back-to-fusedbatchnorm-operation-in-pynative-mode

mindspore/ccsrc/pre_activate/ascend/ascend_backend_optimization.cc

@@ -238,11 +238,16 @@ void AscendBackendIRFusionOptimization(const std::shared_ptr<session::KernelGrap
   }
   auto optimizer = std::make_shared<GraphOptimizer>();
   auto ir_fusion_pm = std::make_shared<PassManager>("ir_fusion_pm");
-  ir_fusion_pm->AddPass(std::make_shared<BatchNormGradSplit>());
-  ir_fusion_pm->AddPass(std::make_shared<LayerNormGradSplit>());
-  ir_fusion_pm->AddPass(std::make_shared<FusedBatchNormFusion>());
-  ir_fusion_pm->AddPass(std::make_shared<FusedBatchNormMixPrecisionFusion0>());
-  ir_fusion_pm->AddPass(std::make_shared<FusedBatchNormMixPrecisionFusion1>());
+  if (context_ptr->execution_mode() == kPynativeMode) {
+    ir_fusion_pm->AddPass(std::make_shared<BnSplit>());
+    ir_fusion_pm->AddPass(std::make_shared<BnGradSplit>());
+  } else {
+    ir_fusion_pm->AddPass(std::make_shared<BatchNormGradSplit>());
+    ir_fusion_pm->AddPass(std::make_shared<LayerNormGradSplit>());
+    ir_fusion_pm->AddPass(std::make_shared<FusedBatchNormFusion>());
+    ir_fusion_pm->AddPass(std::make_shared<FusedBatchNormMixPrecisionFusion0>());
+    ir_fusion_pm->AddPass(std::make_shared<FusedBatchNormMixPrecisionFusion1>());
+  }
   ir_fusion_pm->AddPass(std::make_shared<InsertPadForNMSWithMask>());
   if (context_ptr->ir_fusion_flag()) {
     AddAscendBackendOptionalIRFusion(ir_fusion_pm.get());
@@ -282,11 +287,8 @@ void RunOpAscendBackendIRFusionOptimization(const std::shared_ptr<session::Kerne
   }
   auto optimizer = std::make_shared<GraphOptimizer>();
   auto ir_fusion_pm = std::make_shared<PassManager>("ir_fusion_pm");
-  ir_fusion_pm->AddPass(std::make_shared<BatchNormGradSplit>());
+  ir_fusion_pm->AddPass(std::make_shared<BnSplit>());
   ir_fusion_pm->AddPass(std::make_shared<LayerNormGradSplit>());
-  ir_fusion_pm->AddPass(std::make_shared<FusedBatchNormFusion>());
-  ir_fusion_pm->AddPass(std::make_shared<FusedBatchNormMixPrecisionFusion0>());
-  ir_fusion_pm->AddPass(std::make_shared<FusedBatchNormMixPrecisionFusion1>());
   ir_fusion_pm->AddPass(std::make_shared<TopKSplit>());
   ir_fusion_pm->AddPass(std::make_shared<AddnFission>());
   ir_fusion_pm->AddPass(std::make_shared<InsertPadForNMSWithMask>());

mindspore/nn/layer/normalization.py

@@ -84,13 +84,14 @@ class _BatchNorm(Cell):
         self.dtype = P.DType()
         self.reshape = P.Reshape()
         self.is_ascend = context.get_context("device_target") == "Ascend"
+        self.is_graph_mode = context.get_context("mode") == context.GRAPH_MODE
         self.momentum = 1.0 - momentum
         if context.get_context("enable_ge"):
             self.is_ge_backend = True
         else:
             self.is_ge_backend = False
 
-        if self.is_ge_backend or self.is_ascend:
+        if self.is_graph_mode and (self.is_ge_backend or self.is_ascend):
             self.bn_train = P.BatchNorm(is_training=True,
                                         epsilon=self.eps)
         else:
@@ -152,7 +153,7 @@ class _BatchNorm(Cell):
             if self.is_ge_backend and self.is_global:
                 axes, re_shape = _shape_infer(F.shape(x), self.num_features)
                 y = self._global_sync(x, axes, re_shape)
-            elif self.is_ge_backend or self.is_ascend:
+            elif self.is_graph_mode and (self.is_ge_backend or self.is_ascend):
                 if self.is_global:
                     axes, re_shape = _shape_infer(F.shape(x), self.num_features)
                     y = self._global_sync(x, axes, re_shape)

tests/st/pynative/test_ascend_lenet.py → tests/st/pynative/test_pynative_lenet.py

@@ -157,4 +157,5 @@ def test_ascend_pynative_lenet():
         total_time = total_time + cost_time
 
         print("======epoch: ", epoch, " loss: ", loss_output.asnumpy(), " cost time: ", cost_time)
-    assert loss_output.asnumpy() < 0.1
+    assert loss_output.asnumpy() < 0.004
+    assert loss_output.asnumpy() > 0.003

tests/st/pynative/test_pynative_resnet50.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 time
+import random
+import numpy as np
+import pytest
+
+import mindspore.common.dtype as mstype
+import mindspore.dataset as ds
+import mindspore.dataset.transforms.c_transforms as C
+import mindspore.dataset.transforms.vision.c_transforms as vision
+import mindspore.nn as nn
+import mindspore.ops.functional as F
+
+from mindspore import Tensor
+from mindspore import context
+from mindspore import ParameterTuple
+from mindspore.nn import Cell
+from mindspore.ops import operations as P
+from mindspore.ops import composite as CP
+from mindspore.nn.optim.momentum import Momentum
+from mindspore.common.initializer import initializer
+from mindspore.nn.wrap.cell_wrapper import WithLossCell
+
+random.seed(1)
+np.random.seed(1)
+ds.config.set_seed(1)
+
+
+def weight_variable(shape):
+    return initializer('XavierUniform', shape=shape, dtype=mstype.float32)
+
+
+def weight_variable_uniform(shape):
+    return initializer('Uniform', shape=shape, dtype=mstype.float32)
+
+
+def weight_variable_0(shape):
+    zeros = np.zeros(shape).astype(np.float32)
+    return Tensor(zeros)
+
+
+def weight_variable_1(shape):
+    ones = np.ones(shape).astype(np.float32)
+    return Tensor(ones)
+
+
+def conv3x3(in_channels, out_channels, stride=1, padding=0):
+    """3x3 convolution """
+    weight_shape = (out_channels, in_channels, 3, 3)
+    weight = weight_variable(weight_shape)
+    return nn.Conv2d(in_channels, out_channels,
+                     kernel_size=3, stride=stride, padding=padding, weight_init=weight, has_bias=False, pad_mode="same")
+
+
+def conv1x1(in_channels, out_channels, stride=1, padding=0):
+    """1x1 convolution"""
+    weight_shape = (out_channels, in_channels, 1, 1)
+    weight = weight_variable(weight_shape)
+    return nn.Conv2d(in_channels, out_channels,
+                     kernel_size=1, stride=stride, padding=padding, weight_init=weight, has_bias=False, pad_mode="same")
+
+
+def conv7x7(in_channels, out_channels, stride=1, padding=0):
+    """1x1 convolution"""
+    weight_shape = (out_channels, in_channels, 7, 7)
+    weight = weight_variable(weight_shape)
+    return nn.Conv2d(in_channels, out_channels,
+                     kernel_size=7, stride=stride, padding=padding, weight_init=weight, has_bias=False, pad_mode="same")
+
+
+def bn_with_initialize(out_channels):
+    shape = (out_channels)
+    mean = weight_variable_0(shape)
+    var = weight_variable_1(shape)
+    beta = weight_variable_0(shape)
+    gamma = weight_variable_uniform(shape)
+    bn = nn.BatchNorm2d(out_channels, momentum=0.99, eps=0.00001, gamma_init=gamma,
+                        beta_init=beta, moving_mean_init=mean, moving_var_init=var)
+    return bn
+
+
+def bn_with_initialize_last(out_channels):
+    shape = (out_channels)
+    mean = weight_variable_0(shape)
+    var = weight_variable_1(shape)
+    beta = weight_variable_0(shape)
+    gamma = weight_variable_uniform(shape)
+    bn = nn.BatchNorm2d(out_channels, momentum=0.99, eps=0.00001, gamma_init=gamma,
+                        beta_init=beta, moving_mean_init=mean, moving_var_init=var)
+    return bn
+
+
+def fc_with_initialize(input_channels, out_channels):
+    weight_shape = (out_channels, input_channels)
+    weight = weight_variable(weight_shape)
+    bias_shape = (out_channels)
+    bias = weight_variable_uniform(bias_shape)
+    return nn.Dense(input_channels, out_channels, weight, bias)
+
+
+class ResidualBlock(nn.Cell):
+    expansion = 4
+
+    def __init__(self,
+                 in_channels,
+                 out_channels,
+                 stride=1):
+        super(ResidualBlock, self).__init__()
+
+        out_chls = out_channels // self.expansion
+        self.conv1 = conv1x1(in_channels, out_chls, stride=stride, padding=0)
+        self.bn1 = bn_with_initialize(out_chls)
+
+        self.conv2 = conv3x3(out_chls, out_chls, stride=1, padding=0)
+        self.bn2 = bn_with_initialize(out_chls)
+
+        self.conv3 = conv1x1(out_chls, out_channels, stride=1, padding=0)
+        self.bn3 = bn_with_initialize_last(out_channels)
+
+        self.relu = P.ReLU()
+        self.add = P.TensorAdd()
+
+    def construct(self, x):
+        identity = x
+
+        out = self.conv1(x)
+        out = self.bn1(out)
+        out = self.relu(out)
+
+        out = self.conv2(out)
+        out = self.bn2(out)
+        out = self.relu(out)
+
+        out = self.conv3(out)
+        out = self.bn3(out)
+
+        out = self.add(out, identity)
+        out = self.relu(out)
+
+        return out
+
+
+class ResidualBlockWithDown(nn.Cell):
+    expansion = 4
+
+    def __init__(self,
+                 in_channels,
+                 out_channels,
+                 stride=1,
+                 down_sample=False):
+        super(ResidualBlockWithDown, self).__init__()
+
+        out_chls = out_channels // self.expansion
+        self.conv1 = conv1x1(in_channels, out_chls, stride=stride, padding=0)
+        self.bn1 = bn_with_initialize(out_chls)
+
+        self.conv2 = conv3x3(out_chls, out_chls, stride=1, padding=0)
+        self.bn2 = bn_with_initialize(out_chls)
+
+        self.conv3 = conv1x1(out_chls, out_channels, stride=1, padding=0)
+        self.bn3 = bn_with_initialize_last(out_channels)
+
+        self.relu = P.ReLU()
+        self.downSample = down_sample
+
+        self.conv_down_sample = conv1x1(in_channels, out_channels, stride=stride, padding=0)
+        self.bn_down_sample = bn_with_initialize(out_channels)
+        self.add = P.TensorAdd()
+
+    def construct(self, x):
+        identity = x
+
+        out = self.conv1(x)
+        out = self.bn1(out)
+        out = self.relu(out)
+
+        out = self.conv2(out)
+        out = self.bn2(out)
+        out = self.relu(out)
+
+        out = self.conv3(out)
+        out = self.bn3(out)
+
+        identity = self.conv_down_sample(identity)
+        identity = self.bn_down_sample(identity)
+
+        out = self.add(out, identity)
+        out = self.relu(out)
+
+        return out
+
+
+class MakeLayer0(nn.Cell):
+
+    def __init__(self, block, in_channels, out_channels, stride):
+        super(MakeLayer0, self).__init__()
+        self.a = ResidualBlockWithDown(in_channels, out_channels, stride=1, down_sample=True)
+        self.b = block(out_channels, out_channels, stride=stride)
+        self.c = block(out_channels, out_channels, stride=1)
+
+    def construct(self, x):
+        x = self.a(x)
+        x = self.b(x)
+        x = self.c(x)
+
+        return x
+
+
+class MakeLayer1(nn.Cell):
+
+    def __init__(self, block, in_channels, out_channels, stride):
+        super(MakeLayer1, self).__init__()
+        self.a = ResidualBlockWithDown(in_channels, out_channels, stride=stride, down_sample=True)
+        self.b = block(out_channels, out_channels, stride=1)
+        self.c = block(out_channels, out_channels, stride=1)
+        self.d = block(out_channels, out_channels, stride=1)
+
+    def construct(self, x):
+        x = self.a(x)
+        x = self.b(x)
+        x = self.c(x)
+        x = self.d(x)
+
+        return x
+
+
+class MakeLayer2(nn.Cell):
+
+    def __init__(self, block, in_channels, out_channels, stride):
+        super(MakeLayer2, self).__init__()
+        self.a = ResidualBlockWithDown(in_channels, out_channels, stride=stride, down_sample=True)
+        self.b = block(out_channels, out_channels, stride=1)
+        self.c = block(out_channels, out_channels, stride=1)
+        self.d = block(out_channels, out_channels, stride=1)
+        self.e = block(out_channels, out_channels, stride=1)
+        self.f = block(out_channels, out_channels, stride=1)
+
+    def construct(self, x):
+        x = self.a(x)
+        x = self.b(x)
+        x = self.c(x)
+        x = self.d(x)
+        x = self.e(x)
+        x = self.f(x)
+
+        return x
+
+
+class MakeLayer3(nn.Cell):
+
+    def __init__(self, block, in_channels, out_channels, stride):
+        super(MakeLayer3, self).__init__()
+        self.a = ResidualBlockWithDown(in_channels, out_channels, stride=stride, down_sample=True)
+        self.b = block(out_channels, out_channels, stride=1)
+        self.c = block(out_channels, out_channels, stride=1)
+
+    def construct(self, x):
+        x = self.a(x)
+        x = self.b(x)
+        x = self.c(x)
+
+        return x
+
+
+class ResNet(nn.Cell):
+
+    def __init__(self, block, num_classes=100, batch_size=32):
+        super(ResNet, self).__init__()
+        self.batch_size = batch_size
+        self.num_classes = num_classes
+
+        self.conv1 = conv7x7(3, 64, stride=2, padding=0)
+
+        self.bn1 = bn_with_initialize(64)
+        self.relu = P.ReLU()
+        self.maxpool = P.MaxPoolWithArgmax(ksize=3, strides=2, padding="SAME")
+
+        self.layer1 = MakeLayer0(block, in_channels=64, out_channels=256, stride=1)
+        self.layer2 = MakeLayer1(block, in_channels=256, out_channels=512, stride=2)
+        self.layer3 = MakeLayer2(block, in_channels=512, out_channels=1024, stride=2)
+        self.layer4 = MakeLayer3(block, in_channels=1024, out_channels=2048, stride=2)
+
+        self.pool = P.ReduceMean(keep_dims=True)
+        self.squeeze = P.Squeeze(axis=(2, 3))
+        self.fc = fc_with_initialize(512 * block.expansion, num_classes)
+
+    def construct(self, x):
+        x = self.conv1(x)
+        x = self.bn1(x)
+        x = self.relu(x)
+        x = self.maxpool(x)[0]
+
+        x = self.layer1(x)
+        x = self.layer2(x)
+        x = self.layer3(x)
+        x = self.layer4(x)
+
+        x = self.pool(x, (2, 3))
+        x = self.squeeze(x)
+        x = self.fc(x)
+        return x
+
+
+def resnet50(batch_size, num_classes):
+    return ResNet(ResidualBlock, num_classes, batch_size)
+
+
+def create_dataset(repeat_num=1, training=True, batch_size=32):
+    data_home = "/home/workspace/mindspore_dataset"
+    data_dir = data_home + "/cifar-10-batches-bin"
+    if not training:
+        data_dir = data_home + "/cifar-10-verify-bin"
+    data_set = ds.Cifar10Dataset(data_dir)
+
+    resize_height = 224
+    resize_width = 224
+    rescale = 1.0 / 255.0
+    shift = 0.0
+
+    # define map operations
+    random_crop_op = vision.RandomCrop((32, 32), (4, 4, 4, 4))  # padding_mode default CONSTANT
+    random_horizontal_op = vision.RandomHorizontalFlip()
+    # interpolation default BILINEAR
+    resize_op = vision.Resize((resize_height, resize_width))
+    rescale_op = vision.Rescale(rescale, shift)
+    normalize_op = vision.Normalize((0.4465, 0.4822, 0.4914), (0.2010, 0.1994, 0.2023))
+    changeswap_op = vision.HWC2CHW()
+    type_cast_op = C.TypeCast(mstype.int32)
+
+    c_trans = []
+    if training:
+        c_trans = [random_crop_op, random_horizontal_op]
+    c_trans += [resize_op, rescale_op, normalize_op,
+                changeswap_op]
+
+    # apply map operations on images
+    data_set = data_set.map(input_columns="label", operations=type_cast_op)
+    data_set = data_set.map(input_columns="image", operations=c_trans)
+
+    # apply shuffle operations
+    data_set = data_set.shuffle(buffer_size=1000)
+
+    # apply batch operations
+    data_set = data_set.batch(batch_size=batch_size, drop_remainder=True)
+
+    # apply repeat operations
+    data_set = data_set.repeat(repeat_num)
+
+    return data_set
+
+
+class CrossEntropyLoss(nn.Cell):
+    def __init__(self):
+        super(CrossEntropyLoss, self).__init__()
+        self.cross_entropy = P.SoftmaxCrossEntropyWithLogits()
+        self.mean = P.ReduceMean()
+        self.one_hot = P.OneHot()
+        self.one = Tensor(1.0, mstype.float32)
+        self.zero = Tensor(0.0, mstype.float32)
+
+    def construct(self, logits, label):
+        label = self.one_hot(label, F.shape(logits)[1], self.one, self.zero)
+        loss = self.cross_entropy(logits, label)[0]
+        loss = self.mean(loss, (-1,))
+        return loss
+
+
+class GradWrap(Cell):
+    """ GradWrap definition """
+
+    def __init__(self, network):
+        super(GradWrap, self).__init__()
+        self.network = network
+        self.weights = ParameterTuple(network.trainable_params())
+
+    def construct(self, x, label):
+        weights = self.weights
+        return CP.grad_by_list(self.network, weights)(x, label)
+
+
+@pytest.mark.level0
+@pytest.mark.platform_arm_ascend_training
+@pytest.mark.platform_x86_ascend_training
+@pytest.mark.env_onecard
+def test_pynative_resnet50():
+    context.set_context(mode=context.PYNATIVE_MODE, device_target="Ascend")
+
+    batch_size = 32
+    num_classes = 10
+    net = resnet50(batch_size, num_classes)
+    criterion = CrossEntropyLoss()
+    optimizer = Momentum(learning_rate=0.01, momentum=0.9,
+                         params=filter(lambda x: x.requires_grad, net.get_parameters()))
+
+    net_with_criterion = WithLossCell(net, criterion)
+    net_with_criterion.set_grad()
+    train_network = GradWrap(net_with_criterion)
+    train_network.set_train()
+
+    step = 0
+    max_step = 20
+    data_set = create_dataset(repeat_num=1, training=True, batch_size=batch_size)
+    for element in data_set.create_dict_iterator():
+        step = step + 1
+        if step > max_step:
+            break
+        start_time = time.time()
+        input_data = Tensor(element["image"])
+        input_label = Tensor(element["label"])
+        loss_output = net_with_criterion(input_data, input_label)
+        grads = train_network(input_data, input_label)
+        optimizer(grads)
+        end_time = time.time()
+        cost_time = end_time - start_time
+        print("======step: ", step, " loss: ", loss_output.asnumpy(), " cost time: ", cost_time)
+        if step > 1:
+            assert cost_time < 0.5
+        
\ No newline at end of file