[TOPI] Add generic batch norm (#9694)

* Add topi batch norm and tests

* Handle none values correctly

* Return correct nun outputs for onnx

* Use moving var/mean and update tests

* Add a test for batch norm folding

* Fix comment

* Format with black

* Re-order test args to match interface

* Call fold constant manually

Author: michalpiszczek

python/tvm/relay/frontend/onnx.py

@@ -474,7 +474,10 @@ def _impl_v1(cls, inputs, attr, params):
             op_name="batch_norm",
             ignores=["spatial", "is_test", "consumed_inputs", "momentum", "training_mode"],
         )(inputs, attr, params)
-        return out[0]
+        # We only support test mode, so we return data, moving_mean, moving_var,
+        # and then moving_mean and moving_var again as placeholders for
+        # the expected "saved_mean", "saved_var".
+        return _expr.TupleWrapper(_expr.Tuple((*out, out[1], out[2])), 5)
 
 
 class InstanceNorm(OnnxOpConverter):

python/tvm/relay/op/nn/_nn.py

@@ -152,6 +152,11 @@ def legalize_batch_matmul(attrs, inputs, types):
 reg.register_pattern("nn.batch_matmul", reg.OpPattern.OUT_ELEMWISE_FUSABLE)
 
 
+# batch_norm
+reg.register_strategy("nn.batch_norm", strategy.batch_norm_strategy)
+reg.register_pattern("nn.batch_norm", reg.OpPattern.OUT_ELEMWISE_FUSABLE)
+
+
 # sparse_dense
 @reg.register_compute("nn.sparse_dense")
 def compute_sparse_dense(attrs, inputs, out_type):

python/tvm/relay/op/strategy/generic.py

@@ -848,6 +848,29 @@ def batch_matmul_strategy(attrs, inputs, out_type, target):
     return strategy
 
 
+# batch_norm
+def wrap_compute_batch_norm(topi_compute):
+    """wrap batch_norm topi compute"""
+
+    def _compute_batch_norm(attrs, inputs, out_type):
+        return topi_compute(*inputs, attrs.axis, attrs.epsilon, attrs.center, attrs.scale)
+
+    return _compute_batch_norm
+
+
+@override_native_generic_func("batch_norm_strategy")
+def batch_norm_strategy(attrs, inputs, out_type, target):
+    """batch_norm generic strategy"""
+    logger.warning("batch_norm is not optimized for this platform.")
+    strategy = _op.OpStrategy()
+    strategy.add_implementation(
+        wrap_compute_batch_norm(topi.nn.batch_norm),
+        wrap_topi_schedule(topi.generic.schedule_batch_norm),
+        name="batch_norm.generic",
+    )
+    return strategy
+
+
 # sparse dense
 def wrap_compute_sparse_dense(topi_compute):
     """wrap sparse dense topi compute"""

python/tvm/topi/generic/nn.py

@@ -815,6 +815,23 @@ def schedule_batch_matmul(outs):
     return _default_schedule(outs, False)
 
 
+def schedule_batch_norm(outs):
+    """Schedule for batch_norm
+
+    Parameters
+    ----------
+    outs: Array of Tensor
+          The computation graph description of sparse_transpose
+          in the format of an array of tensors.
+
+    Returns
+    -------
+    sch: Schedule
+        The computation schedule for the op.
+    """
+    return _default_schedule(outs, False)
+
+
 def schedule_correlation_nchw(outs):
     """Schedule for correlation_nchw
 

python/tvm/topi/nn/__init__.py

@@ -42,6 +42,7 @@
 from .bitserial_conv2d import *
 from .bitserial_dense import *
 from .batch_matmul import *
+from .batch_norm import *
 from .sparse import *
 from .pad import *
 from .fifo_buffer import *

python/tvm/topi/nn/batch_norm.py

@@ -0,0 +1,110 @@
+# Licensed to the Apache Software Foundation (ASF) under one
+# or more contributor license agreements.  See the NOTICE file
+# distributed with this work for additional information
+# regarding copyright ownership.  The ASF licenses this file
+# to you 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.
+"""Batch normalization."""
+import typing
+
+from tvm import te
+from tvm import topi
+
+
+def batch_norm(
+    data: te.Tensor,
+    gamma: te.Tensor,
+    beta: te.Tensor,
+    moving_mean: te.Tensor,
+    moving_var: te.Tensor,
+    axis: typing.Optional[int] = None,
+    epsilon: typing.Optional[float] = None,
+    center: typing.Optional[bool] = None,
+    scale: typing.Optional[bool] = None,
+) -> typing.List[te.Tensor]:
+    """Batch normalization layer (Ioffe and Szegedy, 2014).
+
+    Normalizes the input at each batch, i.e. applies a transformation
+    that maintains the mean activation close to 0 and the activation
+    standard deviation close to 1.
+
+    Parameters
+    ----------
+    data : tvm.te.Tensor
+        Input to be batch-normalized.
+
+    gamma : tvm.te.Tensor
+        Scale factor to be applied to the normalized tensor.
+
+    beta : tvm.te.Tensor
+        Offset to be applied to the normalized tensor.
+
+    moving_mean : tvm.te.Tensor
+        Running mean of input.
+
+    moving_var : tvm.te.Tensor
+        Running variance of input.
+
+    axis : int, optional, default=1
+        Specify along which shape axis the normalization should occur.
+
+    epsilon : float, optional, default=1e-5
+        Small float added to variance to avoid dividing by zero.
+
+    center : bool, optional, default=True
+        If True, add offset of beta to normalized tensor, If False,
+        beta is ignored.
+
+    scale : bool, optional, defualt=True
+        If True, scale normalized tensor by gamma. If False, gamma
+        is ignored.
+
+    Returns
+    -------
+    output : list of tvm.te.Tensor
+        Normalized data with same shape as input
+
+    moving_mean : tvm.te.Tensor
+        Running mean of input.
+
+    moving_var : tvm.te.Tensor
+        Running variance of input.
+    """
+    if axis is None:
+        axis = 1
+
+    if epsilon is None:
+        epsilon = 1e-5
+
+    if center is None:
+        center = True
+
+    if scale is None:
+        scale = True
+
+    shape = [1] * len(data.shape)
+    shape[axis] = data.shape[axis]
+
+    moving_mean_rs = topi.reshape(moving_mean, shape)
+    moving_var_rs = topi.reshape(moving_var, shape)
+
+    out = (data - moving_mean_rs) / topi.math.sqrt(moving_var_rs + epsilon)
+
+    if scale:
+        out = out * topi.reshape(gamma, shape)
+    if center:
+        out = out + topi.reshape(beta, shape)
+
+    # Moving mean and var aren't updated during test. To avoid
+    # placeholder reuse, we multiply by 1 and return them.
+    return [out, moving_mean * 1, moving_var * 1]

python/tvm/topi/testing/__init__.py

@@ -49,6 +49,7 @@
 from .gather_nd_python import gather_nd_python
 from .strided_slice_python import strided_slice_python, strided_set_python
 from .batch_matmul import batch_matmul
+from .batch_norm import batch_norm
 from .slice_axis_python import slice_axis_python
 from .sequence_mask_python import sequence_mask
 from .poolnd_python import poolnd_python

python/tvm/topi/testing/batch_norm.py

@@ -0,0 +1,89 @@
+# Licensed to the Apache Software Foundation (ASF) under one
+# or more contributor license agreements.  See the NOTICE file
+# distributed with this work for additional information
+# regarding copyright ownership.  The ASF licenses this file
+# to you 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.
+"""Batch Normalization implemented in Numpy."""
+import numpy as np
+
+
+def batch_norm(
+    x: np.ndarray,
+    gamma: np.ndarray,
+    beta: np.ndarray,
+    moving_mean: np.ndarray,
+    moving_var: np.ndarray,
+    axis: int,
+    epsilon: float,
+    center: bool,
+    scale: bool,
+):
+    """Batch Normalization operator implemented in Numpy.
+
+    Parameters
+    ----------
+    data : np.ndarray
+        Input to be batch-normalized.
+
+    gamma : np.ndarray
+        Scale factor to be applied to the normalized tensor.
+
+    beta : np.ndarray
+        Offset to be applied to the normalized tensor.
+
+    moving_mean : np.ndarray
+        Running mean of input.
+
+    moving_var : np.ndarray
+        Running variance of input.
+
+    axis : int
+        Specify along which shape axis the normalization should occur.
+
+    epsilon : float
+        Small float added to variance to avoid dividing by zero.
+
+    center : bool
+        If True, add offset of beta to normalized tensor, If False,
+        beta is ignored.
+
+    scale : bool
+        If True, scale normalized tensor by gamma. If False, gamma
+        is ignored.
+
+    Returns
+    -------
+    output : np.ndarray
+        Normalized data with same shape as input
+
+    moving_mean : np.ndarray
+        Running mean of input.
+
+    moving_var : np.ndarray
+        Running variance of input.
+    """
+    shape = [1] * len(x.shape)
+    shape[axis] = x.shape[axis]
+
+    moving_mean_rs = moving_mean.reshape(shape)
+    moving_var_rs = moving_var.reshape(shape)
+
+    out = (x - moving_mean_rs) / np.sqrt(moving_var_rs + epsilon)
+
+    if scale:
+        out = out * gamma.reshape(shape)
+    if center:
+        out = out + beta.reshape(shape)
+
+    return [out, moving_mean, moving_var]

src/relay/op/nn/nn.cc

@@ -745,7 +745,8 @@ bool BatchNormRel(const Array<Type>& types, int num_inputs, const Attrs& attrs,
   reporter->Assign(types[4], TensorType({axis_size}, data->dtype));
 
   // output is a tuple of the normed data (same shape as input), new running mean,
-  // and new running average (the latter two are both vectors of length dim)
+  // new running variance, saved mean and saved variance (the latter are all
+  // vectors of length dim)
   std::vector<Type> fields;
   auto vec_ty = TensorType(Array<IndexExpr>({data->shape[axis]}), data->dtype);
   fields.push_back(TensorType(data->shape, data->dtype));

src/topi/schedule.cc

@@ -230,6 +230,9 @@ TVM_REGISTER_GENERIC_FUNC(schedule_dense)
 TVM_REGISTER_GENERIC_FUNC(schedule_batch_matmul)
     .set_default(WrapSchedule(topi::generic::default_schedule));
 
+TVM_REGISTER_GENERIC_FUNC(schedule_batch_norm)
+    .set_default(WrapSchedule(topi::generic::default_schedule));
+
 TVM_REGISTER_GENERIC_FUNC(schedule_pool)
     .set_default(WrapSchedule(topi::generic::default_schedule))
     .register_func({"cpu"}, WrapSchedule(topi::x86::default_schedule))