# Copyright (c) 2018 Graphcore Ltd. All rights reserved. import os from tempfile import TemporaryDirectory import popart import test_util as tu import numpy as np import onnx import pytest def getDevice(): # A cpu device return tu.create_test_device() def test_basic(): builder = popart.Builder() i1 = builder.addInputTensor(popart.TensorInfo("FLOAT", [1, 2, 32, 32])) i2 = builder.addInputTensor(popart.TensorInfo("FLOAT", [1, 2, 32, 32])) old_o = "" o = builder.aiOnnx.abs([i1]) assert old_o != o assert builder.getTensorShape(o) == [1, 2, 32, 32] old_o = o o = builder.aiOnnx.acos([i1]) assert old_o != o assert builder.getTensorShape(o) == [1, 2, 32, 32] old_o = o o = builder.aiOnnx.acosh([i1]) assert old_o != o assert builder.getTensorShape(o) == [1, 2, 32, 32] old_o = o o = builder.aiOnnx.add([i1, i2]) assert old_o != o assert builder.getTensorShape(o) == [1, 2, 32, 32] old_o = o o = builder.aiOnnx.logical_and([i1, i2]) assert old_o != o assert builder.getTensorShape(o) == [1, 2, 32, 32] old_o = o o = builder.aiOnnx.asin([i1]) assert old_o != o assert builder.getTensorShape(o) == [1, 2, 32, 32] old_o = o o = builder.aiOnnx.asinh([i1]) assert old_o != o assert builder.getTensorShape(o) == [1, 2, 32, 32] old_o = o o = builder.aiOnnx.atan([i1]) assert old_o != o assert builder.getTensorShape(o) == [1, 2, 32, 32] old_o = o o = builder.aiOnnx.atanh([i1]) assert old_o != o assert builder.getTensorShape(o) == [1, 2, 32, 32] old_o = o o = builder.aiOnnx.ceil([i1]) assert old_o != o assert builder.getTensorShape(o) == [1, 2, 32, 32] old_o = o o = builder.aiOnnx.cos([i1]) assert old_o != o assert builder.getTensorShape(o) == [1, 2, 32, 32] old_o = o o = builder.aiOnnx.cosh([i1]) assert old_o != o assert builder.getTensorShape(o) == [1, 2, 32, 32] old_o = o o = builder.aiOnnx.div([i1, i2]) assert old_o != o assert builder.getTensorShape(o) == [1, 2, 32, 32] old_o = o o = builder.aiOnnx.elu([i1]) assert old_o != o assert builder.getTensorShape(o) == [1, 2, 32, 32] old_o = o o = builder.aiOnnx.equal([i1, i2]) assert old_o != o assert builder.getTensorShape(o) == [1, 2, 32, 32] old_o = o o = builder.aiOnnx.exp([i1]) assert old_o != o assert builder.getTensorShape(o) == [1, 2, 32, 32] old_o = o o = builder.aiOnnx.floor([i1]) assert old_o != o assert builder.getTensorShape(o) == [1, 2, 32, 32] old_o = o o = builder.aiOnnx.greater([i1, i2]) assert old_o != o assert builder.getTensorShape(o) == [1, 2, 32, 32] old_o = o o = builder.aiOnnx.identity([i1]) assert old_o != o assert builder.getTensorShape(o) == [1, 2, 32, 32] old_o = o o = builder.aiOnnx.less([i1, i2]) assert old_o != o assert builder.getTensorShape(o) == [1, 2, 32, 32] old_o = o o = builder.aiOnnx.log([i1]) assert old_o != o assert builder.getTensorShape(o) == [1, 2, 32, 32] old_o = o o = builder.aiOnnx.logsoftmax([i1]) assert old_o != o assert builder.getTensorShape(o) == [1, 2, 32, 32] old_o = o o = builder.aiOnnx.max([i1, i2]) assert old_o != o assert builder.getTensorShape(o) == [1, 2, 32, 32] old_o = o o = builder.aiOnnx.mean([i1, i2]) assert old_o != o assert builder.getTensorShape(o) == [1, 2, 32, 32] old_o = o o = builder.aiOnnx.min([i1, i2]) assert old_o != o assert builder.getTensorShape(o) == [1, 2, 32, 32] old_o = o o = builder.aiOnnx.mul([i1, i2]) assert old_o != o assert builder.getTensorShape(o) == [1, 2, 32, 32] old_o = o o = builder.aiOnnx.neg([i1]) assert old_o != o assert builder.getTensorShape(o) == [1, 2, 32, 32] old_o = o o = builder.aiOnnx.logical_not([i1]) assert old_o != o assert builder.getTensorShape(o) == [1, 2, 32, 32] old_o = o o = builder.aiOnnx.logical_or([i1, i2]) assert old_o != o assert builder.getTensorShape(o) == [1, 2, 32, 32] old_o = o o = builder.aiOnnx.pow([i1, i2]) assert old_o != o assert builder.getTensorShape(o) == [1, 2, 32, 32] old_o = o o = builder.aiOnnx.reciprocal([i1]) assert old_o != o assert builder.getTensorShape(o) == [1, 2, 32, 32] old_o = o o = builder.aiOnnx.relu([i1]) assert old_o != o assert builder.getTensorShape(o) == [1, 2, 32, 32] old_o = o o = builder.aiOnnx.sigmoid([i1]) assert old_o != o assert builder.getTensorShape(o) == [1, 2, 32, 32] old_o = o o = builder.aiOnnx.sin([i1]) assert old_o != o assert builder.getTensorShape(o) == [1, 2, 32, 32] old_o = o o = builder.aiOnnx.sinh([i1]) assert old_o != o assert builder.getTensorShape(o) == [1, 2, 32, 32] old_o = o o = builder.aiOnnx.softsign([i1]) assert old_o != o assert builder.getTensorShape(o) == [1, 2, 32, 32] old_o = o o = builder.aiOnnx.softmax([i1]) assert old_o != o assert builder.getTensorShape(o) == [1, 2, 32, 32] old_o = o o = builder.aiOnnx.sqrt([i1]) assert old_o != o assert builder.getTensorShape(o) == [1, 2, 32, 32] old_o = o o = builder.aiOnnx.sub([i1, i2]) assert old_o != o assert builder.getTensorShape(o) == [1, 2, 32, 32] old_o = o o = builder.aiOnnx.sum([i1, i2]) assert old_o != o assert builder.getTensorShape(o) == [1, 2, 32, 32] old_o = o o = builder.aiOnnx.tan([i1]) assert old_o != o assert builder.getTensorShape(o) == [1, 2, 32, 32] old_o = o o = builder.aiOnnx.tanh([i1]) assert old_o != o assert builder.getTensorShape(o) == [1, 2, 32, 32] old_o = o o = builder.aiOnnx.logical_xor([i1, i2]) assert builder.getTensorShape(o) == [1, 2, 32, 32] assert old_o != o with pytest.raises(popart.popart_exception) as e_info: builder.aiOnnx.abs([]) assert e_info.value.args[0].startswith("Abs has invalid number of") with pytest.raises(popart.popart_exception) as e_info: builder.aiOnnx.acos([]) assert e_info.value.args[0].startswith("Acos has invalid number of") with pytest.raises(popart.popart_exception) as e_info: builder.aiOnnx.acosh([]) assert e_info.value.args[0].startswith("Acosh has invalid number of") with pytest.raises(popart.popart_exception) as e_info: builder.aiOnnx.add([]) assert e_info.value.args[0].startswith("Add has invalid number of") with pytest.raises(popart.popart_exception) as e_info: builder.aiOnnx.logical_and([]) assert e_info.value.args[0].startswith("And has invalid number of") with pytest.raises(popart.popart_exception) as e_info: builder.aiOnnx.asin([]) assert e_info.value.args[0].startswith("Asin has invalid number of") with pytest.raises(popart.popart_exception) as e_info: builder.aiOnnx.asinh([]) assert e_info.value.args[0].startswith("Asinh has invalid number of") with pytest.raises(popart.popart_exception) as e_info: builder.aiOnnx.atan([]) assert e_info.value.args[0].startswith("Atan has invalid number of") with pytest.raises(popart.popart_exception) as e_info: builder.aiOnnx.atanh([]) assert e_info.value.args[0].startswith("Atanh has invalid number of") with pytest.raises(popart.popart_exception) as e_info: builder.aiOnnx.ceil([]) assert e_info.value.args[0].startswith("Ceil has invalid number of") with pytest.raises(popart.popart_exception) as e_info: builder.aiOnnx.cos([]) assert e_info.value.args[0].startswith("Cos has invalid number of") with pytest.raises(popart.popart_exception) as e_info: builder.aiOnnx.cosh([]) assert e_info.value.args[0].startswith("Cosh has invalid number of") with pytest.raises(popart.popart_exception) as e_info: builder.aiOnnx.div([]) assert e_info.value.args[0].startswith("Div has invalid number of") with pytest.raises(popart.popart_exception) as e_info: builder.aiOnnx.elu([]) assert e_info.value.args[0].startswith("Elu has invalid number of") with pytest.raises(popart.popart_exception) as e_info: builder.aiOnnx.equal([]) assert e_info.value.args[0].startswith("Equal has invalid number of") with pytest.raises(popart.popart_exception) as e_info: builder.aiOnnx.exp([]) assert e_info.value.args[0].startswith("Exp has invalid number of") with pytest.raises(popart.popart_exception) as e_info: builder.aiOnnx.floor([]) assert e_info.value.args[0].startswith("Floor has invalid number of") with pytest.raises(popart.popart_exception) as e_info: builder.aiOnnx.greater([]) assert e_info.value.args[0].startswith("Greater has invalid number of") with pytest.raises(popart.popart_exception) as e_info: builder.aiOnnx.identity([]) assert e_info.value.args[0].startswith("Identity has invalid number of") with pytest.raises(popart.popart_exception) as e_info: builder.aiOnnx.less([]) assert e_info.value.args[0].startswith("Less has invalid number of") with pytest.raises(popart.popart_exception) as e_info: builder.aiOnnx.log([]) assert e_info.value.args[0].startswith("Log has invalid number of") with pytest.raises(popart.popart_exception) as e_info: builder.aiOnnx.max([]) assert e_info.value.args[0].startswith("Max has invalid number of") with pytest.raises(popart.popart_exception) as e_info: builder.aiOnnx.mean([]) assert e_info.value.args[0].startswith("Mean has invalid number of") with pytest.raises(popart.popart_exception) as e_info: builder.aiOnnx.min([]) assert e_info.value.args[0].startswith("Min has invalid number of") with pytest.raises(popart.popart_exception) as e_info: builder.aiOnnx.mul([]) assert e_info.value.args[0].startswith("Mul has invalid number of") with pytest.raises(popart.popart_exception) as e_info: builder.aiOnnx.neg([]) assert e_info.value.args[0].startswith("Neg has invalid number of") with pytest.raises(popart.popart_exception) as e_info: builder.aiOnnx.logical_not([]) assert e_info.value.args[0].startswith("Not has invalid number of") with pytest.raises(popart.popart_exception) as e_info: builder.aiOnnx.logical_or([]) assert e_info.value.args[0].startswith("Or has invalid number of") with pytest.raises(popart.popart_exception) as e_info: builder.aiOnnx.pow([]) assert e_info.value.args[0].startswith("Pow has invalid number of") with pytest.raises(popart.popart_exception) as e_info: builder.aiOnnx.reciprocal([]) assert e_info.value.args[0].startswith("Reciprocal has invalid number of") with pytest.raises(popart.popart_exception) as e_info: builder.aiOnnx.relu([]) assert e_info.value.args[0].startswith("Relu has invalid number of") with pytest.raises(popart.popart_exception) as e_info: builder.aiOnnx.sigmoid([]) assert e_info.value.args[0].startswith("Sigmoid has invalid number of") with pytest.raises(popart.popart_exception) as e_info: builder.aiOnnx.sin([]) assert e_info.value.args[0].startswith("Sin has invalid number of") with pytest.raises(popart.popart_exception) as e_info: builder.aiOnnx.sinh([]) assert e_info.value.args[0].startswith("Sinh has invalid number of") with pytest.raises(popart.popart_exception) as e_info: builder.aiOnnx.softsign([]) assert e_info.value.args[0].startswith("Softsign has invalid number of") with pytest.raises(popart.popart_exception) as e_info: builder.aiOnnx.sqrt([]) assert e_info.value.args[0].startswith("Sqrt has invalid number of") with pytest.raises(popart.popart_exception) as e_info: builder.aiOnnx.sub([]) assert e_info.value.args[0].startswith("Sub has invalid number of") with pytest.raises(popart.popart_exception) as e_info: builder.aiOnnx.sum([]) assert e_info.value.args[0].startswith("Sum has invalid number of") with pytest.raises(popart.popart_exception) as e_info: builder.aiOnnx.tan([]) assert e_info.value.args[0].startswith("Tan has invalid number of") with pytest.raises(popart.popart_exception) as e_info: builder.aiOnnx.tanh([]) assert e_info.value.args[0].startswith("Tanh has invalid number of") with pytest.raises(popart.popart_exception) as e_info: builder.aiOnnx.logical_xor([]) assert e_info.value.args[0].startswith("Xor has invalid number of") proto = builder.getModelProto() assert len(proto) > 0 assert len(i1) > 0 assert len(i2) > 0 assert len(o) > 0 assert i1 != i2 assert i2 != o with pytest.raises(TypeError) as e_info: builder.aiOnnx.add(0, 0) assert e_info.value.args[0].startswith("add(): incompatible function") def test_add_pad(): builder = popart.Builder() i1 = builder.addInputTensor(popart.TensorInfo("FLOAT", [10])) o = builder.aiOnnx.pad([i1], [2, 3], "constant", 0.0) builder.addOutputTensor(o) proto = builder.getModelProto() assert len(proto) > 0 assert builder.getTensorShape(o) == [15] with pytest.raises(popart.popart_exception) as e_info: builder.aiOnnx.pad([i1], [2, 3, 4, 5], "constant", 0.0) assert e_info.value.args[0].startswith("Padding vector (length 4) ") def test_add_subsample(): builder = popart.Builder() i1 = builder.addInputTensor(popart.TensorInfo("FLOAT", [10, 9, 8, 7])) o = builder.aiGraphcore.subsample([i1], strides=[2, 3, 4, 5]) builder.addOutputTensor(o) proto = builder.getModelProto() assert len(proto) > 0 assert builder.getTensorShape(o) == [5, 3, 2, 1] def test_add_constant(): builder = popart.Builder() c = builder.aiOnnx.constant(np.array([[1, 6], [4, 5]], dtype=np.float32)) builder.addOutputTensor(c) proto = builder.getModelProto() assert len(proto) > 0 assert builder.getTensorShape(c) == [2, 2] def test_add_conv(): builder = popart.Builder() i1 = builder.addInputTensor(popart.TensorInfo("FLOAT", [1, 2, 32, 32])) i2 = builder.addInputTensor(popart.TensorInfo("FLOAT", [4, 2, 3, 3])) o = builder.aiOnnx.conv( [i1, i2], dilations=[1, 1], pads=[0, 0, 0, 0], strides=[1, 1] ) builder.addOutputTensor(o) proto = builder.getModelProto() assert len(proto) > 0 assert len(i1) > 0 assert len(i2) > 0 assert len(o) > 0 assert i1 != i2 assert i2 != o assert builder.getTensorShape(o) == [1, 4, 30, 30] with pytest.raises(TypeError) as e_info: builder.aiOnnx.conv() assert e_info.value.args[0].startswith("conv(): incompatible function") with pytest.raises(popart.popart_exception) as e_info: builder.aiOnnx.conv([i1, i2], dilations=[1, 1], pads=[0, 0, 0, 0], strides=[1]) assert e_info.value.args[0].startswith("Length of strides vector 1 != number") with pytest.raises(popart.popart_exception) as e_info: builder.aiOnnx.conv([i1, i2], dilations=[1, 1], pads=[0, 0], strides=[1, 1]) assert e_info.value.args[0].startswith( "Padding vector (length 2) does not have 2 values for each spatial" ) with pytest.raises(popart.popart_exception) as e_info: builder.aiOnnx.conv([i1, i2], dilations=[1], pads=[0, 0, 0, 0], strides=[1, 1]) assert e_info.value.args[0].startswith( "Length of dilations vector 1 != number of spatial d" ) def test_add_conv_and_bias(): builder = popart.Builder() i1 = builder.addInputTensor(popart.TensorInfo("FLOAT", [1, 2, 32, 32])) i2 = builder.addInputTensor(popart.TensorInfo("FLOAT", [4, 2, 3, 3])) i3 = builder.addInputTensor(popart.TensorInfo("FLOAT", [4])) o = builder.aiOnnx.conv( [i1, i2, i3], dilations=[1, 1], pads=[0, 0, 0, 0], strides=[1, 1] ) builder.addOutputTensor(o) proto = builder.getModelProto() assert len(proto) > 0 assert len(i1) > 0 assert len(i2) > 0 assert len(i3) > 0 assert len(o) > 0 assert i1 != i2 assert i2 != o assert builder.getTensorShape(o) == [1, 4, 30, 30] with pytest.raises(TypeError) as e_info: builder.aiOnnx.conv() assert e_info.value.args[0].startswith("conv(): incompatible") def test_add_gemm(): builder = popart.Builder() i1 = builder.addInputTensor(popart.TensorInfo("FLOAT", [12, 8])) i2 = builder.addInputTensor(popart.TensorInfo("FLOAT", [8, 16])) i3 = builder.addInputTensor(popart.TensorInfo("FLOAT", [16])) o = builder.aiOnnx.gemm([i1, i2, i3], 1.0, 1.0, 0, 0) builder.addOutputTensor(o) proto = builder.getModelProto() assert len(proto) > 0 assert len(i1) > 0 assert len(i2) > 0 assert len(o) > 0 assert i1 != i2 assert i2 != o assert builder.getTensorShape(o) == [12, 16] with pytest.raises(TypeError) as e_info: builder.aiOnnx.gemm(0, 0, 0, 0, 0, 0, 0) assert e_info.value.args[0].startswith("gemm(): incompatible function") def test_add_identityloss(): builder = popart.Builder() i = builder.addInputTensor(popart.TensorInfo("FLOAT", [10, 2, 3])) o1 = builder.aiGraphcore.identityloss([i], reduction=popart.ReductionType.Mean) o2 = builder.aiGraphcore.identityloss([i], reduction=popart.ReductionType.Sum) o3 = builder.aiGraphcore.identityloss( [i], reduction=popart.ReductionType.NoReduction ) assert builder.getTensorShape(o1) == [] assert builder.getTensorShape(o2) == [] assert builder.getTensorShape(o3) == [10, 2, 3] assert builder.getTensorDtypeString(o1) == "float32" assert builder.getTensorDtypeString(o2) == "float32" assert builder.getTensorDtypeString(o3) == "float32" def test_add_l1loss(): builder = popart.Builder() i = builder.addInputTensor(popart.TensorInfo("FLOAT", [10, 2, 3])) o1 = builder.aiGraphcore.l1loss([i], 1.0, reduction=popart.ReductionType.Mean) o2 = builder.aiGraphcore.l1loss([i], 1.0, reduction=popart.ReductionType.Sum) o3 = builder.aiGraphcore.l1loss( [i], 1.0, reduction=popart.ReductionType.NoReduction ) assert builder.getTensorShape(o1) == [] assert builder.getTensorShape(o2) == [] assert builder.getTensorShape(o3) == [10, 2, 3] assert builder.getTensorDtypeString(o1) == "float32" assert builder.getTensorDtypeString(o2) == "float32" assert builder.getTensorDtypeString(o3) == "float32" def test_add_matmul(): builder = popart.Builder() i1 = builder.addInputTensor(popart.TensorInfo("FLOAT", [2, 3])) i2 = builder.addInputTensor(popart.TensorInfo("FLOAT", [3, 4])) o = builder.aiOnnx.matmul([i1, i2]) builder.addOutputTensor(o) proto = builder.getModelProto() assert len(proto) > 0 assert len(i1) > 0 assert len(i2) > 0 assert len(o) > 0 assert i1 != i2 assert i2 != o assert builder.getTensorShape(o) == [2, 4] with pytest.raises(TypeError) as e_info: builder.aiOnnx.matmul(0, 0, 0, 0, 0, 0, 0) assert e_info.value.args[0].startswith("matmul(): incompatible function") def test_add_nllloss(): builder = popart.Builder() i1 = builder.addInputTensor(popart.TensorInfo("FLOAT", [10, 2, 3])) i2 = builder.addInputTensor(popart.TensorInfo("INT32", [2, 3])) o1 = builder.aiGraphcore.nllloss([i1, i2], reduction=popart.ReductionType.Mean) o2 = builder.aiGraphcore.nllloss([i1, i2], reduction=popart.ReductionType.Sum) o3 = builder.aiGraphcore.nllloss( [i1, i2], reduction=popart.ReductionType.NoReduction ) assert builder.getTensorShape(o1) == [] assert builder.getTensorShape(o2) == [] assert builder.getTensorShape(o3) == [2, 3] assert builder.getTensorDtypeString(o1) == "float32" assert builder.getTensorDtypeString(o2) == "float32" assert builder.getTensorDtypeString(o3) == "float32" def test_add_reshape(): builder = popart.Builder() i1 = builder.addInputTensor(popart.TensorInfo("FLOAT", [2, 3])) i2 = builder.addInitializedInputTensor(np.array([1, 6], dtype=np.int64)) o = builder.aiOnnx.reshape([i1, i2]) builder.addOutputTensor(o) assert builder.getTensorShape(o) == [1, 6] def test_add_reshape_const(): builder = popart.Builder() i1 = builder.addInputTensor(popart.TensorInfo("FLOAT", [2, 3])) o = builder.reshape_const(builder.aiOnnx, [i1], [1, 6]) builder.addOutputTensor(o) assert builder.getTensorShape(o) == [1, 6] def test_initialized_input_fp32(): builder1 = popart.Builder() builder1.addInputTensor(popart.TensorInfo("FLOAT", [100, 100])) proto1 = builder1.getModelProto() assert len(proto1) > 0 builder2 = popart.Builder() builder2.addInitializedInputTensor(np.ones([100, 100], np.float32)) proto2 = builder2.getModelProto() assert len(proto2) > 0 assert len(proto2) > len(proto1) def test_initialized_input_int32(): builder1 = popart.Builder() builder1.addInputTensor(popart.TensorInfo("INT32", [100, 100])) proto1 = builder1.getModelProto() assert len(proto1) > 0 builder2 = popart.Builder() builder2.addInitializedInputTensor(np.ones([100, 100], np.int32)) proto2 = builder2.getModelProto() assert len(proto2) > 0 assert len(proto2) > len(proto1) def test_initialized_input_fp16(): builder1 = popart.Builder() builder1.addInputTensor(popart.TensorInfo("FLOAT16", [100, 100])) proto1 = builder1.getModelProto() assert len(proto1) > 0 builder2 = popart.Builder() builder2.addInitializedInputTensor(np.ones([100, 100], np.float16)) proto2 = builder2.getModelProto() assert len(proto2) > 0 assert len(proto2) > len(proto1) def test_initialized_input_bool(): builder1 = popart.Builder() builder1.addInputTensor(popart.TensorInfo("BOOL", [100, 100])) proto1 = builder1.getModelProto() assert len(proto1) > 0 builder2 = popart.Builder() builder2.addInitializedInputTensor(np.ones([100, 100], np.bool)) proto2 = builder2.getModelProto() assert len(proto2) > 0 assert len(proto2) > len(proto1) def test_inout_tensor_info(): builder = popart.Builder() i1 = builder.addInputTensor(popart.TensorInfo("FLOAT", [1, 2, 32, 32])) i2 = builder.addInputTensor(popart.TensorInfo("FLOAT", [4, 2, 3, 3])) i3 = builder.addInputTensor(popart.TensorInfo("FLOAT", [4, 2, 3, 3])) x = builder.aiOnnx.conv( [i1, i2], dilations=[1, 1], pads=[0, 0, 0, 0], strides=[1, 1] ) y = builder.aiOnnx.conv( [i1, i3], dilations=[1, 1], pads=[0, 0, 0, 0], strides=[1, 1] ) o = builder.aiOnnx.add([x, y]) builder.addOutputTensor(o) assert builder.getInputTensorIds() == [i1, i2, i3] assert builder.getOutputTensorIds() == [o] assert builder.getValueTensorIds() == [x, y, o] assert builder.getTensorShape(i1) == [1, 2, 32, 32] assert builder.getTensorShape(i2) == [4, 2, 3, 3] assert builder.getTensorShape(i3) == [4, 2, 3, 3] assert builder.getTensorShape(x) == [1, 4, 30, 30] assert builder.getTensorShape(y) == [1, 4, 30, 30] assert builder.getTensorShape(o) == [1, 4, 30, 30] with pytest.raises(popart.popart_exception) as e_info: builder.getTensorShape("NotAnId") assert e_info.value.args[0].find("is not an input or output tensor. Must be") def test_set_virtual_graph(): builder = popart.Builder() i1 = builder.addInputTensor(popart.TensorInfo("FLOAT", [1, 2, 32, 32])) i2 = builder.addInputTensor(popart.TensorInfo("FLOAT", [4, 2, 3, 3])) o = builder.aiOnnx.conv( [i1, i2], dilations=[1, 1], pads=[0, 0, 0, 0], strides=[1, 1] ) builder.virtualGraph(o, 1) builder.addOutputTensor(o) # Test that tha attribute has been set val = 1 res = builder.getVirtualGraph(o) assert res == val def test_set_weights_from_host(): # Run the first builder builder = popart.Builder() shape = popart.TensorInfo("FLOAT", [2]) i1 = builder.addInputTensor(shape) data = np.array([1, 2], dtype=np.float32) i2 = builder.addInitializedInputTensor(data) o = builder.aiOnnx.add([i1, i2]) loss = builder.aiGraphcore.identityloss([o]) proto = builder.getModelProto() dataFlow = popart.DataFlow(1, {o: popart.AnchorReturnType("All")}) optimizer = popart.ConstSGD(0.01) with getDevice() as device: session = popart.TrainingSession( fnModel=proto, dataFlow=dataFlow, loss=loss, optimizer=optimizer, deviceInfo=device, ) anchors = session.initAnchorArrays() session.prepareDevice() inputs = {i1: np.array([1, 2], dtype=np.float32)} stepio = popart.PyStepIO(inputs, anchors) with pytest.raises(popart.popart_exception) as e_info: session.run(stepio) assert e_info.value.args[0].startswith( "Must call weightsFromHost before run as the" ) def test_add_int_attribute(): builder = popart.Builder() i1 = builder.addInputTensor(popart.TensorInfo("FLOAT", [1, 2, 32, 32])) i2 = builder.addInputTensor(popart.TensorInfo("FLOAT", [4, 2, 3, 3])) o = builder.aiOnnx.conv( [i1, i2], dilations=[1, 1], pads=[0, 0, 0, 0], strides=[1, 1] ) builder.addOutputTensor(o) # Set then get val = 100 builder.addNodeAttribute("test", val, set([o])) res = builder.getInt64NodeAttribute("test", set([o])) assert res == val def test_add_int_vector_attribute(): builder = popart.Builder() i1 = builder.addInputTensor(popart.TensorInfo("FLOAT", [1, 2, 32, 32])) i2 = builder.addInputTensor(popart.TensorInfo("FLOAT", [4, 2, 3, 3])) o = builder.aiOnnx.conv( [i1, i2], dilations=[1, 1], pads=[0, 0, 0, 0], strides=[1, 1] ) builder.addOutputTensor(o) # Set then get val = [100, 200, -1] builder.addNodeAttribute("test", val, set([o])) res = builder.getInt64VectorNodeAttribute("test", set([o])) assert res == val def test_add_float_attribute(): builder = popart.Builder() i1 = builder.addInputTensor(popart.TensorInfo("FLOAT", [1, 2, 32, 32])) i2 = builder.addInputTensor(popart.TensorInfo("FLOAT", [4, 2, 3, 3])) o = builder.aiOnnx.conv( [i1, i2], dilations=[1, 1], pads=[0, 0, 0, 0], strides=[1, 1] ) builder.addOutputTensor(o) # Set then get val = 0.1 builder.addNodeAttribute("test", val, set([o])) res = builder.getFloatNodeAttribute("test", set([o])) assert res == pytest.approx(val) def test_add_float_vector_attribute(): builder = popart.Builder() i1 = builder.addInputTensor(popart.TensorInfo("FLOAT", [1, 2, 32, 32])) i2 = builder.addInputTensor(popart.TensorInfo("FLOAT", [4, 2, 3, 3])) o = builder.aiOnnx.conv( [i1, i2], dilations=[1, 1], pads=[0, 0, 0, 0], strides=[1, 1] ) builder.addOutputTensor(o) # Set then get val = [100.0, -0.1, 100.0, 10.0] builder.addNodeAttribute("test", val, set([o])) res = builder.getFloatVectorNodeAttribute("test", set([o])) assert res == pytest.approx(val) def test_add_string_attribute(): builder = popart.Builder() i1 = builder.addInputTensor(popart.TensorInfo("FLOAT", [1, 2, 32, 32])) i2 = builder.addInputTensor(popart.TensorInfo("FLOAT", [4, 2, 3, 3])) o = builder.aiOnnx.conv( [i1, i2], dilations=[1, 1], pads=[0, 0, 0, 0], strides=[1, 1] ) builder.addOutputTensor(o) # Set then get val = "test" builder.addNodeAttribute("test", val, set([o])) res = builder.getStringNodeAttribute("test", set([o])) assert res == val def test_add_string_vector_attribute(): builder = popart.Builder() i1 = builder.addInputTensor(popart.TensorInfo("FLOAT", [1, 2, 32, 32])) i2 = builder.addInputTensor(popart.TensorInfo("FLOAT", [4, 2, 3, 3])) o = builder.aiOnnx.conv( [i1, i2], dilations=[1, 1], pads=[0, 0, 0, 0], strides=[1, 1] ) builder.addOutputTensor(o) # Set then get val = ["test", "test2", "test"] builder.addNodeAttribute("test", val, set([o])) res = builder.getStringVectorNodeAttribute("test", set([o])) assert res == val def test_add_attribute_missing_node(): builder = popart.Builder() val = 100 with pytest.raises(popart.popart_exception) as e_info: builder.addNodeAttribute("test", val, set(("i", "j"))) assert e_info.value.args[0].find("Could not find a node with outputs i, j.") def test_has_attribute(): builder = popart.Builder() i1 = builder.addInputTensor(popart.TensorInfo("FLOAT", [1, 2, 32, 32])) i2 = builder.addInputTensor(popart.TensorInfo("FLOAT", [4, 2, 3, 3])) o = builder.aiOnnx.conv( [i1, i2], dilations=[1, 1], pads=[0, 0, 0, 0], strides=[1, 1] ) builder.addOutputTensor(o) # Set then get val = 100 builder.addNodeAttribute("test", val, set([o])) res = builder.nodeHasAttribute("test", set([o])) assert res res = builder.nodeHasAttribute("test2", set([o])) assert not res def test_get_all_attribute_names(): builder = popart.Builder() i1 = builder.addInputTensor(popart.TensorInfo("FLOAT", [1, 2, 32, 32])) o = builder.aiOnnx.add((i1, i1)) builder.addOutputTensor(o) builder.addNodeAttribute("test", 100.0, set([o])) builder.addNodeAttribute("test2", -1, set([o])) builder.addNodeAttribute("test3", "abba", set([o])) res = builder.getAllNodeAttributeNames(set([o])) assert set(res) == set(("test", "test2", "test3", "__debug_info_id")) res = builder.getFloatNodeAttribute("test", set([o])) assert res == pytest.approx(100.0) res = builder.getInt64NodeAttribute("test2", set([o])) assert res == -1 res = builder.getStringNodeAttribute("test3", set([o])) assert res == "abba" def test_get_conv_strides_attribute(): strides = [1, 1] builder = popart.Builder() i1 = builder.addInputTensor(popart.TensorInfo("FLOAT", [1, 2, 32, 32])) i2 = builder.addInputTensor(popart.TensorInfo("FLOAT", [4, 2, 3, 3])) o = builder.aiOnnx.conv( [i1, i2], dilations=[1, 1], pads=[0, 0, 0, 0], strides=strides ) builder.addOutputTensor(o) res = builder.getInt64VectorNodeAttribute("strides", set([o])) assert res == strides def test_dont_override_attribute(): builder = popart.Builder() i1 = builder.addInputTensor(popart.TensorInfo("FLOAT", [1, 2, 32, 32])) o = builder.aiOnnx.add((i1, i1)) builder.addOutputTensor(o) builder.addNodeAttribute("test", 100.0, set([o])) with pytest.raises(popart.popart_exception) as e_info: builder.addNodeAttribute("test", 100, set([o])) assert e_info.value.args[0] == "Node already has attribute test." def test_get_attribute_doesnt_exist(): builder = popart.Builder() i1 = builder.addInputTensor(popart.TensorInfo("FLOAT", [1, 2, 32, 32])) o = builder.aiOnnx.add((i1, i1)) builder.addOutputTensor(o) builder.addNodeAttribute("test", 100.0, set([o])) with pytest.raises(popart.popart_exception) as e_info: builder.getFloatNodeAttribute("test1", set([o])) assert e_info.value.args[0] == "Node does not have an attribute test1." def test_remove_attribute(): builder = popart.Builder() i1 = builder.addInputTensor(popart.TensorInfo("FLOAT", [1, 2, 32, 32])) i2 = builder.addInputTensor(popart.TensorInfo("FLOAT", [4, 2, 3, 3])) o = builder.aiOnnx.conv( [i1, i2], dilations=[1, 1], pads=[0, 0, 0, 0], strides=[1, 1] ) builder.addOutputTensor(o) # Set then get val = 100 builder.addNodeAttribute("test", val, set([o])) res = builder.nodeHasAttribute("test", set([o])) assert res builder.removeNodeAttribute("test", set([o])) res = builder.nodeHasAttribute("test", set([o])) assert not res def test_remove_attribute_doesnt_exist(): builder = popart.Builder() i1 = builder.addInputTensor(popart.TensorInfo("FLOAT", [1, 2, 32, 32])) i2 = builder.addInputTensor(popart.TensorInfo("FLOAT", [4, 2, 3, 3])) o = builder.aiOnnx.conv( [i1, i2], dilations=[1, 1], pads=[0, 0, 0, 0], strides=[1, 1] ) builder.addOutputTensor(o) # Set then get val = 100 builder.addNodeAttribute("test", val, set([o])) res = builder.nodeHasAttribute("test", set([o])) assert res with pytest.raises(popart.popart_exception) as e_info: builder.removeNodeAttribute("test1", set([o])) assert e_info.value.args[0] == "Cannot remove attribute test1 as it does not exist." def test_get_attribute_wrong_type_int(): builder = popart.Builder() i1 = builder.addInputTensor(popart.TensorInfo("FLOAT", [1, 2, 32, 32])) o = builder.aiOnnx.add((i1, i1)) builder.addOutputTensor(o) builder.addNodeAttribute("test", 100.0, set([o])) with pytest.raises(popart.popart_exception) as e_info: builder.getInt64NodeAttribute("test", set([o])) assert e_info.value.args[0].find("Node test is not an integer.") def test_get_attribute_wrong_type_int_vector(): builder = popart.Builder() i1 = builder.addInputTensor(popart.TensorInfo("FLOAT", [1, 2, 32, 32])) o = builder.aiOnnx.add((i1, i1)) builder.addOutputTensor(o) builder.addNodeAttribute("test", 100.0, set([o])) with pytest.raises(popart.popart_exception) as e_info: builder.getInt64VectorNodeAttribute("test", set([o])) assert e_info.value.args[0].find("Node test is not an integer vector.") def test_get_attribute_wrong_type_float(): builder = popart.Builder() i1 = builder.addInputTensor(popart.TensorInfo("FLOAT", [1, 2, 32, 32])) o = builder.aiOnnx.add((i1, i1)) builder.addOutputTensor(o) builder.addNodeAttribute("test", 100, set([o])) with pytest.raises(popart.popart_exception) as e_info: builder.getFloatNodeAttribute("test", set([o])) assert e_info.value.args[0].find("Node test is not a float.") def test_get_attribute_wrong_type_float_vector(): builder = popart.Builder() i1 = builder.addInputTensor(popart.TensorInfo("FLOAT", [1, 2, 32, 32])) o = builder.aiOnnx.add((i1, i1)) builder.addOutputTensor(o) builder.addNodeAttribute("test", 100, set([o])) with pytest.raises(popart.popart_exception) as e_info: builder.getFloatVectorNodeAttribute("test", set([o])) assert e_info.value.args[0].find("Node test is not a float vector.") def test_get_attribute_wrong_type_string(): builder = popart.Builder() i1 = builder.addInputTensor(popart.TensorInfo("FLOAT", [1, 2, 32, 32])) o = builder.aiOnnx.add((i1, i1)) builder.addOutputTensor(o) builder.addNodeAttribute("test", 100, set([o])) with pytest.raises(popart.popart_exception) as e_info: builder.getStringNodeAttribute("test", set([o])) assert e_info.value.args[0].find("Node test is not a string.") def test_get_attribute_wrong_type_string_vector(): builder = popart.Builder() i1 = builder.addInputTensor(popart.TensorInfo("FLOAT", [1, 2, 32, 32])) o = builder.aiOnnx.add((i1, i1)) builder.addOutputTensor(o) builder.addNodeAttribute("test", 100, set([o])) with pytest.raises(popart.popart_exception) as e_info: builder.getStringVectorNodeAttribute("test", set([o])) assert e_info.value.args[0].find("Node test is not a string vector.") def test_load_onnx_model_from_other_builder(): # Run the first builder builder = popart.Builder() shape = popart.TensorInfo("FLOAT", [2]) i1 = builder.addInputTensor(shape) i2 = builder.addInputTensor(shape) o = builder.aiOnnx.add([i1, i2]) builder.addOutputTensor(o) proto = builder.getModelProto() dataFlow = popart.DataFlow(1, {o: popart.AnchorReturnType("All")}) with getDevice() as device: session = popart.InferenceSession( fnModel=proto, dataFlow=dataFlow, deviceInfo=device ) anchors = session.initAnchorArrays() session.prepareDevice() inputs = { i1: np.array([1, 2], dtype=np.float32), i2: np.array([3, 4], dtype=np.float32), } stepio = popart.PyStepIO(inputs, anchors) session.run(stepio) assert np.array_equal(anchors[o], [4, 6]) # Run a builder that imports the model of the other builder and check the # output is still the same _ = popart.Builder(proto) dataFlow = popart.DataFlow(1, {o: popart.AnchorReturnType("All")}) proto2 = builder.getModelProto() with getDevice() as device: session = popart.InferenceSession( fnModel=proto2, dataFlow=dataFlow, deviceInfo=device ) anchors = session.initAnchorArrays() session.prepareDevice() inputs = { i1: np.array([1, 2], dtype=np.float32), i2: np.array([3, 4], dtype=np.float32), } stepio = popart.PyStepIO(inputs, anchors) session.run(stepio) assert np.array_equal(anchors[o], [4, 6]) def test_load_onnx_model_from_file(tmpdir): # Create a builder, store the model in a file and load it into a different # builder. builder = popart.Builder() shape = popart.TensorInfo("FLOAT", [2]) i1 = builder.addInputTensor(shape) i2 = builder.addInputTensor(shape) o = builder.aiOnnx.add([i1, i2]) builder.addOutputTensor(o) filename = tmpdir + "/model.onnx" with open(filename, "wb") as out: out.write(builder.getModelProto()) builder2 = popart.Builder(str(filename)) dataFlow = popart.DataFlow(1, {o: popart.AnchorReturnType("All")}) _ = popart.ConstSGD(0.01) proto = builder2.getModelProto() with getDevice() as device: session = popart.InferenceSession( fnModel=proto, dataFlow=dataFlow, deviceInfo=device ) anchors = session.initAnchorArrays() session.prepareDevice() inputs = { i1: np.array([1, 2], dtype=np.float32), i2: np.array([3, 4], dtype=np.float32), } stepio = popart.PyStepIO(inputs, anchors) session.run(stepio) assert np.array_equal(anchors[o], [4, 6]) def test_convert_initializers_to_constants(): builder = popart.Builder() i1 = builder.addInputTensor(popart.TensorInfo("FLOAT", [2, 3])) i2 = builder.addInitializedInputTensor(np.array([1, 6], dtype=np.int64)) o = builder.aiOnnx.reshape([i1, i2]) builder.addOutputTensor(o) graph_transformer = popart.GraphTransformer(builder.getModelProto()) graph_transformer.convertInitializersToConstants([i2]) with pytest.raises(popart.popart_exception) as e_info: graph_transformer.convertInitializersToConstants(["unknown"]) assert e_info.value.args[0] == "TensorId unknown not in the model initalizers" builder = popart.Builder(graph_transformer.getModelProto()) ids = builder.getInputTensorIds() assert i1 in ids assert i2 not in ids def test_convert_all_fixed_point_initializers_to_constants(): builder = popart.Builder() i1 = builder.addInputTensor(popart.TensorInfo("FLOAT", [2, 3])) i2 = builder.addInitializedInputTensor(np.array([1, 6], dtype=np.int64)) o1 = builder.aiOnnx.reshape([i1, i2]) i3 = builder.addInputTensor(popart.TensorInfo("FLOAT", [3, 2])) i4 = builder.addInitializedInputTensor(np.array([1, 6], dtype=np.int64)) o2 = builder.aiOnnx.reshape([i3, i4]) o = builder.aiOnnx.add([o1, o2]) builder.addOutputTensor(o) graph_transformer = popart.GraphTransformer(builder.getModelProto()) graph_transformer.convertAllFixedPointInitializersToConstants() builder = popart.Builder(graph_transformer.getModelProto()) ids = builder.getInputTensorIds() assert i1 in ids assert i2 not in ids assert i3 in ids assert i4 not in ids def test_builder_opsetDefaultVersions(): # This will create a build with the default opsets # This test may fail when we upgrade ops.... builder = popart.Builder() i1 = builder.addInitializedInputTensor(np.array([1, 6], dtype=np.int64)) i2 = builder.addInitializedInputTensor(np.array([1, 6], dtype=np.int64)) # This will work as it will use opset 9 version of add _ = builder.aiOnnx.add([i1, i2], "add_v9") # This will not work as it will use opset 6 version of add with pytest.raises(TypeError) as e_info: _ = builder.aiOnnx.add([i1, i2], 1, 0, "add_v6") assert ( "add(): incompatible function arguments. The following argument types are supported:" in e_info.value.args[0] ) def test_builder_aiOnnxOpsetBuilderInterface(): # The purpose of this test is to detect any potential disconnects # between the Builder Python interface (builder.py) and the # corresponding C++ Builder backend (builder.cpp, builder.cpp.gen, etc.) # # Ops added to a popART builder with e.g. opset version 9 # should first instantiate a Python class called AiOnnx9 # (defined in builder.py) which should invoke functions implemented # on the C++ AiOnnxOpset9 class # Create a builder with opset version 9 builder = popart.Builder(opsets={"ai.onnx": 9, "ai.graphcore": 1}) i1 = builder.addInitializedInputTensor(np.array([1, 5], dtype=np.float32)) i2 = builder.addInitializedInputTensor(np.array([1, 5], dtype=np.float32)) # Assert that Add is rerouted to builder interface class AiOnnx9 _ = builder.aiOnnx.add([i1, i2], "test_add_9") aiOnnxVersionAdd = f"{builder.aiOnnx}" assert aiOnnxVersionAdd == "AiOnnx9" # Create a builder with opset version 7 pbuilder = popart.Builder(opsets={"ai.onnx": 7, "ai.graphcore": 1}) i1 = pbuilder.addInitializedInputTensor(np.array([1, 5], dtype=np.float32)) i2 = pbuilder.addInitializedInputTensor(np.array([1, 1], dtype=np.float32)) # Assert that Pow is rerouted to builder interface class AiOnnx7 _ = pbuilder.aiOnnx.add([i1, i2], "test_pow_7") aiOnnxVersionPow = f"{pbuilder.aiOnnx}" assert aiOnnxVersionPow == "AiOnnx7" def test_builder_opsetDefinesVersions(): # This will create a build with the default opsets # This test may fail when we upgrade ops.... builder = popart.Builder(opsets={"ai.onnx": 6, "ai.graphcore": 1}) i1 = builder.addInitializedInputTensor(np.array([1, 6], dtype=np.int64)) i2 = builder.addInitializedInputTensor(np.array([1, 6], dtype=np.int64)) # This will work as it will use opset 6 version of add _ = builder.aiOnnx.add([i1, i2], 1, 0, "add_v6") # This will not work as it will use opset 9 version of add with pytest.raises(TypeError) as e_info: _ = builder.aiOnnx.add([i1, i2], "add_v9") assert ( "add(): incompatible function arguments. The following argument types are supported:" in e_info.value.args[0] ) def test_builder_opsetVersioning(): builder = popart.Builder(opsets={"ai.onnx": 9, "ai.graphcore": 1}) i1 = builder.addInitializedInputTensor(np.array([1, 6], dtype=np.int64)) i2 = builder.addInitializedInputTensor(np.array([1, 6], dtype=np.int64)) _ = builder.aiOnnx.add([i1, i2], "b") _ = builder.aiOnnxOpset9.add([i1, i2], "a") # This will fail as we have already defined the opset as 9 with pytest.raises(RuntimeError) as e_info: _ = builder.aiOnnxOpset6.add([i1, i2], 1, 0, "a") assert ( "Invalid opset 'aiOnnxOpset6' selected. Opset for domain ai.onnx already defined as 9" in e_info.value.args[0] ) def test_name_scope(): builder = popart.Builder(opsets={"ai.onnx": 9, "ai.graphcore": 1}) i1 = builder.addInitializedInputTensor(np.array([1, 6], dtype=np.int64)) i2 = builder.addInitializedInputTensor(np.array([1, 6], dtype=np.int64)) with builder.nameScope("scope1"): o1 = builder.aiOnnx.add([i1, i2], "myop1") with builder.nameScope("scope2"): o2 = builder.aiOnnx.add([i1, i2], "myop2") with builder.nameScope("embedded_scope"): o3 = builder.aiOnnx.add([o1, o2], "myop3") assert o1.startswith("scope1") assert o2.startswith("scope2") assert o3.startswith("scope2/embedded_scope") def test_tensor_names(): builder = popart.Builder(opsets={"ai.onnx": 9, "ai.graphcore": 1}) i1 = builder.addInitializedInputTensor(np.array([1, 6], dtype=np.int64)) i2 = builder.addInitializedInputTensor(np.array([1, 6], dtype=np.int64), "weights") assert i1 == "init_input" assert i2 == "weights" with builder.nameScope("layer1"): i3 = builder.addInitializedInputTensor(np.array([1, 6], dtype=np.int64)) i4 = builder.addInitializedInputTensor( np.array([1, 6], dtype=np.int64), "weights" ) assert i3 == "layer1/init_input" assert i4 == "layer1/weights" i5 = builder.addInputTensor(popart.TensorInfo("FLOAT", [2, 3])) i6 = builder.addInputTensor(popart.TensorInfo("FLOAT", [2, 3]), "data") with builder.nameScope("layer2"): i7 = builder.addInputTensor(popart.TensorInfo("FLOAT", [2, 3]), "label") assert i5 == "input" assert i6 == "data" assert i7 == "layer2/label" i10 = builder.addInputTensor(popart.TensorInfo("FLOAT", [2, 3])) i11 = builder.addInputTensor(popart.TensorInfo("FLOAT", [2, 3]), "data") assert i10 == "input/1" assert i11 == "data/1" def test_add_untyped_input(): builder = popart.Builder() # Trying to add an untyped input tensor on the top-level # graph should throw an error. with pytest.raises(popart.popart_exception) as e_info: builder.addUntypedInputTensor() assert e_info.value.args[0].startswith( "Can not add untyped tensors to the top-level graph." ) # Adding an untyped input on a subgraph should be fine. sub_builder = builder.createSubgraphBuilder() _ = sub_builder.addUntypedInputTensor() def test_save_model_to_file(): builder = popart.Builder() i = builder.addInputTensor(popart.TensorInfo("FLOAT", [2])) _ = builder.aiOnnx.sqrt([i]) with TemporaryDirectory() as tmpdir: tmpfile = os.path.join(tmpdir, "model.onnx") builder.saveModelProto(tmpfile) # Check file has saved assert os.path.exists(tmpfile) is True # Check that, when re-loaded, its contents are the same as when saved new_builder = popart.Builder(tmpfile) assert new_builder.getModelProto() == builder.getModelProto() def test_get_tensor_type(): builder = popart.Builder() i0 = builder.addInputTensor(popart.TensorInfo("FLOAT", [10, 9, 8, 7])) act0 = builder.aiGraphcore.subsample([i0], strides=[2, 3, 4, 5]) act1 = builder.aiOnnx.cast([act0], "INT32") assert builder.getTensorDtypeString(i0) == "float32" assert builder.getTensorDtypeString(act0) == "float32" assert builder.getTensorDtypeString(act1) == "int32" def test_is_initializer(): builder = popart.Builder() i0 = builder.addInputTensor(popart.TensorInfo("FLOAT", [10, 9, 8, 7])) i1 = builder.addInitializedInputTensor(np.array([1, 6], dtype=np.int64)) assert builder.isInitializer(i0) is False assert builder.isInitializer(i1) is True def test_conv_kernel_shape_mismatch(): builder = popart.Builder() i = builder.addInputTensor(popart.TensorInfo("FLOAT", [1, 64, 56, 56])) w = builder.addInputTensor(popart.TensorInfo("FLOAT", [256, 64, 1, 1])) # kernel shape, as inferred from weights shape = [1, 1] with pytest.raises(popart.popart_exception) as e_info: _ = builder.aiOnnx.conv( [i, w], dilations=[1, 1], kernel_shape=[64, 64], # not [1, 1] strides=[1, 1], pads=[0, 0, 0, 0], ) assert ( e_info.value.args[0] == "kernel_shape, [64 64], does not match inferred shape from weight input '" + w + "', [1 1]" ) def test_conv_invalid_kernel_shape(): builder = popart.Builder() i = builder.addInputTensor(popart.TensorInfo("FLOAT", [1, 64, 56, 56])) w = builder.addInputTensor(popart.TensorInfo("FLOAT", [256, 64, 64, 64])) with pytest.raises(popart.popart_exception) as e_info: # Kernel [64, 64] not valid for spatial input [56, 56], given other # conv window parameters _ = builder.aiOnnx.conv( [i, w], dilations=[1, 1], strides=[1, 1], pads=[0, 0, 0, 0] ) assert ( e_info.value.args[0] == "Window parameter values combine to give invalid spatial output shape: [-7 -7]" ) def _get_onnx_opset_version(proto): onnx_domain = "" m = onnx.load_from_string(proto) opsets = m.opset_import opsets = [i for i in opsets if i.domain == onnx_domain] assert len(opsets) == 1 return opsets[0].version # Test setting the opset version in the constructor def test_opset_version_constructor(): builder = popart.Builder(opsets={"ai.onnx": 9, "ai.graphcore": 1}) a = builder.addInputTensor(popart.TensorInfo("FLOAT", [1]), "a") b = builder.addInputTensor(popart.TensorInfo("FLOAT", [1]), "b") o = builder.aiOnnx.add([a, b]) builder.addOutputTensor(o) proto = builder.getModelProto() assert _get_onnx_opset_version(proto) == 9 # Test using the default opset version def test_opset_version_default(): builder = popart.Builder() a = builder.addInputTensor(popart.TensorInfo("FLOAT", [1]), "a") b = builder.addInputTensor(popart.TensorInfo("FLOAT", [1]), "b") o = builder.aiOnnx.add([a, b]) builder.addOutputTensor(o) proto = builder.getModelProto() assert _get_onnx_opset_version(proto) == popart.defaultAiOnnxOpset # Test choosing an opset version by using an explicit opset def test_opset_version_opset_choice(): builder = popart.Builder() a = builder.addInputTensor(popart.TensorInfo("FLOAT", [1]), "a") b = builder.addInputTensor(popart.TensorInfo("FLOAT", [1]), "b") o = builder.aiOnnxOpset8.add([a, b]) builder.addOutputTensor(o) proto = builder.getModelProto() assert _get_onnx_opset_version(proto) == 8 # Check that an error is thrown when trying to use a different # opset to the one specified in the Builder constructor def test_opset_version_constructor_override(): builder = popart.Builder(opsets={"ai.onnx": 9, "ai.graphcore": 1}) a = builder.addInputTensor(popart.TensorInfo("FLOAT", [1]), "a") b = builder.addInputTensor(popart.TensorInfo("FLOAT", [1]), "b") # Trying to override the constructor chosen opset version. with pytest.raises(RuntimeError) as e_info: _ = builder.aiOnnxOpset8.add([a, b]) assert e_info.value.args[0].startswith( "Invalid opset 'aiOnnxOpset8' selected. Opset " "for domain ai.onnx already defined as 9" ) # Check that an error is thrown when trying to use two different opsets def test_opset_version_opset_choice_override(): builder = popart.Builder() a = builder.addInputTensor(popart.TensorInfo("FLOAT", [1]), "a") b = builder.addInputTensor(popart.TensorInfo("FLOAT", [1]), "b") _ = builder.aiOnnxOpset9.add([a, b]) # Trying to override the chosen opset. with pytest.raises(RuntimeError) as e_info: _ = builder.aiOnnxOpset8.add([a, b]) assert e_info.value.args[0].startswith( "Invalid opset 'aiOnnxOpset8' selected. Opset " "for domain ai.onnx already defined as 9" ) # Check that an error is thrown when trying to override the default opset def test_opset_version_opset_default_choice_override(): builder = popart.Builder() a = builder.addInputTensor(popart.TensorInfo("FLOAT", [1]), "a") b = builder.addInputTensor(popart.TensorInfo("FLOAT", [1]), "b") # This sets the opset strictly to the default opset. _ = builder.aiOnnx.add([a, b]) with pytest.raises(RuntimeError) as e_info: _ = builder.aiOnnxOpset8.add([a, b]) assert e_info.value.args[0].startswith( "Invalid opset 'aiOnnxOpset8' selected. Opset " "for domain ai.onnx already defined as 10" ) # Check that the default opset, `Builder.aiOnnx`, is set to # the explicitly chosen opset def test_opset_version_opset_default_version_changed(): builder = popart.Builder() a = builder.addInputTensor(popart.TensorInfo("FLOAT", [1]), "a") b = builder.addInputTensor(popart.TensorInfo("FLOAT", [1]), "b") _ = builder.aiOnnxOpset8.add([a, b]) # Check that builder.aiOnnx is now adding ops using opset 8. _ = builder.aiOnnx.add([a, b]) proto = builder.getModelProto() assert _get_onnx_opset_version(proto) == 8 # Test that when we construct an op with an invalid input tensor ID # we raise an error. def test_error_on_invalid_input_tensor(): with pytest.raises(popart.popart_exception) as e: b = popart.Builder() b.aiOnnx.relu(["blaaa"]) # should throw error # Check error mentions unknown tensor name. assert "'blaaa'" in str(e) # For some ops we need to specify the number of outputs. # This test checks that an error is thrown in case too many outputs are specified def test_too_many_outputs(): builder = popart.Builder() i = builder.addInputTensor(popart.TensorInfo("FLOAT", [1, 1, 1])) # rnn has only 2 outputs with pytest.raises(popart.popart_exception) as e: builder.aiOnnx.rnn([i, i, i], 3) assert "has an invalid number of outputs" in e.value.args[0]