03 image classification (#5192)

* add batch_norm_layer

* add img_conv_group layer and test

* add check to Tensor.type()

* forward can run

* with backward

* change label data time from int32 to int64

* refine code

* follow comment

paddle/framework/operator.h

@@ -408,6 +408,7 @@ class OperatorWithKernel : public OperatorBase {
   // indicate kernel DataType by input data. Defaultly all input data must be
   // same.
   virtual DataType IndicateDataType(const ExecutionContext& ctx) const {
+    VLOG(3) << "Default IndicateDataType " << this->Type();
     auto& scope = ctx.scope();
     int data_type = -1;
     for (auto& input : this->inputs_) {

paddle/framework/tensor.h

@@ -126,11 +126,16 @@ class Tensor {
   inline Tensor Slice(const int& begin_idx, const int& end_idx) const;
 
   platform::Place place() const {
-    PADDLE_ENFORCE_NOT_NULL(holder_, "Tensor get place() must contains holder");
+    PADDLE_ENFORCE_NOT_NULL(
+        holder_, "Tensor not initialized yet when Tensor::place() is called.");
     return holder_->place();
   }
 
-  std::type_index type() const { return holder_->type(); }
+  std::type_index type() const {
+    PADDLE_ENFORCE_NOT_NULL(
+        holder_, "Tensor not initialized yet when Tensor::type() is called.");
+    return holder_->type();
+  }
 
   size_t memory_size() const;
 

paddle/operators/batch_norm_op.cc

@@ -18,6 +18,7 @@ namespace paddle {
 namespace operators {
 
 using Tensor = framework::Tensor;
+using LoDTensor = framework::LoDTensor;
 template <typename T, int MajorType = Eigen::RowMajor,
           typename IndexType = Eigen::DenseIndex>
 using EigenMatrix = framework::EigenMatrix<T, MajorType, IndexType>;
@@ -64,6 +65,9 @@ class BatchNormOp : public framework::OperatorWithKernel {
         (tensor_format == TensorFormat::NCHW ? x_dims[1]
                                              : x_dims[x_dims.size() - 1]);
 
+    PADDLE_ENFORCE(x_dims.size() >= 3 && x_dims.size() <= 5,
+                   "Input x must have 3 to 5 dimensions.");
+
     PADDLE_ENFORCE_EQ(ctx->GetInputDim("Scale").size(), 1UL);
     PADDLE_ENFORCE_EQ(ctx->GetInputDim("Scale")[0], C);
     PADDLE_ENFORCE_EQ(ctx->GetInputDim("Bias").size(), 1UL);
@@ -108,10 +112,12 @@ class BatchNormOpMaker : public framework::OpProtoAndCheckerMaker {
               "Store the global Variance when training");
     AddOutput("SavedMean",
               "Mean of the current mini batch, "
-              "will apply to output when training");
+              "will apply to output when training")
+        .AsIntermediate();
     AddOutput("SavedVariance",
               "Variance of the current mini batch, "
-              "will apply to output when training");
+              "will apply to output when training")
+        .AsIntermediate();
     AddComment(R"DOC(
 https://arxiv.org/pdf/1502.03167.pdf
 
@@ -135,7 +141,6 @@ class BatchNormKernel<platform::CPUPlace, T> : public framework::OpKernel<T> {
 
     const auto *x = ctx.Input<Tensor>("X");
     const auto &x_dims = x->dims();
-
     PADDLE_ENFORCE(x_dims.size() >= 3 && x_dims.size() <= 5,
                    "The Input dim size should be between 3 and 5");
     const int N = x_dims[0];
@@ -289,6 +294,25 @@ class BatchNormGradOp : public framework::OperatorWithKernel {
     ctx->SetOutputDim(framework::GradVarName("Scale"), {C});
     ctx->SetOutputDim(framework::GradVarName("Bias"), {C});
   }
+
+  framework::DataType IndicateDataType(
+      const framework::ExecutionContext &ctx) const override {
+    VLOG(3) << "IndicateDataType " << this->Type();
+    const auto *var = ctx.InputVar(framework::GradVarName("Y"));
+    if (var == nullptr) {
+      PADDLE_THROW("can't find Y@GRAD");
+    }
+    const Tensor *t = nullptr;
+    if (var->IsType<Tensor>()) {
+      t = &var->Get<Tensor>();
+    } else if (var->IsType<LoDTensor>()) {
+      t = &var->Get<LoDTensor>();
+    }
+    if (t == nullptr) {
+      PADDLE_THROW("can't find Y@GRAD");
+    }
+    return framework::ToDataType(t->type());
+  }
 };
 
 template <typename T>

paddle/operators/reshape_op.cc

@@ -34,13 +34,19 @@ class ReshapeOp : public framework::OperatorWithKernel {
 
     auto shape = ctx->Attrs().Get<std::vector<int>>("shape");
     PADDLE_ENFORCE(shape.size() > 0, "Attr(shape) shouldn't be empty.");
-    for (auto dim : shape) {
+    auto x_dims = ctx->GetInputDim("X");
-      PADDLE_ENFORCE(dim > 0, "Each dimension of shape must be positive.");
+    // TODO(qiao) change batch_size
+    for (int i = 1; i < shape.size(); ++i) {
+      PADDLE_ENFORCE(shape[i] > 0,
+                     "Each dimension of shape "
+                     "must be positiv except the first.");
+    }
+    if (shape[0] < 0) {
+      shape[0] = x_dims[0];
     }
     // capacity check
     int64_t capacity =
         std::accumulate(shape.begin(), shape.end(), 1, std::multiplies<int>());
-    auto x_dims = ctx->GetInputDim("X");
     int64_t in_size = framework::product(x_dims);
     PADDLE_ENFORCE_EQ(capacity, in_size,
                       "The size of Input(X) mismatches with Attr(shape).");

paddle/operators/reshape_op.h

@@ -26,13 +26,8 @@ class ReshapeKernel : public framework::OpKernel<T> {
   void Compute(const framework::ExecutionContext& ctx) const {
     auto* out = ctx.Output<framework::Tensor>("Out");
     auto* in = ctx.Input<framework::Tensor>("X");
+    auto out_dims = out->dims();
     out->mutable_data<T>(ctx.GetPlace());
-
-    auto shape = ctx.Attr<std::vector<int>>("shape");
-    std::vector<int64_t> shape_int64(shape.size(), 0);
-    std::transform(shape.begin(), shape.end(), shape_int64.begin(),
-                   [](int a) { return static_cast<int64_t>(a); });
-    auto out_dims = framework::make_ddim(shape_int64);
     out->CopyFrom(*in, ctx.GetPlace(), ctx.device_context());
     out->Resize(out_dims);
   }

python/paddle/v2/framework/framework.py

@@ -352,7 +352,10 @@ class Block(object):
         return {v for k, v in self.vars.iteritems() if isinstance(v, Parameter)}
 
     def create_var(self, *args, **kwargs):
-        return Variable(self, *args, **kwargs)
+        var = Variable(self, *args, **kwargs)
+        if 'init_attr' in kwargs:
+            self._prepend_initialize_ops_(var, kwargs['init_attr'])
+        return var
 
     def has_var(self, name):
         return name in self.vars

python/paddle/v2/framework/layers.py

@@ -161,6 +161,7 @@ def _create_op_func_(op_type):
 _create_op_func_('mean')
 _create_op_func_('mul')
 _create_op_func_('dropout')
+_create_op_func_('reshape')
 
 
 def cast(x, data_type, program=None):
@@ -308,6 +309,96 @@ def pool2d(input,
     return pool_out
 
 
+def batch_norm(input,
+               act=None,
+               is_test=False,
+               momentum=0.9,
+               epsilon=1e05,
+               param_attr=None,
+               bias_attr=None,
+               data_layout='NCHW',
+               program=None,
+               init_program=None):
+    helper = LayerHelper('batch_norm', **locals())
+    dtype = helper.input_dtype()
+
+    input_shape = input.shape
+    if data_layout == 'NCHW':
+        channel_num = input_shape[1]
+    else:
+        if data_layout == 'NHWC':
+            channel_num = input_shape[-1]
+        else:
+            raise ValueError("unsupported data layout:" + data_layout)
+
+    def get_init_attr(value):
+        if not isinstance(value, float):
+            raise ValueError("attr value should be a float")
+        return {'type': 'fill_constant', 'value': value}
+
+    def prepend_init_op(var, init_attr):
+        assert isinstance(var, Variable)
+        op_type = init_attr['type']
+        init_attr['shape'] = var.shape
+        init_attr['data_type'] = int(var.data_type)
+        op = var.block.prepend_op(
+            type=op_type, inputs=None, outputs={'Out': [var]}, attrs=init_attr)
+        return op
+
+    def create_persistable_var(dtype, shape, init_attr=None):
+        name = unique_name(".".join([helper.name, "xxxx"]))
+        var = init_program.global_block().create_var(
+            dtype=dtype, shape=shape, name=name, persistable=True)
+        if 'init_attr' is not None:
+            prepend_init_op(var, init_attr)
+        return program.global_block().create_var(
+            name=name, dtype=dtype, shape=shape, persistable=True)
+
+    param_shape = [channel_num]
+
+    # create parameter
+    scale = helper.create_parameter(
+        attr=helper.param_attr, shape=param_shape, dtype=dtype)
+    bias = helper.create_parameter(
+        attr=helper.param_attr, shape=param_shape, dtype=dtype)
+
+    # create input
+    mean = create_persistable_var(dtype, param_shape, get_init_attr(0.0))
+    variance = create_persistable_var(dtype, param_shape, get_init_attr(1.0))
+
+    # create output
+    # mean and mean_out share the same memory
+    mean_out = mean
+    # variance and variance out share the same memory
+    variance_out = variance
+    saved_mean = helper.create_tmp_variable(dtype)
+    saved_variance = helper.create_tmp_variable(dtype)
+
+    batch_norm_out = helper.create_tmp_variable(dtype)
+
+    helper.append_op(
+        type="batch_norm",
+        inputs={
+            "X": input,
+            "Scale": scale,
+            "Bias": bias,
+            "Mean": mean,
+            "Variance": variance
+        },
+        outputs={
+            "Y": batch_norm_out,
+            "MeanOut": mean_out,
+            "VarianceOut": variance_out,
+            "SavedMean": saved_mean,
+            "SavedVariance": saved_variance
+        },
+        attrs={"momentum": momentum,
+               "epsilon": epsilon,
+               "is_test": is_test})
+
+    return helper.append_activation(batch_norm_out)
+
+
 class BlockGuard(object):
     """
     BlockGuard used to create sub-block in program by using Python `with` 

python/paddle/v2/framework/nets.py

@@ -7,6 +7,7 @@ def simple_img_conv_pool(input,
                          pool_size,
                          pool_stride,
                          act,
+                         pool_type='max',
                          program=None,
                          init_program=None):
     conv_out = layers.conv2d(
@@ -20,7 +21,75 @@ def simple_img_conv_pool(input,
     pool_out = layers.pool2d(
         input=conv_out,
         pool_size=pool_size,
-        pool_type='max',
+        pool_type=pool_type,
+        pool_stride=pool_stride,
+        program=program,
+        init_program=init_program)
+    return pool_out
+
+
+def img_conv_group(input,
+                   conv_num_filter,
+                   pool_size,
+                   conv_padding=1,
+                   conv_filter_size=3,
+                   conv_act=None,
+                   conv_with_batchnorm=False,
+                   conv_batchnorm_drop_rate=None,
+                   pool_stride=1,
+                   pool_type=None,
+                   program=None,
+                   init_program=None):
+    """
+    Image Convolution Group, Used for vgg net.
+    """
+    tmp = input
+    assert isinstance(conv_num_filter, list) or \
+           isinstance(conv_num_filter, tuple)
+
+    def __extend_list__(obj):
+        if not hasattr(obj, '__len__'):
+            return [obj] * len(conv_num_filter)
+        else:
+            return obj
+
+    conv_padding = __extend_list__(conv_padding)
+    conv_filter_size = __extend_list__(conv_filter_size)
+    conv_with_batchnorm = __extend_list__(conv_with_batchnorm)
+    conv_batchnorm_drop_rate = __extend_list__(conv_batchnorm_drop_rate)
+
+    for i in xrange(len(conv_num_filter)):
+        local_conv_act = conv_act
+        if conv_with_batchnorm[i]:
+            local_conv_act = None
+
+        tmp = layers.conv2d(
+            input=tmp,
+            num_filters=conv_num_filter[i],
+            filter_size=conv_filter_size[i],
+            padding=conv_padding[i],
+            act=local_conv_act,
+            program=program,
+            init_program=init_program)
+
+        if conv_with_batchnorm[i]:
+            tmp = layers.batch_norm(
+                input=tmp,
+                act=conv_act,
+                program=program,
+                init_program=init_program)
+            drop_rate = conv_batchnorm_drop_rate[i]
+            if abs(drop_rate) > 1e-5:
+                tmp = layers.dropout(
+                    x=tmp,
+                    dropout_prob=drop_rate,
+                    program=program,
+                    init_program=init_program)
+
+    pool_out = layers.pool2d(
+        input=tmp,
+        pool_size=pool_size,
+        pool_type=pool_type,
         pool_stride=pool_stride,
         program=program,
         init_program=init_program)

python/paddle/v2/framework/tests/test_image_classification_layer.py

@@ -0,0 +1,75 @@
+import unittest
+
+import paddle.v2.framework.layers as layers
+import paddle.v2.framework.nets as nets
+from paddle.v2.framework.framework import Program
+
+
+def conv_block(input,
+               num_filter,
+               groups,
+               dropouts,
+               program=None,
+               init_program=None):
+    return nets.img_conv_group(
+        input=input,
+        pool_size=2,
+        pool_stride=2,
+        conv_num_filter=[num_filter] * groups,
+        conv_filter_size=3,
+        conv_act='relu',
+        conv_with_batchnorm=True,
+        conv_batchnorm_drop_rate=dropouts,
+        pool_type='max',
+        program=program,
+        init_program=init_program)
+
+
+class TestLayer(unittest.TestCase):
+    def test_batch_norm_layer(self):
+        program = Program()
+        init_program = Program()
+        images = layers.data(
+            name='pixel',
+            shape=[3, 48, 48],
+            data_type='float32',
+            program=program)
+        layers.batch_norm(
+            input=images, program=program, init_program=init_program)
+
+        #print str(program)
+
+    def test_dropout_layer(self):
+        program = Program()
+        init_program = Program()
+        images = layers.data(
+            name='pixel',
+            shape=[3, 48, 48],
+            data_type='float32',
+            program=program)
+        layers.dropout(
+            x=images,
+            dropout_prob=0.5,
+            program=program,
+            init_program=init_program)
+
+        #print str(program)
+
+    def test_img_conv_group(self):
+        program = Program()
+        init_program = Program()
+
+        images = layers.data(
+            name='pixel',
+            shape=[3, 48, 48],
+            data_type='float32',
+            program=program,
+            init_program=init_program)
+        conv1 = conv_block(images, 64, 2, [0.3, 0], program, init_program)
+        conv2 = conv_block(conv1, 256, 3, [0.4, 0.4, 0], program, init_program)
+
+        # print str(program)
+
+
+if __name__ == '__main__':
+    unittest.main()

python/paddle/v2/framework/tests/test_image_classification_train.py

@@ -0,0 +1,133 @@
+import paddle.v2 as paddle
+import paddle.v2.framework.layers as layers
+import paddle.v2.framework.nets as nets
+import paddle.v2.framework.core as core
+import paddle.v2.framework.optimizer as optimizer
+
+from paddle.v2.framework.framework import Program, g_program
+from paddle.v2.framework.executor import Executor
+
+import numpy as np
+
+
+def vgg16_bn_drop(input, program, init_program):
+    def conv_block(input,
+                   num_filter,
+                   groups,
+                   dropouts,
+                   program=None,
+                   init_program=None):
+        return nets.img_conv_group(
+            input=input,
+            pool_size=2,
+            pool_stride=2,
+            conv_num_filter=[num_filter] * groups,
+            conv_filter_size=3,
+            conv_act='relu',
+            conv_with_batchnorm=True,
+            conv_batchnorm_drop_rate=dropouts,
+            pool_type='max',
+            program=program,
+            init_program=init_program)
+
+    conv1 = conv_block(input, 64, 2, [0.3, 0], program, init_program)
+    conv2 = conv_block(conv1, 128, 2, [0.4, 0], program, init_program)
+    conv3 = conv_block(conv2, 256, 3, [0.4, 0.4, 0], program, init_program)
+    conv4 = conv_block(conv3, 512, 3, [0.4, 0.4, 0], program, init_program)
+    conv5 = conv_block(conv4, 512, 3, [0.4, 0.4, 0], program, init_program)
+
+    drop = layers.dropout(
+        x=conv5, dropout_prob=0.5, program=program, init_program=init_program)
+    fc1 = layers.fc(input=drop,
+                    size=512,
+                    act=None,
+                    program=program,
+                    init_program=init_program)
+    reshape1 = layers.reshape(
+        x=fc1,
+        shape=list(fc1.shape + (1, 1)),
+        program=program,
+        init_program=init_program)
+    bn = layers.batch_norm(
+        input=reshape1, act='relu', program=program, init_program=init_program)
+    drop2 = layers.dropout(
+        x=bn, dropout_prob=0.5, program=program, init_program=init_program)
+    fc2 = layers.fc(input=drop2,
+                    size=512,
+                    act=None,
+                    program=program,
+                    init_program=init_program)
+    return fc2
+
+
+init_program = Program()
+program = Program()
+
+classdim = 10
+data_shape = [3, 32, 32]
+
+images = layers.data(
+    name='pixel', shape=data_shape, data_type='float32', program=program)
+
+label = layers.data(
+    name='label',
+    shape=[1],
+    data_type='int64',
+    program=program,
+    init_program=init_program)
+vgg_net = vgg16_bn_drop(images, program, init_program)
+predict = layers.fc(input=vgg_net,
+                    size=classdim,
+                    act='softmax',
+                    program=program,
+                    init_program=init_program)
+cost = layers.cross_entropy(
+    input=predict, label=label, program=program, init_program=init_program)
+avg_cost = layers.mean(x=cost, program=program, init_program=init_program)
+
+sgd_optimizer = optimizer.SGDOptimizer(learning_rate=0.001)
+opts = sgd_optimizer.minimize(avg_cost)
+
+BATCH_SIZE = 128
+PASS_NUM = 1
+
+train_reader = paddle.batch(
+    paddle.reader.shuffle(
+        paddle.dataset.cifar.train10(), buf_size=128 * 10),
+    batch_size=BATCH_SIZE)
+
+place = core.CPUPlace()
+exe = Executor(place)
+
+exe.run(init_program, feed={}, fetch_list=[])
+
+for pass_id in range(PASS_NUM):
+    batch_id = 0
+    for data in train_reader():
+        img_data = np.array(map(lambda x: x[0].reshape(data_shape),
+                                data)).astype("float32")
+        y_data = np.array(map(lambda x: x[1], data)).astype("int64")
+        batch_size = 1
+        for i in y_data.shape:
+            batch_size = batch_size * i
+        y_data = y_data.reshape([batch_size, 1])
+
+        tensor_img = core.LoDTensor()
+        tensor_y = core.LoDTensor()
+        tensor_img.set(img_data, place)
+        tensor_y.set(y_data, place)
+
+        outs = exe.run(program,
+                       feed={"pixel": tensor_img,
+                             "label": tensor_y},
+                       fetch_list=[avg_cost])
+
+        loss = np.array(outs[0])
+        # print("pass_id:" + str(pass_id) + " batch_id:" + str(batch_id) +
+        #       " loss:" + str(loss))
+        batch_id = batch_id + 1
+
+        if batch_id > 1:
+            # this model is slow, so if we can train two mini batch, we think it works properly.
+            exit(0)
+exit(1)