit_stage):
logger.info("Exiting early due to --exit-stage {0}".format(iter))
return
current_num_jobs = int(0.5 + args.num_jobs_initial +
(args.num_jobs_final - args.num_jobs_initial) *
float(iter) / num_iters)
if args.stage <= iter:
lrate = common_train_lib.get_learning_rate(
iter, current_num_jobs, num_iters, num_archives_processed,
num_archives_to_process, args.initial_effective_lrate,
args.final_effective_lrate)
shrinkage_value = 1.0 - (args.proportional_shrink * lrate)
if shrinkage_value <= 0.5:
raise Exception(
"proportional-shrink={0} is too large, it gives "
"shrink-value={1}".format(args.proportional_shrink,
shrinkage_value))
train_lib.common.train_one_iteration(
dir=args.dir,
iter=iter,
srand=args.srand,
egs_dir=egs_dir,
num_jobs=current_num_jobs,
num_archives_processed=num_archives_processed,
num_archives=num_archives,
learning_rate=lrate,
dropout_edit_string=common_train_lib.get_dropout_edit_string(
args.dropout_schedule,
float(num_archives_processed) / num_archives_to_process,
iter),
minibatch_size_str=args.minibatch_size,
frames_per_eg=args.frames_per_eg,
momentum=args.momentum,
max_param_change=args.max_param_change,
shrinkage_value=shrinkage_value,
shuffle_buffer_size=args.shuffle_buffer_size,
run_opts=run_opts,
get_raw_nnet_from_am=False,
image_augmentation_opts=args.image_augmentation_opts,
use_multitask_egs=use_multitask_egs,
backstitch_training_scale=args.backstitch_training_scale,
backstitch_training_interval=args.backstitch_training_interval)
if args.cleanup:
# do a clean up everything but the last 2 models, under certain
# conditions
common_train_lib.remove_model(args.dir,
iter - 2,
num_iters,
models_to_combine,
args.preserve_model_interval,
get_raw_nnet_from_am=False)
if args.email is not None:
reporting_iter_interval = num_iters * args.reporting_interval
if iter % reporting_iter_interval == 0:
# lets do some reporting
[report, times,
data] = (nnet3_log_parse.generate_acc_logprob_report(
args.dir))
message = report
subject = ("Update : Expt {dir} : "
"Iter {iter}".format(dir=args.dir, iter=iter))
common_lib.send_mail(message, subject, args.email)
num_archives_processed = num_archives_processed + current_num_jobs
if args.stage <= num_iters:
if args.do_final_combination:
logger.info("Doing final combination to produce final.raw")
train_lib.common.combine_models(
dir=args.dir,
num_iters=num_iters,
models_to_combine=models_to_combine,
egs_dir=egs_dir,
minibatch_size_str=args.minibatch_size,
run_opts=run_opts,
get_raw_nnet_from_am=False,
max_objective_evaluations=args.max_objective_evaluations,
use_multitask_egs=use_multitask_egs)
else:
common_lib.force_symlink("{0}.raw".format(num_iters),
"{0}/final.raw".format(args.dir))
if args.compute_average_posteriors and args.stage <= num_iters + 1:
logger.info("Getting average posterior for output-node 'output'.")
train_lib.common.compute_average_posterior(
dir=args.dir,
iter='final',
egs_dir=egs_dir,
num_archives=num_archives,
prior_subset_size=args.prior_subset_size,
run_opts=run_opts,
get_raw_nnet_from_am=False)
if args.cleanup:
logger.info("Cleaning up the experiment directory "
"{0}".format(args.dir))
remove_egs = args.remove_egs
if args.egs_dir is not None:
# this egs_dir was not created by this experiment so we will not
# delete it
remove_egs = False
common_train_lib.clean_nnet_dir(
nnet_dir=args.dir,
num_iters=num_iters,
egs_dir=egs_dir,
preserve_model_interval=args.preserve_model_interval,
remove_egs=remove_egs,
get_raw_nnet_from_am=False)
# do some reporting
outputs_list = common_train_lib.get_outputs_list(
"{0}/final.raw".format(args.dir), get_raw_nnet_from_am=False)
if 'output' in outputs_list:
[report, times,
data] = nnet3_log_parse.generate_acc_logprob_report(args.dir)
if args.email is not None:
common_lib.send_mail(
report, "Update : Expt {0} : "
"complete".format(args.dir), args.email)
with open(
"{dir}/accuracy.{output_name}.report".format(
dir=args.dir, output_name="output"), "w") as f:
f.write(report)
common_lib.execute_command("steps/info/nnet3_dir_info.pl "
"{0}".format(args.dir))
def egs_generate(args, run_opts):
""" The main function for egs_generate.
Args:
args: a Namespace object with the required parameters
obtained from the function process_args()
run_opts: RunOpts object obtained from the process_args()
"""
arg_string = pprint.pformat(vars(args))
logger.info("Arguments for the experiment\n{0}".format(arg_string))
# Check files
chain_lib.check_for_required_files(args.feat_dir, args.tree_dir,
args.lat_dir)
# Set some variables.
num_jobs = common_lib.get_number_of_jobs(args.tree_dir)
feat_dim = common_lib.get_feat_dim(args.feat_dir)
ivector_dim = common_lib.get_ivector_dim(args.online_ivector_dir)
ivector_id = common_lib.get_ivector_extractor_id(args.online_ivector_dir)
# split the training data into parts for individual jobs
# we will use the same number of jobs as that used for alignment
common_lib.execute_command("utils/split_data.sh {0} {1}"
"".format(args.feat_dir, num_jobs))
with open('{0}/num_jobs'.format(args.dir), 'w') as f:
f.write(str(num_jobs))
if args.input_model is None:
config_dir = '{0}/configs'.format(args.dir)
var_file = '{0}/vars'.format(config_dir)
variables = common_train_lib.parse_generic_config_vars_file(var_file)
else:
# If args.input_model is specified, the model left and right contexts
# are computed using input_model.
variables = common_train_lib.get_input_model_info(args.input_model)
# Set some variables.
try:
model_left_context = variables['model_left_context']
model_right_context = variables['model_right_context']
except KeyError as e:
raise Exception("KeyError {0}: Variables need to be defined in "
"{1}".format(str(e), '{0}/configs'.format(args.dir)))
left_context = args.chunk_left_context + model_left_context
right_context = args.chunk_right_context + model_right_context
left_context_initial = (args.chunk_left_context_initial + model_left_context if
args.chunk_left_context_initial >= 0 else -1)
right_context_final = (args.chunk_right_context_final + model_right_context if
args.chunk_right_context_final >= 0 else -1)
egs_left_context = left_context + args.frame_subsampling_factor / 2
egs_right_context = right_context + args.frame_subsampling_factor / 2
# note: the '+ args.frame_subsampling_factor / 2' is to allow for the
# fact that we'll be shifting the data slightly during training to give
# variety to the training data.
egs_left_context_initial = (left_context_initial +
args.frame_subsampling_factor / 2 if
left_context_initial >= 0 else -1)
egs_right_context_final = (right_context_final +
args.frame_subsampling_factor / 2 if
right_context_final >= 0 else -1)
default_egs_dir = '{0}/egs_{1}'.format(args.dir, args.suffix)
if ((args.stage <= -3) and args.egs_dir is None):
logger.info("Generating egs")
if (not os.path.exists("{0}/den.fst".format(args.dir)) or
not os.path.exists("{0}/normalization.fst".format(args.dir)) or
not os.path.exists("{0}/tree".format(args.dir))):
raise Exception("Chain egs generation expects {0}/den.fst, "
"{0}/normalization.fst and {0}/tree "
"to exist.".format(args.dir))
# this is where get_egs.sh is called.
chain_lib.generate_chain_egs(
dir=args.dir, data=args.feat_dir,
lat_dir=args.lat_dir, egs_dir=default_egs_dir,
left_context=egs_left_context,
right_context=egs_right_context,
left_context_initial=egs_left_context_initial,
right_context_final=egs_right_context_final,
run_opts=run_opts,
left_tolerance=args.left_tolerance,
right_tolerance=args.right_tolerance,
frame_subsampling_factor=args.frame_subsampling_factor,
alignment_subsampling_factor=args.alignment_subsampling_factor,
frames_per_eg_str=args.chunk_width,
srand=args.srand,
egs_opts=args.egs_opts,
cmvn_opts=args.cmvn_opts,
online_ivector_dir=args.online_ivector_dir,
frames_per_iter=args.frames_per_iter,
transform_dir=args.transform_dir,
stage=args.egs_stage)
if args.egs_dir is None:
egs_dir = default_egs_dir
else:
egs_dir = args.egs_dir
[egs_left_context, egs_right_context,
frames_per_eg_str, num_archives] = (
common_train_lib.verify_egs_dir(egs_dir, feat_dim,
ivector_dim, ivector_id,
egs_left_context, egs_right_context,
egs_left_context_initial,
egs_right_context_final))
assert(args.chunk_width == frames_per_eg_str)
num_archives_expanded = num_archives * args.frame_subsampling_factor