# Copyright (c) 2020, NVIDIA CORPORATION. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. # Copyright 2018-2020 William Falcon # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. import os import pickle import sys from functools import partial from typing import Callable, Optional import numpy as np import pytest import torch from scipy.stats import entropy from torch.distributions.utils import logits_to_probs from torch.multiprocessing import Pool, set_start_method from torchmetrics import Metric from nemo.collections.common.metrics import GlobalAverageLossMetric, Perplexity NUM_PROCESSES = 2 NUM_BATCHES = 10 BATCH_SIZE = 16 NUM_CLASSES = 5 EXTRA_DIM = 3 THRESHOLD = 0.5 def setup_ddp(rank, world_size): """ Setup ddp enviroment """ os.environ["MASTER_ADDR"] = 'localhost' os.environ['MASTER_PORT'] = '8088' if torch.distributed.is_available() and sys.platform not in ['win32', 'cygwin']: torch.distributed.init_process_group("gloo", rank=rank, world_size=world_size) def _class_test( rank: int, worldsize: int, preds: torch.Tensor, target: torch.Tensor, metric_class: Metric, sk_metric: Callable, dist_sync_on_step: bool, metric_args: dict = {}, check_dist_sync_on_step: bool = True, check_batch: bool = True, atol: float = 1e-8, ): """ Utility function doing the actual comparison between lightning class metric and reference metric. Args: rank: rank of current process worldsize: number of processes preds: torch tensor with predictions target: torch tensor with targets metric_class: lightning metric class that should be tested sk_metric: callable function that is used for comparison dist_sync_on_step: bool, if true will synchronize metric state across processes at each ``forward()`` metric_args: dict with additional arguments used for class initialization check_dist_sync_on_step: bool, if true will check if the metric is also correctly calculated per batch per device (and not just at the end) check_batch: bool, if true will check if the metric is also correctly calculated across devices for each batch (and not just at the end) """ # Instanciate lightning metric metric = metric_class(dist_sync_on_step=dist_sync_on_step, **metric_args) # verify metrics work after being loaded from pickled state pickled_metric = pickle.dumps(metric) metric = pickle.loads(pickled_metric) for i in range(rank, NUM_BATCHES, worldsize): batch_result = metric(preds[i], target[i]) if metric.dist_sync_on_step: if rank == 0: ddp_preds = torch.stack([preds[i + r] for r in range(worldsize)]) ddp_target = torch.stack([target[i + r] for r in range(worldsize)]) sk_batch_result = sk_metric(ddp_preds, ddp_target) # assert for dist_sync_on_step if check_dist_sync_on_step: assert np.allclose(batch_result.numpy(), sk_batch_result, atol=atol) else: sk_batch_result = sk_metric(preds[i], target[i]) # assert for batch if check_batch: assert np.allclose(batch_result.numpy(), sk_batch_result, atol=atol) # check on all batches on all ranks result = metric.compute() assert isinstance(result, torch.Tensor) total_preds = torch.stack([preds[i] for i in range(NUM_BATCHES)]) total_target = torch.stack([target[i] for i in range(NUM_BATCHES)]) sk_result = sk_metric(total_preds, total_target) # assert after aggregation assert np.allclose(result.numpy(), sk_result, atol=atol) def _functional_test( preds: torch.Tensor, target: torch.Tensor, metric_functional: Callable, sk_metric: Callable, metric_args: dict = {}, atol: float = 1e-8, ): """ Utility function doing the actual comparison between lightning functional metric and reference metric. Args: preds: torch tensor with predictions target: torch tensor with targets metric_functional: lightning metric functional that should be tested sk_metric: callable function that is used for comparison metric_args: dict with additional arguments used for class initialization """ metric = partial(metric_functional, **metric_args) for i in range(NUM_BATCHES): lightning_result = metric(preds[i], target[i]) sk_result = sk_metric(preds[i], target[i]) # assert its the same assert np.allclose(lightning_result.numpy(), sk_result, atol=atol) class MetricTester: """ Class used for efficiently run alot of parametrized tests in ddp mode. Makes sure that ddp is only setup once and that pool of processes are used for all tests. All tests should subclass from this and implement a new method called `test_metric_name` where the method `self.run_metric_test` is called inside. """ atol = 1e-8 def setup_class(self): """ Setup the metric class. This will spawn the pool of workers that are used for metric testing and setup_ddp """ try: set_start_method('spawn') except RuntimeError: pass self.poolSize = NUM_PROCESSES self.pool = Pool(processes=self.poolSize) self.pool.starmap(setup_ddp, [(rank, self.poolSize) for rank in range(self.poolSize)]) def teardown_class(self): """ Close pool of workers """ self.pool.close() self.pool.join() def run_functional_metric_test( self, preds: torch.Tensor, target: torch.Tensor, metric_functional: Callable, sk_metric: Callable, metric_args: dict = {}, ): """ Main method that should be used for testing functions. Call this inside testing method Args: preds: torch tensor with predictions target: torch tensor with targets metric_functional: lightning metric class that should be tested sk_metric: callable function that is used for comparison metric_args: dict with additional arguments used for class initialization """ _functional_test( preds=preds, target=target, metric_functional=metric_functional, sk_metric=sk_metric, metric_args=metric_args, atol=self.atol, ) def run_class_metric_test( self, ddp: bool, preds: torch.Tensor, target: torch.Tensor, metric_class: Metric, sk_metric: Callable, dist_sync_on_step: bool, metric_args: dict = {}, check_dist_sync_on_step: bool = True, check_batch: bool = True, ): """ Main method that should be used for testing class. Call this inside testing methods. Args: ddp: bool, if running in ddp mode or not preds: torch tensor with predictions target: torch tensor with targets metric_class: lightning metric class that should be tested sk_metric: callable function that is used for comparison dist_sync_on_step: bool, if true will synchronize metric state across processes at each ``forward()`` metric_args: dict with additional arguments used for class initialization check_dist_sync_on_step: bool, if true will check if the metric is also correctly calculated per batch per device (and not just at the end) check_batch: bool, if true will check if the metric is also correctly calculated across devices for each batch (and not just at the end) """ if ddp: if sys.platform == "win32": pytest.skip("DDP not supported on windows") self.pool.starmap( partial( _class_test, preds=preds, target=target, metric_class=metric_class, sk_metric=sk_metric, dist_sync_on_step=dist_sync_on_step, metric_args=metric_args, check_dist_sync_on_step=check_dist_sync_on_step, check_batch=check_batch, atol=self.atol, ), [(rank, self.poolSize) for rank in range(self.poolSize)], ) else: _class_test( 0, 1, preds=preds, target=target, metric_class=metric_class, sk_metric=sk_metric, dist_sync_on_step=dist_sync_on_step, metric_args=metric_args, check_dist_sync_on_step=check_dist_sync_on_step, check_batch=check_batch, atol=self.atol, ) def reference_perplexity_func(probs): ent = entropy(probs, axis=-1) ppl = np.exp(ent) return ppl.mean() def _perplexity_class_test( rank: int, worldsize: int, probs: Optional[torch.Tensor], logits: Optional[torch.Tensor], dist_sync_on_step: bool, metric_args: dict = {}, check_dist_sync_on_step: bool = True, check_batch: bool = True, atol: float = 1e-8, ): """ Utility function doing the actual comparison between lightning class metric and reference metric. Args: rank: rank of current process worldsize: number of processes probs: torch tensor with probabilities logits: torch tensor with logits. The function checks ``probs`` and ``logits are mutually exclusive for ``Perplexity`` metric. dist_sync_on_step: bool, if true will synchronize metric state across processes at each ``forward()`` metric_args: dict with additional arguments used for class initialization check_dist_sync_on_step: bool, if true will check if the metric is also correctly calculated per batch per device (and not just at the end) check_batch: bool, if true will check if the metric is also correctly calculated across devices for each batch (and not just at the end) """ # Instanciate lightning metric perplexity = Perplexity(dist_sync_on_step=dist_sync_on_step, **metric_args) if (probs is None) == (logits is None): with pytest.raises(ValueError): perplexity(probs, logits) return # verify perplexity works after being loaded from pickled state pickled_metric = pickle.dumps(perplexity) perplexity = pickle.loads(pickled_metric) for i in range(rank, NUM_BATCHES, worldsize): batch_result = perplexity(None if probs is None else probs[i], None if logits is None else logits[i]) if perplexity.dist_sync_on_step: if rank == 0: if probs is not None: ddp_probs = torch.stack([probs[i + r] for r in range(worldsize)]) else: ddp_logits = torch.stack([logits[i + r] for r in range(worldsize)]) ddp_probs = logits_to_probs(ddp_logits, is_binary=False) sk_batch_result = reference_perplexity_func(ddp_probs) # assert for dist_sync_on_step if check_dist_sync_on_step: assert np.allclose(batch_result.numpy(), sk_batch_result, atol=atol) else: if probs is None: p = logits_to_probs(logits[i], is_binary=False) else: p = probs[i] sk_batch_result = reference_perplexity_func(p) # assert for batch if check_batch: assert np.allclose(batch_result.numpy(), sk_batch_result, atol=atol) assert (probs is None) != (logits is None) # check on all batches on all ranks result = perplexity.compute() assert isinstance(result, torch.Tensor) if probs is None: probs = logits_to_probs(logits, is_binary=False) sk_result = reference_perplexity_func(probs) # assert after aggregation assert np.allclose(result.numpy(), sk_result, atol=atol) class PerplexityTester(MetricTester): def run_class_perplexity_test( self, ddp: bool, probs: Optional[torch.Tensor], logits: Optional[torch.Tensor], dist_sync_on_step: bool, metric_args: dict = {}, check_dist_sync_on_step: bool = True, check_batch: bool = True, ): """ Main method that should be used for testing class. Call this inside testing methods. Args: ddp: bool, if running in ddp mode or not probs: torch tensor with probabilities. logits: torch tensor with logits. This test checks that probs and logits are mutually exclusive for ``Perplexity`` metric. dist_sync_on_step: bool, if true will synchronize metric state across processes at each ``forward()`` metric_args: dict with additional arguments used for class initialization check_dist_sync_on_step: bool, if true will check if the metric is also correctly calculated per batch per device (and not just at the end) check_batch: bool, if true will check if the metric is also correctly calculated across devices for each batch (and not just at the end) """ if ddp: if sys.platform == "win32": pytest.skip("DDP not supported on windows") self.pool.starmap( partial( _perplexity_class_test, probs=probs, logits=logits, dist_sync_on_step=dist_sync_on_step, metric_args=metric_args, check_dist_sync_on_step=check_dist_sync_on_step, check_batch=check_batch, atol=self.atol, ), [(rank, self.poolSize) for rank in range(self.poolSize)], ) else: _perplexity_class_test( 0, 1, probs=probs, logits=logits, dist_sync_on_step=dist_sync_on_step, metric_args=metric_args, check_dist_sync_on_step=check_dist_sync_on_step, check_batch=check_batch, atol=self.atol, ) def reference_loss_func(loss_sum_or_avg: torch.Tensor, num_measurements: torch.Tensor, take_avg_loss: bool): """ Returns average loss for data from``loss_sum_or_avg``. This function sums all losses from ``loss_sum_or_avg`` and divides the sum by the sum of ``num_measurements`` elements. If ``take_avg_loss`` is ``True`` then ``loss_sum_or_avg[i]`` elements are mean values of ``num_measurements[i]`` losses. In that case before computing sum of losses each element of ``loss_sum_or_avg`` is multiplied by corresponding element of ``num_measurements``. If ``num_measurements`` sum is zero then the function returns NaN tensor. The function is used for testing ``nemo.collections.common.metrics.GlobalAverageLossMetric`` class. Args: loss_sum_or_avg: a one dimensional float ``torch.Tensor``. Sums or mean values of loss. num_measurements: a one dimensional integer ``torch.Tensor``. Number of values on which sums of means in ``loss_sum_or_avg`` are calculated. take_avg_loss: if ``True`` then ``loss_sum_or_avg`` contains mean losses else ``loss_sum_or_avg`` contains sums of losses. """ loss_sum_or_avg = loss_sum_or_avg.clone().detach() if take_avg_loss: loss_sum_or_avg *= num_measurements nm_sum = num_measurements.sum() if nm_sum.eq(0): return torch.tensor(float('nan')) return loss_sum_or_avg.sum() / nm_sum def _loss_class_test( rank: int, worldsize: int, loss_sum_or_avg: Optional[torch.Tensor], num_measurements: Optional[torch.Tensor], dist_sync_on_step: bool, take_avg_loss: bool, check_dist_sync_on_step: bool = True, check_batch: bool = True, atol: float = 1e-8, ): """ Utility function doing the actual comparison between lightning class metric and reference metric. Args: rank: rank of current process worldsize: number of processes loss_sum_or_avg: a one dimensional float torch tensor with loss sums or means. num_measurements: a one dimensional integer torch tensor with number of values on which sums or means from ``loss_sum_or_avg`` were computed. dist_sync_on_step: bool, if true will synchronize metric state across processes at each call of the method :meth:`forward()` take_avg_loss: dict with additional arguments used for class initialization check_dist_sync_on_step: bool, if true will check if the metric is also correctly calculated per batch per device (and not just at the end) check_batch: bool, if true will check if the metric is also correctly calculated across devices for each batch (and not just at the end) """ # Instantiate lightning metric loss_metric = GlobalAverageLossMetric(dist_sync_on_step=dist_sync_on_step, take_avg_loss=take_avg_loss) # verify loss works after being loaded from pickled state pickled_metric = pickle.dumps(loss_metric) loss_metric = pickle.loads(pickled_metric) for i in range(rank, NUM_BATCHES, worldsize): batch_result = loss_metric(loss_sum_or_avg[i], num_measurements[i]) if loss_metric.dist_sync_on_step: if rank == 0: ddp_loss_sum_or_avg = torch.stack([loss_sum_or_avg[i + r] for r in range(worldsize)]) ddp_num_measurements = torch.stack([num_measurements[i + r] for r in range(worldsize)]) sk_batch_result = reference_loss_func(ddp_loss_sum_or_avg, ddp_num_measurements, take_avg_loss) # assert for dist_sync_on_step if check_dist_sync_on_step: if sk_batch_result.isnan(): assert batch_result.isnan() else: assert np.allclose( batch_result.numpy(), sk_batch_result, atol=atol ), f"batch_result = {batch_result.numpy()}, sk_batch_result = {sk_batch_result}, i = {i}" else: ls = loss_sum_or_avg[i : i + 1] nm = num_measurements[i : i + 1] sk_batch_result = reference_loss_func(ls, nm, take_avg_loss) # assert for batch if check_batch: if sk_batch_result.isnan(): assert batch_result.isnan() else: assert np.allclose( batch_result.numpy(), sk_batch_result, atol=atol ), f"batch_result = {batch_result.numpy()}, sk_batch_result = {sk_batch_result}, i = {i}" # check on all batches on all ranks result = loss_metric.compute() assert isinstance(result, torch.Tensor) sk_result = reference_loss_func(loss_sum_or_avg, num_measurements, take_avg_loss) # assert after aggregation if sk_result.isnan(): assert result.isnan() else: assert np.allclose(result.numpy(), sk_result, atol=atol), f"result = {result.numpy()}, sk_result = {sk_result}" class LossTester(MetricTester): def run_class_loss_test( self, ddp: bool, loss_sum_or_avg: torch.Tensor, num_measurements: torch.Tensor, dist_sync_on_step: bool, take_avg_loss: bool, check_dist_sync_on_step: bool = True, check_batch: bool = True, ): if ddp: if sys.platform == "win32": pytest.skip("DDP not supported on windows") self.pool.starmap( partial( _loss_class_test, loss_sum_or_avg=loss_sum_or_avg, num_measurements=num_measurements, dist_sync_on_step=dist_sync_on_step, take_avg_loss=take_avg_loss, check_dist_sync_on_step=check_dist_sync_on_step, check_batch=check_batch, atol=self.atol, ), [(rank, self.poolSize) for rank in range(self.poolSize)], ) else: _loss_class_test( 0, 1, loss_sum_or_avg=loss_sum_or_avg, num_measurements=num_measurements, dist_sync_on_step=dist_sync_on_step, take_avg_loss=take_avg_loss, check_dist_sync_on_step=check_dist_sync_on_step, check_batch=check_batch, atol=self.atol, )