# Copyright (c) 2025, 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. from dataclasses import dataclass import pytest import torch from nemo.collections.audio.parts.submodules.flow import ConditionalFlowMatchingEulerSampler, OptimalTransportFlow NUM_STEPS = [1, 5, 10, 20, 100] TIMES_MIN = [0.0, 1e-8, 1e-2, 0.1, 0.25, 0.4] TIMES_MAX = [0.5, 0.7, 0.99, 1.0 - 1e-8, 1.0] @pytest.mark.parametrize("num_steps", NUM_STEPS) @pytest.mark.parametrize("estimator_target", ['conditional_vector_field', 'data']) def test_euler_sampler_nfe(num_steps, estimator_target): """ For this specific solver the number of steps should be equal to the number of function (estimator) evaluations """ class IdentityEstimator(torch.nn.Module): def forward(self, input, input_length, condition): return input, input_length @dataclass class ForwardCounterHook: counter: int = 0 def __call__(self, *args, **kwargs): self.counter += 1 estimator = IdentityEstimator() counter_hook = ForwardCounterHook() estimator.register_forward_hook(counter_hook) flow = OptimalTransportFlow() sampler = ConditionalFlowMatchingEulerSampler( estimator=estimator, num_steps=num_steps, estimator_target=estimator_target, flow=flow ) b, c, d, l = 2, 3, 4, 5 lengths = [5, 3] init_state = torch.randn(b, c, d, l) init_state_length = torch.LongTensor(lengths) sampler.forward(state=init_state, estimator_condition=None, state_length=init_state_length) assert counter_hook.counter == sampler.num_steps @pytest.mark.parametrize('time_min', TIMES_MIN) @pytest.mark.parametrize('time_max', TIMES_MAX) def test_time_generation_bounds_optimal_transport(time_min, time_max): """ This test uses a flow with certain time_min and time_max parameters to generate timepoints and checks if timepoints belong in [time_min, time_max] interval. """ rng = torch.Generator(device='cpu') rng.manual_seed(0) flow = OptimalTransportFlow(time_min=time_min, time_max=time_max) time = flow.generate_time(batch_size=1_000, rng=rng) assert torch.all(time >= time_min).item() assert torch.all(time <= time_max).item() @pytest.mark.parametrize('time_min', TIMES_MIN) @pytest.mark.parametrize('time_max', TIMES_MAX) def test_time_generation_bounds_optimal_transport_negative_examples(time_min, time_max): """ This test uses a flow with certain time_min and time_max parameters, widens them, generates timepoints and checks if timepoints belong in [time_min, time_max] interval. Since we widen the interval when initializing the flow, we expect that after taking enough samples some of them will be outside intended interval. """ rng = torch.Generator(device='cpu') rng.manual_seed(0) flow = OptimalTransportFlow(time_min=time_min - 0.1, time_max=time_max + 0.1) time = flow.generate_time(batch_size=1_000, rng=rng) assert not torch.all(time >= time_min).item() assert not torch.all(time <= time_max).item()