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 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.
config_dir = '{0}/configs'.format(args.dir)
am_var_file = '{0}/vars_am'.format(config_dir)
xvec_var_file = '{0}/vars_xvec'.format(config_dir)
am_variables = common_train_lib.parse_generic_config_vars_file(am_var_file)
xvec_variables = common_train_lib.parse_generic_config_vars_file(xvec_var_file)
# Set some variables.
try:
am_model_left_context = am_variables['model_left_context']
am_model_right_context = am_variables['model_right_context']
xvec_model_left_context = xvec_variables['model_left_context']
xvec_model_right_context = xvec_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)))
am_left_context = am_model_left_context
am_right_context = am_model_right_context
xvec_left_context = xvec_model_left_context
xvec_right_context = xvec_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 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))
am_egs_dir = args.am_egs_dir
xvec_egs_dir = args.xvec_egs_dir
am_output_name = args.am_output_name
xvec_output_name = args.xvec_output_name
am_weight = args.am_weight
xvec_weight = args.xvec_weight
feat_dim = int(common_lib.get_command_stdout("cat {0}/info/feat_dim".format(am_egs_dir)))
num_archives = int(common_lib.get_command_stdout("cat {0}/info/num_archives".format(am_egs_dir)))
tmp_feat_dim = int(common_lib.get_command_stdout("cat {0}/info/feat_dim".format(xvec_egs_dir)))
tmp_num_archives = int(common_lib.get_command_stdout("cat {0}/info/num_archives".format(xvec_egs_dir)))
# frames_per_eg is no longer a parameter but load from am_egs/info/frames_per_eg
am_frames_per_eg = int(common_lib.get_command_stdout("cat {0}/info/frames_per_eg".format(am_egs_dir)))
if feat_dim != tmp_feat_dim or num_archives*am_frames_per_eg != tmp_num_archives:
raise Exception('The am egs and xvec egs do not match')
if args.num_jobs_final > num_archives:
raise Exception('num_jobs_final cannot exceed the number of archives '
'in the egs directory')
# # No need to copy files for decoding
# common_train_lib.copy_egs_properties_to_exp_dir(am_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 * am_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_cvector_one_iteration(
dir=args.dir,
iter=iter,
srand=args.srand,
am_output_name=am_output_name,
am_weight=am_weight,
am_egs_dir=am_egs_dir,
xvec_output_name=xvec_output_name,
xvec_weight=xvec_weight,
xvec_egs_dir=xvec_egs_dir,
num_jobs=current_num_jobs,
num_archives_processed=num_archives_processed,
num_archives=num_archives,
learning_rate=lrate,
minibatch_size_str=args.minibatch_size,
momentum=args.momentum,
max_param_change=args.max_param_change,
shuffle_buffer_size=args.shuffle_buffer_size,
run_opts=run_opts,
am_frames_per_eg=am_frames_per_eg,
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,
get_raw_nnet_from_am=False,
backstitch_training_scale=args.backstitch_training_scale,
backstitch_training_interval=args.backstitch_training_interval)
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
# when we do final combination, just use the xvector egs
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=xvec_egs_dir,
minibatch_size_str="64", run_opts=run_opts,
get_raw_nnet_from_am=False,
max_objective_evaluations=args.max_objective_evaluations,
use_egs=True)
# sum_to_one_penalty=args.combine_sum_to_one_penalty,
else:
common_lib.force_symlink("{0}.raw".format(num_iters),
"{0}/final.raw".format(args.dir))
if args.cleanup:
logger.info("Cleaning up the experiment directory "
"{0}".format(args.dir))
remove_egs = False
common_train_lib.clean_nnet_dir(
nnet_dir=args.dir, num_iters=num_iters, egs_dir=am_egs_dir,
preserve_model_interval=args.preserve_model_interval,
remove_egs=remove_egs,
get_raw_nnet_from_am=False)
# TODO: we may trace other output nodes expect for "output"
# 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("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))
# 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))
