From f11abf26b6dd84a8cc325da96d8528c48659c509 Mon Sep 17 00:00:00 2001 From: Siju Samuel Date: Tue, 4 Aug 2020 19:18:05 +0530 Subject: [PATCH] [COREML]Unary ops support added in frontend (#6196) * [COREML]Unary ops support added in frontend * Used coreml enums --- python/tvm/relay/frontend/coreml.py | 30 ++++- tests/python/frontend/coreml/test_forward.py | 189 +++++++++++++++++++++++++++ 2 files changed, 218 insertions(+), 1 deletion(-) diff --git a/python/tvm/relay/frontend/coreml.py b/python/tvm/relay/frontend/coreml.py index 0027c7f..6e1fce9 100644 --- a/python/tvm/relay/frontend/coreml.py +++ b/python/tvm/relay/frontend/coreml.py @@ -14,7 +14,7 @@ # KIND, either express or implied. See the License for the # specific language governing permissions and limitations # under the License. -# pylint: disable=invalid-name, import-self, unused-argument, unused-variable +# pylint: disable=invalid-name, import-self, unused-argument, unused-variable, no-else-return # pylint: disable=inconsistent-return-statements, import-outside-toplevel """CoreML frontend.""" import math @@ -350,6 +350,33 @@ def _MinLayerParams(op, inexpr, etab): return _min +def _UnaryFunctionLayerParams(op, inexpr, etab): + op_type = op.type + if op_type == op.SQRT: + return _op.sqrt(inexpr) + elif op_type == op.RSQRT: + epsilon = _expr.const(op.epsilon) + return _op.rsqrt(inexpr + epsilon) + elif op_type == op.INVERSE: + epsilon = _expr.const(op.epsilon) + return _expr.const(1.0) / (inexpr + epsilon) + elif op_type == op.POWER: + alpha = _expr.const(op.alpha) + return _op.power(inexpr, alpha) + elif op_type == op.EXP: + return _op.exp(inexpr) + elif op_type == op.LOG: + return _op.log(inexpr) + elif op_type == op.ABS: + return _op.abs(inexpr) + elif op_type == op.THRESHOLD: + alpha = _expr.const(op.alpha) + return _op.maximum(inexpr, alpha) + else: + msg = 'Unary Op type value {} is not supported in frontend CoreML.' + raise tvm.error.OpAttributeUnImplemented(msg.format(op_type)) + + _convert_map = { 'NeuralNetworkMeanImage': _NeuralNetworkMeanImage, 'NeuralNetworkImageScaler': _NeuralNetworkImageScaler, @@ -372,6 +399,7 @@ _convert_map = { 'AverageLayerParams': _AverageLayerParams, 'MaxLayerParams': _MaxLayerParams, 'MinLayerParams': _MinLayerParams, + 'UnaryFunctionLayerParams': _UnaryFunctionLayerParams, } # SAME padding: https://www.tensorflow.org/api_guides/python/nn diff --git a/tests/python/frontend/coreml/test_forward.py b/tests/python/frontend/coreml/test_forward.py index 6f058f4..49f0987 100644 --- a/tests/python/frontend/coreml/test_forward.py +++ b/tests/python/frontend/coreml/test_forward.py @@ -337,6 +337,194 @@ def test_forward_min(): verify_min((1, 3, 20, 20)) verify_min((20, 20)) + +def verify_unary_sqrt(input_dim): + dtype = 'float32' + + a_np = np.random.uniform(size=input_dim).astype(dtype) + ref_val = np.sqrt(a_np) + + inputs = [('input', datatypes.Array(*input_dim))] + output = [('output', datatypes.Array(*ref_val.shape))] + builder = NeuralNetworkBuilder(inputs, output) + builder.add_unary(name="sqrt", + input_name='input', + output_name='output', + mode='sqrt') + + model = cm.models.MLModel(builder.spec) + for target, ctx in ctx_list(): + out = run_tvm_graph(model, target, ctx, [a_np], + ['input'], ref_val.shape, dtype) + tvm.testing.assert_allclose(out, ref_val, rtol=1e-5) + + +def verify_unary_rsqrt(input_dim, epsilon=0): + dtype = 'float32' + + a_np = np.random.uniform(size=input_dim).astype(dtype) + ref_val = 1 / np.sqrt(a_np + epsilon) + + inputs = [('input', datatypes.Array(*input_dim))] + output = [('output', datatypes.Array(*ref_val.shape))] + builder = NeuralNetworkBuilder(inputs, output) + builder.add_unary(name="rsqrt", + input_name='input', + output_name='output', + mode='rsqrt', + epsilon=epsilon) + + model = cm.models.MLModel(builder.spec) + for target, ctx in ctx_list(): + out = run_tvm_graph(model, target, ctx, [a_np], + ['input'], ref_val.shape, dtype) + tvm.testing.assert_allclose(out, ref_val, rtol=1e-5) + + +def verify_unary_inverse(input_dim, epsilon=0): + dtype = 'float32' + + a_np = np.random.uniform(size=input_dim).astype(dtype) + ref_val = 1 / (a_np + epsilon) + + inputs = [('input', datatypes.Array(*input_dim))] + output = [('output', datatypes.Array(*ref_val.shape))] + builder = NeuralNetworkBuilder(inputs, output) + builder.add_unary(name="inverse", + input_name='input', + output_name='output', + mode='inverse', + epsilon=epsilon) + + model = cm.models.MLModel(builder.spec) + for target, ctx in ctx_list(): + out = run_tvm_graph(model, target, ctx, [a_np], + ['input'], ref_val.shape, dtype) + tvm.testing.assert_allclose(out, ref_val, rtol=1e-5) + + +def verify_unary_power(input_dim, alpha): + dtype = 'float32' + + a_np = np.random.uniform(size=input_dim).astype(dtype) + ref_val = np.power(a_np, alpha) + + inputs = [('input', datatypes.Array(*input_dim))] + output = [('output', datatypes.Array(*ref_val.shape))] + builder = NeuralNetworkBuilder(inputs, output) + builder.add_unary(name="power", + input_name='input', + output_name='output', + mode='power', + alpha=alpha) + + model = cm.models.MLModel(builder.spec) + for target, ctx in ctx_list(): + out = run_tvm_graph(model, target, ctx, [a_np], + ['input'], ref_val.shape, dtype) + tvm.testing.assert_allclose(out, ref_val, rtol=1e-5) + + +def verify_unary_exp(input_dim): + dtype = 'float32' + + a_np = np.random.uniform(size=input_dim).astype(dtype) + ref_val = np.exp(a_np) + + inputs = [('input', datatypes.Array(*input_dim))] + output = [('output', datatypes.Array(*ref_val.shape))] + builder = NeuralNetworkBuilder(inputs, output) + builder.add_unary(name="exp", + input_name='input', + output_name='output', + mode='exp') + + model = cm.models.MLModel(builder.spec) + for target, ctx in ctx_list(): + out = run_tvm_graph(model, target, ctx, [a_np], + ['input'], ref_val.shape, dtype) + tvm.testing.assert_allclose(out, ref_val, rtol=1e-5) + + +def verify_unary_log(input_dim): + dtype = 'float32' + + a_np = np.random.uniform(size=input_dim).astype(dtype) + ref_val = np.log(a_np) + + inputs = [('input', datatypes.Array(*input_dim))] + output = [('output', datatypes.Array(*ref_val.shape))] + builder = NeuralNetworkBuilder(inputs, output) + builder.add_unary(name="log", + input_name='input', + output_name='output', + mode='log') + + model = cm.models.MLModel(builder.spec) + for target, ctx in ctx_list(): + out = run_tvm_graph(model, target, ctx, [a_np], + ['input'], ref_val.shape, dtype) + tvm.testing.assert_allclose(out, ref_val, rtol=1e-5) + + +def verify_unary_abs(input_dim): + dtype = 'float32' + + a_np = np.random.uniform(-100.0, 100.0, size=input_dim).astype(dtype) + ref_val = np.abs(a_np) + + inputs = [('input', datatypes.Array(*input_dim))] + output = [('output', datatypes.Array(*ref_val.shape))] + builder = NeuralNetworkBuilder(inputs, output) + builder.add_unary(name="abs", + input_name='input', + output_name='output', + mode='abs') + + model = cm.models.MLModel(builder.spec) + for target, ctx in ctx_list(): + out = run_tvm_graph(model, target, ctx, [a_np], + ['input'], ref_val.shape, dtype) + tvm.testing.assert_allclose(out, ref_val, rtol=1e-5) + + +def verify_unary_threshold(input_dim, alpha): + dtype = 'float32' + + a_np = np.random.uniform(-100.0, 100.0, size=input_dim).astype(dtype) + ref_val = np.maximum(a_np, alpha) + + inputs = [('input', datatypes.Array(*input_dim))] + output = [('output', datatypes.Array(*ref_val.shape))] + builder = NeuralNetworkBuilder(inputs, output) + builder.add_unary(name="threshold", + input_name='input', + output_name='output', + mode='threshold', + alpha=alpha) + + model = cm.models.MLModel(builder.spec) + for target, ctx in ctx_list(): + out = run_tvm_graph(model, target, ctx, [a_np], + ['input'], ref_val.shape, dtype) + tvm.testing.assert_allclose(out, ref_val, rtol=1e-5) + + +def test_forward_unary(): + verify_unary_sqrt((1, 3, 20, 20)) + verify_unary_rsqrt((1, 3, 20, 20)) + verify_unary_rsqrt((1, 3, 20, 20), epsilon=1e-6) + verify_unary_inverse((1, 3, 20, 20)) + verify_unary_inverse((1, 3, 20, 20), epsilon=1e-6) + verify_unary_power((1, 3, 20, 20), alpha=0.5) + verify_unary_power((1, 3, 20, 20), alpha=4) + verify_unary_exp((1, 3, 20, 20)) + verify_unary_log((1, 3, 20, 20)) + verify_unary_abs((1, 3, 20, 20)) + verify_unary_threshold((1, 3, 20, 20), alpha=-6.0) + verify_unary_threshold((1, 3, 20, 20), alpha=5.0) + + def verify_image_scaler(input_dim, blue_bias=0.0, green_bias=0.0, red_bias=0.0, image_scale=1.0): dtype = 'float32' a_np = np.random.uniform(size=input_dim).astype(dtype) @@ -413,6 +601,7 @@ if __name__ == '__main__': test_forward_average() test_forward_max() test_forward_min() + test_forward_unary() test_mobilenet_checkonly() test_resnet50_checkonly() test_forward_image_scaler() -- 2.7.4