Table of Contents
Shortcuts

ignite.engine#

class ignite.engine.engine.Engine(process_function)[source]#

Runs a given process_function over each batch of a dataset, emitting events as it goes.

Parameters

process_function (callable) – A function receiving a handle to the engine and the current batch in each iteration, and returns data to be stored in the engine’s state.

state#

object that is used to pass internal and user-defined state between event handlers. It is created with the engine and its attributes (e.g. state.iteration, state.epoch etc) are reset on every run().

Type

State

last_event_name#

last event name triggered by the engine.

Type

Events

Examples

Create a basic trainer

def update_model(engine, batch):
    inputs, targets = batch
    optimizer.zero_grad()
    outputs = model(inputs)
    loss = criterion(outputs, targets)
    loss.backward()
    optimizer.step()
    return loss.item()

trainer = Engine(update_model)

@trainer.on(Events.ITERATION_COMPLETED(every=100))
def log_training(engine):
    batch_loss = engine.state.output
    lr = optimizer.param_groups[0]['lr']
    e = engine.state.epoch
    n = engine.state.max_epochs
    i = engine.state.iteration
    print("Epoch {}/{} : {} - batch loss: {}, lr: {}".format(e, n, i, batch_loss, lr))

trainer.run(data_loader, max_epochs=5)

> Epoch 1/5 : 100 - batch loss: 0.10874069479016124, lr: 0.01
> ...
> Epoch 2/5 : 1700 - batch loss: 0.4217900575859437, lr: 0.01

Create a basic evaluator to compute metrics

from ignite.metrics import Accuracy

def predict_on_batch(engine, batch)
    model.eval()
    with torch.no_grad():
        x, y = prepare_batch(batch, device=device, non_blocking=non_blocking)
        y_pred = model(x)

    return y_pred, y

evaluator = Engine(predict_on_batch)
Accuracy().attach(evaluator, "val_acc")
evaluator.run(val_dataloader)

Compute image mean/std on training dataset

from ignite.metrics import Average

def compute_mean_std(engine, batch):
    b, c, *_ = batch['image'].shape
    data = batch['image'].reshape(b, c, -1).to(dtype=torch.float64)
    mean = torch.mean(data, dim=-1).sum(dim=0)
    mean2 = torch.mean(data ** 2, dim=-1).sum(dim=0)
    return {"mean": mean, "mean^2": mean2}

compute_engine = Engine(compute_mean_std)
img_mean = Average(output_transform=lambda output: output['mean'])
img_mean.attach(compute_engine, 'mean')
img_mean2 = Average(output_transform=lambda output: output['mean^2'])
img_mean2.attach(compute_engine, 'mean2')
state = compute_engine.run(train_loader)
state.metrics['std'] = torch.sqrt(state.metrics['mean2'] - state.metrics['mean'] ** 2)
mean = state.metrics['mean'].tolist()
std = state.metrics['std'].tolist()

Resume engine’s run from a state. User can load a state_dict and run engine starting from loaded state :

# Restore from an epoch
state_dict = {"epoch": 3, "max_epochs": 100, "epoch_length": len(data_loader)}
# or an iteration
# state_dict = {"iteration": 500, "max_epochs": 100, "epoch_length": len(data_loader)}

trainer = Engine(...)
trainer.load_state_dict(state_dict)
trainer.run(data)
add_event_handler(event_name, handler, *args, **kwargs)[source]#

Add an event handler to be executed when the specified event is fired.

Parameters

  • event_name (Any) – An event or a list of events to attach the handler. Valid events are from Events or any event_name added by register_events().

  • handler (callable) – the callable event handler that should be invoked. No restrictions on its signature. The first argument can be optionally engine, the Engine object, handler is bound to.

  • *args – optional args to be passed to handler.

  • **kwargs – optional keyword args to be passed to handler.

Note

Note that other arguments can be passed to the handler in addition to the *args and **kwargs passed here, for example during EXCEPTION_RAISED.

Returns

RemovableEventHandle, which can be used to remove the handler.

Parameters

Example usage:

engine = Engine(process_function)

def print_epoch(engine):
    print("Epoch: {}".format(engine.state.epoch))

engine.add_event_handler(Events.EPOCH_COMPLETED, print_epoch)

events_list = Events.EPOCH_COMPLETED | Events.COMPLETED

def execute_something():
    # do some thing not related to engine
    pass

engine.add_event_handler(events_list, execute_something)

Note

Since v0.3.0, Events become more flexible and allow to pass an event filter to the Engine. See Events for more details.

fire_event(event_name)[source]#

Execute all the handlers associated with given event.

This method executes all handlers associated with the event event_name. This is the method used in run() to call the core events found in Events.

Custom events can be fired if they have been registered before with register_events(). The engine state attribute should be used to exchange “dynamic” data among process_function and handlers.

This method is called automatically for core events. If no custom events are used in the engine, there is no need for the user to call the method.

Parameters

event_name (Any) – event for which the handlers should be executed. Valid events are from Events or any event_name added by register_events().

Return type

None

has_event_handler(handler, event_name=None)[source]#

Check if the specified event has the specified handler.

Parameters

  • handler (callable) – the callable event handler.

  • event_name (Optional[Any]) – The event the handler attached to. Set this to None to search all events.

load_state_dict(state_dict)[source]#

Setups engine from state_dict.

State dictionary should contain keys: iteration or epoch and max_epochs, epoch_length and seed. If engine.state_dict_user_keys contains keys, they should be also present in the state dictionary. Iteration and epoch values are 0-based: the first iteration or epoch is zero.

This method does not remove any custom attributs added by user.

Parameters

state_dict (Mapping) – a dict with parameters

Return type

None

# Restore from the 4rd epoch
state_dict = {"epoch": 3, "max_epochs": 100, "epoch_length": len(data_loader)}
# or 500th iteration
# state_dict = {"iteration": 499, "max_epochs": 100, "epoch_length": len(data_loader)}

trainer = Engine(...)
trainer.load_state_dict(state_dict)
trainer.run(data)
on(event_name, *args, **kwargs)[source]#

Decorator shortcut for add_event_handler.

Parameters

  • event_name – An event to attach the handler to. Valid events are from Events or any event_name added by register_events().

  • *args – optional args to be passed to handler.

  • **kwargs – optional keyword args to be passed to handler.

Example usage:

engine = Engine(process_function)

@engine.on(Events.EPOCH_COMPLETED)
def print_epoch():
    print("Epoch: {}".format(engine.state.epoch))

@engine.on(Events.EPOCH_COMPLETED | Events.COMPLETED)
def execute_something():
    # do some thing not related to engine
    pass
register_events(*event_names, event_to_attr=None)[source]#

Add events that can be fired.

Registering an event will let the user fire these events at any point. This opens the door to make the run() loop even more configurable.

By default, the events from Events are registered.

Parameters

  • *event_names – An object (ideally a string or int) to define the name of the event being supported.

  • event_to_attr (dict, optional) – A dictionary to map an event to a state attribute.

Return type

None

Example usage:

from ignite.engine import Engine, EventEnum

class CustomEvents(EventEnum):
    FOO_EVENT = "foo_event"
    BAR_EVENT = "bar_event"

engine = Engine(process_function)
engine.register_events(*CustomEvents)

Example with State Attribute:

from enum import Enum
from ignite.engine import Engine, EventEnum

class TBPTT_Events(EventEnum):
    TIME_ITERATION_STARTED = "time_iteration_started"
    TIME_ITERATION_COMPLETED = "time_iteration_completed"

TBPTT_event_to_attr = {
    TBPTT_Events.TIME_ITERATION_STARTED: 'time_iteration',
    TBPTT_Events.TIME_ITERATION_COMPLETED: 'time_iteration'
}

engine = Engine(process_function)
engine.register_events(*TBPTT_Events, event_to_attr=TBPTT_event_to_attr)
engine.run(data)
# engine.state contains an attribute time_iteration, which can be accessed using engine.state.time_iteration
remove_event_handler(handler, event_name)[source]#

Remove event handler handler from registered handlers of the engine

Parameters

  • handler (callable) – the callable event handler that should be removed

  • event_name (Any) – The event the handler attached to.

run(data, max_epochs=None, epoch_length=None, seed=None)[source]#

Runs the process_function over the passed data.

Engine has a state and the following logic is applied in this function:

  • At the first call, new state is defined by max_epochs, epoch_length, seed if provided. A timer for

    total and per-epoch time is initialized when Events.STARTED is handled.

  • If state is already defined such that there are iterations to run until max_epochs and no input arguments

    provided, state is kept and used in the function.

  • If state is defined and engine is “done” (no iterations to run until max_epochs), a new state is defined.

  • If state is defined, engine is NOT “done”, then input arguments if provided override defined state.

Parameters

  • data (Iterable) – Collection of batches allowing repeated iteration (e.g., list or DataLoader).

  • max_epochs (int, optional) – Max epochs to run for (default: None). If a new state should be created (first run or run again from ended engine), it’s default value is 1. If run is resuming from a state, provided max_epochs will be taken into account and should be larger than engine.state.max_epochs.

  • epoch_length (int, optional) – Number of iterations to count as one epoch. By default, it can be set as len(data). If data is an iterator and epoch_length is not set, then it will be automatically determined as the iteration on which data iterator raises StopIteration. This argument should not change if run is resuming from a state.

  • seed (int, optional) – Deprecated argument. Please, use torch.manual_seed or manual_seed().

Returns

output state.

Return type

State

Note

User can dynamically preprocess input batch at ITERATION_STARTED and store output batch in engine.state.batch. Latter is passed as usually to process_function as argument:

trainer = ...

@trainer.on(Events.ITERATION_STARTED)
def switch_batch(engine):
    engine.state.batch = preprocess_batch(engine.state.batch)
set_data(data)[source]#

Method to set data. After calling the method the next batch passed to processing_function is from newly provided data. Please, note that epoch length is not modified.

Parameters

data (Iterable) – Collection of batches allowing repeated iteration (e.g., list or DataLoader).

Example usage:

User can switch data provider during the training:

data1 = ...
data2 = ...

switch_iteration = 5000

def train_step(e, batch):
    # when iteration <= switch_iteration
    # batch is from data1
    # when iteration > switch_iteration
    # batch is from data2
    ...

trainer = Engine(train_step)

@trainer.on(Events.ITERATION_COMPLETED(once=switch_iteration))
def switch_dataloader():
    trainer.set_data(data2)

trainer.run(data1, max_epochs=100)
state_dict()[source]#

Returns a dictionary containing engine’s state: “seed”, “epoch_length”, “max_epochs” and “iteration” and other state values defined by engine.state_dict_user_keys

engine = Engine(...)
engine.state_dict_user_keys.append("alpha")
engine.state_dict_user_keys.append("beta")
...

@engine.on(Events.STARTED)
def init_user_value(_):
     engine.state.alpha = 0.1
     engine.state.beta = 1.0

@engine.on(Events.COMPLETED)
def save_engine(_):
    state_dict = engine.state_dict()
    assert "alpha" in state_dict and "beta" in state_dict
    torch.save(state_dict, "/tmp/engine.pt")
Returns

a dictionary containing engine’s state

Return type

OrderedDict

terminate()[source]#

Sends terminate signal to the engine, so that it terminates completely the run after the current iteration.

Return type

None

terminate_epoch()[source]#

Sends terminate signal to the engine, so that it terminates the current epoch after the current iteration.

Return type

None

class ignite.engine.engine.Engine(process_function)[source]#

Runs a given process_function over each batch of a dataset, emitting events as it goes.

Parameters

process_function (callable) – A function receiving a handle to the engine and the current batch in each iteration, and returns data to be stored in the engine’s state.

state#

object that is used to pass internal and user-defined state between event handlers. It is created with the engine and its attributes (e.g. state.iteration, state.epoch etc) are reset on every run().

Type

State

last_event_name#

last event name triggered by the engine.

Type

Events

Examples

Create a basic trainer

def update_model(engine, batch):
    inputs, targets = batch
    optimizer.zero_grad()
    outputs = model(inputs)
    loss = criterion(outputs, targets)
    loss.backward()
    optimizer.step()
    return loss.item()

trainer = Engine(update_model)

@trainer.on(Events.ITERATION_COMPLETED(every=100))
def log_training(engine):
    batch_loss = engine.state.output
    lr = optimizer.param_groups[0]['lr']
    e = engine.state.epoch
    n = engine.state.max_epochs
    i = engine.state.iteration
    print("Epoch {}/{} : {} - batch loss: {}, lr: {}".format(e, n, i, batch_loss, lr))

trainer.run(data_loader, max_epochs=5)

> Epoch 1/5 : 100 - batch loss: 0.10874069479016124, lr: 0.01
> ...
> Epoch 2/5 : 1700 - batch loss: 0.4217900575859437, lr: 0.01

Create a basic evaluator to compute metrics

from ignite.metrics import Accuracy

def predict_on_batch(engine, batch)
    model.eval()
    with torch.no_grad():
        x, y = prepare_batch(batch, device=device, non_blocking=non_blocking)
        y_pred = model(x)

    return y_pred, y

evaluator = Engine(predict_on_batch)
Accuracy().attach(evaluator, "val_acc")
evaluator.run(val_dataloader)

Compute image mean/std on training dataset

from ignite.metrics import Average

def compute_mean_std(engine, batch):
    b, c, *_ = batch['image'].shape
    data = batch['image'].reshape(b, c, -1).to(dtype=torch.float64)
    mean = torch.mean(data, dim=-1).sum(dim=0)
    mean2 = torch.mean(data ** 2, dim=-1).sum(dim=0)
    return {"mean": mean, "mean^2": mean2}

compute_engine = Engine(compute_mean_std)
img_mean = Average(output_transform=lambda output: output['mean'])
img_mean.attach(compute_engine, 'mean')
img_mean2 = Average(output_transform=lambda output: output['mean^2'])
img_mean2.attach(compute_engine, 'mean2')
state = compute_engine.run(train_loader)
state.metrics['std'] = torch.sqrt(state.metrics['mean2'] - state.metrics['mean'] ** 2)
mean = state.metrics['mean'].tolist()
std = state.metrics['std'].tolist()

Resume engine’s run from a state. User can load a state_dict and run engine starting from loaded state :

# Restore from an epoch
state_dict = {"epoch": 3, "max_epochs": 100, "epoch_length": len(data_loader)}
# or an iteration
# state_dict = {"iteration": 500, "max_epochs": 100, "epoch_length": len(data_loader)}

trainer = Engine(...)
trainer.load_state_dict(state_dict)
trainer.run(data)
add_event_handler(event_name, handler, *args, **kwargs)[source]#

Add an event handler to be executed when the specified event is fired.

Parameters

  • event_name (Any) – An event or a list of events to attach the handler. Valid events are from Events or any event_name added by register_events().

  • handler (callable) – the callable event handler that should be invoked. No restrictions on its signature. The first argument can be optionally engine, the Engine object, handler is bound to.

  • *args – optional args to be passed to handler.

  • **kwargs – optional keyword args to be passed to handler.

Note

Note that other arguments can be passed to the handler in addition to the *args and **kwargs passed here, for example during EXCEPTION_RAISED.

Returns

RemovableEventHandle, which can be used to remove the handler.

Parameters

Example usage:

engine = Engine(process_function)

def print_epoch(engine):
    print("Epoch: {}".format(engine.state.epoch))

engine.add_event_handler(Events.EPOCH_COMPLETED, print_epoch)

events_list = Events.EPOCH_COMPLETED | Events.COMPLETED

def execute_something():
    # do some thing not related to engine
    pass

engine.add_event_handler(events_list, execute_something)

Note

Since v0.3.0, Events become more flexible and allow to pass an event filter to the Engine. See Events for more details.

fire_event(event_name)[source]#

Execute all the handlers associated with given event.

This method executes all handlers associated with the event event_name. This is the method used in run() to call the core events found in Events.

Custom events can be fired if they have been registered before with register_events(). The engine state attribute should be used to exchange “dynamic” data among process_function and handlers.

This method is called automatically for core events. If no custom events are used in the engine, there is no need for the user to call the method.

Parameters

event_name (Any) – event for which the handlers should be executed. Valid events are from Events or any event_name added by register_events().

Return type

None

has_event_handler(handler, event_name=None)[source]#

Check if the specified event has the specified handler.

Parameters

  • handler (callable) – the callable event handler.

  • event_name (Optional[Any]) – The event the handler attached to. Set this to None to search all events.

load_state_dict(state_dict)[source]#

Setups engine from state_dict.

State dictionary should contain keys: iteration or epoch and max_epochs, epoch_length and seed. If engine.state_dict_user_keys contains keys, they should be also present in the state dictionary. Iteration and epoch values are 0-based: the first iteration or epoch is zero.

This method does not remove any custom attributs added by user.

Parameters

state_dict (Mapping) – a dict with parameters

Return type

None

# Restore from the 4rd epoch
state_dict = {"epoch": 3, "max_epochs": 100, "epoch_length": len(data_loader)}
# or 500th iteration
# state_dict = {"iteration": 499, "max_epochs": 100, "epoch_length": len(data_loader)}

trainer = Engine(...)
trainer.load_state_dict(state_dict)
trainer.run(data)
on(event_name, *args, **kwargs)[source]#

Decorator shortcut for add_event_handler.

Parameters

  • event_name – An event to attach the handler to. Valid events are from Events or any event_name added by register_events().

  • *args – optional args to be passed to handler.

  • **kwargs – optional keyword args to be passed to handler.

Example usage:

engine = Engine(process_function)

@engine.on(Events.EPOCH_COMPLETED)
def print_epoch():
    print("Epoch: {}".format(engine.state.epoch))

@engine.on(Events.EPOCH_COMPLETED | Events.COMPLETED)
def execute_something():
    # do some thing not related to engine
    pass
register_events(*event_names, event_to_attr=None)[source]#

Add events that can be fired.

Registering an event will let the user fire these events at any point. This opens the door to make the run() loop even more configurable.

By default, the events from Events are registered.

Parameters

  • *event_names – An object (ideally a string or int) to define the name of the event being supported.

  • event_to_attr (dict, optional) – A dictionary to map an event to a state attribute.

Return type

None

Example usage:

from ignite.engine import Engine, EventEnum

class CustomEvents(EventEnum):
    FOO_EVENT = "foo_event"
    BAR_EVENT = "bar_event"

engine = Engine(process_function)
engine.register_events(*CustomEvents)

Example with State Attribute:

from enum import Enum
from ignite.engine import Engine, EventEnum

class TBPTT_Events(EventEnum):
    TIME_ITERATION_STARTED = "time_iteration_started"
    TIME_ITERATION_COMPLETED = "time_iteration_completed"

TBPTT_event_to_attr = {
    TBPTT_Events.TIME_ITERATION_STARTED: 'time_iteration',
    TBPTT_Events.TIME_ITERATION_COMPLETED: 'time_iteration'
}

engine = Engine(process_function)
engine.register_events(*TBPTT_Events, event_to_attr=TBPTT_event_to_attr)
engine.run(data)
# engine.state contains an attribute time_iteration, which can be accessed using engine.state.time_iteration
remove_event_handler(handler, event_name)[source]#

Remove event handler handler from registered handlers of the engine

Parameters

  • handler (callable) – the callable event handler that should be removed

  • event_name (Any) – The event the handler attached to.

run(data, max_epochs=None, epoch_length=None, seed=None)[source]#

Runs the process_function over the passed data.

Engine has a state and the following logic is applied in this function:

  • At the first call, new state is defined by max_epochs, epoch_length, seed if provided. A timer for

    total and per-epoch time is initialized when Events.STARTED is handled.

  • If state is already defined such that there are iterations to run until max_epochs and no input arguments

    provided, state is kept and used in the function.

  • If state is defined and engine is “done” (no iterations to run until max_epochs), a new state is defined.

  • If state is defined, engine is NOT “done”, then input arguments if provided override defined state.

Parameters

  • data (Iterable) – Collection of batches allowing repeated iteration (e.g., list or DataLoader).

  • max_epochs (int, optional) – Max epochs to run for (default: None). If a new state should be created (first run or run again from ended engine), it’s default value is 1. If run is resuming from a state, provided max_epochs will be taken into account and should be larger than engine.state.max_epochs.

  • epoch_length (int, optional) – Number of iterations to count as one epoch. By default, it can be set as len(data). If data is an iterator and epoch_length is not set, then it will be automatically determined as the iteration on which data iterator raises StopIteration. This argument should not change if run is resuming from a state.

  • seed (int, optional) – Deprecated argument. Please, use torch.manual_seed or manual_seed().

Returns

output state.

Return type

State

Note

User can dynamically preprocess input batch at ITERATION_STARTED and store output batch in engine.state.batch. Latter is passed as usually to process_function as argument:

trainer = ...

@trainer.on(Events.ITERATION_STARTED)
def switch_batch(engine):
    engine.state.batch = preprocess_batch(engine.state.batch)
set_data(data)[source]#

Method to set data. After calling the method the next batch passed to processing_function is from newly provided data. Please, note that epoch length is not modified.

Parameters

data (Iterable) – Collection of batches allowing repeated iteration (e.g., list or DataLoader).

Example usage:

User can switch data provider during the training:

data1 = ...
data2 = ...

switch_iteration = 5000

def train_step(e, batch):
    # when iteration <= switch_iteration
    # batch is from data1
    # when iteration > switch_iteration
    # batch is from data2
    ...

trainer = Engine(train_step)

@trainer.on(Events.ITERATION_COMPLETED(once=switch_iteration))
def switch_dataloader():
    trainer.set_data(data2)

trainer.run(data1, max_epochs=100)
state_dict()[source]#

Returns a dictionary containing engine’s state: “seed”, “epoch_length”, “max_epochs” and “iteration” and other state values defined by engine.state_dict_user_keys

engine = Engine(...)
engine.state_dict_user_keys.append("alpha")
engine.state_dict_user_keys.append("beta")
...

@engine.on(Events.STARTED)
def init_user_value(_):
     engine.state.alpha = 0.1
     engine.state.beta = 1.0

@engine.on(Events.COMPLETED)
def save_engine(_):
    state_dict = engine.state_dict()
    assert "alpha" in state_dict and "beta" in state_dict
    torch.save(state_dict, "/tmp/engine.pt")
Returns

a dictionary containing engine’s state

Return type

OrderedDict

terminate()[source]#

Sends terminate signal to the engine, so that it terminates completely the run after the current iteration.

Return type

None

terminate_epoch()[source]#

Sends terminate signal to the engine, so that it terminates the current epoch after the current iteration.

Return type

None

ignite.engine.events#

class ignite.engine.events.Events(value)[source]#

Events that are fired by the Engine during execution. Built-in events:

  • STARTED : triggered when engine’s run is started

  • EPOCH_STARTED : triggered when the epoch is started

  • GET_BATCH_STARTED : triggered before next batch is fetched

  • GET_BATCH_COMPLETED : triggered after the batch is fetched

  • ITERATION_STARTED : triggered when an iteration is started

  • ITERATION_COMPLETED : triggered when the iteration is ended

  • DATALOADER_STOP_ITERATION : engine’s specific event triggered when dataloader has no more data to provide

  • EXCEPTION_RAISED : triggered when an exception is encountered

  • TERMINATE_SINGLE_EPOCH : triggered when the run is about to end the current epoch, after receiving terminate_epoch() call.

  • TERMINATE : triggered when the run is about to end completely, after receiving terminate() call.

  • EPOCH_COMPLETED : triggered when the epoch is ended

  • COMPLETED : triggered when engine’s run is completed

Since v0.3.0, Events become more flexible and allow to pass an event filter to the Engine:

engine = Engine()

# a) custom event filter
def custom_event_filter(engine, event):
    if event in [1, 2, 5, 10, 50, 100]:
        return True
    return False

@engine.on(Events.ITERATION_STARTED(event_filter=custom_event_filter))
def call_on_special_event(engine):
    # do something on 1, 2, 5, 10, 50, 100 iterations

# b) "every" event filter
@engine.on(Events.ITERATION_STARTED(every=10))
def call_every(engine):
    # do something every 10th iteration

# c) "once" event filter
@engine.on(Events.ITERATION_STARTED(once=50))
def call_once(engine):
    # do something on 50th iteration

Event filter function event_filter accepts as input engine and event and should return True/False. Argument event is the value of iteration or epoch, depending on which type of Events the function is passed.

Since v0.4.0, user can also combine events with |-operator:

events = Events.STARTED | Events.COMPLETED | Events.ITERATION_STARTED(every=3)
engine = ...

@engine.on(events)
def call_on_events(engine):
    # do something
class ignite.engine.events.State(**kwargs)[source]#

An object that is used to pass internal and user-defined state between event handlers. By default, state contains the following attributes:

state.iteration         # 1-based, the first iteration is 1
state.epoch             # 1-based, the first epoch is 1
state.seed              # seed to set at each epoch
state.dataloader        # data passed to engine
state.epoch_length      # optional length of an epoch
state.max_epochs        # number of epochs to run
state.batch             # batch passed to `process_function`
state.output            # output of `process_function` after a single iteration
state.metrics           # dictionary with defined metrics if any
state.times             # dictionary with total and per-epoch times fetched on
                        # keys: Events.EPOCH_COMPLETED.name and Events.COMPLETED.name
class ignite.engine.events.RemovableEventHandle(event_name, handler, engine)[source]#

A weakref handle to remove a registered event.

A handle that may be used to remove a registered event handler via the remove method, with-statement, or context manager protocol. Returned from add_event_handler().

Parameters

  • event_name (Union[CallableEventWithFilter, Enum, EventsList]) – Registered event name.

  • handler (Callable) – Registered event handler, stored as weakref.

  • engine – Target engine, stored as weakref.

Example usage:

engine = Engine()

def print_epoch(engine):
    print("Epoch: {}".format(engine.state.epoch))

with engine.add_event_handler(Events.EPOCH_COMPLETED, print_epoch):
    # print_epoch handler registered for a single run
    engine.run(data)

# print_epoch handler is now unregistered
remove()[source]#

Remove handler from engine.

Return type

None

ignite.engine.deterministic#

class ignite.engine.deterministic.DeterministicEngine(process_function)[source]#

Deterministic engine derived from Engine.

“Deterministic” run is done by adding additional handlers to synchronize the dataflow and overriding some methods of Engine:

for e in range(num_epochs):
    set_seed(seed_offset + e)
    if resume:
        setup_saved_rng_states()
    do_single_epoch_iterations(dataloader)

If input data provider is DataLoader, its batch sampler is replaced by ReproducibleBatchSampler.

for e in range(num_epochs):
    set_seed(seed_offset + e)
    setup_sampling(dataloader)
    if resume:
        setup_saved_rng_states()
    do_single_epoch_iterations(dataloader)

Internally, torch.backends.cudnn.deterministic = True and torch.backends.cudnn.benchmark = False are also applied.

For more details about dataflow synchronization, please see “Concepts/Dataflow synchronization”.

Note

This class can produce exactly the same dataflow when resuming the run from an epoch (or more precisely from dataflow restart) and using torch DataLoader with num_workers > 1 as data provider.

Parameters

process_function (Callable) –

state_dict()[source]#

Returns a dictionary containing engine’s state: “seed”, “epoch_length”, “max_epochs” and “iteration” and other state values defined by engine.state_dict_user_keys

engine = Engine(...)
engine.state_dict_user_keys.append("alpha")
engine.state_dict_user_keys.append("beta")
...

@engine.on(Events.STARTED)
def init_user_value(_):
     engine.state.alpha = 0.1
     engine.state.beta = 1.0

@engine.on(Events.COMPLETED)
def save_engine(_):
    state_dict = engine.state_dict()
    assert "alpha" in state_dict and "beta" in state_dict
    torch.save(state_dict, "/tmp/engine.pt")
Returns

a dictionary containing engine’s state

Return type

OrderedDict

class ignite.engine.deterministic.ReproducibleBatchSampler(batch_sampler, start_iteration=None)[source]#

Reproducible batch sampler. This class internally iterates and stores indices of the input batch sampler. This helps to start providing data batches from an iteration in a deterministic way.

Usage:

Setup dataloader with ReproducibleBatchSampler and start providing data batches from an iteration:

from ignite.engine.deterministic import update_dataloader

dataloader = update_dataloader(dataloader, ReproducibleBatchSampler(dataloader.batch_sampler))
# rewind dataloader to a specific iteration:
dataloader.batch_sampler.start_iteration = start_iteration
Parameters
ignite.engine.deterministic.keep_random_state(func)[source]#

Helper decorator to keep random state of torch, numpy and random intact while executing a function. For more details on usage, please see “Concepts/Dataflow synchronization”.

Parameters

func (callable) – function to decorate

ignite.engine.deterministic.update_dataloader(dataloader, new_batch_sampler)[source]#

Helper function to replace current batch sampler of the dataloader by a new batch sampler. Function returns new dataloader with new batch sampler.

Parameters
Returns

DataLoader

Return type

DataLoader