Merge branch 'fix-deconv_inputs' into 'master'

deconv support variable output shape  and support Maximum op

See merge request !681

mace/kernels/deconv_2d.h

@@ -231,6 +231,7 @@ struct Deconv2dFunctor : Deconv2dFunctorBase {
   MaceStatus operator()(const Tensor *input,   // NCHW
                   const Tensor *filter,  // OIHW
                   const Tensor *bias,
+                  const Tensor *output_shape_tensor,
                   Tensor *output,
                   StatsFuture *future) {
     MACE_UNUSED(future);
@@ -240,10 +241,20 @@ struct Deconv2dFunctor : Deconv2dFunctorBase {
 
     if (!from_caffe_) {  // tensorflow
       std::vector<index_t> output_shape(4);
-      output_shape[0] = output_shape_[0];
-      output_shape[1] = output_shape_[3];
-      output_shape[2] = output_shape_[1];
-      output_shape[3] = output_shape_[2];
+      if (output_shape_.size() == 4) {
+        output_shape[0] = output_shape_[0];
+        output_shape[1] = output_shape_[3];
+        output_shape[2] = output_shape_[1];
+        output_shape[3] = output_shape_[2];
+      } else {
+        MACE_CHECK_NOTNULL(output_shape_tensor);
+        MACE_CHECK(output_shape_tensor->size() == 4);
+        Tensor::MappingGuard output_shape_mapper(output_shape_tensor);
+        auto output_shape_data =
+            output_shape_tensor->data<int32_t>();
+        output_shape =
+            std::vector<index_t>(output_shape_data, output_shape_data + 4);
+      }
       paddings_.clear();
       paddings_ = std::vector<int>(2, 0);
       CalcDeconvPaddingAndInputSize(
@@ -326,6 +337,7 @@ struct Deconv2dFunctor<DeviceType::GPU, T> : Deconv2dFunctorBase {
   MaceStatus operator()(const Tensor *input,
                   const Tensor *filter,
                   const Tensor *bias,
+                  const Tensor *output_shape_tensor,
                   Tensor *output,
                   StatsFuture *future);
 

mace/kernels/opencl/deconv_2d_opencl.cc

@@ -167,6 +167,7 @@ MaceStatus Deconv2dFunctor<DeviceType::GPU, T>::operator()(
     const Tensor *input,
     const Tensor *filter,
     const Tensor *bias,
+    const Tensor *output_shape_tensor,
     Tensor *output,
     StatsFuture *future) {
   MACE_CHECK_NOTNULL(input);
@@ -174,6 +175,15 @@ MaceStatus Deconv2dFunctor<DeviceType::GPU, T>::operator()(
   MACE_CHECK_NOTNULL(output);
 
   if (!from_caffe_) {
+    if (output_shape_.size() != 4) {
+      MACE_CHECK_NOTNULL(output_shape_tensor);
+      MACE_CHECK(output_shape_tensor->size() == 4);
+      Tensor::MappingGuard output_shape_mapper(output_shape_tensor);
+      auto output_shape_data =
+          output_shape_tensor->data<int32_t>();
+      output_shape_ =
+          std::vector<index_t>(output_shape_data, output_shape_data + 4);
+    }
     paddings_.clear();
     paddings_ = std::vector<int>(2, 0);
     CalcDeconvPaddingAndInputSize(input->shape().data(), filter->shape().data(),

mace/ops/deconv_2d.h

@@ -40,17 +40,19 @@ class Deconv2dOp : public ConvPool2dOpBase<D, T> {
   MaceStatus Run(StatsFuture *future) override {
     const Tensor *input = this->Input(INPUT);
     const Tensor *filter = this->Input(FILTER);
-    const Tensor *bias = this->InputSize() >= 3 ? this->Input(BIAS) : nullptr;
+    const Tensor *output_shape =
+        this->InputSize() >= 3 ? this->Input(OUTPUT_SHAPE) : nullptr;
+    const Tensor *bias = this->InputSize() >= 4 ? this->Input(BIAS) : nullptr;
     Tensor *output = this->Output(OUTPUT);
 
-    return functor_(input, filter, bias, output, future);
+    return functor_(input, filter, bias, output_shape, output, future);
   }
 
  private:
   kernels::Deconv2dFunctor<D, T> functor_;
 
  protected:
-  MACE_OP_INPUT_TAGS(INPUT, FILTER, BIAS);
+  MACE_OP_INPUT_TAGS(INPUT, FILTER, OUTPUT_SHAPE,  BIAS);
   MACE_OP_OUTPUT_TAGS(OUTPUT);
 };
 

mace/ops/shape.cc

@@ -23,6 +23,19 @@ void Register_Shape(OperatorRegistryBase *op_registry) {
                                           .TypeConstraint<float>("T")
                                           .Build(),
                          ShapeOp<DeviceType::CPU, float>);
+#ifdef MACE_ENABLE_OPENCL
+  MACE_REGISTER_OPERATOR(op_registry, OpKeyBuilder("Shape")
+                                          .Device(DeviceType::GPU)
+                                          .TypeConstraint<float>("T")
+                                          .Build(),
+                         ShapeOp<DeviceType::GPU, float>);
+
+  MACE_REGISTER_OPERATOR(op_registry, OpKeyBuilder("Shape")
+                                          .Device(DeviceType::GPU)
+                                          .TypeConstraint<half>("T")
+                                          .Build(),
+                         ShapeOp<DeviceType::GPU, half>);
+#endif
 }
 
 }  // namespace ops

mace/python/tools/converter_tool/tensorflow_converter.py

@@ -151,6 +151,7 @@ class TensorflowConverter(base_converter.ConverterInterface):
             TFOpType.Div.name: self.convert_elementwise,
             TFOpType.Min.name: self.convert_elementwise,
             TFOpType.Max.name: self.convert_elementwise,
+            TFOpType.Maximum.name: self.convert_elementwise,
             TFOpType.Neg.name: self.convert_elementwise,
             TFOpType.Abs.name: self.convert_elementwise,
             TFOpType.Pow.name: self.convert_elementwise,
@@ -372,19 +373,21 @@ class TensorflowConverter(base_converter.ConverterInterface):
                 dilation_val = [1, 1]
             dilation_arg.ints.extend(dilation_val)
         else:
+            mace_check(len(tf_op.inputs) >= 3,
+                       "deconv should have (>=) 3 inputs.")
             output_shape_arg = op.arg.add()
             output_shape_arg.name = MaceKeyword.mace_output_shape_str
-            if len(tf_op.inputs) >= 3:
-                del op.input[1:]
-                output_shape_value =\
+            if tf_op.inputs[0].op.type == TFOpType.Const.name:
+                output_shape_value = \
                     tf_op.inputs[0].eval().astype(np.int32).flat
                 output_shape_arg.ints.extend(output_shape_value)
-                self._skip_tensor.add(tf_op.inputs[0].name)
-                del op.input[0]
-                op.input.extend([tf_op.inputs[2].name, tf_op.inputs[1].name])
             else:
-                output_shape_value = tf_op.get_attr(tf_strides_str)
+                output_shape_value = {}
                 output_shape_arg.ints.extend(output_shape_value)
+            del op.input[:]
+            op.input.extend([tf_op.inputs[2].name,
+                             tf_op.inputs[1].name,
+                             tf_op.inputs[0].name])
 
     def convert_elementwise(self, tf_op):
         op = self.convert_general_op(tf_op)

mace/python/tools/converter_tool/transformer.py

@@ -980,8 +980,10 @@ class Transformer(base_converter.ConverterInterface):
             if op.type == MaceOp.Conv2D.name \
                     or op.type == MaceOp.Deconv2D.name:
                 self.buffer_to_image(op, 1, OpenCLBufferType.CONV2D_FILTER)
-                if len(op.input) >= 3:
+                if len(op.input) >= 3 and op.type == MaceOp.Conv2D.name:
                     self.buffer_to_image(op, 2, OpenCLBufferType.ARGUMENT)
+                elif len(op.input) >= 4 and op.type == MaceOp.Deconv2D.name:
+                    self.buffer_to_image(op, 3, OpenCLBufferType.ARGUMENT)
             elif op.type == MaceOp.DepthwiseConv2d.name:
                 self.buffer_to_image(op, 1, OpenCLBufferType.DW_CONV2D_FILTER)
                 if len(op.input) >= 3:

mace/python/tools/memory_optimizer.py

@@ -24,6 +24,7 @@ class MemoryOptimizer(object):
         self.op_mem = {}  # op_name->mem_id
         self.mem_block = {}  # mem_id->[size] or mem_id->[x, y]
         self.total_mem_count = 0
+        self.total_cpu_mem_count = 0
         self.input_ref_counter = {}
         self.mem_ref_counter = {}
 
@@ -184,6 +185,15 @@ class GPUMemoryOptimizer(MemoryOptimizer):
             for arg in op.arg:
                 if arg.name == 'mode' and arg.i == 0:
                     return False
+        elif op.type == 'Shape':
+            for i in range(len(op.output)):
+                mem_id = self.total_cpu_mem_count
+                self.total_cpu_mem_count += 1
+                op_mem_block = self.get_op_mem_block(
+                    op.type,
+                    op.output_shape[i].dims)
+                self.mem_block[mem_id] = op_mem_block
+            return False
         return op.type != 'ImageToBuffer'
 
     def get_op_mem_block(self, op_type, output_shape):
@@ -191,13 +201,18 @@ class GPUMemoryOptimizer(MemoryOptimizer):
         if op_type == 'WinogradTransform' or op_type == 'MatMul':
             mem_block[0] = output_shape[2]
             mem_block[1] = output_shape[0] * int((output_shape[1] + 3) / 4)
+        elif op_type == 'Shape':
+            mem_block[0] = output_shape[0]
+            mem_block[1] = 1
         else:
             if len(output_shape) == 2:  # only support fc/softmax
                 mem_block[0] = int((output_shape[1] + 3) / 4)
                 mem_block[1] = output_shape[0]
-            else:
+            elif len(output_shape) == 4:
                 mem_block[0] = output_shape[2] * int((output_shape[3] + 3) / 4)
                 mem_block[1] = output_shape[0] * output_shape[1]
+            else:
+                raise Exception('output shape dim size is not 2 or 4.')
         return mem_block
 
     def mem_size(self, memory_block):