def _test_distrib_integration(device, tol=1e-6):
import numpy as np
from ignite.engine import Engine
rank = idist.get_rank()
n_iters = 100
s = 10
offset = n_iters * s
y_true = torch.arange(0,
offset * idist.get_world_size(),
dtype=torch.float).to(device)
y_preds = torch.ones(offset * idist.get_world_size(),
dtype=torch.float).to(device)
def update(engine, i):
return (
y_preds[i * s + offset * rank:(i + 1) * s + offset * rank],
y_true[i * s + offset * rank:(i + 1) * s + offset * rank],
)
engine = Engine(update)
m = MeanSquaredError()
m.attach(engine, "mse")
data = list(range(n_iters))
engine.run(data=data, max_epochs=1)
assert "mse" in engine.state.metrics
res = engine.state.metrics["mse"]
true_res = np.mean(np.power((y_true - y_preds).cpu().numpy(), 2.0))
assert pytest.approx(res, rel=tol) == true_res
def test_ignite_evaluator_reporting_metrics():
try:
from ignite.metrics import MeanSquaredError
except ImportError:
pytest.skip('pytorch-ignite is not installed')
# This tests verifies that either, usuer manually reported metrics
# and ignite calculated ones are correctly reflected in the reporter
# observation
model = IgniteDummyModel()
n_data = 10
x = torch.randn((n_data, 2), requires_grad=True)
y = torch.randn((n_data, 2))
dataset = torch.utils.data.TensorDataset(x, y)
loader = torch.utils.data.DataLoader(dataset, batch_size=3)
evaluator = create_dummy_evaluator(model)
# Attach metrics to the evaluator
metric = MeanSquaredError()
metric.attach(evaluator, 'mse')
evaluator_ignite_ext = ppe.training.extensions.IgniteEvaluator(
evaluator, loader, model, progress_bar=False)
reporter = ppe.reporting.Reporter()
with reporter:
result = evaluator_ignite_ext()
# Internally reported metrics
assert result['main/x'] == 1.5
# Ignite calculated metric
assert result['val/mse'] == 0.0
def _test(metric_device):
engine = Engine(update)
m = MeanSquaredError(device=metric_device)
m.attach(engine, "mse")
data = list(range(n_iters))
engine.run(data=data, max_epochs=1)
assert "mse" in engine.state.metrics
res = engine.state.metrics["mse"]
true_res = np.mean(np.power((y_true - y_preds).cpu().numpy(), 2.0))
assert pytest.approx(res, rel=tol) == true_res
