def remap_var_list(self, var_list):
"""Map old vars in checkpoint to new vars in current session.
Args:
var_list (dict): var names mapped to variables (or some related
quantity, such as variable shapes).
Returns:
dict: New var names mapped to the corresponding restored var.
Examples:
>>>var_list
{'Weights': <tf.Variable>}
>>>self.load_param_dict
{'Weights': 'Filters'}
>>>self.remap_var_list(var_list)
{'Filters': <tf.Variable>}
"""
if self.load_param_dict is None:
log.info('No variable mapping specified.')
return var_list
for old_name, new_name in self.load_param_dict.items():
for name in var_list:
if old_name == name:
var_list[old_name] = var_list.pop(old_name)
break
return var_list
def create_train_tpu_config(model_dir,
model_params,
tpu_name,
gcp_project,
steps_per_checkpoint,
tpu_zone=DEFAULT_TPU_ZONE,
num_shards=DEFAULT_NUM_SHARDS,
keep_checkpoint_max=5,
iterations_per_loop=DEFAULT_ITERATIONS_PER_LOOP):
tpu_cluster_resolver = (tpu_cluster_resolver_lib.TPUClusterResolver(
tpu=[tpu_name], zone=tpu_zone, project=gcp_project))
if iterations_per_loop == -1 or (
steps_per_checkpoint is not None
and steps_per_checkpoint < iterations_per_loop):
log.info(
'Setting iterations_per_loop ({}) to be the same as steps_per_checkpoint ({}).'
.format(iterations_per_loop, steps_per_checkpoint))
iterations_per_loop = steps_per_checkpoint
model_params['iterations_per_loop'] = iterations_per_loop
config = tpu_config_lib.RunConfig(
cluster=tpu_cluster_resolver,
model_dir=model_dir,
save_checkpoints_steps=steps_per_checkpoint,
save_checkpoints_secs=None,
keep_checkpoint_max=keep_checkpoint_max,
log_step_count_steps=iterations_per_loop,
tpu_config=tpu_config_lib.TPUConfig(
iterations_per_loop=iterations_per_loop, num_shards=num_shards))
return config
def get_restore_vars(self, save_file, all_vars=None):
"""Create the `var_list` init argument to tf.Saver from save_file.
Extracts the subset of variables from tf.global_variables that match the
name and shape of variables saved in the checkpoint file, and returns these
as a list of variables to restore.
To support multi-model training, a model prefix is prepended to all
tf global_variable names, although this prefix is stripped from
all variables before they are saved to a checkpoint. Thus,
Args:
save_file: path of tf.train.Saver checkpoint.
Returns:
dict: checkpoint variables.
"""
reader = tf.train.NewCheckpointReader(save_file)
var_shapes = reader.get_variable_to_shape_map()
log.info('Saved Vars:\n' + str(var_shapes.keys()))
var_shapes = { # Strip the prefix off saved var names.
strip_prefix_from_name(self.params['model_params']['prefix'], name): shape
for name, shape in var_shapes.items()}
# Map old vars from checkpoint to new vars via load_param_dict.
mapped_var_shapes = self.remap_var_list(var_shapes)
log.info('Saved shapes:\n' + str(mapped_var_shapes))
if all_vars is None:
all_vars = tf.global_variables() + tf.local_variables() # get list of all variables
all_vars = strip_prefix(self.params['model_params']['prefix'], all_vars)
# Specify which vars are to be restored vs. reinitialized.
if self.load_param_dict is None:
restore_vars = {name: var for name, var in all_vars.items() if name in mapped_var_shapes}
else:
# associate checkpoint names with actual variables
load_var_dict = {}
for ckpt_var_name, curr_var_name in self.load_param_dict.items():
for curr_name, curr_var in all_vars.items():
if curr_name == curr_var_name:
load_var_dict[ckpt_var_name] = curr_var
break
restore_vars = load_var_dict
restore_vars = self.filter_var_list(restore_vars)
# Ensure the vars to restored have the correct shape.
var_list = {}
for name, var in restore_vars.items():
var_shape = var.get_shape().as_list()
if var_shape == mapped_var_shapes[name]:
var_list[name] = var
return var_list
def get_restore_vars(self, save_file):
"""Create the `var_list` init argument to tf.Saver from save_file.
Extracts the subset of variables from tf.global_variables that match the
name and shape of variables saved in the checkpoint file, and returns these
as a list of variables to restore.
To support multi-model training, a model prefix is prepended to all
tf global_variable names, although this prefix is stripped from
all variables before they are saved to a checkpoint. Thus,
Args:
save_file: path of tf.train.Saver checkpoint.
Returns:
dict: checkpoint variables.
"""
reader = tf.train.NewCheckpointReader(save_file)
var_shapes = reader.get_variable_to_shape_map()
# Map old vars from checkpoint to new vars via load_param_dict.
log.info('Saved vars and shapes:\n' + str(var_shapes))
# Specify which vars are to be restored vs. reinitialized.
all_vars = self.var_list
if not self.load_param_dict:
restore_vars = {
name: var for name, var in all_vars.items() \
if name in var_shapes}
else:
# associate checkpoint names with actual variables
load_var_dict = {}
for ckpt_var_name, curr_var_name in self.load_param_dict.items():
if curr_var_name in all_vars:
load_var_dict[ckpt_var_name] = all_vars[curr_var_name]
restore_vars = load_var_dict
restore_vars = self.filter_var_list(restore_vars)
# Ensure the vars to restored have the correct shape.
var_list = {}
for name, var in restore_vars.items():
var_shape = var.get_shape().as_list()
if var_shape == var_shapes[name]:
var_list[name] = var
else:
log.info('Shape mismatch for %s' % name \
+ str(var_shape) \
+ str(var_shapes[name]))
return var_list
def initialize(self, no_scratch=False):
"""Fetch record then uses tf's saver.restore."""
if self.do_restore:
# First, determine which checkpoint to use.
if self.from_ckpt is not None:
# Use a cached checkpoint file.
ckpt_filename = self.from_ckpt
log.info('Restoring variables from checkpoint %s ...' % ckpt_filename)
else:
# Otherwise, use a database checkpoint.
self.load_rec() if self.load_data is None else None
if self.load_data is not None:
rec, ckpt_filename = self.load_data
log.info('Restoring variables from record %s (step %d)...' %
(str(rec['_id']), rec['step']))
else:
# No db checkpoint to load.
ckpt_filename = None
if ckpt_filename is not None:
all_vars = tf.global_variables() + tf.local_variables() # get list of all variables
self.all_vars = strip_prefix(self.params['model_params']['prefix'], all_vars)
# Next, determine which vars should be restored from the specified checkpoint.
restore_vars = self.get_restore_vars(ckpt_filename, self.all_vars)
restore_stripped = strip_prefix(self.params['model_params']['prefix'], list(restore_vars.values()))
restore_names = [name for name, var in restore_stripped.items()]
# Actually load the vars.
log.info('Restored Vars:\n' + str(restore_names))
tf_saver_restore = tf.train.Saver(restore_vars)
tf_saver_restore.restore(self.sess, ckpt_filename)
log.info('... done restoring.')
# Reinitialize all other, unrestored vars.
unrestored_vars = [var for name, var in self.all_vars.items() if name not in restore_names]
unrestored_var_names = [name for name, var in self.all_vars.items() if name not in restore_names]
log.info('Unrestored Vars:\n' + str(unrestored_var_names))
self.sess.run(tf.variables_initializer(unrestored_vars)) # initialize variables not restored
assert len(self.sess.run(tf.report_uninitialized_variables())) == 0, (
self.sess.run(tf.report_uninitialized_variables()))
if not self.do_restore or (self.load_data is None and self.from_ckpt is None):
init_op_global = tf.global_variables_initializer()
self.sess.run(init_op_global)
init_op_local = tf.local_variables_initializer()
self.sess.run(init_op_local)
def version_check_and_info(module):
"""Return either git info or standard module version if not a git repo.
Args:
module (module): python module object to get info for.
Returns:
dict: dictionary of info
"""
srcpath = inspect.getsourcefile(module)
try:
repo = git.Repo(srcpath, search_parent_directories=True)
except git.InvalidGitRepositoryError:
log.info('module %s not in a git repo, checking package version' %
module.__name__)
info = version_info(module)
else:
info = git_info(repo)
info['source_path'] = srcpath
return info
def do_tpu_validation():
log.info('Starting to evaluate.')
eval_results = eval_cls.evaluate(input_fn=valid_fn,
hooks=valid_hooks,
steps=valid_steps)
log.info('Saving eval results to database.')
trarg['dbinterface'].save(valid_res={valid_k: eval_results},
validation_only=True)
log.info('Done saving eval results to database.')
return eval_results
def test_estimator(cls_dict, param, ttarg):
# load params query stores path to checkpoint
if param['load_params']['do_restore'] and (param['load_params']['query']
is not None):
# path to specific checkpoint
load_dir = param['load_params']['query']
else:
# gets latest checkpoint from model_dir
load_dir = None
ttarg['dbinterface'] = DBInterface(sess=None,
params=param,
save_params=param['save_params'],
load_params=param['load_params'])
ttarg['dbinterface'].start_time_step = time.time()
m_predictions = {}
for valid_k in cls_dict.keys():
cls = cls_dict[valid_k]
validation_data_params = param['validation_params'][valid_k][
'data_params']
# can use to filter particular params to save, if not there will set to None and all saved
filter_keys = param['validation_params'][valid_k].get('keys_to_save')
session_hooks = param['validation_params'][valid_k].get('hooks')
valid_fn = validation_data_params['func']
log.info('Starting to evaluate ({}).'.format(valid_k))
eval_results = cls.predict(input_fn=valid_fn,
predict_keys=filter_keys,
hooks=session_hooks,
checkpoint_path=load_dir)
m_predictions[valid_k] = list(eval_results)
log.info('Saving eval results to database.')
# set validation only to be True to just save the results and not filters
ttarg['dbinterface'].save(valid_res=m_predictions, validation_only=True)
log.info('Done saving eval results to database.')
# sync with hosts
res = []
ttarg['dbinterface'].sync_with_host()
res.append(trarg['dbinterface'].outrecs)
# returning final eval results for convenience
return eval_results, res
def train_estimator(train_cls, eval_cls, param, trarg):
if eval_cls is None:
eval_cls = train_cls
model_dir = param['save_params'].get('cache_dir', '')
train_steps = param['train_params']['num_steps']
# only single targets during eval mode
need_val = len(param['validation_params'].keys()) > 0
steps_per_eval = param['save_params'].get('save_valid_freq')
if need_val:
valid_k = param['validation_params'].keys()[0]
validation_data_params = param['validation_params'][valid_k][
'data_params']
valid_steps = param['validation_params'][valid_k]['num_steps']
valid_fn = validation_data_params['func']
if steps_per_eval is None:
steps_per_eval = param['save_params']['save_filters_freq']
else:
save_filters_freq = param['save_params'].get('save_filters_freq')
if save_filters_freq is not None:
# these need to be the same right now because estimator loads
# from last checkpoint after validating
assert (steps_per_eval == save_filters_freq)
else:
param['save_params']['save_filters_freq'] = steps_per_eval
train_fn = param['train_params']['data_params']['func']
model_params = param['model_params']
iterations_per_loop = model_params.get('iterations_per_loop',
DEFAULT_ITERATIONS_PER_LOOP)
if (steps_per_eval is
None) or (steps_per_eval < iterations_per_loop
): # eval steps cannot be less than TPU iterations
log.info(
'Setting save_valid_freq ({}) to be the same as iterations_per_loop ({}).'
.format(steps_per_eval, iterations_per_loop))
steps_per_eval = iterations_per_loop
train_hooks = param['train_params'].get('hooks')
if need_val:
valid_hooks = param['validation_params'][valid_k].get('hooks')
else:
valid_hooks = None
current_step = estimator._load_global_step_from_checkpoint_dir(model_dir)
# initialize db here (currently no support for loading and saving to different places. May need to modify init so load_params can load from different dir, estimator interface limited
# when loading and saving to different paths, may need to create a new config)
trarg['dbinterface'] = DBInterface(sess=None,
params=param,
global_step=current_step,
save_params=param['save_params'],
load_params=param['load_params'],
cache_dir=model_dir)
log.info('Training beginning ...')
log.info('Training for %d steps. Current '
'step %d' % (train_steps, current_step))
trarg['dbinterface'].start_time_step = time.time()
tpu_validate_first = param['train_params'].get('tpu_validate_first', False)
def do_tpu_validation():
log.info('Starting to evaluate.')
eval_results = eval_cls.evaluate(input_fn=valid_fn,
hooks=valid_hooks,
steps=valid_steps)
log.info('Saving eval results to database.')
trarg['dbinterface'].save(valid_res={valid_k: eval_results},
validation_only=True)
log.info('Done saving eval results to database.')
return eval_results
if tpu_validate_first:
eval_results = do_tpu_validation()
while current_step < train_steps:
next_eval = min(current_step + steps_per_eval, train_steps)
log.info('Training until step %d' % next_eval)
train_cls.train(input_fn=train_fn,
max_steps=next_eval,
hooks=train_hooks)
current_step = next_eval
if need_val:
eval_results = do_tpu_validation()
# sync with hosts
res = []
trarg['dbinterface'].sync_with_host()
res.append(trarg['dbinterface'].outrecs)
# returning final eval results for convenience
return eval_results, res
def model_fn(features, labels, mode, params):
model_params = params['model_params']
opt_params = params['opt_params']
loss_agg_func = params['loss_agg_func']
loss_per_case_func = params['loss_per_case_func']
loss_func_kwargs = params['loss_func_kwargs']
loss_agg_func_kwargs = params['loss_agg_func_kwargs']
lr_params = params['lr_params']
model_params['train'] = (mode == tf.estimator.ModeKeys.TRAIN)
if use_tpu:
model_params['batch_size'] = params[
'batch_size'] # per shard batch_size
model_func = model_params.pop('func')
outputs = model_func(inputs=features, **model_params)
if isinstance(outputs, dict):
logit_key = model_params.get('logit_key', 'logits')
logits = outputs[logit_key]
else:
logits = outputs
loss_args = (outputs, labels)
loss = loss_per_case_func(*loss_args, **loss_func_kwargs)
loss = loss_agg_func(loss, **loss_agg_func_kwargs)
if isinstance(loss, list):
optimizer_loss = loss
spec_loss = tf.add_n(loss)
else:
optimizer_loss = loss
spec_loss = loss
global_step = tf.train.get_global_step()
lr_func = lr_params.pop('func')
learning_rate = lr_func(global_step=global_step, **lr_params)
if mode == tf.estimator.ModeKeys.TRAIN:
opt_func = opt_params.pop('optimizer', ClipOptimizer)
# For deprecated parameter func
old_opt_func = opt_params.pop('func', None)
if old_opt_func:
log.info('func in optimizer_params is deprecated, ' + \
'please use optimizer')
opt_func = old_opt_func
log.info('Passing optimizer class to CrossShardMultiOptimizer')
optimizer_base = CrossShardMultiOptimizer(
opt_func(learning_rate=learning_rate, **opt_params))
update_ops = tf.get_collection(tf.GraphKeys.UPDATE_OPS)
with tf.control_dependencies(update_ops):
train_op = optimizer_base.minimize(optimizer_loss, global_step)
else:
train_op = None
eval_metrics = None
if mode == tf.estimator.ModeKeys.EVAL:
num_valid_targets = len(validation_params.keys())
metric_fn_kwargs = {'labels': labels, 'logits': logits}
if use_tpu:
assert (
num_valid_targets == 1
) # tpu estimators currently only support single targets :(
first_valid = validation_params.keys()[0]
valid_target = validation_params[first_valid]['targets']
metric_fn = valid_target['func']
if isinstance(outputs, dict):
for kw in outputs.keys():
if kw != logit_key:
kw_val = outputs[kw]
new_kw = kw
if isinstance(new_kw, int):
new_kw = 'i%i' % new_kw
metric_fn_kwargs.update({new_kw: kw_val})
for kw in valid_target.keys():
v = valid_target[kw]
if isinstance(v, dict):
for kw1 in v.keys():
# add any additional kwargs
kw_val = v[kw1]
metric_fn_kwargs.update({kw1: kw_val})
#metric_fn_kwargs[kw] = kw_val
eval_metrics = (metric_fn, metric_fn_kwargs)
else:
# normal estimators expect dicts and can support multiple targets (but same dataset and eval_steps etc)
eval_dict = {}
for k in validation_params.keys():
k_metric_fn_kwargs = metric_fn_kwargs
k_target = k['targets']
for kw in k_target.keys():
if kw != 'func':
# add any additional kwargs
kw_val = k_target[kw]
k_metric_fn_kwargs[kw] = kw_val
eval_dict[k] = (k_target['func'], k_metric_fn_kwargs)
eval_metrics = eval_dict
if use_tpu:
return tpu_estimator_lib.TPUEstimatorSpec(
mode=mode,
loss=spec_loss,
train_op=train_op,
eval_metrics=eval_metrics)
else:
return tf.estimator.EstimatorSpec(mode=mode,
loss=spec_loss,
train_op=train_op,
eval_metric_ops=eval_metrics)
def train_from_params(
save_params,
model_params,
train_params,
loss_params=None,
learning_rate_params=None,
optimizer_params=None,
validation_params=None,
load_params=None,
log_device_placement=DEFAULT_PARAMS[
'log_device_placement'], # advanced
dont_run=DEFAULT_PARAMS['dont_run'], # advanced
skip_check=DEFAULT_PARAMS['skip_check'], # advanced
use_estimator=False):
"""
Main training interface function.
Args:
save_params (dict):
Describing the parameters used to construct the save database, and
control saving. These include:
- host (str)
Hostname where database connection lives
- port (int)
Port where database connection lives
- dbname (str)
Name of database for storage
- collname (str)
Name of collection for storage
- exp_id (str)
Experiment id descriptor
NOTE: the variables host/port/dbname/coll/exp_id control
the location of the saved data for the run, in order of
increasing specificity. When choosing these, note that:
- If a given host/port/dbname/coll/exp_id already has saved checkpoints,\
then any new call to start training with these same location variables\
will start to train from the most recent saved checkpoint. If you mistakenly\
try to start training a new model with different variable names, or structure,\
from that existing checkpoint, an error will be raised, as the model will be\
incompatiable with the saved variables.
- When choosing what dbname, coll, and exp_id, to use, keep in mind that mongodb\
queries only operate over a single collection. So if you want to analyze\
results from a bunch of experiments together using mongod queries, you should\
put them all in the same collection, but with different exp_ids. If, on the\
other hand, you never expect to analyze data from two experiments together,\
you can put them in different collections or different databases. Choosing\
between putting two experiments in two collections in the same database\
or in two totally different databases will depend on how you want to organize\
your results and is really a matter of preference.
- do_save (bool, default: True)
Whether to save to database
- save_initial_filters (bool, default: True)
Whether to save initial model filters at step = 0,
- save_metrics_freq (int, default: 5)
How often to store train results to database
- save_valid_freq (int, default: 3000)
How often to calculate and store validation results
to database
- save_filters_freq (int, default: 30000)
How often to save filter values to database
- cache_filters_freq (int, default: 3000)
How often to cache filter values locally and save
to ___RECENT database
- cache_max_num (int, default: 6)
Maximal number of cached filters to keep in __RECENT database
- cache_dir (str, default: None)
Path where caches will be saved locally. If None, will default to
~/.tfutils/<host:post>/<dbname>/<collname>/<exp_id>.
model_params (dict): Containing function that produces model and arguments to that function.
- model_params['func']
The function producing the model.
The function's signature is:
Args:
- ``inputs``: data object
- ``train`` (boolean): if in training or testing
- ``seed`` (int): seed for use in random generation
Returns:
- ``outputs`` (tf.Operations): train output tensorflow nodes
- Additional configurations you want to store in database
- Remaining items in model_params are dictionary of arguments passed to func.
train_params (dict): Containing params for data sources and targets in training.
- train_params['data_params']
This contains params for the data
- ``train_params['data_params']['func']`` is the function that constructs the data:
The function's signature is:
Args:
- ``batch_size``: Batch size for input data
Returns:
- ``inputs``: A dictionary of tensors that will be sent to model function
- ``train_params['data_params']['batch_size']`` batch size of the data, will be sent to func
- Remainder of ``train_params['data_params']`` are kwargs passed to func
- train_params['targets'] (optional)
contains params for additional train targets
- ``train_params['targets']['func']`` is a function that produces tensorflow nodes as training targets:
The function's signature is:
Args:
- ``inputs``: returned values of ``train_params['data_params']['func']``
- ``output``: first returned value of ``train_params['model_params']['func']``
Returns:
A dictionary of tensors that will be computed and stored in the database
- Remainder of ``train_parms['targets']`` are arguments to func.
- train_params['validate_first'] (optional, bool, default is True):
controls whether validating before training
- train_params['thres_loss'] (optional, float, default: 100):
If loss exceeds this during training, HiLossError is thrown
- train_params['num_steps'] (int or None, default: None):
How many total steps of the optimization are run.
If None, train is run until process is cancelled.
loss_params (dict): Parameters for helper.get_loss_base function to build loss.
- loss_params['pred_targets'] (a string or a list of strings):
contain the names of inputs nodes that will be sent into the loss function
- loss_params['loss_func']:
the function used to calculate the loss. Must be provided.
- loss_params['loss_func_kwargs'] (dict):
Keyword parameters sent to ``loss_params['loss_func']``. Default is {}.
- loss_params['agg_func']:
The aggregate function, default is None.
- loss_params['agg_func_kwargs']:
Keyword parameters sent to ``loss_params['agg_func']``. Default is {}.
- loss_params['loss_per_case_func'] (Deprecated):
Deprecated parameter, the same as ``loss_params['loss_func']``.
- loss_params['targets'] (Deprecated):
Deprecated parameter, the same as ``loss_params['targets']``.
learning_rate_params (dict): Parameters for specifying learning_rate.
- learning_rate_params['func']:
The function producing tensorflow node acting as learning rate.
This function must accept argument ``global_step``.
- remainder of learning_rate_params are arguments to func.
optimizer_params (dict): Parameters for creating optimizer.
- optimizer_params['optimizer']:
A class producing an optimizer object,
which should have function ``compute_gradients`` and ``apply_gradients``.
The signatures of these two functions are similar as tensorflow basic optimizer classes.
Must accept:
- "learning_rate" -- the result of the learning_rate_func call
- Remainder of optimizer_params (aside form "optimizer") are arguments
to the optimizer func
- optimizer_params['func'] (Deprecated):
Deprecated parameter, the same as ``optimizer_params['optimizer']``.
validation_params (dict): Dictionary of validation sources. The structure if this dictionary is:
{
<validation_target_name_1>: {
data: {
'func': (callable) data source function for this validation,
<kwarg1>: <value1> for 'func',
...
},
targets: {
'func': (callable) returning targets,
<kwarg1>: <value1> for 'func',
...
},
num_steps (int):
number of batches of validation source to compute,
agg_func (optional, callable):
how to aggregate validation results
across batches after computation. Signature is:
- one input argument: the list of validation batch results
- one output: aggregated version
Default is ``utils.identity_func``
online_agg_func (optional, callable):
how to aggregate validation results
on a per-batch basis. Siganture is:
- three input arguments: (current aggregate, new result, step)
- one output: new aggregated result
On first step, current aggregate passed in is None.
The final result is passed to the "agg_func".
Default is ``utils.append_and_return``
},
<validation_target_name_2>: ...
}
For each validation_target_name key, the targets are computed and then added to
the output dictionary to be computed every so often -- unlike train_targets which
are computed on each time step, these are computed on a basic controlled by the
valid_save_freq specific in the save_params.
load_params (dict):
Similar to save_params, if you want loading to happen from a different
location than where saving occurs. Parameters include:
- host (str)
Hostname where database connection lives
- port (int)
Port where database connection lives
- dbname (str)
Name of database for storage
- collname (str)
Name of collection for storage
- exp_id (str)
Experiment id descriptor
- do_restore (bool, default: True)
Whether to restore from saved model
- query (dict)
mongodb query describing how to load from loading database
- from_ckpt (string)
Path to load from a TensorFlow checkpoint (instead of from the db)
- to_restore (list of strings or a regex/callable which returns strings)
Specifies which variables should be loaded from the checkpoint.
Any variables not specified here will be reinitialized.
- load_param_dict (dict)
A dictionary whose keys are the names of the variables that are to be loaded
from the checkpoint, and the values are the names of the variables of the model
that you want to restore with the value of the corresponding checkpoint variable.
log_device_placement (bool, default is False):
Advanced parameter. Whether to log device placement in tensorflow session
dont_run (bool, default is False):
Advanced parameter. Whether returning everything, not actually training
skip_check (bool, default is False):
Advanced parameter. Whether skipping github check, could be useful when working in detached head
"""
# use tpu only if a tpu_name has been specified and not a multi-model
if isinstance(model_params, list): # multi-model mode
use_tpu = (model_params[0].get('tpu_name', None) is not None)
assert (use_tpu is False)
else:
use_tpu = (model_params.get('tpu_name', None) is not None)
if use_tpu:
log.info('Using tpu: %s' % model_params['tpu_name'])
params, train_args = parse_params(
'train',
model_params,
dont_run=dont_run,
skip_check=skip_check,
load_params=load_params,
loss_params=loss_params,
save_params=save_params,
train_params=train_params,
optimizer_params=optimizer_params,
validation_params=validation_params,
learning_rate_params=learning_rate_params,
log_device_placement=log_device_placement,
use_tpu=use_tpu or use_estimator)
if use_estimator or use_tpu:
return tpu_train_from_params(params, train_args, use_tpu=use_tpu)
else:
with tf.Graph().as_default(), tf.device(DEFAULT_HOST):
# For convenience, use list of dicts instead of dict of lists
_params = [{key: value[i]
for (key, value) in params.items()}
for i in range(len(params['model_params']))]
_trargs = [{key: value[i]
for (key, value) in train_args.items()}
for i in range(len(params['model_params']))]
# Use a single dataprovider for all models.
data_params = _params[0]['train_params']['data_params']
_params[0]['train_params']['data_params'], inputs \
= get_data(**data_params)
# Build a graph for each distinct model.
var_manager_list = []
for param, trarg in zip(_params, _trargs):
_, _, param, trarg, var_manager \
= get_model(inputs,
param['model_params'],
param=param,
trarg=trarg)
trarg['validation_targets'], _ = \
get_valid_targets_dict(
var_manager=var_manager,
**param)
var_manager_list.append(var_manager)
# Create session.
gpu_options = tf.GPUOptions(allow_growth=True)
sess = tf.Session(config=tf.ConfigProto(
allow_soft_placement=True,
gpu_options=gpu_options,
log_device_placement=log_device_placement,
))
# Initialize variables here
init_op_global = tf.global_variables_initializer()
sess.run(init_op_global)
init_op_local = tf.local_variables_initializer()
sess.run(init_op_local)
log.info('Initialized from scratch first')
# Build database interface for each model
# This interface class will handle the records saving, model saving, and
# model restoring.
for param, trarg, var_manager in zip(_params, _trargs,
var_manager_list):
trarg['dbinterface'] = DBInterface(
sess=sess,
params=param,
var_manager=var_manager,
global_step=trarg['global_step'],
save_params=param['save_params'],
load_params=param['load_params'])
## Model will be restored from saved database here
trarg['dbinterface'].initialize()
# Convert back to a dictionary of lists
params = {
key: [param[key] for param in _params]
for key in _params[0].keys()
}
train_args = {
key: [trarg[key] for trarg in _trargs]
for key in _trargs[0].keys()
}
if dont_run:
return train_args
return train(sess, **train_args)
def load_from_db(self,
query,
cache_filters=False,
collfs=None,
collfs_recent=None):
"""Load checkpoint from the database.
Checks the recent and regular checkpoint fs to find the latest one
matching the query. Returns the GridOut obj corresponding to the
record.
Args:
query: dict expressing MongoDB query
"""
if collfs is None:
collfs = self.collfs
coll = collfs._GridFS__files
if collfs_recent is None:
collfs_recent = self.collfs_recent
coll_recent = collfs_recent._GridFS__files
query['saved_filters'] = True
count = collfs.find(query).count()
if count > 0: # get latest that matches query
ckpt_record = coll.find(query, sort=[('uploadDate', -1)])[0]
loading_from = coll
else:
ckpt_record = None
try:
count_recent = collfs_recent.find(query).count()
except Exception as inst:
raise er.OperationFailure(
inst.args[0] +
"\n Is your dbname too long? Mongo requires that dbnames be no longer than 64 characters."
)
if count_recent > 0: # get latest that matches query
ckpt_record_recent = coll_recent.find(query,
sort=[('uploadDate', -1)])[0]
# use the record with latest timestamp
if ckpt_record is None or ckpt_record_recent[
'uploadDate'] > ckpt_record['uploadDate']:
loading_from = coll_recent
ckpt_record = ckpt_record_recent
if count + count_recent == 0: # no matches for query
log.warning('No matching checkpoint for query "{}"'.format(
repr(query)))
return
database = loading_from._Collection__database
log.info('Loading checkpoint from %s' % loading_from.full_name)
if cache_filters:
filename = os.path.basename(ckpt_record['filename'])
cache_filename = os.path.join(self.cache_dir, filename)
# check if there is no local copy
if not os.path.isfile(cache_filename):
log.info('No cache file at %s, loading from DB' %
cache_filename)
# create new file to write from gridfs
load_dest = open(cache_filename, "w+")
load_dest.close()
load_dest = open(cache_filename, 'rwb+')
fsbucket = gridfs.GridFSBucket(
database, bucket_name=loading_from.name.split('.')[0])
fsbucket.download_to_stream(ckpt_record['_id'], load_dest)
load_dest.close()
if ckpt_record[
'_saver_write_version'] == saver_pb2.SaverDef.V2:
assert cache_filename.endswith('.tar')
tar = tarfile.open(cache_filename)
tar.extractall(path=self.cache_dir)
tar.close()
cache_filename = os.path.splitext(cache_filename)[0]
verify_pb2_v2_files(cache_filename, ckpt_record)
else:
if ckpt_record[
'_saver_write_version'] == saver_pb2.SaverDef.V2:
cache_filename = os.path.splitext(cache_filename)[0]
verify_pb2_v2_files(cache_filename, ckpt_record)
log.info('Cache file found at %s, using that to load' %
cache_filename)
else:
cache_filename = None
return ckpt_record, cache_filename
def initialize(self):
"""Fetch record then uses tf's saver.restore."""
if self.do_restore:
# First, determine which checkpoint to use.
if self.from_ckpt is not None:
# Use a cached checkpoint file.
ckpt_filename = self.from_ckpt
log.info('Restoring variables from checkpoint %s ...' \
% ckpt_filename)
else:
# Otherwise, use a database checkpoint.
self.load_rec() if self.load_data is None else None
if self.load_data is not None:
rec, ckpt_filename = self.load_data
log.info('Restoring variables from record %s (step %d)...' \
% (str(rec['_id']), rec['step']))
else:
# No db checkpoint to load.
ckpt_filename = None
if ckpt_filename is not None:
# Determine which vars should be restored from the specified checkpoint.
restore_vars = self.get_restore_vars(ckpt_filename)
restore_names = [name for name, var in restore_vars.items()]
# remap the actually restored names to the new ones
if self.load_param_dict:
for each_old_name in self.load_param_dict.keys():
if each_old_name in restore_names:
restore_names.remove(each_old_name)
restore_names.append(
self.load_param_dict[each_old_name])
# Actually load the vars.
log.info('Restored Vars (in ckpt, in graph):\n' +
str(restore_names))
tf_saver_restore = tf.train.Saver(restore_vars)
tf_saver_restore.restore(self.sess, ckpt_filename)
log.info('... done restoring.')
# Run post init_ops if needed
if self.var_manager:
self.sess.run(
tf.group(*self.var_manager.get_post_init_ops()))
# Reinitialize all other, unrestored vars.
unrestored_vars = [\
var \
for name, var in self.var_list.items() \
if name not in restore_names]
unrestored_var_names = [\
name \
for name, var in self.var_list.items() \
if (name not in restore_names) and not(any([name.endswith(s) for s in OPTIMIZER_NAMES]))]
log.info('Unrestored Vars (in graph, not in ckpt):\n' +
str(unrestored_var_names))
self.sess.run(tf.variables_initializer(
unrestored_vars)) # initialize variables not restored
assert len(self.sess.run(
tf.report_uninitialized_variables())) == 0, (self.sess.run(
tf.report_uninitialized_variables()))
if not self.do_restore \
or (self.load_data is None and self.from_ckpt is None):
init_op_global = tf.global_variables_initializer()
self.sess.run(init_op_global)
init_op_local = tf.local_variables_initializer()
self.sess.run(init_op_local)
if self.var_manager:
self.sess.run(tf.group(*self.var_manager.get_post_init_ops()))
def tpu_train_from_params(params, train_args, use_tpu=False):
"""
Main tpu training interface function, called by train_from_params in tfutils.train.
See the doc string there for info.
"""
# use this for tpu and estimator logging
tf.logging.set_verbosity(tf.logging.INFO)
# For convenience, use list of dicts instead of dict of lists
_params = [{key: value[i]
for (key, value) in params.items()}
for i in range(len(params['model_params']))]
_trargs = [{key: value[i]
for (key, value) in train_args.items()}
for i in range(len(params['model_params']))]
param = _params[0]
trarg = _trargs[0]
# Support only single model
assert (len(_params) == 1)
train_data_params = param['train_params']['data_params']
model_params = param['model_params']
lr_params = param['learning_rate_params']
opt_params = param['optimizer_params']
loss_params = param['loss_params']
validation_params = param['validation_params']
save_params = param['save_params']
# set up estimator func
estimator_fn, params_to_pass = create_train_estimator_fn(
use_tpu=use_tpu,
model_params=model_params,
lr_params=lr_params,
opt_params=opt_params,
loss_params=loss_params,
validation_params=validation_params)
if use_tpu:
if len(param['validation_params'].keys()) > 0:
valid_k = param['validation_params'].keys()[0]
validation_data_params = param['validation_params'][valid_k][
'data_params']
eval_batch_size = validation_data_params['batch_size']
else:
eval_batch_size = None
# grab tpu name and gcp, etc from model params
train_m_config = create_train_tpu_config(
model_dir=save_params.get('cache_dir', ''),
tpu_name=model_params.get('tpu_name', None),
gcp_project=model_params.get('gcp_project', None),
steps_per_checkpoint=save_params.get('save_filters_freq', None),
tpu_zone=model_params.get('tpu_zone', DEFAULT_TPU_ZONE),
num_shards=model_params.get('num_shards', DEFAULT_NUM_SHARDS),
keep_checkpoint_max=save_params.get('checkpoint_max', 5),
iterations_per_loop=model_params.get('iterations_per_loop',
DEFAULT_ITERATIONS_PER_LOOP),
model_params=model_params)
train_estimator_classifier = tpu_estimator_lib.TPUEstimator(
use_tpu=True,
model_fn=estimator_fn,
config=train_m_config,
train_batch_size=train_data_params['batch_size'],
eval_batch_size=eval_batch_size,
params=params_to_pass)
val_estimator_classifier = None
if model_params.get('num_shards', DEFAULT_NUM_SHARDS) > 8:
log.info("You are training in pod mode")
log.info(
"Setting up validation on a single independent TPU device")
assert model_params.get('val_tpu_name') is not None
val_m_config = create_train_tpu_config(
model_dir=save_params.get('cache_dir', ''),
tpu_name=model_params.get('val_tpu_name', None),
gcp_project=model_params.get('gcp_project', None),
steps_per_checkpoint=save_params.get('save_filters_freq',
None),
tpu_zone=model_params.get('val_tpu_zone', DEFAULT_TPU_ZONE),
num_shards=8,
keep_checkpoint_max=save_params.get('checkpoint_max', 5),
iterations_per_loop=model_params.get(
'iterations_per_loop', DEFAULT_ITERATIONS_PER_LOOP),
model_params=model_params)
val_estimator_classifier = tpu_estimator_lib.TPUEstimator(
use_tpu=True,
model_fn=estimator_fn,
config=val_m_config,
train_batch_size=train_data_params['batch_size'],
eval_batch_size=eval_batch_size,
params=params_to_pass)
else:
train_estimator_classifier = tf.estimator.Estimator(
model_fn=estimator_fn, params=params_to_pass)
return train_estimator(train_cls=train_estimator_classifier,
eval_cls=val_estimator_classifier,
param=param,
trarg=trarg)
def save(self,
train_res=None,
valid_res=None,
step=None,
validation_only=False):
"""Actually save record into DB and makes local filter caches."""
if train_res is None:
train_res = {}
if valid_res is None:
valid_res = {}
if (not validation_only) and (step is None):
if not hasattr(self.global_step, 'eval'):
raise NoGlobalStepError(
'If step is none, you must pass global_step'
' tensorflow operation to the saver.')
step = self.global_step.eval(session=self.sess)
train_res = copy.copy(train_res)
valid_res = {_k: copy.copy(_v) for _k, _v in valid_res.items()}
duration = time.time() - self.start_time_step
if self.rec_to_save is None:
rec = {
'exp_id': self.exp_id,
'params': self.sonified_params,
'saved_filters': False,
'duration': duration
}
self.rec_to_save = rec
else:
rec = self.rec_to_save
rec['step'] = step
if len(train_res) > 0:
# TODO: also include error rate of the train set to monitor overfitting
message = 'Step {} ({:.0f} ms) -- '.format(step, 1000 * duration)
# If ndarray found, get the mean of it
for k, v in train_res.items():
if k not in ['optimizer', '__grads__'] and \
isinstance(v, np.ndarray) and len(v) > 1:
train_res[k] = np.mean(v)
msg2 = [
'{}: {:.4f}'.format(k, v) for k, v in train_res.items()
if k not in ['optimizer', '__grads__']
and k not in self.save_to_gfs
]
message += ', '.join(msg2)
log.info(message)
if '__grads__' in train_res:
del train_res['__grads__']
if 'optimizer' in train_res:
del train_res['optimizer']
if 'train_results' not in rec:
rec['train_results'] = []
rec['train_results'].append(train_res)
# print validation set performance
if len(valid_res) > 0:
rec['validation_results'] = valid_res
message = 'Validation -- '
message += ', '.join('{}: {}'.format(
k,
{_k: _v
for _k, _v in v.items() if _k not in self.save_to_gfs})
for k, v in valid_res.items())
log.info(message)
if validation_only:
rec['validates'] = self.load_data[0]['_id']
save_filters_permanent = save_filters_tmp = False
need_to_save = True
else:
save_filters_permanent = (
(step % self.save_filters_freq == 0)
and (step > 0 or
(self.save_initial_filters and not self.load_data)))
save_filters_tmp = (
(step % self.cache_filters_freq == 0)
and (step > 0 or
(self.save_initial_filters and not self.load_data)))
save_metrics_now = step % self.save_metrics_freq == 0
save_valid_now = step % self.save_valid_freq == 0
need_to_save = save_filters_permanent or save_filters_tmp or save_metrics_now or save_valid_now
need_to_save = self.do_save and need_to_save
if need_to_save:
self.rec_to_save = None
self.sync_with_host()
save_to_gfs = {}
for _k in self.save_to_gfs:
if train_res:
if 'train_results' not in save_to_gfs:
save_to_gfs['train_results'] = {}
if _k in train_res:
save_to_gfs['train_results'][_k] = [
r.pop(_k) for r in rec['train_results'] if _k in r
]
if len(save_to_gfs['train_results'][_k]) == 1:
save_to_gfs['train_results'][_k] == save_to_gfs[
'train_results'][_k][0]
if valid_res:
if 'validation_results' not in save_to_gfs:
save_to_gfs['validation_results'] = {}
for _vk in valid_res:
if _vk not in save_to_gfs['validation_results']:
save_to_gfs['validation_results'][_vk] = {}
if _k in valid_res[_vk]:
save_to_gfs['validation_results'][_vk][
_k] = valid_res[_vk].pop(_k)
save_rec = sonify(rec, skip=self._skip_check)
make_mongo_safe(save_rec)
coord = tf.train.Coordinator()
thread = CoordinatedThread(coord=coord,
target=self._save_thread,
args=(save_filters_permanent,
save_filters_tmp, save_rec, step,
save_to_gfs))
thread.daemon = True
thread.start()
self.checkpoint_thread = thread
self.checkpoint_coord = coord
def test_from_params(load_params,
model_params,
validation_params,
log_device_placement=False,
save_params=None,
dont_run=False,
skip_check=False,
use_estimator=False):
"""
Main testing interface function.
Same as train_from_parameters; but just performs testing without training.
For documentation, see argument descriptions in train_from_params.
"""
# use tpu only if a tpu_name has been specified and not a multi-model
if isinstance(model_params, list): # multi-model mode
use_tpu = (model_params[0].get('tpu_name', None) is not None)
assert (use_tpu is False)
else:
use_tpu = (model_params.get('tpu_name', None) is not None)
if use_tpu:
log.info('Using tpu: %s' % model_params['tpu_name'])
params, test_args = parse_params('test',
model_params,
dont_run=dont_run,
skip_check=skip_check,
save_params=save_params,
load_params=load_params,
validation_params=validation_params,
log_device_placement=log_device_placement,
use_tpu=use_tpu)
# do not need to create sess with estimator interface
if use_estimator or use_tpu:
return tpu_test_from_params(params, test_args, use_tpu=use_tpu)
else:
with tf.Graph().as_default(), tf.device(DEFAULT_HOST):
# create session
gpu_options = tf.GPUOptions(allow_growth=True)
sess = tf.Session(config=tf.ConfigProto(
allow_soft_placement=True,
gpu_options=gpu_options,
log_device_placement=log_device_placement,
))
init_op_global = tf.global_variables_initializer()
sess.run(init_op_global)
init_op_local = tf.local_variables_initializer()
sess.run(init_op_local)
log.info('Initialized from scratch first')
# For convenience, use list of dicts instead of dict of lists
_params = [{key: value[i]
for (key, value) in params.items()}
for i in range(len(params['model_params']))]
_ttargs = [{key: value[i]
for (key, value) in test_args.items()}
for i in range(len(params['model_params']))]
# Build a graph for each distinct model.
for param, ttarg in zip(_params, _ttargs):
print(param['load_params'])
from_ckpt = param['load_params'].get('from_ckpt')
use_ckpt = (from_ckpt is not None)
if not 'cache_dir' in load_params:
temp_cache_dir = save_params.get('cache_dir', None)
load_params['cache_dir'] = temp_cache_dir
log.info('cache_dir not found in load_params, '\
+ 'using cache_dir ({}) from save_params'.format(
temp_cache_dir))
ttarg['dbinterface'] = DBInterface(
var_manager=None,
params=param,
load_params=param['load_params'])
if not use_ckpt:
ttarg['dbinterface'].load_rec()
ld = ttarg['dbinterface'].load_data
assert ld is not None, "No load data found for query, aborting"
ld = ld[0]
# TODO: have option to reconstitute model_params entirely from
# saved object ("revivification")
param['model_params']['seed'] = ld['params'][
'model_params']['seed']
cfg_final = ld['params']['model_params']['cfg_final']
else:
cfg_final = param['model_params'].get('cfg_final', {})
ttarg['validation_targets'], var_manager \
= get_valid_targets_dict(
loss_params=None,
cfg_final=cfg_final,
**param)
param['load_params']['do_restore'] = True
param['model_params']['cfg_final'] = cfg_final
# Build database interface class, loading model
ttarg['dbinterface'] = DBInterface(
sess=sess,
params=param,
var_manager=var_manager,
load_params=param['load_params'],
save_params=param['save_params'])
ttarg['dbinterface'].initialize()
ttarg['save_intermediate_freq'] \
= param['save_params'].get('save_intermediate_freq')
# Convert back to a dictionary of lists
params = {
key: [param[key] for param in _params]
for key in _params[0].keys()
}
test_args = {
key: [ttarg[key] for ttarg in _ttargs]
for key in _ttargs[0].keys()
}
if dont_run:
return test_args
res = test(sess, **test_args)
sess.close()
return res
def _save_thread(self, save_filters_permanent, save_filters_tmp, save_rec,
step, save_to_gfs):
if save_filters_permanent or save_filters_tmp:
save_rec['saved_filters'] = True
save_path = os.path.join(self.cache_dir, 'checkpoint')
log.info('Saving model with path prefix %s ... ' % save_path)
saved_path = self.tf_saver.save(self.sess,
save_path=save_path,
global_step=step,
write_meta_graph=False)
log.info('... done saving with path prefix %s' % saved_path)
putfs = self.collfs if save_filters_permanent else self.collfs_recent
log.info('Putting filters into %s database' % repr(putfs))
save_rec['_saver_write_version'] = self.tf_saver._write_version
if self.tf_saver._write_version == saver_pb2.SaverDef.V2:
file_data = get_saver_pb2_v2_files(saved_path)
save_rec['_saver_num_data_files'] = file_data['num_data_files']
tarfilepath = saved_path + '.tar'
tar = tarfile.open(tarfilepath, 'w')
for _f in file_data['files']:
tar.add(_f, arcname=os.path.split(_f)[1])
tar.close()
with open(tarfilepath, 'rb') as _fp:
outrec = putfs.put(_fp, filename=tarfilepath, **save_rec)
else:
with open(saved_path, 'rb') as _fp:
outrec = putfs.put(_fp, filename=saved_path, **save_rec)
log.info('... done putting filters into database.')
if not save_filters_permanent:
recent_gridfs_files = self.collfs_recent._GridFS__files
recent_query_result = recent_gridfs_files.find(
{'saved_filters': True}, sort=[('uploadDate', 1)])
num_cached_filters = recent_query_result.count()
cache_max_num = self.cache_max_num
if num_cached_filters > cache_max_num:
log.info('Cleaning up cached filters')
fsbucket = gridfs.GridFSBucket(
recent_gridfs_files._Collection__database,
bucket_name=recent_gridfs_files.name.split('.')[0])
for del_indx in xrange(0,
num_cached_filters - cache_max_num):
#log.info(recent_query_result[del_indx]['uploadDate'])
fsbucket.delete(recent_query_result[del_indx]['_id'])
if not save_filters_permanent:
save_rec['saved_filters'] = False
log.info('Inserting record into database.')
outrec = self.collfs._GridFS__files.insert_one(save_rec)
if not isinstance(outrec, ObjectId):
outrec = outrec.inserted_id
if save_to_gfs:
idval = str(outrec)
save_to_gfs_path = idval + "_fileitems"
self.collfs.put(cPickle.dumps(save_to_gfs),
filename=save_to_gfs_path,
item_for=outrec)
sys.stdout.flush() # flush the stdout buffer
self.outrecs.append(outrec)
def train(sess,
dbinterface,
train_loop,
train_targets,
global_step,
num_minibatches=1,
num_steps=float('inf'),
thres_loss=DEFAULT_TRAIN_THRES_LOSS,
validate_first=True,
validation_targets=None):
"""Actually runs the training evaluation loop.
Args:
sess (tesorflow.Session):
Object in which to run calculations.
dbinterface (DBInterface object): Saver through which to save results.
train_loop (callable withs args: sess and train_targets):
Callable that specifies a custom training loop
train_targets (dict of tensorflow nodes): Targets to train.
One item in this dict must be "optimizer" or similar
to make anything happen
num_minibatches (int): How many minibatches to use to before applying gradient update.
num_steps (int): How many steps to train to before quitting
validation_targets (dict of tensorflow objects, default: None):
Objects on which validation will be computed
thres_loss (float, default: 100):
If loss exceeds this during training, HiLossError is thrown
"""
# Collect args in a dict of lists
train_args = {
'num_steps': num_steps,
'thres_loss': thres_loss,
'train_loop': train_loop,
'global_step': global_step,
'dbinterface': dbinterface,
'train_targets': train_targets,
'validate_first': validate_first,
'num_minibatches': num_minibatches,
'validation_targets': validation_targets
}
# Convert to a list of dicts
trargs = [{key: value[i]
for (key, value) in train_args.items()}
for i in range(len(train_targets))]
num_steps = [t['num_steps'] for t in trargs]
steps = [t['global_step'].eval(session=sess) for t in trargs]
# Start initial validation
for (step, trarg) in zip(steps, trargs):
if step >= trarg['num_steps']:
log.info('Training cancelled since step ({}) is >= num_steps ({})'.
format(step, trarg['num_steps']))
return
log.info('Training beginning ...')
if step == 0:
trarg['dbinterface'].start_time_step = time.time()
if trarg['validate_first']:
valid_res = run_all_validations(
sess,
trarg['validation_targets'],
dbinterface=trarg['dbinterface'])
train_loop = train_args['train_loop'][0]
train_targets = train_args['train_targets']
# Run training
while any(step < num_step for (step, num_step) in zip(steps, num_steps)):
start_time_step = time.time()
train_results = train_loop(sess,
train_targets,
num_minibatches=trarg['num_minibatches'])
for (step, trarg, train_res) in zip(steps, trargs, train_results):
old_step = step
step = trarg['global_step'].eval(session=sess)
if step <= old_step:
raise NoChangeError(\
'Your optimizer should have incremented the global step,'
' but did not: old_step=%d, new_step=%d' \
% (old_step, step))
if np.isnan(train_res['loss']):
raise NanLossError(\
'Loss has become NaN')
if train_res['loss'] > trarg['thres_loss']:
raise HiLossError(\
'Loss {:.2f} exceeded the threshold {:.2f}'.format(
train_res['loss'],
trarg['thres_loss']))
# Validation
vtargs = trarg['validation_targets'] \
if step % trarg['dbinterface'].save_valid_freq == 0 else {}
valid_res = run_all_validations(sess, vtargs)
# Save
trarg['dbinterface'].start_time_step = start_time_step
trarg['dbinterface'].save(train_res=train_res,
valid_res=valid_res,
validation_only=False)
steps = [t['global_step'].eval(session=sess) for t in trargs]
# Sync and close the session
res = []
for trarg in trargs:
trarg['dbinterface'].sync_with_host()
res.append(trarg['dbinterface'].outrecs)
sess.close()
return res
def test_from_params(load_params,
model_params,
validation_params,
log_device_placement=False,
save_params=None,
dont_run=False,
skip_check=False,
):
"""
Main testing interface function.
Same as train_from_parameters; but just performs testing without training.
For documentation, see argument descriptions in train_from_params.
"""
params, test_args = parse_params(
'test',
model_params,
dont_run=dont_run,
skip_check=skip_check,
save_params=save_params,
load_params=load_params,
validation_params=validation_params,
log_device_placement=log_device_placement,
)
with tf.Graph().as_default(), tf.device(DEFAULT_HOST):
# create session
sess = tf.Session(
config=tf.ConfigProto(
allow_soft_placement=True,
log_device_placement=log_device_placement,
))
init_op_global = tf.global_variables_initializer()
sess.run(init_op_global)
init_op_local = tf.local_variables_initializer()
sess.run(init_op_local)
log.info('Initialized from scratch first')
# For convenience, use list of dicts instead of dict of lists
_params = [{key: value[i] for (key, value) in params.items()}
for i in range(len(params['model_params']))]
_ttargs = [{key: value[i] for (key, value) in test_args.items()}
for i in range(len(params['model_params']))]
# Build a graph for each distinct model.
for param, ttarg in zip(_params, _ttargs):
if not 'cache_dir' in load_params:
temp_cache_dir = save_params.get('cache_dir', None)
load_params['cache_dir'] = temp_cache_dir
log.info('cache_dir not found in load_params, using cache_dir ({}) from save_params'.format(temp_cache_dir))
ttarg['dbinterface'] = DBInterface(params=param, load_params=param['load_params'])
ttarg['dbinterface'].load_rec()
ld = ttarg['dbinterface'].load_data
assert ld is not None, "No load data found for query, aborting"
ld = ld[0]
# TODO: have option to reconstitute model_params entirely from
# saved object ("revivification")
param['model_params']['seed'] = ld['params']['model_params']['seed']
cfg_final = ld['params']['model_params']['cfg_final']
ttarg['validation_targets'] = \
get_valid_targets_dict(
loss_params=None,
cfg_final=cfg_final,
**param)
# tf.get_variable_scope().reuse_variables()
param['load_params']['do_restore'] = True
param['model_params']['cfg_final'] = cfg_final
prefix = param['model_params']['prefix'] + '/'
all_vars = variables._all_saveable_objects()
var_list = strip_prefix(prefix, all_vars)
ttarg['dbinterface'] = DBInterface(sess=sess,
params=param,
var_list=var_list,
load_params=param['load_params'],
save_params=param['save_params'])
ttarg['dbinterface'].initialize(no_scratch=True)
ttarg['save_intermediate_freq'] = param['save_params'].get('save_intermediate_freq')
# Convert back to a dictionary of lists
params = {key: [param[key] for param in _params]
for key in _params[0].keys()}
test_args = {key: [ttarg[key] for ttarg in _ttargs]
for key in _ttargs[0].keys()}
if dont_run:
return test_args
res = test(sess, **test_args)
sess.close()
return res
