returnn.tf.network

Defines the TFNetwork and ExternData.

exception returnn.tf.network.DataNotFound[source]

When accessing non-existing ExternData data key

class returnn.tf.network.ExternData(data=None, default_input='data', default_target='classes')[source]

This holds Data instances for every data-key of external data from the dataset, i.e. the description such as shape and sparsity, etc.

Parameters:data (None|dict[str,dict[str]]) – optional init kwargs for Data
init_from_config(config, auto_create_placeholders=True)[source]
Parameters:
  • config (Config.Config) –
  • auto_create_placeholders (bool) –
classmethod data_kwargs_from_dataset_key(dataset, key)[source]
Parameters:
  • dataset (Dataset.Dataset) –
  • key (str) –
Return type:

dict[str]

init_from_dataset(dataset, auto_create_placeholders=True)[source]
Parameters:
init_batch_info()[source]

Initializes and sets the batch info on the extern data, i.e. sets Data.batch. See BatchInfo.

check_matched_dataset(dataset, used_data_keys=None)[source]
Parameters:
  • dataset (Dataset.Dataset) –
  • used_data_keys (set[str]|list[str]) –
Returns:

nothing, will assert the check

register_data_from_dict(data)[source]
Parameters:data (dict[str,dict[str]]) – init kwargs for Data
register_data(data)[source]
Parameters:data (Data) – will use data.name as the key
has_data(name)[source]
Parameters:name (str) –
Return type:bool
get_data(name)[source]
Parameters:name (str) –
Return type:Data
get_default_input_data()[source]
Return type:Data
get_default_target_data()[source]
Return type:Data
get_data_description()[source]
Returns:str describing the data
Return type:str
get_queue_args(with_batch_dim, fixed_batch_dim=None)[source]
Parameters:
  • with_batch_dim (bool) –
  • fixed_batch_dim (int|None) –
Returns:

kwargs for tf.Queue.__init__

Return type:

dict[str,list]

get_sorted_data_items()[source]
Return type:list[(str,Data)]
get_all_dimension_tags(allow_same_feature_dim=False)[source]
Parameters:allow_same_feature_dim (bool) –
Return type:list[DimensionTag]
get_batch_info()[source]
Return type:returnn.tf.util.data.BatchInfo
class returnn.tf.network.TFNetwork(config=None, extern_data=None, rnd_seed=None, train_flag=None, eval_flag=None, search_flag=None, parent_layer=None, parent_net=None, extra_parent_net=None, extra_name_prefix=None, inside_rec_time_dim=None, over_rec_time_dim=None, over_rec_time_dim_subs=None, control_flow_ctx=None, absolute_name_prefix=None, name=None)[source]

The main neural network, i.e. collection of interconnected layers, i.e. computation graph with trainable params.

Parameters:
  • config (returnn.config.Config) – only needed to init extern_data if not specified explicitly
  • extern_data (ExternData|None) –
  • rnd_seed (int|None) –
  • train_flag (bool|tf.Tensor) – True if we want to use this model in training, False if in eval, or dynamic
  • eval_flag (bool) – whether to calculate losses. if train_flag is not False, this will be set to True
  • search_flag (bool) – whether we perform a beam-search. see usage
  • parent_layer (returnn.tf.layers.base.LayerBase|None) –
  • parent_net (TFNetwork|None) –
  • extra_parent_net (TFNetwork|None) – we are on the same level (not really a child), but an “extra” net of extra_parent_net
  • extra_name_prefix (str|None) –
  • inside_rec_time_dim (DimensionTag|None) – dim tag of outer rec layer, when run inside the loop (not optimized)
  • over_rec_time_dim (DimensionTag|None) – dim tag of outer rec layer, when optimized out of the loop
  • over_rec_time_dim_subs (set[DimensionTag]|None) – outer rec layer, out of loop, potential shorter
  • control_flow_ctx (returnn.tf.util.data.ControlFlowContext) –
  • absolute_name_prefix (str|None) –
  • name (str) – only for debugging
get_root_network()[source]
Return type:TFNetwork
get_root_ctx_network()[source]
Returns:in contrast to get_root_network(), stop where we have is_root_in_ctx set, and return that network, together with the prefix
Return type:(TFNetwork, str)
get_control_flow_ctx()[source]
Return type:returnn.tf.util.data.ControlFlowContext|None
is_extra_internal_template_construction()[source]
Return type:LayerBase|None
get_absolute_name_scope_prefix()[source]
Returns:TF scope name, always with “/” at the end, or “”
Return type:str
get_absolute_name_prefix()[source]
Returns:name, always with “/” at the end, or “”
Return type:str
construct_from_dict(net_dict, get_layer=None)[source]
Parameters:
  • net_dict (dict[str,dict[str]]) –
  • get_layer (((str)->LayerBase)|None) –
make_extra_net(prefix_name, net_name=None, only_template=False, boundary=False)[source]
Parameters:
  • prefix_name (str) – “extra.Whatever”
  • net_name (str|None) –
  • only_template (bool) –
  • boundary (bool) –
Return type:

TFNetwork

construct_extra_net(net_dict, layer_list, search_flag=None, dep_layers_in_extra=False, check_existing=False, net_name=None, prefix_name=None, base_get_layer=None, base_add_layer=None)[source]

The purpose is to create another net like self but with different flags, e.g. with search_flag = True. That extra_net can have different losses, which will be added. Layers in layer_list will be explicitly re-created in the extra net. Other layers are taken from self.

The creation of the extra net and layers in the extra net can be triggered explicitly by referring to another layer as e.g. "extra.search:layer". When done this way, all the dependencies of it are created in self again; unless you explicitly have called another layer like "extra.search:dep". See test_extra_search() for an example.

Parameters:
  • net_dict (dict[str,dict[str]]) –
  • layer_list (list[str]) –
  • search_flag (bool|None) –
  • dep_layers_in_extra (bool) – layers not in layer_list, but which are not yet created, will be part of the extra net, not self.
  • check_existing (bool) –
  • net_name (str|None) –
  • prefix_name (str|None) – e.g. “extra.search”, such that layers would be called like “extra.search:layer”
  • base_get_layer – like in construct_layer
  • base_add_layer – like in construct_layer
Returns:

the layers created via layer_list (all in extra net)

Return type:

list[LayerBase]

construct_layer(net_dict, name, get_layer=None, add_layer=None, check_existing=True)[source]

This triggers the construction of the layer name if it is not constructed yet. Every construction trigger corresponds to add_layer call (which by default does the actual construction). This can recursively also get/construct other layers (via get_layer).

Parameters:
  • net_dict (dict[str,dict[str]]) –
  • name (str) – layer name
  • -> LayerBase)|None get_layer (((str)) –

    optional, for source layers, for transform_config_dict. By default, this wraps self.construct_layer(). I.e. the name might be misleading, as this should return an existing layer, or construct it if it does not exist yet.

    Note on custom nested/wrapped get_layer:
    This is tricky. When an outer get_layer calls an inner get_layer, then the inner get_layer might construct the layer, and this construction never can get back to the outer get_layer again. This is fine when this is anyway not allowed (e.g. to “base:…”, where the base net is not allowed to access this parent net). But otherwise, this is not an option!
  • LayerBase, dict) -> LayerBase) | None add_layer (((str,) – by default self.add_layer
  • check_existing (bool) – check self.get_layer. (self.layers will be checked in any case)
Return type:

LayerBase

layer_creation_scope(name)[source]
Parameters:name (str) –
Returns:ctx
add_layer(name, layer_class, **layer_desc)[source]

This will construct the layer given the layer_desc arguments, and add it to the network.

Parameters:
  • name (str) –
  • layer_class ((()->LayerBase)|LayerBase) –
  • layer_desc – contains the kwargs for the layer class. the args should have been transformed via layer_class.transform_config_dict before (see construct_layer). must not contain “name” and “network”, which will be automatically added here. should not contain “output”, which will be initialized to layer_class.get_out_data_from_opts. the layer_class will usually then define the layer.output and its placeholder. there is one notable exception: the InternalLayer, where you predefine the output.
get_extern_data(key, mark_data_key_as_used=True)[source]

Returns Data and add the key to self.used_data_keys if mark_data_key_as_used. :param str key: e.g. “data” or “classes” :param bool mark_data_key_as_used: :rtype: Data

get_used_data_keys(exclude_extra_added=True)[source]
Parameters:exclude_extra_added (bool) –
Return type:set[str]
get_seq_tags(mark_data_key_as_used=True, beam=None)[source]
Parameters:
  • mark_data_key_as_used (bool) – for extern_data
  • beam (returnn.tf.util.data.SearchBeam|None) –
Returns:

tensor of shape (batch,) of dtype string, via extern_data

Return type:

tf.Tensor

make_subnet(name, opts)[source]
Parameters:
  • name (str) –
  • opts (dict[str]) –
Return type:

Subnetwork

get_losses_initialized(reduce_func=None, with_total=False)[source]
Parameters:
  • reduce_func (((tf.Tensor)->tf.Tensor)|None) – as in get_losses. e.g. TFUtil.identity
  • with_total (bool) – whether to return total loss / constraints
Returns:

loss name (e.g. “output” or “rec_layer/output” or so) -> LossHolder (initialized, i.e. layer set), and optionally total loss and total constraints (if with_total)

Return type:

(dict[str,LossHolder], tf.Tensor|int|None, tf.Tensor|int|None)

maybe_construct_objective()[source]

Construct self.total_object.

get_objective()[source]
Return type:int|tf.Tensor
Returns:0 if no loss, or tf.Tensor, scalar. loss + constraints. will be used for the updater.
get_total_loss()[source]
Return type:int|tf.Tensor
Returns:0 if no loss, or tf.Tensor, scalar. without constraints. will be used for the updater
get_total_constraints()[source]
Return type:int|tf.Tensor
Returns:0 if no constraints, or tf.Tensor, scalar. will be used for the updater
get_fetches_dict(config=None, should_train=None, should_eval=None, with_summary=False, with_size=False, horovod_collected_reduce_inputs=None)[source]
Parameters:
  • config (Config.Config|None) –
  • should_train (bool|None) –
  • should_eval (bool|None) –
  • with_summary (bool) –
  • with_size (bool) –
  • horovod_collected_reduce_inputs (dict[str,(tf.Tensor,tf.Tensor)]|None) – will write into. see below
Returns:

values and actions which should be calculated and executed in self.run() by the TF session for each step

Return type:

dict[str,tf.Tensor|tf.Operation]

get_used_targets()[source]
Returns:sorted list of targets
Return type:list[str]
get_default_target()[source]
Returns:e.g. “classes”
Return type:str
get_output_layers()[source]
Return type:list[LayerBase]
get_default_output_layer_name()[source]
Return type:str|None
Returns:default output layer name if there is one, or None
get_default_output_layer(must_exist=True)[source]
Parameters:must_exist (bool) – if it does not exist, will raise an exception
Return type:LayerBase|None
Returns:the default output layer
get_layer(layer_name)[source]

Normally just self.layers[layer_name] but with some extra logic added, such as resolving “base:” prefix to the parent network. Raises LayerNotFound if the layer is not found.

Parameters:layer_name (str) –
Return type:LayerBase
get_all_layers_shallow()[source]
Returns:layers, including extra net, not including sub layers
Return type:list[LayerBase]
get_all_layers_deep()[source]
Returns:all layers, including extra net, including sub layers. duplicates are made unique. It might exclude internal layers. We ensure that layers are unique by their absolute name.
Return type:list[LayerBase]
get_params_list()[source]
Returns:list of model variables, i.e. from all the layers, excluding auxiliary vars like global_step
Return type:list[tf.Variable]
get_saveable_param_replace_dict()[source]
Returns:params and saveable_param_replace resolved, union of all layers
Return type:dict[tf.Variable,tensorflow.python.training.saver.BaseSaverBuilder.SaveableObject]
get_saveable_params_list()[source]
Returns:list of model variables or SaveableObject, to save/restore
Return type:list[tf.Variable|tensorflow.python.training.saver.BaseSaverBuilder.SaveableObject]
get_trainable_params()[source]
Returns:list of variables
Return type:list[tf.Variable]
declare_train_params(hidden_layer_selection=None, with_output=None, global_trainable=None)[source]
Parameters:
  • hidden_layer_selection (list[str]|None) –
  • with_output (bool|None) –
  • global_trainable (bool|None) –
get_num_params()[source]
Returns:number of model parameters, i.e. total dimension
Return type:int
initialize_params(session)[source]
Parameters:session (tf.compat.v1.Session) –

Note: This will create a new node to the graph for each call! And it will overwrite also the already initialized variables. So you should call this only once after network construction and before you maybe load some of the params from external sources. If you know that you will load all params explicitly, you would not need to call this function.

get_var_assigner(var)[source]
Parameters:var (tf.Variable) –
get_param_values_dict(session)[source]
Parameters:session (tf.compat.v1.Session) –
Returns:dict: layer_name -> param_name -> variable numpy array
Return type:dict[str,dict[str,numpy.ndarray]]

Note that this excludes auxiliary params.

set_param_values_by_dict(values_dict, ignore_non_existing=False, **kwargs)[source]
Parameters:
  • values_dict (dict[str,dict[str,numpy.ndarray]]) –
  • ignore_non_existing (bool) –
  • kwargs – passed to LayerBase.set_param_values_by_dict()

Note that this excludes auxiliary params.

get_auxiliary_params()[source]
Return type:list[tf.Variable]
get_params_serialized(session)[source]
Parameters:session (tf.compat.v1.Session) –
Return type:TFNetworkParamsSerialized
set_params_by_serialized(serialized, session, **kwargs)[source]
Parameters:
set_global_train_step(step, session)[source]
Parameters:
  • step (int) –
  • session (tf.compat.v1.Session) –
get_global_train_step(session)[source]
Parameters:session (tf.compat.v1.Session) –
Return type:int
get_epoch_step()[source]
Returns:int64
Return type:tf.Tensor
reset_saver()[source]

Resets the tf.train.Saver object which will be used for load_params_from_file() and save_params_to_file(). Warning: Don’t repeat that too often as it will always create new ops in the computation graph.

save_params_to_file(filename, session)[source]

Will save the model parameters to the filename. Note that the model parameters live inside the current TF session.

Parameters:
  • filename (str) –
  • session (tf.compat.v1.Session) –
load_params_from_file(filename, session)[source]

Will load the model parameters from the filename. Note that the model parameters live inside the current TF session.

Parameters:
  • filename (str) –
  • session (tf.compat.v1.Session) –
print_network_info(name='Network')[source]
Parameters:name (str) –
Returns:nothing, prints very brief net topology on log
cond_on_train(fn_train, fn_eval)[source]

Uses fn_train() or fn_eval() base on self.train_flag. It will be a branched evaluation.

Parameters:
  • fn_train (()->tf.Tensor) –
  • fn_eval (()->tf.Tensor) –
Returns:

fn_train() if self.train_flag else fn_eval()

Return type:

tf.Tensor

get_search_choices(sources=None, src=None, base_search_choice=None, _layer_to_search_choices=None, debug_stream=None)[source]

Recursively searches through all sources, and if there is a ChoiceLayer / any layer with search_choices, returns it. Could also go to the parent network. If there are multiple, it assumes they are on the same search-sequence in the search-tree and it will return the last one.

Parameters:
  • src (LayerBase|None) –
  • base_search_choice (LayerBase|None) –
  • sources (list[LayerBase]|None) –
  • _layer_to_search_choices (dict[LayerBase]|None) – keep track of visited layers in case there are circular deps
  • debug_stream (typing.TextIO|None) – if given, will print additional debug info into it
Returns:

(direct or indirect) source LayerBase which has search_choices, or None

Return type:

LayerBase|None

debug_search_choices(base_search_choice)[source]
Parameters:base_search_choice (LayerBase) –
Returns:nothing, by intention, such that constructs like assert …, debug_search_choices(…) or (…) work
get_data_batch_dim()[source]

Get the batch-dim size, i.e. amount of sequences in the current batch. Consider that the data tensor is usually of shape [batch, time, dim], this would return shape(data)[0].

The code currently assumes that the batch-dim can be taken from the extern data. If it does not have that available for some reason (e.g. some subnetwork), it will try some alternative sources and assumes that they have the correct batch-dim.

Note that the batch-dim usually stays always the same across the whole network and also every individual batch sequence will stay related. One notable exception of this is the choice layer, where the batch-dim will get expanded by the beam search if search is used, as well as in all following layers, until there is a decide layer.

Returns:int scalar tensor which states the batch-dim
Return type:int|tf.Tensor
get_global_batch_info()[source]
Returns:global batch info from root network from extern data
Return type:returnn.tf.util.data.BatchInfo
set_rec_step_info(i, end_flag=None, end_flag_source=None, seq_lens=None)[source]

Used by _SubnetworkRecCell.

Parameters:
  • i (tf.Tensor) – scalar, int32, current step (time)
  • end_flag (tf.Tensor|None) – (batch,), bool, says that the current sequence has ended
  • end_flag_source (LayerBase|None) –
  • seq_lens (tf.Tensor|None) – (batch,) int32, seq lens
is_inside_rec_layer(inside_loop=True)[source]
Parameters:inside_loop (bool) – only True if we are inside the loop of the most recent rec layer
Returns:whether we are inside a RecLayer (with inside_loop: and not optimized out-of-the-loop). At template construction inside a rec layer, this is always true, but the rec layer itself does not exist yet.
Return type:bool

Also see get_inside_rec_time_dim() and get_rec_parent_layer().

get_inside_rec_time_dim(inside_loop=True)[source]
Parameters:inside_loop (bool) – only True if we are inside the loop of the most recent rec layer
Returns:when the net is inside a rec loop (RecLayer and not optimized out of the loop), this returns the dim tag the rec layer iterates over
Return type:DimensionTag|None
get_all_rec_time_dims()[source]
Returns:all rec time dims, moved out or not, including all parents
Return type:set[DimensionTag]
get_rec_parent_layer(inside_loop=True)[source]
Parameters:inside_loop (bool) – only return if the network is constructed within the loop (not moved out) of the most recent parent rec layer
Returns:if we are a subnet of a RecLayer, will return the RecLayer instance. At template construction time, this is always None.
Return type:returnn.tf.layers.rec.RecLayer|None
have_rec_step_info()[source]
Return type:bool
get_rec_step_info(must_exist=True)[source]
Parameters:must_exist (bool) – if True, will throw exception if not available
Return type:returnn.tf.layers.rec.RecStepInfoLayer|None
get_rec_step_index()[source]

Assumes that have_rec_step_info is True.

Return type:tf.Tensor
Returns:scalar, int32
get_config(consider_global_config=True, fallback_dummy_config=True)[source]
Parameters:
  • consider_global_config (bool) – if no config is set, check for global config
  • fallback_dummy_config (bool) – if no config, return a new empty Config, otherwise return None
Return type:

returnn.config.Config|None

static register_post_control_dependencies(deps)[source]

Will register the control dependencies or globally for a session run on this network. This can e.g. be called inside self.post_init. We use UPDATE_OPS, as that is also e.g. used by batchnorm. See:

Parameters:deps (list[tf.Tensor|tf.Operation]) –
Returns:nothing
static get_post_control_dependencies()[source]
Return type:list[tf.Operation]
register_graph_reset_callback(cb)[source]

Note: These callbacks are not called automatically. You explicitly have to call call_graph_reset_callbacks().

Note: We don’t store this in the graph itself (e.g. via tf.get_collection), as we don’t want to serialize this (which would also lead to an error, because it cannot be serialized).

Note: Currently these callbacks might get called multiple times, so make sure that this is not a problem. Also make sure that the network/session is still in a valid state after this has been called, e.g. such that further session runs would still work correctly.

Note: These callbacks will only be called if there was not any error.

Parameters:cb (function|()->None) –
get_graph_reset_callbacks()[source]
Return type:list[()->None]
call_graph_reset_callbacks()[source]

Calls any callbacks registered via register_graph_reset_callback().

set_run_opts(epoch, dataset_name)[source]

The run options are valid during one loop over some dataset.

Contrary to epoch_step, train_flag, etc, we do not provide these as TF placeholders, for convenience, because it is not needed right now. If it is needed, it probably is easier to introduce auxiliary TF variables (on CPU) instead and just set them once here.

Parameters:
  • epoch (int) –
  • dataset_name (str|None) –
get_run_opts()[source]
Return type:dict[str]
register_run_finished_callback(cb)[source]
Parameters:cb (function|()->None) –
set_run_finished(error_occurred=False)[source]

Maybe calls any callbacks registered via register_run_finished_callback() (if no error occurred) and cleans up the run opts.

Parameters:error_occurred (bool) –
classmethod get_network_stack()[source]
Return type:list[TFNetwork]
classmethod get_current_network(must_exist=True)[source]
Parameters:must_exist (bool) –
Return type:TFNetwork|None
register_network_scope()[source]

Registers a ref to this network inside the current TF computation graph.

get_search_choices_from_beam(beam)[source]

Currently we have somewhat redundant information in returnn.tf.util.data.SearchBeam (which is totally independent from other things in RETURNN (which is good)) and returnn.tf.layers.base.SearchChoices (which is more dependent on the RETURNN layers,

and has some more info).

The Data (which is also independent from other things in RETURNN (which is also good)) only knows about returnn.tf.util.data.SearchBeam but not about returnn.tf.layers.base.SearchChoices. Thus there are situations where we only have a ref to the former, but like to get a ref to the latter.

Note that this might (hopefully) get cleaned up at some point…

Parameters:beam (returnn.tf.util.data.SearchBeam) –
Return type:returnn.tf.layers.base.SearchChoices|None
register_search_choices_for_beam(beam, search_choices)[source]
Parameters:
class returnn.tf.network.Subnetwork(parent_net, name, opts=None)[source]

Represents a subnetwork.

Despite the different namespace, optionally some variable sharing, and optionally some custom input data, layers behave just as in the root network, with the same dependency resolution (both ways). I.e. a layer outside can depend only on a single sub layer and not the whole subnetwork (in contrast to LayerBase.get_sub_layer()).

This is usually used with SubnetworkLayer, via LayerBase:cls_get_sub_network().

This works for custom calls on TFNetwork.construct_layer() with custom get_layer or add_layer e.g. in template construction from the RecLayer subnetwork and doesn’t require extra logic for this.

This has also a mode to start its own template construction, for the case this layer is embedded in another layer (e.g. CondLayer or MaskedComputationLayer, in contrast to SubnetworkLayer). This is triggered by a special type of extra parent network with extra_only_template set. This implies that the parent (non-extra) network can not directly access the sub network, which is important for the template construction here (see _construct_template_subnet()).

A special extra parent can also have the extra_boundary flag set, which triggers that we have our own construction code (but not using templates, but constructing the real layers). This is used also for the embedded case (e.g. MaskedComputationLayer). This is needed when the parent (non-extra) network cannot directly access this sub network.

Parameters:
  • parent_net (TFNetwork) –
  • name (str) –
  • opts (dict[str]|None) –
get_sub_layer_func(base_get_layer)[source]
Parameters:base_get_layer (((str)->LayerBase)|None) –
Return type:(str)->LayerBase
get_layer_func(get_layer)[source]
Parameters:get_layer ((str)->LayerBase) –
Return type:(str)->LayerBase
construct_layer(name, parent_get_layer=None)[source]

With default parent_get_layer, this will not trigger recursive constructions in the parent net, but any recursive construction in this subnet.

Parameters:
  • name (str) –
  • parent_get_layer (((str)->LayerBase)|None) –
Return type:

LayerBase

construct_all(parent_get_layer=None)[source]

Trigger the standard construction of all layers in the net dict.

Parameters:parent_get_layer (((str)->LayerBase)|None) –
complete_construction_parent_subnet_layer(parent_get_layer=None)[source]
Parameters:parent_get_layer (((str)->LayerBase)|None) –
Return type:returnn.tf.layers.basic.SubnetworkLayer
have_layer(name)[source]
Parameters:name (str) –
Return type:bool
get_layer_desc(name)[source]
Parameters:name (str) –
Return type:dict[str]
get_layer_class(name)[source]
Parameters:name (str) –
Return type:type[LayerBase]
class returnn.tf.network.TFNetworkParamsSerialized(values_dict, global_train_step)[source]

Holds all the params as numpy arrays, including auxiliary params.

Parameters:
  • values_dict (dict[str,dict[str,numpy.ndarray]]) – dict: layer_name -> param_name -> variable numpy array
  • global_train_step (int) –
class returnn.tf.network.LossHolder(name, loss, layer_output, reduce_func=None, layer=None, loss_value=None, error_value=None, norm_factor=None, only_on_eval=None, network=None)[source]

This object just keeps a reference to the loss/error value, and does the necessary logic to collect it, and also the normalization logic. Every new computation (nodes in the computation graph) must be constructed on demand, to allow first to collect all possible losses without calculating them, and then calculating them in the right context (e.g. inside a while_loop, or so).

After construction, you should call init() before usage, in case you do not provide layer here.

Parameters:
  • name (str) – The name uniquely identifies the loss. Earlier, this was the same as the layer name. This is still true for simple cases, but for losses coming from a subnetwork or other extended losses, it can be something else. It could look like “output”, or “output/sublayer”.
  • layer (LayerBase) – We can always point to a layer where this comes from (either in the subnet, or the parent layer).
  • layer_output (Data) – template describing the layer output
  • network (TFNetwork) – for which network to create this LossHolder. might be different from layer.network
  • loss (returnn.tf.layers.base.Loss) –
  • reduce_func (((tf.Tensor)->tf.Tensor)|None) – if given, will overwrite the reduce func for the loss. By default, every loss_value and error_value is a scalar (sum or average over the batches, and over the frames for frame-wise losses). However, if you provide reduce_func = TFUtil.identity, you can get the unreduced tensor.
  • loss_value (tf.Tensor|None) –
  • error_value (tf.Tensor|None) –
  • norm_factor (tf.Tensor) –
  • only_on_eval (bool) –
init(layer)[source]

It will just set the layer. The LossHolder is initialized if the layer is set.

Parameters:layer (LayerBase) –
Returns:self
Return type:LossHolder
get_layer()[source]
Returns:layer. assumes that it is set
Return type:LayerBase
get_only_on_eval()[source]
Returns:only_on_eval flag. assumes that it is set
Return type:bool
get_tf_name()[source]
Returns:name which can be used for a TF op, thus contains no “/” or other special chars
Return type:str
get_loss_value()[source]
Returns:loss value. scalar
Return type:tf.Tensor|None
get_loss_value_for_fetch()[source]
Returns:loss value for fetch. scalar. same as loss_value, but maybe with additional checks
Return type:tf.Tensor|None
get_loss_value_for_objective()[source]
Returns:loss value for objective. scalar. might be scaled (scale) and/or normalized (use_normalized_loss)
Return type:tf.Tensor|None
get_error_value()[source]
Returns:error value for fetch. scalar
Return type:tf.Tensor|None
get_norm_factor()[source]
Returns:norm factor for loss and error. scalar
Return type:tf.Tensor
get_normalized_loss_value_per_seq()[source]
Returns:(batch,) or None if loss is None
Return type:tf.Tensor|None
get_normalized_error_value_per_seq()[source]
Returns:(batch,) or None if error is None
Return type:tf.Tensor|None
get_loss_value_per_pos()[source]
Returns:(batch,time) or None if loss is None
Return type:tf.Tensor|None
get_error_value_per_pos()[source]
Returns:(batch,time) or None if error is None
Return type:tf.Tensor|None
copy_new_base(name=None, layer=None, network=None, reduce_func=None)[source]
Parameters:
  • layer (LayerBase) –
  • network (TFNetwork) –
  • name (str) –
  • reduce_func (((tf.Tensor)->tf.Tensor)|None) –
Returns:

new copy of LossHolder

Return type:

LossHolder

exception returnn.tf.network.NetworkLayerException(message, layer_name, network, net_dict=None)[source]

Some exception by the network, e.g. during construction.

Parameters:
  • message (str) –
  • layer_name (str) –
  • network (TFNetwork) –
  • net_dict (dict[str]|None) –
exception returnn.tf.network.NetworkConstructionDependencyLoopException(network, layer_name, constructing_layers, net_dict)[source]

This is raised when there is a dependency loop in the network construction.

Parameters:
  • network (TFNetwork) –
  • layer_name (str) –
  • constructing_layers (list[str]) –
  • net_dict (dict[str,dict[str]]) –
exception returnn.tf.network.LayerNotFound(message, layer_name, network, net_dict=None)[source]

Via TFNetwork.get_layer().

Parameters:
  • message (str) –
  • layer_name (str) –
  • network (TFNetwork) –
  • net_dict (dict[str]|None) –
returnn.tf.network.help_on_tf_exception(session, exception, fetches, feed_dict=None, meta_step_info=None, extern_data=None, file=<_io.TextIOWrapper name='<stdout>' mode='w' encoding='UTF-8'>)[source]

Generic debugging helper, on any TF exception (or even any other exception as well). Will try to provide as much helpful context information as possible. (This is not in TFUtil because it depends on ExternData, which is only defined here.)

Parameters:
  • session (tf.compat.v1.Session) –
  • exception (tf.errors.OpError|BaseException) –
  • fetches (tf.Tensor|list[tf.Tensor]|dict[str,tf.Tensor]|object|None) –
  • feed_dict (dict[tf.Tensor,numpy.ndarray]|None) –
  • meta_step_info (dict[str]|None) –
  • extern_data (ExternData|None) –
  • file (typing.IO[str]|io.TextIOBase|io.StringIO) –
class returnn.tf.network.CustomCheckpointLoader(filename, saveable_params, params_prefix='', load_if_prefix='', ignore_missing=False, ignore_params=(), ignore_params_prefixes=(), var_name_mapping=None, network=None)[source]

This uses tf.train.NewCheckpointReader.

It would do automatic conversions if needed, e.g. between different LSTM implementations. However, be careful that for some LSTM implementation, there is an additional forget_bias option, which is an additional scalar which gets added (not to the param, but to the forget value directly). When we convert the parameters, this is ignored, and you must take care about that explicitly to make sure you get the same results.

It tries to automatically resolve renames, similar to this:

Also see:

Parameters:
  • filename (str) – filepattern for NewCheckpointReader or .index/.meta file path
  • saveable_params (list[tf.Variable|tensorflow.python.training.saver.BaseSaverBuilder.SaveableObject]) –
  • params_prefix (str) – expect that all vars in saveable_params have this prefix, and remove it
  • load_if_prefix (str) – if given, only load variables with a name containing this string. the variables in the file are expected to have the same name but without this string.
  • ignore_missing (bool) – any vars in the model, which are not found in the checkpoint, will be ignored. however, if there is no single var in the checkpoint, this is still an error.
  • ignore_params (typing.Container[str]) – these param (by name) will not be loaded
  • ignore_params_prefixes (typing.Iterable[str]) – these param (by prefix name) will not be loaded
  • var_name_mapping (dict[str,str]) – defines a custom mapping (new_name -> name_in_checkpoint) for renamed vars in the checkpoint
  • network (TFNetwork) –
class CustomParamImporter(layer, checkpoint_loader)[source]

Helper class for custom param loading.

Parameters:
assign_var(var, session)[source]
Parameters:
  • var (tf.Variable) –
  • session (tf.compat.v1.Session) –
class VariableValue(value=None, custom_param_importer=None)[source]

Helper to assign some variable.

Parameters:
  • value (numpy.ndarray|None) –
  • custom_param_importer (CustomCheckpointLoader.CustomParamImporter|None) –
assign_var(var, session)[source]
Parameters:
  • var (tf.Variable) –
  • session (tf.compat.v1.Session) –
get_variable_value_map()[source]
Returns:var -> numpy array
Return type:dict[tf.Variable,CustomCheckpointLoader.VariableValue]
load_now(session)[source]
Parameters:session (tf.compat.v1.Session) –
Returns:nothing, will assign the variables in the session
set_as_custom_init()[source]

Make sure that this loader is used during initialization.

returnn.tf.network.set_custom_post_init(var, func)[source]

It registers the provided func such that it gets called for this variable in TFNetwork.initialize_params().

Parameters:
  • var (tf.Variable) –
  • func ((tf.compat.v1.Session)->None) –
returnn.tf.network.have_custom_post_init(var)[source]
Parameters:var (tf.Variable) –
Returns:whether set_custom_post_init() was called on this var, i.e. we have custom init
Return type:bool