!3147 add RNNTLoss and RandomCategorical ops for aicpu

Merge pull request !3147 from yanzhenxiang2020/add_rnnt_cate_open.new

mindspore/ops/_grad/grad_nn_ops.py

@@ -567,6 +567,16 @@ def get_bprop_l2_loss(self):
     return bprop
 
 
+@bprop_getters.register(P.RNNTLoss)
+def get_bprop_rnnt_loss(self):
+    """Grad definition for `RNNTLoss` operation."""
+
+    def bprop(acts, labels, act_lens, label_lens, out, dout):
+        grad = out[1]
+        return grad, zeros_like(labels), zeros_like(act_lens), zeros_like(label_lens)
+    return bprop
+
+
 @bprop_getters.register(P.PReLU)
 def get_bprop_prelu(self):
     """Grad definition for `PReLU` operation."""

mindspore/ops/_op_impl/aicpu/__init__.py

@@ -30,3 +30,5 @@ from .ctcloss import _ctcloss_aicpu
 from .reverse_sequence import _reverse_sequence_aicpu
 from .crop_and_resize import _crop_and_resize_aicpu
 from .end_of_sequence import _end_of_sequence_aicpu
+from .rnnt_loss import _rnnt_loss_aicpu
+from .random_categorical import _random_categorical_aicpu

mindspore/ops/_op_impl/aicpu/random_categorical.py

+# Copyright 2020 Huawei Technologies Co., Ltd
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+# ============================================================================
+
+"""RandomCategorical op"""
+from mindspore.ops.op_info_register import op_info_register, AiCPURegOp, DataType
+
+random_categorical_op_info = AiCPURegOp("RandomCategorical") \
+    .fusion_type("OPAQUE") \
+    .input(0, "logits", "required") \
+    .input(1, "num_sample", "required") \
+    .input(2, "seed", "required") \
+    .output(0, "output", "required") \
+    .dtype_format(DataType.F16_Default, DataType.I32_Default, DataType.I32_Default, DataType.I16_Default) \
+    .dtype_format(DataType.F32_Default, DataType.I32_Default, DataType.I32_Default, DataType.I16_Default) \
+    .dtype_format(DataType.F64_Default, DataType.I32_Default, DataType.I32_Default, DataType.I16_Default) \
+    .dtype_format(DataType.F16_Default, DataType.I32_Default, DataType.I32_Default, DataType.I32_Default) \
+    .dtype_format(DataType.F32_Default, DataType.I32_Default, DataType.I32_Default, DataType.I32_Default) \
+    .dtype_format(DataType.F64_Default, DataType.I32_Default, DataType.I32_Default, DataType.I32_Default) \
+    .dtype_format(DataType.F16_Default, DataType.I32_Default, DataType.I32_Default, DataType.I64_Default) \
+    .dtype_format(DataType.F32_Default, DataType.I32_Default, DataType.I32_Default, DataType.I64_Default) \
+    .dtype_format(DataType.F64_Default, DataType.I32_Default, DataType.I32_Default, DataType.I64_Default) \
+    .dtype_format(DataType.F16_Default, DataType.I64_Default, DataType.I64_Default, DataType.I16_Default) \
+    .dtype_format(DataType.F32_Default, DataType.I64_Default, DataType.I64_Default, DataType.I16_Default) \
+    .dtype_format(DataType.F64_Default, DataType.I64_Default, DataType.I64_Default, DataType.I16_Default) \
+    .dtype_format(DataType.F16_Default, DataType.I64_Default, DataType.I64_Default, DataType.I32_Default) \
+    .dtype_format(DataType.F32_Default, DataType.I64_Default, DataType.I64_Default, DataType.I32_Default) \
+    .dtype_format(DataType.F64_Default, DataType.I64_Default, DataType.I64_Default, DataType.I32_Default) \
+    .dtype_format(DataType.F16_Default, DataType.I64_Default, DataType.I64_Default, DataType.I64_Default) \
+    .dtype_format(DataType.F32_Default, DataType.I64_Default, DataType.I64_Default, DataType.I64_Default) \
+    .dtype_format(DataType.F64_Default, DataType.I64_Default, DataType.I64_Default, DataType.I64_Default) \
+    .get_op_info()
+
+@op_info_register(random_categorical_op_info)
+def _random_categorical_aicpu():
+    """RandomCategorical AiCPU register"""
+    return

mindspore/ops/_op_impl/aicpu/rnnt_loss.py

+# Copyright 2020 Huawei Technologies Co., Ltd
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+# ============================================================================
+
+"""RNNTLoss op"""
+from mindspore.ops.op_info_register import op_info_register, AiCPURegOp, DataType
+
+rnnt_loss_op_info = AiCPURegOp("RNNTLoss") \
+    .fusion_type("OPAQUE") \
+    .input(0, "acts", "required") \
+    .input(1, "labels", "required") \
+    .input(2, "input_lengths", "required") \
+    .input(3, "label_lengths", "required") \
+    .output(0, "costs", "required") \
+    .output(1, "grads", "required") \
+    .attr("blank_label", "int") \
+    .dtype_format(DataType.F32_NCHW, DataType.I32_NCHW, DataType.I32_NCHW, DataType.I32_NCHW, DataType.F32_NCHW,
+                  DataType.F32_NCHW) \
+    .dtype_format(DataType.F32_Default, DataType.I32_Default, DataType.I32_Default, DataType.I32_Default,
+                  DataType.F32_Default, DataType.F32_Default) \
+    .get_op_info()
+
+@op_info_register(rnnt_loss_op_info)
+def _rnnt_loss_aicpu():
+    """RNNTLoss AiCPU register"""
+    return

mindspore/ops/operations/__init__.py

@@ -55,7 +55,7 @@ from .math_ops import (Abs, ACos, Asin, Asinh, AddN, AccumulateNV2, AssignAdd, A
                        Sin, Sqrt, Rsqrt, BesselI0e, BesselI1e, TruncateDiv, TruncateMod,
                        Square, Sub, TensorAdd, Sign, Round, SquareSumAll, Atan, Atanh, Cosh, Sinh, Eps, Tan)
 
-from .random_ops import (RandomChoiceWithMask, Normal)
+from .random_ops import (RandomChoiceWithMask, Normal, RandomCategorical)
 from .nn_ops import (LSTM, SGD, Adam, SparseApplyAdam, SparseApplyLazyAdam, ApplyMomentum, BatchNorm,
                      BiasAdd, Conv2D,
                      DepthwiseConv2dNative,
@@ -70,6 +70,7 @@ from .nn_ops import (LSTM, SGD, Adam, SparseApplyAdam, SparseApplyLazyAdam, Appl
                      ResizeBilinear, Sigmoid,
                      SigmoidCrossEntropyWithLogits,
                      SmoothL1Loss, Softmax, Softsign, Softplus, LRN,
+                     RNNTLoss,
                      SoftmaxCrossEntropyWithLogits, ROIAlign,
                      SparseSoftmaxCrossEntropyWithLogits, Tanh,
                      TopK, BinaryCrossEntropy, SparseApplyAdagrad, LARSUpdate, ApplyFtrl, SparseApplyFtrl,
@@ -171,6 +172,7 @@ __all__ = [
     'Tanh',
     'RandomChoiceWithMask',
     'Normal',
+    'RandomCategorical',
     'ResizeBilinear',
     'ScalarSummary',
     'ImageSummary',
@@ -202,6 +204,7 @@ __all__ = [
     'SmoothL1Loss',
     'L2Loss',
     'CTCLoss',
+    'RNNTLoss',
     'ReduceAll',
     'ScalarToArray',
     'ScalarToTensor',

mindspore/ops/operations/nn_ops.py

@@ -1736,6 +1736,62 @@ class DataFormatDimMap(PrimitiveWithInfer):
         return x_type
 
 
+class RNNTLoss(PrimitiveWithInfer):
+    """
+    Computes the RNNTLoss and its gradient with respect to the softmax outputs.
+
+    Args:
+        blank_label (int): blank label. Default: 0.
+
+    Inputs:
+        - **acts** (Tensor[float32]) - Tensor of shape :math:`(B, T, U, V)`.
+        - **labels** (Tensor[int32]) - Tensor of shape :math:`(B, U-1)`.
+        - **input_lengths** (Tensor[int32]) - Tensor of shape :math:`(B,)`.
+        - **label_lebgths** (Tensor[int32]) - Tensor of shape :math:`(B,)`.
+
+    Outputs:
+        - **costs** (Tensor[int32]) - Tensor of shape :math:`(B,)`.
+        - **grads** (Tensor[int32]) - Has the same shape as `acts`.
+
+    Examples:
+        >>> B, T, U, V = 1, 2, 3, 5
+        >>> acts = np.random.random((B, T, U, V)).astype(np.float32)
+        >>> labels = np.array([[1, 2]]).astype(np.int32)
+        >>> input_length = np.array([T] * B).astype(np.int32)
+        >>> label_length = np.array([len(l) for l in labels]).astype(np.int32)
+        >>> rnnt_loss = P.RNNTLoss(blank_label=blank)
+        >>> costs, grads = rnnt_loss(Tensor(acts), Tensor(labels), Tensor(input_length), Tensor(label_length))
+    """
+    @prim_attr_register
+    def __init__(self, blank_label=0):
+        validator.check_value_type('blank_label', blank_label, [int], self.name)
+        self.init_prim_io_names(inputs=['acts', 'labels', 'input_length', 'label_length'],
+                                outputs=['costs', 'grads'])
+
+    def infer_shape(self, acts_shape, labels_shape, input_length_shape, label_length_shape):
+        validator.check_integer('acts_rank', len(acts_shape), 4, Rel.EQ, self.name)
+        validator.check_integer('labels_rank', len(labels_shape), 2, Rel.EQ, self.name)
+        validator.check_integer('input_length_rank', len(input_length_shape), 1, Rel.EQ, self.name)
+        validator.check_integer('label_length_rank', len(label_length_shape), 1, Rel.EQ, self.name)
+        validator.check('labels shape[0]', labels_shape[0], 'acts shape[0]', acts_shape[0], Rel.EQ, self.name)
+        validator.check('labels shape[1]', labels_shape[1], 'acts shape[2]-1', acts_shape[2]-1, Rel.EQ, self.name)
+        validator.check('input_length size', input_length_shape[0], 'acts shape[0]', acts_shape[0], Rel.EQ, self.name)
+        validator.check('label_length size', label_length_shape[0], 'acts shape[0]', acts_shape[0], Rel.EQ, self.name)
+        costs_shape = (acts_shape[0],)
+        return (costs_shape, acts_shape)
+
+    def infer_dtype(self, acts_type, labels_type, input_length_type, label_length_type):
+        validator.check_subclass("acts_type", acts_type, mstype.tensor, self.name)
+        validator.check_subclass("labels_type", labels_type, mstype.tensor, self.name)
+        validator.check_subclass("input_length_type", input_length_type, mstype.tensor, self.name)
+        validator.check_subclass("label_length_type", label_length_type, mstype.tensor, self.name)
+        validator.check_tensor_type_same({"acts_type": acts_type}, [mstype.float32], self.name)
+        validator.check_tensor_type_same({"labels_type": labels_type}, [mstype.int32], self.name)
+        validator.check_tensor_type_same({"input_length_type": input_length_type}, [mstype.int32], self.name)
+        validator.check_tensor_type_same({"label_length_type": label_length_type}, [mstype.int32], self.name)
+        return (acts_type, acts_type)
+
+
 class SGD(PrimitiveWithInfer):
     """
     Computes stochastic gradient descent (optionally with momentum).

mindspore/ops/operations/random_ops.py

@@ -108,3 +108,61 @@ class Normal(PrimitiveWithInfer):
                "dtype": mstype.float32,
                "value": None}
         return out
+
+
+class RandomCategorical(PrimitiveWithInfer):
+    """
+    Generates random samples from a given categorical distribution tensor.
+
+    Args:
+        dtype (mindspore.dtype): The type of output. Its value should be one of [mindspore.int16,
+            mindspore.int32, mindspore.int64]. Default: mindspore.int64.
+
+    Inputs:
+        - **logits** (Tensor) - The input tensor. 2-D Tensor with shape [batch_size, num_classes].
+        - **num_sample** (int) - Number of sample to be drawn. Only constant values is allowed.
+        - **seed** (int) - Random seed. Default: 0. Only constant values is allowed.
+
+    Outputs:
+        - **output** (Tensor) - The output Tensor with shape [batch_size, num_samples].
+
+    Examples:
+        >>> class Net(nn.Cell):
+        >>>   def __init__(self, num_sample):
+        >>>     super(Net, self).__init__()
+        >>>     self.random_categorical = P.RandomCategorical(mindspore.int64)
+        >>>     self.num_sample = num_sample
+        >>>   def construct(self, logits, seed=0):
+        >>>     return self.random_categorical(logits, self.num_sample, seed)
+        >>>
+        >>> x = np.random.random((10, 5)).astype(np.float32)
+        >>> net = Net(8)
+        >>> output = net(Tensor(x))
+    """
+    @prim_attr_register
+    def __init__(self, dtype=mstype.int64):
+        """Init RandomCategorical"""
+        self.dtype = dtype
+
+        valid_values = (mstype.int32, mstype.int16, mstype.int64)
+        validator.check_type_name("dtype", dtype, valid_values, self.name)
+        self.init_prim_io_names(inputs=['logits', 'num_samples', 'seed'],
+                                outputs=['output'])
+
+    def __infer__(self, logits, num_samples, seed):
+        logits_dtype = logits['dtype']
+        valid_types = (mstype.float32, mstype.float16, mstype.float64)
+        validator.check_tensor_type_same({'logits': logits_dtype}, valid_types, self.name)
+        num_samples_v = num_samples['value']
+        seed_v = seed['value']
+        validator.check_value_type('num_samples', num_samples_v, (int,), self.name)
+        validator.check_value_type('seed', seed_v, (int,), self.name)
+        validator.check_integer("num_samples", num_samples_v, 0, Rel.GT, self.name)
+        x_shape = list(logits['shape'])
+        if len(x_shape) != 2:
+            raise ValueError("RandomCategorical shape should be 2-dimension.")
+        ndim = len(x_shape) - 1
+        x_shape[ndim] = num_samples_v
+        return {'shape': (x_shape),
+                'dtype': (self.dtype),
+                'value': None}

tests/st/ops/ascend/test_aicpu_ops/test_random_categorical.py

+# Copyright 2020 Huawei Technologies Co., Ltd
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+# ============================================================================
+import numpy as np
+import mindspore
+import mindspore.nn as nn
+import mindspore.context as context
+from mindspore import Tensor
+from mindspore.ops import operations as P
+
+context.set_context(mode=context.GRAPH_MODE, device_target="Ascend")
+class Net(nn.Cell):
+    def __init__(self, num_sample):
+        super(Net, self).__init__()
+        self.random_categorical = P.RandomCategorical(mindspore.int64)
+        self.num_sample = num_sample
+
+    def construct(self, logits, seed=0):
+        return self.random_categorical(logits, self.num_sample, seed)
+
+def test_net():
+    x = np.random.random((10, 5)).astype(np.float32)
+    net = Net(8)
+    output = net(Tensor(x))
+    print(x)
+    print(output.asnumpy())
+    #print(output.dtype())

tests/st/ops/ascend/test_aicpu_ops/test_rnnt_loss.py

+# Copyright 2020 Huawei Technologies Co., Ltd
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+# ============================================================================
+import numpy as np
+import mindspore.nn as nn
+import mindspore.context as context
+from mindspore import Tensor
+from mindspore.ops import operations as P
+
+context.set_context(mode=context.GRAPH_MODE, device_target="Ascend")
+class Net(nn.Cell):
+    def __init__(self):
+        super(Net, self).__init__()
+        self.rnnt_loss = P.RNNTLoss(blank_label=0)
+
+    def construct(self, acts, labels, act_lens, label_lens):
+        return self.rnnt_loss(acts, labels, act_lens, label_lens)
+
+
+def test_net():
+    B, T, U, V = 1, 2, 3, 5
+    acts = np.random.random((B, T, U, V)).astype(np.float32)
+    labels = np.array([[np.random.randint(1, V-1) for _ in range(U-1)]]).astype(np.int32)
+    input_length = np.array([T] * B).astype(np.int32)
+    label_length = np.array([len(l) for l in labels]).astype(np.int32)
+    rnnt_loss = Net()
+    costs, grads = rnnt_loss(Tensor(acts), Tensor(labels), Tensor(input_length), Tensor(label_length))
+    print(Tensor(acts), Tensor(labels), Tensor(input_length), Tensor(label_length))
+    print(costs.asnumpy())
+    print(grads.asnumpy())