Merge branch 'develop' into develop

python/tools/mdl2fluid/float2halffloat.py

+# encoding:utf-8
+import math
+import re
+
+
+def Real2HalfFloat(data):
+    MINNUM = -65536
+    MAXNUM = 65535
+    FloatVal = 0
+    if data:
+        if data < MINNUM:
+            data = MINNUM
+        if data > MAXNUM:
+            data = MAXNUM
+
+        sign = 0
+        if data < 0:
+            sign = 1
+            data = -data
+
+        exp = math.floor((math.log2(data)))
+        expout = exp + 16
+
+        Mantial = round(data / pow(2, exp - 10)) - 1024
+
+        if expout <= 0:
+            FloatVal = 0
+        else:
+            FloatVal = sign * 32768 + expout * 1024 + Mantial
+    return FloatVal
+
+
+def ReadCfloatData(sourcefile):
+    input = []
+    with open(sourcfile, 'r') as f:
+        for line in f.readlines():
+            line = line.strip()
+            line = re.sub('\s+', ' ', line)  # 两个数字间多个空格
+            input.append(line.split(' '))
+    destfile = sourcefile.replace('.dat', '')
+    destfile = destfile.replace('.txt', '')
+    destfile += 'Out.dat'
+    with open(destfile, 'w') as fw:
+        for i in range(len(input)):
+            if len(input[i]) == 2:
+                real = Real2HalfFloat(float(input[i][0]))
+                imag = Real2HalfFloat(float(input[i][1]))
+                result = real * 65536 + imag
+                if imag and not real:
+                    fw.write('0x0000' + "%X" % result + '\n')
+                elif not imag and not real:
+                    fw.write('0x00000000' + '\n')
+                else:
+                    fw.write('0x' + "%X" % result + '\n')
+            elif len(input[i]) == 1:
+                result = Real2HalfFloat(float(input[i][0]))
+                if result:
+                    fw.write('0x' + "%X" % result + '\n')
+                else:
+                    fw.write('0x0000' + '\n')
+
+
+if __name__ == '__main__':
+    print('Tips: Input number 0 if you want to exit!\n')
+    while True:
+        sourcfile = input("input source file:\n")
+        if sourcfile is '0':
+            break
+        ReadCfloatData(sourcfile)
+        print('Transfer Success!')

python/tools/mdl2fluid/framework.proto

+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved.
+
+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. */
+
+syntax = "proto2";
+option optimize_for = LITE_RUNTIME;
+package paddle_mobile.framework.proto;
+
+enum AttrType {
+  INT = 0;
+  FLOAT = 1;
+  STRING = 2;
+  INTS = 3;
+  FLOATS = 4;
+  STRINGS = 5;
+  BOOLEAN = 6;
+  BOOLEANS = 7;
+  BLOCK = 8;
+  LONG = 9;
+}
+
+// OpDesc describes an instance of a C++ framework::OperatorBase
+// derived class type.
+message OpDesc {
+
+  message Attr {
+    required string name = 1;
+    required AttrType type = 2;
+    optional int32 i = 3;
+    optional float f = 4;
+    optional string s = 5;
+    repeated int32 ints = 6;
+    repeated float floats = 7;
+    repeated string strings = 8;
+    optional bool b = 10;
+    repeated bool bools = 11;
+    optional int32 block_idx = 12;
+    optional int64 l = 13;
+  };
+
+  message Var {
+    required string parameter = 1;
+    repeated string arguments = 2;
+  };
+
+  required string type = 3;
+  repeated Var inputs = 1;
+  repeated Var outputs = 2;
+  repeated Attr attrs = 4;
+  optional bool is_target = 5 [ default = false ];
+};
+
+// OpProto describes a C++ framework::OperatorBase derived class.
+message OpProto {
+
+  // VarProto describes the C++ type framework::Variable.
+  message Var {
+    required string name = 1;
+    required string comment = 2;
+
+    optional bool duplicable = 3 [ default = false ];
+    optional bool intermediate = 4 [ default = false ];
+    optional bool dispensable = 5 [ default = false ];
+  }
+
+  // AttrProto describes the C++ type Attribute.
+  message Attr {
+    required string name = 1;
+    required AttrType type = 2;
+    required string comment = 3;
+    // If that attribute is generated, it means the Paddle third
+    // language binding has responsibility to fill that
+    // attribute. End-User should not set that attribute.
+    optional bool generated = 4 [ default = false ];
+  }
+
+  required string type = 1;
+  repeated Var inputs = 2;
+  repeated Var outputs = 3;
+  repeated Attr attrs = 4;
+  required string comment = 5;
+}
+
+message VarType {
+  enum Type {
+    // Pod Types
+    BOOL = 0;
+    INT16 = 1;
+    INT32 = 2;
+    INT64 = 3;
+    FP16 = 4;
+    FP32 = 5;
+    FP64 = 6;
+
+    // Other types that may need additional descriptions
+    LOD_TENSOR = 7;
+    SELECTED_ROWS = 8;
+    FEED_MINIBATCH = 9;
+    FETCH_LIST = 10;
+    STEP_SCOPES = 11;
+    LOD_RANK_TABLE = 12;
+    LOD_TENSOR_ARRAY = 13;
+    PLACE_LIST = 14;
+    READER = 15;
+    CHANNEL = 16;
+    // Any runtime decided variable type is raw
+    // raw variables should manage their own allocations
+    // in operators like nccl_op
+    RAW = 17;
+    TUPLE = 18;
+  }
+
+  required Type type = 1;
+
+  message TensorDesc {
+    // Should only be PODType. Is enforced in C++
+    required Type data_type = 1;
+    repeated int64 dims = 2; // [UNK, 640, 480] is saved as [-1, 640, 480]
+  }
+  optional TensorDesc selected_rows = 2;
+
+  message LoDTensorDesc {
+    required TensorDesc tensor = 1;
+    optional int32 lod_level = 2 [ default = 0 ];
+  }
+  optional LoDTensorDesc lod_tensor = 3;
+
+  message LoDTensorArrayDesc {
+    required TensorDesc tensor = 1;
+    optional int32 lod_level = 2 [ default = 0 ];
+  }
+  optional LoDTensorArrayDesc tensor_array = 4;
+
+  message ReaderDesc { repeated LoDTensorDesc lod_tensor = 1; }
+  optional ReaderDesc reader = 5;
+
+  message ChannelDesc {
+    required Type data_type = 1;
+    required int64 capacity = 2;
+  }
+  optional ChannelDesc channel = 6;
+
+  message Tuple { repeated Type element_type = 1; }
+  optional Tuple tuple = 7;
+}
+
+message VarDesc {
+  required string name = 1;
+  required VarType type = 2;
+  optional bool persistable = 3 [ default = false ];
+}
+
+message BlockDesc {
+  required int32 idx = 1;
+  required int32 parent_idx = 2;
+  repeated VarDesc vars = 3;
+  repeated OpDesc ops = 4;
+  optional int32 forward_block_idx = 5 [ default = -1 ];
+}
+
+// Please refer to
+// https://github.com/PaddlePaddle/Paddle/blob/develop/doc/design/program.md
+// for more details.
+// TODO(panyx0718): A model can have multiple programs. Need a
+// way to distinguish them. Maybe ID or name?
+message ProgramDesc { repeated BlockDesc blocks = 1; }

python/tools/mdl2fluid/loader.py

+import datetime
+import json
+import os
+
+import google.protobuf as pbg
+import framework_pb2 as framework_pb2
+
+
+def loadmdl(json_path):
+    print('mdl json path : ' + json_path)
+    with open(json_path, 'r') as f:
+        json_dick = json.load(f)
+        # print(json_dick)
+        layers = (json_dick['layer'])
+        for layer in layers:
+            print(layer)
+
+

python/tools/mdl2fluid/mdl2fluid.py

+import json
+import os
+
+import framework_pb2 as framework_pb2
+import op_types as types
+from swicher import Swichter
+import shutil
+
+
+def load_mdl(mdl_json_path):
+    # print('mdl json path : ' + mdl_json_path)
+    with open(mdl_json_path, 'r') as f:
+        return json.load(f)
+
+
+class Converter:
+    'convert mdlmodel to fluidmodel'
+
+    def __init__(self, mdl_json_path):
+        self.mdl_json_path = mdl_json_path
+        print mdl_json_path
+        self.mdl_json = load_mdl(self.mdl_json_path)
+        self.program_desc = framework_pb2.ProgramDesc()
+        self.weight_list_ = []
+        self.deepwise_weight_list_ = []
+        # print(json_dick)
+        # layers = (json_dick['layer'])
+        # for layer in layers:
+        #     print(layer)
+
+    def convert(self):
+        print 'convert begin.....'
+        # add block_desc
+        block_desc = self.program_desc.blocks.add()
+        block_desc.idx = 0
+        block_desc.parent_idx = -1
+        self.package_ops(block_desc)
+        self.package_vars(block_desc)
+        print 'blocks: '
+        print self.program_desc.blocks
+        print 'convert end.....'
+        desc_serialize_to_string = self.program_desc.SerializeToString()
+        shutil.rmtree('newyolo/')
+        shutil.copytree('multiobjects/float32s_nchw_with_head', 'newyolo/')
+
+        f = open("newyolo/__model__", "wb")
+        f.write(desc_serialize_to_string)
+        f.close()
+
+    def package_ops(self, block_desc):
+
+        self.add_op_feed(block_desc)
+
+        # add ops with layer
+        if 'layer' in self.mdl_json:
+
+            layers_ = self.mdl_json['layer']
+            for layer in layers_:
+                desc_ops_add = block_desc.ops.add()
+
+                # print layer
+                # for i in layer:
+                #     print i
+                if 'name' in layer:
+                    l_name = layer['name']
+                if 'type' in layer:
+                    self.package_ops_type(desc_ops_add, layer)
+
+                if 'weight' in layer:
+                    self.package_ops_weight2inputs(desc_ops_add, layer)
+
+                if 'output' in layer:
+                    self.package_ops_outputs(desc_ops_add, layer)
+
+                if 'input' in layer:
+                    self.package_ops_inputs(desc_ops_add, layer)
+
+                self.package_ops_attrs(desc_ops_add, layer)
+
+        self.add_op_fetch(block_desc)
+
+    def add_op_feed(self, block_desc):
+        desc_ops_add = block_desc.ops.add()
+        inputs_add = desc_ops_add.inputs.add()
+        inputs_add.parameter = 'X'
+        inputs_add.arguments.append('feed')
+        desc_ops_add.type = 'feed'
+        outputs_add = desc_ops_add.outputs.add()
+        outputs_add.parameter = 'Out'
+        outputs_add.arguments.append('data')
+        attrs_add = desc_ops_add.attrs.add()
+        attrs_add.name = 'col'
+        # boolean
+        attrs_add.type = 0
+        attrs_add.i = 0
+
+    def add_op_fetch(self, block_desc):
+        desc_ops_add = block_desc.ops.add()
+        inputs_add = desc_ops_add.inputs.add()
+        inputs_add.parameter = 'X'
+        inputs_add.arguments.append('conv_pred_87')
+        desc_ops_add.type = 'fetch'
+        outputs_add = desc_ops_add.outputs.add()
+        outputs_add.parameter = 'Out'
+        outputs_add.arguments.append('fetch')
+        attrs_add = desc_ops_add.attrs.add()
+        attrs_add.name = 'col'
+        # boolean
+        attrs_add.type = 0
+        attrs_add.i = 0
+
+    @staticmethod
+    def package_ops_attrs(desc_ops_add, layer):
+        # print l_params
+        # print desc_ops_add.type
+        if desc_ops_add.type == types.op_fluid_fusion_conv_add:
+            Converter.pack_fusion_conv_add_attr(desc_ops_add, layer)
+        elif desc_ops_add.type == types.op_fluid_relu:
+            # fusion_conv_add : attrs
+            attrs_add = desc_ops_add.attrs.add()
+            attrs_add.name = 'use_mkldnn'
+            # boolean
+            attrs_add.type = 6
+            attrs_add.b = 0
+
+    @staticmethod
+    def pack_fusion_conv_add_attr(desc_ops_add, layer):
+
+        # fusion_conv_add : attrs
+        attrs_add = desc_ops_add.attrs.add()
+        attrs_add.name = 'workspace_size_MB'
+        # 0-->INT
+        attrs_add.type = 0
+        attrs_add.i = 4096
+
+        attrs_add = desc_ops_add.attrs.add()
+        attrs_add.name = 'data_format'
+        # 2-->STRING
+        attrs_add.type = 2
+        attrs_add.s = 'AnyLayout'
+
+        attrs_add = desc_ops_add.attrs.add()
+        attrs_add.name = 'use_mkldnn'
+        # boolean
+        attrs_add.type = 6
+        attrs_add.b = 0
+
+        attrs_add = desc_ops_add.attrs.add()
+        attrs_add.name = 'use_cudnn'
+        # boolean
+        attrs_add.type = 6
+        attrs_add.b = 1
+
+        attrs_add = desc_ops_add.attrs.add()
+        attrs_add.name = 'dilations'
+        # ints
+        attrs_add.type = 3
+        attrs_add.ints.append(1)
+        attrs_add.ints.append(1)
+
+        attrs_add = desc_ops_add.attrs.add()
+        attrs_add.name = 'axis'
+        # int
+        attrs_add.type = 0
+        attrs_add.i = 1
+
+        if 'param' in layer:
+            l_params = layer['param']
+
+            attrs_add = desc_ops_add.attrs.add()
+            attrs_add.name = 'paddings'
+            # ints
+            attrs_add.type = 3
+            attrs_add.ints.append(l_params[types.fusion_conv_add_attrs_dict.get('paddings')])
+            attrs_add.ints.append(l_params[types.fusion_conv_add_attrs_dict.get('paddings')])
+
+            attrs_add = desc_ops_add.attrs.add()
+            attrs_add.name = 'strides'
+            # ints
+            attrs_add.type = 3
+            attrs_add.ints.append(l_params[types.fusion_conv_add_attrs_dict.get('strides')])
+            attrs_add.ints.append(l_params[types.fusion_conv_add_attrs_dict.get('strides')])
+
+            attrs_add = desc_ops_add.attrs.add()
+            attrs_add.name = 'groups'
+            # int
+            attrs_add.type = 0
+            attrs_add.i = l_params[types.fusion_conv_add_attrs_dict.get('groups')]
+            # attrs_add.i = 1
+
+        #
+        # op_attrs_tupl = types.op_io_dict.get(desc_ops_add.type) \
+        #     .get(types.mdl_attrs_key)
+        #
+        #
+        #
+        #
+        # # group stride padding
+        # print '----------------------'
+        # for i, val in enumerate(op_attrs_tupl):
+        #     attrs_add = desc_ops_add.attrs.add()
+        #     attr_name = op_attrs_tupl[i]
+        #     print attr_name
+        #     attrs_add.name = attr_name
+        #     attrs_add.type = types.fluid_attrs_type_dict.get(attr_name)
+        #     attrs_add.
+        #     print l_params[types.fusion_conv_add_attrs_dict.get(attr_name)]
+
+        # for p in l_params:
+        #     attrs_add = desc_ops_add.attrs.add()
+
+    @staticmethod
+    def package_ops_inputs(desc_ops_add, layer):
+        l_inputs = layer['input']
+        for i in l_inputs:
+            inputs_add = desc_ops_add.inputs.add()
+            # print i
+            inputs_add.parameter = types.op_io_dict.get(desc_ops_add.type).get(types.mdl_inputs_key)
+            inputs_add.arguments.append(i)
+
+    @staticmethod
+    def package_ops_outputs(desc_ops_add, layer):
+        l_outputs = layer['output']
+        for o in l_outputs:
+            # print o
+            outputs_add = desc_ops_add.outputs.add()
+            outputs_add.parameter = types.op_io_dict.get(desc_ops_add.type).get(types.mdl_outputs_key)
+            outputs_add.arguments.append(o)
+
+    def package_ops_weight2inputs(self, desc_ops_add, layer):
+        l_weights = layer['weight']
+        for w in l_weights:
+            self.weight_list_.append(w)
+
+        if layer['type'] == 'DepthwiseConvolutionLayer':
+            # print l_weights[0]
+            self.deepwise_weight_list_.append(l_weights[0])
+
+        op_weight_tup = types.op_io_dict.get(desc_ops_add.type).get(types.mdl_weight_key)
+        # print len(op_weight_tup)
+        for i, val in enumerate(op_weight_tup):
+            # print i
+            # print val
+            inputs_add = desc_ops_add.inputs.add()
+            inputs_add.parameter = op_weight_tup[i]
+            inputs_add.arguments.append(l_weights[i])
+
+        # for w in l_weights:
+        #     inputs_add = desc_ops_add.inputs.add()
+        #     # print w
+        #     inputs_add.parameter = op_weight_tup[0]
+        #     inputs_add.arguments.append(w)
+
+    @staticmethod
+    def package_ops_type(desc_ops_add, layer):
+        l_type = layer['type']
+        # print l_type
+        # print mdl2fluid_op_layer_dict.get(l_type)
+        desc_ops_add.type = types.mdl2fluid_op_layer_dict.get(l_type)
+
+    def package_vars(self, block_desc):
+        vars_add = block_desc.vars.add()
+        vars_add.name = 'feed'
+        vars_add.type.type = 9  # 9 is FEED_MINIBATCH
+        vars_add.persistable = 1
+        # fetch
+        vars_add = block_desc.vars.add()
+        vars_add.name = 'fetch'
+        vars_add.type.type = 10  # 10 is fetch list
+        vars_add.persistable = 1
+
+        json_matrix_ = self.mdl_json['matrix']
+        # print json_matrix_
+        for j in json_matrix_:
+            vars_add = block_desc.vars.add()
+            vars_add.name = j
+            vars_add.type.type = 7  # 7 is lodtensor
+            # print j
+            tensor = vars_add.type.lod_tensor.tensor
+            tensor.data_type = 5  # 5 is FP32
+
+            # print json_matrix_
+
+            dims_of_matrix = json_matrix_.get(j)
+            # dims_size = len(dims_of_matrix)
+            # print dims_size
+
+            # if dims_size == 4:
+            #     tensor.dims.append(dims_of_matrix[0])  # N
+            #     tensor.dims.append(dims_of_matrix[3])  # C
+            #     tensor.dims.append(dims_of_matrix[1])  # H
+            #     tensor.dims.append(dims_of_matrix[2])  # W
+            # else:
+
+            # issues in mdl model filter swich n and c
+            if j in self.deepwise_weight_list_ and len(dims_of_matrix) == 4:
+                print j
+                tensor.dims.append(dims_of_matrix[1])
+                tensor.dims.append(dims_of_matrix[0])
+                tensor.dims.append(dims_of_matrix[2])
+                tensor.dims.append(dims_of_matrix[3])
+                print tensor.dims
+            else:
+                for dims in dims_of_matrix:
+                    # print dims
+                    tensor.dims.append(dims)
+
+            if j in self.weight_list_:
+                vars_add.persistable = 1
+                dims_size = len(dims_of_matrix)
+                # print dims_size
+                if dims_size == 4:
+                    # convert weight from nhwc to nchw
+                    Swichter().nhwc2nchw_one_slice_add_head(
+                        '/Users/xiebaiyuan/PaddleProject/paddle-mobile/python/tools/mdl2fluid/multiobjects/float32s_nhwc/' + j + '.bin',
+                        '/Users/xiebaiyuan/PaddleProject/paddle-mobile/python/tools/mdl2fluid/multiobjects/float32s_nchw_with_head/' + j,
+                        '/Users/xiebaiyuan/PaddleProject/paddle-mobile/python/tools/mdl2fluid/multiobjects/float32s_nchw/' + j + '.tmp',
+                        dims_of_matrix[0],
+                        dims_of_matrix[1],
+                        dims_of_matrix[2],
+                        dims_of_matrix[3]
+                    )
+                else:
+                    Swichter().copy_add_head(
+                        '/Users/xiebaiyuan/PaddleProject/paddle-mobile/python/tools/mdl2fluid/multiobjects/float32s_nhwc/' + j + '.bin',
+                        '/Users/xiebaiyuan/PaddleProject/paddle-mobile/python/tools/mdl2fluid/multiobjects/float32s_nchw_with_head/' + j,
+                        '/Users/xiebaiyuan/PaddleProject/paddle-mobile/python/tools/mdl2fluid/multiobjects/float32s_nchw/' + j + '.tmp'
+                    )
+            else:
+                vars_add.persistable = 0
+
+
+mdl_path = "/Users/xiebaiyuan/PaddleProject/paddle-mobile/python/tools/mdl2fluid/multiobjects/YOLO_Universal.json"
+converter = Converter(mdl_path)
+converter.convert()

python/tools/mdl2fluid/model_reader.py

+import os
+
+import framework_pb2 as framework_pb2
+
+
+def read_model(model_path):
+    print('read_model.')
+    path_8 = unicode(model_path, 'utf8')
+
+    try:
+        with open(path_8, "rb") as f_model:
+            print get_file_size(model_path)
+            desc = framework_pb2.ProgramDesc()
+            desc.ParseFromString(f_model.read())
+            print desc
+            # print desc.blocks
+
+    except IOError:
+        print ": File not found.  Creating a new file."
+
+
+def get_file_size(file_path):
+    file_path = unicode(file_path, 'utf8')
+    fsize = os.path.getsize(file_path)
+    fsize = fsize / float(1024 * 1024)
+    return round(fsize, 2)
+
+
+path = "newyolo/__model__"
+read_model(path)

python/tools/mdl2fluid/op_types.py

+# coding=utf-8
+
+# mdl layers
+layer_mdl_conv = 'ConvolutionLayer'
+layer_mdl_deepwise_conv = 'DepthwiseConvolutionLayer'
+layer_mdl_relu = 'ReluLayer'
+layer_mdl_pointwise_add = 'PointwiseConvolutionLayer'
+
+# fluid ops
+op_fluid_fusion_conv_add = 'fusion_conv_add'
+op_fluid_relu = 'relu'
+
+# dict mdk layer ---  fluid op
+mdl2fluid_op_layer_dict = {
+    layer_mdl_conv: op_fluid_fusion_conv_add,
+    layer_mdl_deepwise_conv: op_fluid_fusion_conv_add,
+    layer_mdl_relu: op_fluid_relu,
+    layer_mdl_pointwise_add: op_fluid_fusion_conv_add
+}
+
+mdl_outputs_key = "outputs"
+mdl_inputs_key = "inputs"
+mdl_weight_key = "weights"
+mdl_attrs_key = "params"
+
+# dict of mdl-input _out param  to fluid input out attrs
+fusion_conv_add_dict = {
+    mdl_inputs_key: 'Input',
+    mdl_outputs_key: 'Out',
+    mdl_weight_key: ('Filter', 'Y'),
+    mdl_attrs_key: (
+        # 'workspace_size_MB', 'use_mkldnn', 'use_cudnn', 'data_format','dilations',
+        # dilations =  [1,1]
+        'groups', 'paddings', 'strides'
+        # 'axis'
+    )
+}
+
+relu_dict = {
+    mdl_inputs_key: 'X',
+    mdl_outputs_key: 'Out',
+    mdl_weight_key: ()
+
+}
+# mdl layers  ---  fluid ops
+op_io_dict = {
+    'fusion_conv_add': fusion_conv_add_dict,
+    'relu': relu_dict
+}
+
+# fluid attr key  ---  mdl params key
+fusion_conv_add_attrs_dict = {
+    'paddings': 'pad',
+    'strides': 'stride',
+    'groups': 'group'
+}
+# fluid attr key  ---  mdl params key
+fluid_attrs_type_dict = {
+    'paddings': 0,
+    'strides': 6,
+    'groups': 6
+}
+
+# '': "bias_term",    是不是要add   目前 yolo的模型都是 bias_term = 1
+
+
+# attrs {
+#       name: "axis"
+#       type: INT
+#       i: 1
+#     }
+
+
+# attrs_name = {
+#     'name': "workspace_size_MB",
+#     'type': 'INT',
+#     'i': '4096'
+# }
+# attrs
+# {
+#     name: "data_format"
+#     type: STRING
+#     s: "AnyLayout"
+# }
+# attrs
+# {
+#     name: "use_mkldnn"
+#     type: BOOLEAN
+#     b: false
+# }
+# attrs
+# {
+#     name: "use_cudnn"
+#     type: BOOLEAN
+#     b: true
+# }
+# attrs
+# {
+#     name: "dilations"
+#     type: INTS
+#     ints: 1
+#     ints: 1
+# }
+# attrs
+# {
+#     name: "groups"
+#     type: INT
+#     i: 1
+# }
+# attrs
+# {
+#     name: "paddings"
+#     type: INTS
+#     ints: 0
+#     ints: 0
+# }
+# attrs
+# {
+#     name: "strides"
+#     type: INTS
+#     ints: 1
+#     ints: 1
+# }

python/tools/mdl2fluid/swicher.py

+from array import array
+
+
+class Swichter:
+    def __init__(self):
+        pass
+
+    def nhwc2nchw_one_slice(self, from_file_name, to_file_name, batch, channel, height, width):
+        from_file = open(from_file_name, "rb")
+        to_file = open(to_file_name, "wb")
+
+        float_array = array("f")
+        float_array.fromfile(from_file, width * height * batch * channel)
+        float_write_array = array("f")
+
+        for b in range(batch):
+            for c in range(channel):
+                for h in range(height):
+                    for w in range(width):
+                        float_value = float_array[b * channel * width * height
+                                                  + channel * (h * width + w) + c]
+
+                        float_write_array.append(float_value)
+
+        float_write_array.tofile(to_file)
+        from_file.close()
+        to_file.close()
+
+    def copy(self, from_file_name, to_file_name):
+        from_file = open(from_file_name, "rb")
+        to_file = open(to_file_name, "wb")
+
+        to_file.write(from_file.read())
+        from_file.close()
+        to_file.close()
+
+    def nhwc2nchw_one_slice_add_head(self, from_file_name, to_file_name, tmp_file_name, batch, channel, height, width):
+        from_file = open(from_file_name, "rb")
+        tmp_file = open(tmp_file_name, "wb+")
+        float_array = array("f")
+        float_array.fromfile(from_file, width * height * batch * channel)
+        float_write_array = array("f")
+
+        for b in range(batch):
+            for c in range(channel):
+                for h in range(height):
+                    for w in range(width):
+                        float_value = float_array[b * channel * width * height
+                                                  + channel * (h * width + w) + c]
+
+                        float_write_array.append(float_value)
+
+        float_write_array.tofile(tmp_file)
+        tmp_file.close()
+        from_file.close()
+
+        tmp_file = open(tmp_file_name, "rb")
+        to_file = open(to_file_name, "wb")
+
+        tmp = tmp_file.read()
+        head = self.read_head('/Users/xiebaiyuan/PaddleProject/paddle-mobile/python/tools/mdl2fluid/yolo/conv1_biases')
+        to_file.write(head)
+        to_file.write(tmp)
+        tmp_file.close()
+        to_file.close()
+
+    def read_head(self, head_file):
+        from_file = open(head_file, "rb")
+        read = from_file.read(20)
+        # print read
+        from_file.close()
+        # print read
+        return read
+
+    def copy_add_head(self, from_file_name, to_file_name, tmp_file_name):
+        from_file = open(from_file_name, "rb")
+        to_file = open(to_file_name, "wb")
+        # tmp_file = open(tmp_file_name, "wb")
+
+        head = self.read_head('/Users/xiebaiyuan/PaddleProject/paddle-mobile/python/tools/mdl2fluid/yolo/conv1_biases')
+        to_file.write(head)
+        to_file.write(from_file.read())
+        from_file.close()
+        to_file.close()
+        pass
+
+
+# Swichter().nhwc2nchw_one_slice(
+#     '/Users/xiebaiyuan/PaddleProject/paddle-mobile/python/tools/mdl2fluid/multiobjects/float32s_nhwc/conv5_6_dw_0.bin',
+#     '/Users/xiebaiyuan/PaddleProject/paddle-mobile/python/tools/mdl2fluid/multiobjects/float32s_nchw/conv5_6_dw_0', 1,
+#     512, 3, 3)
+Swichter().read_head('/Users/xiebaiyuan/PaddleProject/paddle-mobile/python/tools/mdl2fluid/yolo/conv1_biases')

src/fpga/api.cpp

@@ -17,17 +17,21 @@ limitations under the License. */
 #include <sys/ioctl.h>
 #include <sys/mman.h>
 #include <algorithm>
+#include <map>
 #include "bias_scale.h"
 #include "filter.h"
 #include "image.h"
 #define FPGA_TEST_MODE
-//#define PADDLE_MOBILE_OS_LINUX
+#define PADDLE_MOBILE_OS_LINUX
 
 namespace paddle_mobile {
 namespace fpga {
 
 static int fd = -1;
 static const char *device_path = "/dev/fpgadrv0";
+#ifdef PADDLE_MOBILE_OS_LINUX
+static std::map<void *, size_t> memory_map;
+#endif
 
 static inline int do_ioctl(int req, const void *arg) {
 #ifdef PADDLE_MOBILE_OS_LINUX
@@ -48,10 +52,13 @@ int open_device() {
 
 // memory management;
 void *fpga_malloc(size_t size) {
-  DLOG << size << " bytes allocated";
+  static uint64_t counter = 0;
+  counter += size;
+  DLOG << size << " bytes allocated. Total " << counter << " bytes";
 #ifdef PADDLE_MOBILE_OS_LINUX
-  return reinterpret_cast<void *>(
-      mmap64(NULL, size, PROT_READ | PROT_WRITE, MAP_SHARED, fd, 0));
+  auto ptr = mmap64(nullptr, size, PROT_READ | PROT_WRITE, MAP_SHARED, fd, 0);
+  memory_map.insert(std::make_pair(ptr, size));
+  return ptr;
 #else
   return malloc(size);
 #endif
@@ -59,7 +66,16 @@ void *fpga_malloc(size_t size) {
 
 void fpga_free(void *ptr) {
 #ifdef PADDLE_MOBILE_OS_LINUX
-  munmap(ptr, 0);
+  static uint64_t counter = 0;
+  size_t size = 0;
+  auto iter = memory_map.find(ptr);  // std::map<void *, size_t>::iterator
+  if (iter != memory_map.end()) {
+    size = iter->second;
+    munmap(ptr, size);
+    memory_map.erase(iter);
+  }
+  counter += size;
+  DLOG << size << " bytes freed. Total " << counter << " bytes";
 #else
   free(ptr);
 #endif