"""Parity of the state bridge against the unchanged OpenTau scan.""" import copy,unittest import torch try: from robottt_pi05.reference_core_adapter import ReferenceLayer except ModuleNotFoundError as error: if error.name!='einops':raise ReferenceLayer=None from robottt_pi05.optimization import learning_rate @unittest.skipIf(ReferenceLayer is None,'Install the optional reference extra') class ReferenceParityTests(unittest.TestCase): def test_sequential_bridge_equals_native_scan_and_gradients(self): torch.manual_seed(9);wrapped=ReferenceLayer(width=16).double();native=copy.deepcopy(wrapped) x=torch.randn(1,3*26,16,dtype=torch.float64) state=wrapped.initial_state(1);outputs=[] for t in range(3): y,state=wrapped.step(x[:,t*26:(t+1)*26],state,differentiable=True);outputs.append(y) sequence=torch.cat(outputs,1) y,fast=native.core(x,mini_batch_size=26) expected=native.gate(x,y) torch.testing.assert_close(sequence,expected,rtol=1e-12,atol=1e-12) for a,b in zip(state.weights,(fast.w1,fast.b1,fast.w2,fast.b2)):torch.testing.assert_close(a,b,rtol=1e-12,atol=1e-12) sequence.square().mean().backward();expected.square().mean().backward() for (an,ap),(bn,bp) in zip(wrapped.named_parameters(),native.named_parameters()): self.assertEqual(an,bn);self.assertIsNotNone(ap.grad) torch.testing.assert_close(ap.grad,bp.grad,rtol=1e-9,atol=1e-11) self.assertEqual(state.updates.tolist(),[3]) def test_explicit_schedule(self): self.assertAlmostEqual(learning_rate(0,'opentau_3000'),1e-4/101) self.assertAlmostEqual(learning_rate(3000,'opentau_3000'),1e-5) self.assertGreater(learning_rate(99,'opentau_3000'),learning_rate(0,'opentau_3000')) self.assertEqual(learning_rate(10000,'opentau_3000'),learning_rate(3000,'opentau_3000'))