# Copyright (c) 2020 Graphcore Ltd. All rights reserved. import numpy as np import popart import torch def test_log1p_0(op_tester): input_data = np.array([0.01, 0.1, 0.5, 1, 2, 12345], dtype=np.float32) def init_builder(builder): i1 = builder.addInputTensor(input_data) o = builder.aiGraphcore.log1p([i1]) builder.addOutputTensor(o) return [o] def reference(_): # ref_data is an unused argument torch_test_data = torch.tensor(input_data, requires_grad=False) xp1 = torch.add(torch_test_data, 1.0) out = torch.log(xp1) return [out] op_tester.run(init_builder, reference, "infer") def test_log1p_1(op_tester): input_data = np.linspace(0, 1e9, 100, dtype=np.float32) def init_builder(builder): i1 = builder.addInputTensor(input_data) o = builder.aiGraphcore.log1p([i1]) builder.addOutputTensor(o) return [o] def reference(_): # ref_data is an unused argument torch_test_data = torch.tensor(input_data, requires_grad=False) out = torch.log1p(torch_test_data) return [out] op_tester.run(init_builder, reference, "infer") def test_log1p_nan(op_tester): input_data = np.array([-10000.1, -123.1, -2.0, -1.55, -1.0, 10.0], dtype=np.float32) def init_builder(builder): i1 = builder.addInputTensor(input_data) o = builder.aiGraphcore.log1p([i1]) builder.addOutputTensor(o) return [o] def reference(_): # ref_data is an unused argument torch_test_data = torch.tensor(input_data, requires_grad=False) out = torch.log1p(torch_test_data) return [out] op_tester.equal_nan = True op_tester.run(init_builder, reference, "infer") # Similar to gelu_test.py def test_log1p_inplace_0(op_tester): input_data = np.array( [-0.99, -0.5, 0.01, 0.1, 0.333, 0.5, 0.789, 1, 2, 8.8, 1212.1, 12345.6], dtype=np.float32, ) def init_builder(builder): i1 = builder.addInputTensor(input_data) o = builder.aiGraphcore.log1p([i1]) builder.addOutputTensor(o) return [o] def reference(_): # ref_data is an unused argument torch_test_data = torch.tensor(input_data, requires_grad=False) out = torch.log1p(torch_test_data) return [out] op_tester.setPatterns(["InPlace"], enableRuntimeAsserts=False) op_tester.run(init_builder, reference, "infer") # Similar to exp_test.py def test_log1p_inplace_1(op_tester): """ Test of both Log1pInplace (priority > 0) and Log1p (priority <= 0) """ d2 = np.array([1.0, 1.0, 1.0, 1.0], dtype=np.float32) d3 = np.array([1.0, 1.0, 1.0, 1.0], dtype=np.float32) for inplace_priority in [-100.0, +100.0]: d1 = np.random.rand(4).astype(np.float32) def init_builder(builder): i1 = builder.addInputTensor(d1) i2 = builder.addInputTensor(d2) i3 = builder.addInputTensor(d3) a1 = builder.aiOnnx.sub([i1, i2]) o0 = builder.aiGraphcore.log1p([a1]) builder.setInplacePreferences(o0, {"Log1pInplace": inplace_priority}) o1 = builder.aiOnnx.exp([o0]) a2 = builder.aiOnnx.sub([o1, i3]) o2 = builder.aiGraphcore.log1p([a2]) builder.setInplacePreferences(o2, {"Log1pInplace": inplace_priority}) o3 = builder.aiOnnx.exp([o2]) builder.addOutputTensor(o3) return [o3] def reference(_): # ref_data is an unused argument """ exp(logp1(sub(exp(logp1(sub))))(x) = x """ a = torch.tensor(d1, requires_grad=True) return [a * 1.0] op_tester.run(init_builder, reference, "infer") def test_log1p_inplace_2(op_tester): """ Test of both 1) 1 Log1pInplace and 2 Log1p (priority > 0) and 2) 3 Log1p (priority <= 0) """ for inplace_priority in [-100.0, +100.0]: d1 = np.random.rand(4).astype(np.float32) def init_builder(builder): i1 = builder.addInputTensor(d1) # log1p 1 o1 = builder.aiGraphcore.log1p([i1]) builder.setInplacePreferences(o1, {"Log1pInplace": inplace_priority}) # log1p 2 o2 = builder.aiGraphcore.log1p([i1]) builder.setInplacePreferences(o2, {"Log1pInplace": inplace_priority}) # log1p 3 o3 = builder.aiGraphcore.log1p([i1]) builder.setInplacePreferences(o3, {"Log1pInplace": inplace_priority}) o4 = builder.aiOnnx.sum([o1, o2, o3]) return [o4] def reference(_): # ref_data is an unused argument """ 3*log1p(in) """ a = torch.tensor(3 * np.log1p(d1), requires_grad=True) return [a] op_tester.run(init_builder, reference, "infer") def test_log1p_grad_0(op_tester): d1 = np.array( [-0.99, -0.5, 0.01, 0.1, 0.333, 0.5, 0.789, 1, 2, 8.8, 1212.1, 12345.6], dtype=np.float32, ) def init_builder(builder): i1 = builder.addInputTensor(d1) o = builder.aiGraphcore.log1p([i1]) builder.addOutputTensor(o) return [ o, popart.reservedGradientPrefix() + i1, popart.reservedGradientPrefix() + o, ] def reference(ref_data): a = torch.tensor(d1, requires_grad=True) b = torch.log1p(a) d__o = ref_data.getOutputTensorGrad(0) b.backward(torch.tensor(d__o)) return [b, a.grad, None] op_tester.setPatterns(["PreUniRepl", "Log1pGradOp"], enableRuntimeAsserts=False) op_tester.run(init_builder, reference, "train") def test_log1p_grad_1(op_tester): d1 = np.linspace(0, 1e9, 100, dtype=np.float32) def init_builder(builder): i1 = builder.addInputTensor(d1) o = builder.aiGraphcore.log1p([i1]) builder.addOutputTensor(o) return [ o, popart.reservedGradientPrefix() + i1, popart.reservedGradientPrefix() + o, ] def reference(ref_data): a = torch.tensor(d1, requires_grad=True) b = torch.log1p(a) d__o = ref_data.getOutputTensorGrad(0) b.backward(torch.tensor(d__o)) return [b, a.grad, None] op_tester.setPatterns(["PreUniRepl", "Log1pGradOp"], enableRuntimeAsserts=False) op_tester.run(init_builder, reference, "train") def test_log1p_grad_nan(op_tester): d1 = np.array([-10000.1, -123.1, -2.0, -1.55, -1.0, 10.0], dtype=np.float32) def init_builder(builder): i1 = builder.addInputTensor(d1) o = builder.aiGraphcore.log1p([i1]) builder.addOutputTensor(o) return [ o, popart.reservedGradientPrefix() + i1, popart.reservedGradientPrefix() + o, ] def reference(ref_data): a = torch.tensor(d1, requires_grad=True) b = torch.log1p(a) d__o = ref_data.getOutputTensorGrad(0) b.backward(torch.tensor(d__o)) return [b, a.grad, None] op_tester.setPatterns(["PreUniRepl", "Log1pGradOp"], enableRuntimeAsserts=False) op_tester.equal_nan = True op_tester.run(init_builder, reference, "train")