import tempfile import torch from diffusers import PNDMScheduler from .test_schedulers import SchedulerCommonTest class PNDMSchedulerTest(SchedulerCommonTest): scheduler_classes = (PNDMScheduler,) forward_default_kwargs = (("num_inference_steps", 50),) def get_scheduler_config(self, **kwargs): config = { "num_train_timesteps": 1000, "beta_start": 0.0001, "beta_end": 0.02, "beta_schedule": "linear", } config.update(**kwargs) return config def check_over_configs(self, time_step=0, **config): kwargs = dict(self.forward_default_kwargs) num_inference_steps = kwargs.pop("num_inference_steps", None) sample = self.dummy_sample residual = 0.1 * sample dummy_past_residuals = [residual + 0.2, residual + 0.15, residual + 0.1, residual + 0.05] for scheduler_class in self.scheduler_classes: scheduler_config = self.get_scheduler_config(**config) scheduler = scheduler_class(**scheduler_config) scheduler.set_timesteps(num_inference_steps) # copy over dummy past residuals scheduler.ets = dummy_past_residuals[:] with tempfile.TemporaryDirectory() as tmpdirname: scheduler.save_config(tmpdirname) new_scheduler = scheduler_class.from_pretrained(tmpdirname) new_scheduler.set_timesteps(num_inference_steps) # copy over dummy past residuals new_scheduler.ets = dummy_past_residuals[:] output = scheduler.step_prk(residual, time_step, sample, **kwargs).prev_sample new_output = new_scheduler.step_prk(residual, time_step, sample, **kwargs).prev_sample assert torch.sum(torch.abs(output - new_output)) < 1e-5, "Scheduler outputs are not identical" output = scheduler.step_plms(residual, time_step, sample, **kwargs).prev_sample new_output = new_scheduler.step_plms(residual, time_step, sample, **kwargs).prev_sample assert torch.sum(torch.abs(output - new_output)) < 1e-5, "Scheduler outputs are not identical" def test_from_save_pretrained(self): pass def check_over_forward(self, time_step=0, **forward_kwargs): kwargs = dict(self.forward_default_kwargs) num_inference_steps = kwargs.pop("num_inference_steps", None) sample = self.dummy_sample residual = 0.1 * sample dummy_past_residuals = [residual + 0.2, residual + 0.15, residual + 0.1, residual + 0.05] for scheduler_class in self.scheduler_classes: scheduler_config = self.get_scheduler_config() scheduler = scheduler_class(**scheduler_config) scheduler.set_timesteps(num_inference_steps) # copy over dummy past residuals (must be after setting timesteps) scheduler.ets = dummy_past_residuals[:] with tempfile.TemporaryDirectory() as tmpdirname: scheduler.save_config(tmpdirname) new_scheduler = scheduler_class.from_pretrained(tmpdirname) # copy over dummy past residuals new_scheduler.set_timesteps(num_inference_steps) # copy over dummy past residual (must be after setting timesteps) new_scheduler.ets = dummy_past_residuals[:] output = scheduler.step_prk(residual, time_step, sample, **kwargs).prev_sample new_output = new_scheduler.step_prk(residual, time_step, sample, **kwargs).prev_sample assert torch.sum(torch.abs(output - new_output)) < 1e-5, "Scheduler outputs are not identical" output = scheduler.step_plms(residual, time_step, sample, **kwargs).prev_sample new_output = new_scheduler.step_plms(residual, time_step, sample, **kwargs).prev_sample assert torch.sum(torch.abs(output - new_output)) < 1e-5, "Scheduler outputs are not identical" def full_loop(self, **config): scheduler_class = self.scheduler_classes[0] scheduler_config = self.get_scheduler_config(**config) scheduler = scheduler_class(**scheduler_config) num_inference_steps = 10 model = self.dummy_model() sample = self.dummy_sample_deter scheduler.set_timesteps(num_inference_steps) for i, t in enumerate(scheduler.prk_timesteps): residual = model(sample, t) sample = scheduler.step_prk(residual, t, sample).prev_sample for i, t in enumerate(scheduler.plms_timesteps): residual = model(sample, t) sample = scheduler.step_plms(residual, t, sample).prev_sample return sample def test_step_shape(self): kwargs = dict(self.forward_default_kwargs) num_inference_steps = kwargs.pop("num_inference_steps", None) for scheduler_class in self.scheduler_classes: scheduler_config = self.get_scheduler_config() scheduler = scheduler_class(**scheduler_config) sample = self.dummy_sample residual = 0.1 * sample if num_inference_steps is not None and hasattr(scheduler, "set_timesteps"): scheduler.set_timesteps(num_inference_steps) elif num_inference_steps is not None and not hasattr(scheduler, "set_timesteps"): kwargs["num_inference_steps"] = num_inference_steps # copy over dummy past residuals (must be done after set_timesteps) dummy_past_residuals = [residual + 0.2, residual + 0.15, residual + 0.1, residual + 0.05] scheduler.ets = dummy_past_residuals[:] output_0 = scheduler.step_prk(residual, 0, sample, **kwargs).prev_sample output_1 = scheduler.step_prk(residual, 1, sample, **kwargs).prev_sample self.assertEqual(output_0.shape, sample.shape) self.assertEqual(output_0.shape, output_1.shape) output_0 = scheduler.step_plms(residual, 0, sample, **kwargs).prev_sample output_1 = scheduler.step_plms(residual, 1, sample, **kwargs).prev_sample self.assertEqual(output_0.shape, sample.shape) self.assertEqual(output_0.shape, output_1.shape) def test_timesteps(self): for timesteps in [100, 1000]: self.check_over_configs(num_train_timesteps=timesteps) def test_steps_offset(self): for steps_offset in [0, 1]: self.check_over_configs(steps_offset=steps_offset) scheduler_class = self.scheduler_classes[0] scheduler_config = self.get_scheduler_config(steps_offset=1) scheduler = scheduler_class(**scheduler_config) scheduler.set_timesteps(10) assert torch.equal( scheduler.timesteps, torch.LongTensor( [901, 851, 851, 801, 801, 751, 751, 701, 701, 651, 651, 601, 601, 501, 401, 301, 201, 101, 1] ), ) def test_betas(self): for beta_start, beta_end in zip([0.0001, 0.001], [0.002, 0.02]): self.check_over_configs(beta_start=beta_start, beta_end=beta_end) def test_schedules(self): for schedule in ["linear", "squaredcos_cap_v2"]: self.check_over_configs(beta_schedule=schedule) def test_prediction_type(self): for prediction_type in ["epsilon", "v_prediction"]: self.check_over_configs(prediction_type=prediction_type) def test_time_indices(self): for t in [1, 5, 10]: self.check_over_forward(time_step=t) def test_inference_steps(self): for t, num_inference_steps in zip([1, 5, 10], [10, 50, 100]): self.check_over_forward(num_inference_steps=num_inference_steps) def test_pow_of_3_inference_steps(self): # earlier version of set_timesteps() caused an error indexing alpha's with inference steps as power of 3 num_inference_steps = 27 for scheduler_class in self.scheduler_classes: sample = self.dummy_sample residual = 0.1 * sample scheduler_config = self.get_scheduler_config() scheduler = scheduler_class(**scheduler_config) scheduler.set_timesteps(num_inference_steps) # before power of 3 fix, would error on first step, so we only need to do two for i, t in enumerate(scheduler.prk_timesteps[:2]): sample = scheduler.step_prk(residual, t, sample).prev_sample def test_inference_plms_no_past_residuals(self): with self.assertRaises(ValueError): scheduler_class = self.scheduler_classes[0] scheduler_config = self.get_scheduler_config() scheduler = scheduler_class(**scheduler_config) scheduler.step_plms(self.dummy_sample, 1, self.dummy_sample).prev_sample def test_full_loop_no_noise(self): sample = self.full_loop() result_sum = torch.sum(torch.abs(sample)) result_mean = torch.mean(torch.abs(sample)) assert abs(result_sum.item() - 198.1318) < 1e-2 assert abs(result_mean.item() - 0.2580) < 1e-3 def test_full_loop_with_v_prediction(self): sample = self.full_loop(prediction_type="v_prediction") result_sum = torch.sum(torch.abs(sample)) result_mean = torch.mean(torch.abs(sample)) assert abs(result_sum.item() - 67.3986) < 1e-2 assert abs(result_mean.item() - 0.0878) < 1e-3 def test_full_loop_with_set_alpha_to_one(self): # We specify different beta, so that the first alpha is 0.99 sample = self.full_loop(set_alpha_to_one=True, beta_start=0.01) result_sum = torch.sum(torch.abs(sample)) result_mean = torch.mean(torch.abs(sample)) assert abs(result_sum.item() - 230.0399) < 1e-2 assert abs(result_mean.item() - 0.2995) < 1e-3 def test_full_loop_with_no_set_alpha_to_one(self): # We specify different beta, so that the first alpha is 0.99 sample = self.full_loop(set_alpha_to_one=False, beta_start=0.01) result_sum = torch.sum(torch.abs(sample)) result_mean = torch.mean(torch.abs(sample)) assert abs(result_sum.item() - 186.9482) < 1e-2 assert abs(result_mean.item() - 0.2434) < 1e-3