ignite.engine#

Main module of the library containing:

Engine - abstraction that loops provided data, executes a processing function and returns a result
Events - events triggered by the Engine during execution
State - object to pass internal and user-defined data between event handlers

and helper methods:

create_supervised_trainer() - creates single model/optimizer/criterion supervised trainer
create_supervised_evaluator() - creates single model supervised evaluation engine

More details about those structures can be found in Concepts.

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

event_name (Any) –
handler (Callable) –

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 trigger 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 (iterable) – Defines the name of the event being supported. New events can be a str or an object derived from EventEnum. See example below.
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, Events, EventEnum

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

def process_function(e, batch):
    # ...
    trainer.fire_event("bwd_event")
    loss.backward()
    # ...
    trainer.fire_event("opt_event")
    optimizer.step()

trainer = Engine(process_function)
trainer.register_events(*CustomEvents)
trainer.register_events("bwd_event", "opt_event")

@trainer.on(Events.EPOCH_COMPLETED)
def trigger_custom_event():
    if required(...):
        trainer.fire_event(CustomEvents.FOO_EVENT)
    else:
        trainer.fire_event(CustomEvents.BAR_EVENT)

@trainer.on(CustomEvents.FOO_EVENT)
def do_foo_op():
    # ...

@trainer.on(CustomEvents.BAR_EVENT)
def do_bar_op():
    # ...

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.create_supervised_trainer(model, optimizer, loss_fn, device=None, non_blocking=False, prepare_batch=<function _prepare_batch>, output_transform=<function <lambda>>, deterministic=False)[source]#

Factory function for creating a trainer for supervised models.

Parameters

model (torch.nn.Module) – the model to train.
optimizer (torch.optim.Optimizer) – the optimizer to use.
loss_fn (torch.nn loss function) – the loss function to use.
device (str, optional) – device type specification (default: None). Applies to batches after starting the engine. Model will not be moved. Device can be CPU, GPU or TPU.
non_blocking (bool, optional) – if True and this copy is between CPU and GPU, the copy may occur asynchronously with respect to the host. For other cases, this argument has no effect.
prepare_batch (callable, optional) – function that receives batch, device, non_blocking and outputs tuple of tensors (batch_x, batch_y).
output_transform (callable, optional) – function that receives ‘x’, ‘y’, ‘y_pred’, ‘loss’ and returns value to be assigned to engine’s state.output after each iteration. Default is returning loss.item().
deterministic (bool, optional) – if True, returns deterministic engine of type DeterministicEngine, otherwise Engine (default: False).

Return type

Engine

Note

engine.state.output for this engine is defined by output_transform parameter and is the loss of the processed batch by default.

Warning

The internal use of device has changed. device will now only be used to move the input data to the correct device. The model should be moved by the user before creating an optimizer. For more information see:

Returns

a trainer engine with supervised update function.

Return type

Engine

Parameters

model (Module) –
optimizer (Optimizer) –
loss_fn (Union[Callable, Module]) –
device (Optional[Union[device, str]]) –
non_blocking (bool) –
prepare_batch (Callable) –
output_transform (Callable) –
deterministic (bool) –

ignite.engine.create_supervised_evaluator(model, metrics=None, device=None, non_blocking=False, prepare_batch=<function _prepare_batch>, output_transform=<function <lambda>>)[source]#

Factory function for creating an evaluator for supervised models.

Parameters

model (torch.nn.Module) – the model to train.
metrics (dict of str - Metric) – a map of metric names to Metrics.
device (str, optional) – device type specification (default: None). Applies to batches after starting the engine. Model will not be moved.
non_blocking (bool, optional) – if True and this copy is between CPU and GPU, the copy may occur asynchronously with respect to the host. For other cases, this argument has no effect.
prepare_batch (callable, optional) – function that receives batch, device, non_blocking and outputs tuple of tensors (batch_x, batch_y).
output_transform (callable, optional) – function that receives ‘x’, ‘y’, ‘y_pred’ and returns value to be assigned to engine’s state.output after each iteration. Default is returning (y_pred, y,) which fits output expected by metrics. If you change it you should use output_transform in metrics.

Return type

Engine

Note

engine.state.output for this engine is defind by output_transform parameter and is a tuple of (batch_pred, batch_y) by default.

Warning

The internal use of device has changed. device will now only be used to move the input data to the correct device. The model should be moved by the user before creating an optimizer.

For more information see:

Returns

an evaluator engine with supervised inference function.

Return type

Engine

Parameters

model (Module) –
metrics (Optional[Dict[str, Metric]]) –
device (Optional[Union[device, str]]) –
non_blocking (bool) –
prepare_batch (Callable) –
output_transform (Callable) –

Resuming the training#

It is possible to resume the training from a checkpoint and approximately reproduce original run’s behaviour. Using Ignite, this can be easily done using Checkpoint handler. Engine provides two methods to serialize and deserialize its internal state state_dict() and load_state_dict(). In addition to serializing model, optimizer, lr scheduler etc user can store the trainer and then resume the training. For example:

from ignite.engine import Engine, Events
from ignite.handlers import Checkpoint, DiskSaver

trainer = ...
model = ...
optimizer = ...
lr_scheduler = ...
data_loader = ...

to_save = {'trainer': trainer, 'model': model, 'optimizer': optimizer, 'lr_scheduler': lr_scheduler}
handler = Checkpoint(to_save, DiskSaver('/tmp/training', create_dir=True))
trainer.add_event_handler(Events.EPOCH_COMPLETED, handler)
trainer.run(data_loader, max_epochs=100)

ls /tmp/training
> "checkpoint_50000.pt"

We can then restore the training from the last checkpoint.

from ignite.handlers import Checkpoint

trainer = ...
model = ...
optimizer = ...
lr_scheduler = ...
data_loader = ...

to_load = {'trainer': trainer, 'model': model, 'optimizer': optimizer, 'lr_scheduler': lr_scheduler}
checkpoint = torch.load(checkpoint_file)
Checkpoint.load_objects(to_load=to_load, checkpoint=checkpoint)

trainer.run(train_loader, max_epochs=100)

It is also possible to store checkpoints every N iterations and continue the training from one of these checkpoints, i.e from iteration.

Complete examples that resumes the training from a checkpoint can be found here:

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

Since v0.4.0, custom events defined by user should inherit from EventEnum :

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

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#

Deterministic training#

In general, it is rather difficult task to achieve deterministic and reproducible trainings as it relies on multiple aspects, e.g. data version, code version, software environment, hardware etc. According to PyTorch documentation: there are some steps to take in order to make computations deterministic on your specific problem on one specific platform and PyTorch release:

setup random state seed
set cudnn to deterministic if applicable

By default, these two options can be enough to run and rerun experiments in a deterministic way. Ignite’s engine does not impact this behaviour.

In this module we provide helper methods and classes to make additional “Dataflow synchronization” to ensure that model sees the same data for a given epoch:

DeterministicEngine
ReproducibleBatchSampler

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 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

batch_sampler (torch.utils.data.sampler.BatchSampler) – batch sampler same as used with torch.utils.data.DataLoader
start_iteration (int, optional) – optional start iteration

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 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

dataloader (torch.utils.data.DataLoader) – input dataloader
new_batch_sampler (torch.utils.data.sampler.BatchSampler) – new batch sampler to use

Returns

DataLoader

Return type

DataLoader

Dataflow synchronization#

Ignite provides an option to control the dataflow by synchronizing random state on epochs. In this way, for a given iteration/epoch the dataflow can be the same for a given seed. More precisely it is roughly looks like:

for e in range(num_epochs):
    set_seed(seed + e)
    do_single_epoch_iterations(dataloader)

In addition, if data provider is torch.utils.data.DataLoader, batch data indices can be made completely deterministic. Here is a trivial example of usage:

import torch
from torch.utils.data import DataLoader
from ignite.engine import DeterministicEngine, Events
from ignite.utils import manual_seed


def random_train_data_loader(size):
    data = torch.arange(0, size)
    return DataLoader(data, batch_size=4, shuffle=True)


def print_train_data(engine, batch):
    i = engine.state.iteration
    e = engine.state.epoch
    print("train", e, i, batch.tolist())

trainer = DeterministicEngine(print_train_data)

print("Original Run")
manual_seed(56)
trainer.run(random_train_data_loader(40), max_epochs=2, epoch_length=5)

print("Resumed Run")
# Resume from 2nd epoch
trainer.load_state_dict({"epoch": 1, "epoch_length": 5, "max_epochs": 2, "rng_states": None})
manual_seed(56)
trainer.run(random_train_data_loader(40))

Original Run
train 1 1 [31, 13, 3, 4]
train 1 2 [23, 18, 6, 16]
train 1 3 [10, 8, 33, 36]
train 1 4 [1, 37, 19, 9]
train 1 5 [20, 30, 14, 26]
train 2 6 [29, 35, 38, 34]
train 2 7 [7, 22, 12, 17]
train 2 8 [25, 21, 24, 15]
train 2 9 [39, 5, 2, 28]
train 2 10 [27, 11, 32, 0]
Resumed Run
train 2 6 [29, 35, 38, 34]
train 2 7 [7, 22, 12, 17]
train 2 8 [25, 21, 24, 15]
train 2 9 [39, 5, 2, 28]
train 2 10 [27, 11, 32, 0]

We can see that the data samples are exactly the same between original and resumed runs.

Complete examples that simulates a crash on a defined iteration and resumes the training from a checkpoint can be found here:

Note

In case when input data is torch.utils.data.DataLoader, previous batches are skipped and the first provided batch corresponds to the batch after the checkpoint iteration. Internally, while resuming, previous datapoint indices are just skipped without fetching the data.

Warning

However, while resuming from iteration, random data augmentations are not synchronized in the middle of the epoch and thus batches remaining until the end of the epoch can be different of those from the initial run.

Warning

However, please, keep in mind that there can be an issue with dataflow synchronization on every epoch if user’s handler synchronizes the random state, for example, by calling periodically torch.manual_seed(seed) during the run. This can have an impact on the dataflow:

def random_train_data_generator():
    while True:
        yield torch.randint(0, 100, size=(1, ))

trainer = DeterministicEngine(print_train_data)

@trainer.on(Events.ITERATION_COMPLETED(every=3))
def user_handler():
    # handler synchronizes the random state
    torch.manual_seed(12)
    a = torch.rand(1)

trainer.run(random_train_data_generator(), max_epochs=3, epoch_length=5);

train 1 1 [32]
train 1 2 [29]
train 1 3 [40]
train 1 4 [3]  <---
train 1 5 [22]
train 2 6 [77]
train 2 7 [3]  <---
train 2 8 [22]
train 2 9 [77]
train 2 10 [3] <---
train 3 11 [22]
train 3 12 [77]
train 3 13 [3] <---
train 3 14 [22]
train 3 15 [77]

Initially, the function random_train_data_generator() generates randomly data batches using the random state set up by trainer. This is intended behaviour until user_handler() is called. After user_handler() execution, random state is altered and thus random_train_data_generator() will produce random batches based on altered random state.

We provide helper decorator keep_random_state() to save and restore random states for torch, numpy and random. Therefore, we can deal with described issue using this decorator:

from ignite.engine.deterministic import keep_random_state

@trainer.on(Events.ITERATION_COMPLETED(every=3))
@keep_random_state
def user_handler():
    # handler synchronizes the random state
    torch.manual_seed(12)
    a = torch.rand(1)