class TopKAccuracy(object):
r"""Top K accuracy of an output.
Counts a prediction as correct if the target value is in the top ``k``
predictions, false otherwise, and returns the number of correct
instances relative to total instances (0.0 to 100.0).
Args:
topk (int, optional): The number of top predictions to consider.
Default: ``1``
"""
def __init__(self, topk=1):
self.topk = topk
self.reset()
def __call__(self, output, target):
"""Computes the precision@k for the specified values of k
Args:
output (:obj:`torch.Tensor`): Predictions of a model
target (:obj:`torch.Tensor`): Target labels
Example:
>>> m = nn.Softmax()
>>> input = torch.randn(10, 50)
>>> preds = m(input)
>>> targets = torch.randint(0, 1, (10,50))
>>> topk = TopKAccuracy(5)
>>> precision = topk(preds, targets)
Returns:
float
"""
batch_size = target.size(0)
_, pred = output.topk(self.topk, 1, True, True)
pred = pred.t()
correct = pred.eq(target.view(1, -1).expand_as(pred))
correct_k = correct[:self.topk].view(-1).float().sum(0, keepdim=True)
return correct_k.mul_(100.0 / batch_size)
def reset(self):
"""Reset metric tracking stats"""
self.top = AverageMeter()
def update(self, prec, size):
"""Add new measurement to running stats"""
self.top.update(prec, size)
def average(self):
"""Average stats."""
return global_average(self.top.sum, self.top.count)
@property
def name(self):
"""str: Name of this metric."""
return "Prec@{}".format(self.topk)
def validation_round(
session,
validation_set_init_op,
loss_op,
metrics,
batch_size,
num_batches_per_epoch_for_validation,
tracker,
):
""" Handles one full iteration of validation on the whole validation set.
Args:
session (obj): The tensorflow session
validation_set_init_op (obj): The trainset initialisation tf operation
loss_op (obj): The tensorflow loss operation
metrics (list): List of metrics to track
batch_size (int): The batch size
num_batches_per_epoch_for_validation (int): Maximum number of batches to validate
for per epoch, default: `None` (all batches)
tracker (`obj`:mlbench_core.utils.Tracker): Tracker object to use
"""
session.run(validation_set_init_op)
tracker.validation()
loss_meter = AverageMeter()
metrics_meter = [AverageMeter() for _ in metrics]
for i_batch in range(num_batches_per_epoch_for_validation):
out = session.run({
"metrics": [m.metric_op for m in metrics],
"loss": loss_op
})
# Update tracker
loss_meter.update(out["loss"], n=batch_size)
for meter, o in zip(metrics_meter, out["metrics"]):
meter.update(o, n=batch_size)
logging.debug("{}/{} Validation loss={:10.3e} | metrics: [{}]".format(
tracker.current_epoch,
i_batch,
loss_meter.avg,
",".join([format(m.avg, "10.3e") for m in metrics_meter]),
))
tracker.record_loss(loss_meter.avg, log_to_api=True)
if tracker.rank == 0:
tracker.record_stat("global_loss", loss_meter.avg, log_to_api=True)
for i, metric, meter in zip(range(len(metrics)), metrics, metrics_meter):
tracker.record_metric(metric, meter.avg, log_to_api=True)
if tracker.rank == 0:
tracker.record_stat("global_{}".format(metric.name),
meter.avg,
log_to_api=True)
def train_one_epoch(self, tracker):
"""Train a model for an epoch and use tracker to log stats."""
logging.info("Initialize training dataset.")
self.sess.run(self.train_set_init_op)
loss_meter = AverageMeter()
metrics_meter = [AverageMeter() for _ in self.metrics]
for i_batch in range(self.num_batches_per_epoch_for_train):
# for i_batch in range(1):
tracker.batch_stats = [("start", time.time())]
out = self.sess.run({
"metrics": [m['value'] for m in self.metrics],
"loss": self.loss,
"train_op": self.train_op,
})
tracker.batch_stats.append(('end', time.time()))
# Update tracker
loss_meter.update(out["loss"], n=self.batch_size)
for meter, o in zip(metrics_meter, out['metrics']):
meter.update(o, n=self.batch_size)
# Print logging information.
logging.debug(
"E{}:B{}/{} loss={:10.3e} | metrics: [{}] | best epoch {} ({:10.3e})"
.format(
tracker.current_epoch, i_batch,
self.num_batches_per_epoch_for_train, loss_meter.avg,
",".join([format(m.avg, "10.3e") for m in metrics_meter]),
tracker.best_epoch, tracker.best_epoch_value))
# Record training loss and metrics.
tracker.cumu_time_train.append(tracker.batch_stats[-1][1] -
tracker.batch_stats[0][1])
LogMetrics.log(self.run_id,
self.rank,
tracker.current_epoch,
'train_loss',
loss_meter.avg,
tracker=tracker,
time=sum(tracker.cumu_time_train))
for metric, meter in zip(self.metrics, metrics_meter):
LogMetrics.log(self.run_id,
self.rank,
tracker.current_epoch,
'train_' + metric['name'],
meter.avg,
tracker=tracker,
time=sum(tracker.cumu_time_train))
logging.info("Finish training for one epoch.")
def _validate(
dataloader,
model,
loss_function,
metrics,
dtype,
transform_target_type=None,
use_cuda=False,
max_batch_per_epoch=None,
):
"""Evaluate the model on the test dataset.
Args:
dataloader (:obj:`torch.utils.data.DataLoader`): The validation set
model (`obj`:torch.nn.Module): The model to train
loss_function (`obj`:torch.nn.Module): The loss function
metrics (list): List of metrics to track
dtype (str): The datatype to use, one of `fp32`or `fp64`
transform_target_type (str): Datatype to convert data to, default: `None`
use_cuda (bool): Whether to use GPU for training, default: `False`
max_batch_per_epoch (int): Maximum number of batches tot rain for per epoch,
default: `None` (all batches)
"""
# Initialize the accumulators for loss and metrics
losses = AverageMeter()
for metric in metrics:
metric.reset()
# Each worker computer their own losses and metrics
with torch.no_grad():
data_iter = iterate_dataloader(
dataloader, dtype, max_batch_per_epoch, use_cuda, transform_target_type
)
for data, target in data_iter:
# Inference
output = model(data)
# Compute loss
loss = loss_function(output, target)
# Update loss
losses.update(loss.item(), data.size(0))
# Update metrics
for metric in metrics:
metric_value = metric(loss, output, target)
metric.update(metric_value, data.size(0))
# Aggregate metrics and loss for all workers
metrics_averages = {metric: metric.average().item() for metric in metrics}
loss_average = global_average(losses.sum, losses.count).item()
return metrics_averages, loss_average
def valid_one_epoch(self, tracker):
self.sess.run(self.validation_set_init_op)
loss_meter = AverageMeter()
metrics_meter = [AverageMeter() for _ in self.metrics]
for i_batch in range(self.num_batches_per_epoch_for_validation):
out = self.sess.run({
"metrics": [m['value'] for m in self.metrics],
"loss": self.loss
})
# Update tracker
loss_meter.update(out["loss"], n=self.batch_size)
for meter, o in zip(metrics_meter, out['metrics']):
meter.update(o, n=self.batch_size)
logging.debug(
"{}/{} Validation loss={:10.3e} | metrics: [{}]".format(
tracker.current_epoch, i_batch, loss_meter.avg,
",".join([format(m.avg, "10.3e") for m in metrics_meter])))
LogMetrics.log(self.run_id,
self.rank,
tracker.current_epoch,
'val_loss',
loss_meter.avg,
tracker=tracker,
time=sum(tracker.cumu_time_train))
for i, metric, meter in zip(range(len(self.metrics)), self.metrics,
metrics_meter):
metric_name = 'val_' + metric['name']
# Here we implicitly assume the larger metrics value means better results
if ((i == 0) and
(len([r for r in tracker.records if r['name'] == metric_name])
== 0 or meter.avg > max([
float(r.value)
for r in tracker.records if r['name'] == metric_name
]))):
tracker.best_epoch = tracker.current_epoch
tracker.best_epoch_value = meter.avg
LogMetrics.log(self.run_id,
self.rank,
tracker.current_epoch,
metric_name,
meter.avg,
tracker=tracker,
time=sum(tracker.cumu_time_train))
def validation_round(
val_loader,
metrics,
model,
loss_func,
iter_size,
translator,
tracker=None,
use_cuda=False,
):
# Set tracker and model in eval mode
model.eval()
if tracker:
tracker.validation()
tracker.validation_start()
losses = AverageMeter()
# Reset metrics
for metric in metrics:
metric.reset()
with torch.no_grad():
for (data, target) in val_loader:
data, target = prepare_batch(data, target, use_cuda=use_cuda)
output = compute_model_output(model, data, target)
# Compute loss
loss, loss_per_token = compute_loss(data, target, output,
loss_func, iter_size)
# Update loss
losses.update(loss_per_token, 1)
# Update metrics
translated, targets = translator.translate(data, target)
for metric in metrics:
metric_value = metric(translated, targets)
size = data[0].shape[1]
metric.update(metric_value, size)
metrics_averages = {metric: metric.average().item() for metric in metrics}
loss_average = global_average(losses.sum, losses.count).item()
if tracker:
tracker.validation_end()
return metrics_averages, loss_average
class MLBenchMetric(object):
def __init__(self):
self.average_meter = AverageMeter()
@abstractmethod
def __call__(self, loss, output, target):
pass
def reset(self):
self.average_meter = AverageMeter()
def update(self, perc, size):
self.average_meter.update(perc, size)
def average(self):
return global_average(self.average_meter.sum, self.average_meter.count)
def validation_round(loader, metrics, criterion, translator, tracker, use_cuda=False):
"""Performs one round of validation for the Transformer model
Args:
loader (:obj:`torch.utils.data.DataLoader`): Data loader
metrics (list): List of metrics for evaluation
criterion (:obj:`torch.nn.Module): Loss function
translator (:obj:`mlbench_core.models.pytorch.transformer.SequenceGenerator`): Translator module
tracker (:obj:`mlbench_core.utils.Tracker`): Current Tracker
use_cuda (bool): Use GPU acceleration. Default: `False`.
Returns:
(dict of :obj:`mlbench_core.evaluation.pytorch.MLBenchMetric`: float, float):
The metrics averages over all workers, and the loss average.
"""
model = translator.model
model.eval()
tracker.validation()
tracker.validation_start()
losses = AverageMeter()
for metric in metrics:
metric.reset()
with torch.no_grad():
for batch in loader:
batch = prepare_batch(batch, use_cuda=use_cuda)
output = model(**batch["net_input"])
loss, sample_size = compute_loss(batch, output, criterion)
losses.update(loss.item() / sample_size, 1)
translated, targets = translator.translate_batch(batch)
for metric in metrics:
metric_value = metric(translated, targets)
size = batch["target"].size(0) # Number of translated sentences
metric.update(metric_value, size)
metric_averages = {metric: metric.average().item() for metric in metrics}
loss_average = global_average(losses.sum, losses.count)
tracker.validation_end()
return metric_averages, loss_average
def validate(self, dataloader):
r"""Validate the quality of the model in terms of loss and metrics.
Args:
dataloader (:obj:`torch.utils.data.DataLoader`): The validation set
"""
# Turn on evaluation mode for the model
self.model.eval()
# Initialize the accumulators for loss and metrics
losses = AverageMeter()
for metric in self.metrics:
metric.reset()
# Each worker computer their own losses and metrics
with torch.no_grad():
data_iter = iterate_dataloader(dataloader, self.dtype,
self.max_batch_per_epoch,
self.use_cuda,
self.transform_target_type)
for data, target in data_iter:
# Inference
output = self.model(data)
# Compute loss
loss = self.loss_function(output, target)
# Update loss
losses.update(loss.item(), data.size(0))
# Update metrics
for metric in self.metrics:
metric_value = metric(output, target)
metric.update(metric_value, data.size(0))
# Aggregate metrics and loss for all workers
metrics_averages = {
metric.name: metric.average().item()
for metric in self.metrics
}
loss_average = global_average(losses.sum, losses.count).item()
return metrics_averages, loss_average
def validation_round(loader,
metrics,
criterion,
translator,
tracker=None,
use_cuda=False):
model = translator.model
model.eval()
if tracker:
tracker.validation()
tracker.validation_start()
losses = AverageMeter()
for metric in metrics:
metric.reset()
with torch.no_grad():
for batch in loader:
batch = prepare_batch(batch, use_cuda=use_cuda)
output = model(**batch["net_input"])
loss, sample_size = compute_loss(batch, output, criterion)
losses.update(loss.item() / sample_size, 1)
translated, targets = translator.translate_batch(batch)
for metric in metrics:
metric_value = metric(loss.item(), translated, targets)
size = batch["target"].size(
0) # Number of translated sentences
metric.update(metric_value, size)
metric_averages = {metric: metric.average().item() for metric in metrics}
loss_average = global_average(losses.sum, losses.count)
if tracker:
tracker.validation_end()
return metric_averages, loss_average
def train_round(
session,
train_set_init_op,
train_op,
loss_op,
metrics,
batch_size,
num_batches_per_epoch_for_train,
tracker,
lr_scheduler_level=None,
lr_tensor=None,
):
""" Performs num_batches_per_epoch_for_train batches of training (or full trainset if
not specified)
Args:
session (obj): The tensorflow session
train_set_init_op (obj): The trainset initialisation tf operation
train_op (obj): The tensorflow training operation
loss_op (obj): The tensorflow loss operation
metrics (list): List of metrics to track
batch_size (int): The batch size
num_batches_per_epoch_for_train (int): Maximum number of batches tot rain for per epoch,
default: `None` (all batches)
tracker (`obj`:mlbench_core.utils.Tracker): Tracker object to use
lr_scheduler_level (str): Learning Rate scheduler mode, one of `batch` or `epoch`
lr_tensor (obj): The learningrate schedule tensorflow operation
"""
logging.info("Initialize training dataset.")
session.run(train_set_init_op)
tracker.train()
loss_meter = AverageMeter()
metrics_meter = [AverageMeter() for _ in metrics]
if lr_scheduler_level == "epoch" and lr_tensor is not None:
lr = session.run(lr_tensor)
logging.debug("Epoch {} Learning Rate : {:10.3e}".format(
tracker.current_epoch, lr))
for i_batch in range(num_batches_per_epoch_for_train):
# for i_batch in range(1):
tracker.batch_start()
if lr_scheduler_level == "batch" and lr_tensor is not None:
lr = session.run(lr_tensor)
logging.debug("Epoch {} Learning Rate : {:10.3e}".format(
tracker.current_epoch, lr))
out = session.run({
"metrics": [m.metric_op for m in metrics],
"loss": loss_op,
"train_op": train_op,
})
tracker.batch_end()
# Update tracker
loss_meter.update(out["loss"], n=batch_size)
tracker.record_loss(loss_meter.avg, log_to_api=True)
for metric, meter, o in zip(metrics, metrics_meter, out["metrics"]):
meter.update(o, n=batch_size)
tracker.record_metric(metric, meter.avg, log_to_api=True)
# Print logging information.
progress = i_batch / num_batches_per_epoch_for_train
progress += tracker.current_epoch
status = "Epoch {:5.2f} Batch {:4}: ".format(progress, i_batch)
logging.info(status + str(tracker))
# Record training loss and metrics.
tracker.record_loss(loss_meter.avg, log_to_api=True)
for metric, meter in zip(metrics, metrics_meter):
tracker.record_metric(metric, meter.avg, log_to_api=True)
logging.info("Finish training for one epoch.")
def validation_round(
dataloader,
model,
loss_function,
metrics,
dtype,
tracker=None,
transform_target_type=False,
use_cuda=False,
max_batches=None,
):
"""Evaluate the model on the test dataset.
Args:
dataloader (`obj`:torch.utils.data.DataLoader): The validation set
model (`obj`:torch.nn.Module): The model to train
loss_function (`obj`:torch.nn.Module): The loss function
metrics (list): List of metrics to track
dtype (str): The datatype to use, one of `fp32`or `fp64`
tracker (`obj`:mlbench_core.utils.Tracker | None): Tracker object to use.
transform_target_type (bool): Convert target to `dtype`. Default `False`
use_cuda (bool): Whether to use GPU for training, default: `False`
max_batches (int | None): Maximum number of batches to validate on
Returns:
(dict, float): Dictionary of average of each metric, and average validation loss
"""
model.eval()
if tracker:
tracker.validation()
tracker.validation_start()
# Initialize the accumulators for loss and metrics
losses = AverageMeter()
for metric in metrics:
metric.reset()
# Each worker computer their own losses and metrics
with torch.no_grad():
data_iter = iterate_dataloader(dataloader, dtype, max_batches,
use_cuda, transform_target_type)
for data, target in data_iter:
output = model(data)
# Compute loss
loss = loss_function(output, target)
# Update loss
losses.update(loss.item(), data.size(0))
# Update metrics
for metric in metrics:
metric_value = metric(output, target)
metric.update(metric_value, data.size(0))
# Aggregate metrics and loss for all workers
metrics_averages = {metric: metric.average().item() for metric in metrics}
loss_average = global_average(losses.sum, losses.count).item()
if tracker:
tracker.validation_end()
return metrics_averages, loss_average
def validation_round(
val_set, model, batch_size, metrics, loss_function, tracker, use_cuda=False
):
"""Performs a validation round
Args:
val_set (:obj:): Validation set
model (:obj:`torch.nn.Module`): Model to evaluate
batch_size (int): Validation batch size
metrics (list): List of metrics to compute
loss_function (:obj:`torch.nn.Module`): Loss function
tracker (:obj:`mlbench_core.utils.Tracker`): Tracker object
use_cuda (bool): Use GPU acceleration
Returns:
(dict, float): Metric averages and total loss average
"""
# finish one epoch training and to decide if we want to val our model.
tracker.validation()
tracker.validation_start()
# each worker finish one epoch training.
model.eval()
losses = AverageMeter()
for metric in metrics:
metric.reset()
# Each worker computer their own losses and metrics
with torch.no_grad():
hidden = model.init_hidden(batch_size)
num_batches = val_set.num_batches()
for batch_idx in range(num_batches):
data, target = val_set.get_batch(batch_idx, cuda=use_cuda)
batch_seq_len = data.size(0)
# Inference
output, hidden = model(data, hidden)
# Compute loss
loss = loss_function(output, target)
# Update loss
losses.update(loss.item(), batch_seq_len)
hidden = repackage_hidden(hidden)
# Update metrics
for metric in metrics:
metric_value = metric(output, target)
metric.update(metric_value, 1)
# Aggregate metrics and loss for all workers
loss_average = global_average(losses.sum, losses.count)
metrics_averages = {
metric: torch.exp(loss_average).item()
if metric.name == "Perplexity"
else metric.average().item()
for metric in metrics
}
logger.info(
"Got loss {}, avg metric={}".format(
loss_average,
[m.average().item() for m in metrics if m.name == "Perplexity"][0],
)
)
tracker.validation_end()
return metrics_averages, loss_average.item()