fix conflix

mace/kernels/opencl/addn.cc

@@ -6,6 +6,7 @@
 #include "mace/core/runtime/opencl/opencl_runtime.h"
 #include "mace/kernels/opencl/helper.h"
 #include "mace/utils/utils.h"
+#include "mace/utils/tuner.h"
 
 namespace mace {
 namespace kernels {
@@ -33,8 +34,6 @@ static void AddN(const std::vector<const Tensor *> &input_tensors,
   built_options.emplace("-DINPUT_NUM=" + ToString(input_tensors.size()));
   auto addn_kernel = runtime->BuildKernel("addn", "addn", built_options);
 
-  const uint32_t lws = runtime->GetKernelMaxWorkGroupSize(addn_kernel);
-
   uint32_t idx = 0;
   for (auto input : input_tensors) {
   addn_kernel.setArg(idx++,
@@ -42,12 +41,47 @@ static void AddN(const std::vector<const Tensor *> &input_tensors,
   }
   addn_kernel.setArg(idx++, *(static_cast<cl::Image2D *>(output->buffer())));
 
+  const uint32_t gws[2] = {
+      static_cast<uint32_t>(width_pixels),
+      static_cast<uint32_t>(batch_height_pixels)
+  };
+  const uint32_t kwg_size = runtime->GetKernelMaxWorkGroupSize(addn_kernel);
+  std::vector<uint32_t> lws = {64, 16};
+  auto params_generator = [&]() -> std::vector<std::vector<uint32_t>> {
+    uint32_t local_ws[2];
+    local_ws[0] = std::min<uint32_t>(width_pixels, kwg_size);
+    local_ws[1] = std::min<uint32_t>(batch_height_pixels, kwg_size / local_ws[0]);
+    return {{local_ws[0], local_ws[1]},
+            {kwg_size / 16, 16},
+            {kwg_size / 32, 32},
+            {kwg_size / 64, 64},
+            {kwg_size / 128, 128},
+            {kwg_size / 256, 256},
+            {kwg_size, 1},
+            {1, kwg_size}
+    };
+  };
+  auto func = [&](const std::vector<uint32_t> &params) -> cl_int {
     cl_int error = runtime->command_queue().enqueueNDRangeKernel(
         addn_kernel, cl::NullRange,
-      cl::NDRange(width_pixels, batch_height_pixels),
-      cl::NDRange(64, 16),  // TODO fix this
-      nullptr, OpenCLRuntime::Get()->GetDefaultEvent());
-  MACE_CHECK(error == CL_SUCCESS) << "error code: " << error;
+        cl::NDRange(gws[0], gws[1]),
+        cl::NDRange(params[0], params[1]),
+        NULL, OpenCLRuntime::Get()->GetDefaultEvent());
+
+    MACE_CHECK(error == CL_SUCCESS) << "Error code: " << error;
+    return error;
+  };
+  std::stringstream ss;
+  ss << "addn_opencl_kernel_"
+     << output->dim(0) << "_"
+     << output->dim(1) << "_"
+     << output->dim(2) << "_"
+     << output->dim(3);
+  Tuner<uint32_t>::Get()->template TuneOrRun<cl_int>(ss.str(),
+                                                     lws,
+                                                     params_generator,
+                                                     func);
+
 }
 
 template <typename T>

mace/kernels/opencl/batch_norm_opencl.cc

@@ -48,8 +48,13 @@ void BatchNormFunctor<DeviceType::OPENCL, T>::operator()(
                            static_cast<uint32_t>(height * batch)};
   const std::vector<uint32_t> lws = {8, 16, 8};
   const uint32_t kwg_size = runtime->GetKernelMaxWorkGroupSize(bm_kernel);
-  auto params_generator = [&kwg_size]() -> std::vector<std::vector<uint32_t>> {
+  auto params_generator = [&]() -> std::vector<std::vector<uint32_t>> {
+    std::vector<uint32_t> local_ws(3, 0);
+    local_ws[0] = std::min<uint32_t>(channel_blocks, kwg_size);
+    local_ws[1] = std::min<uint32_t>(width, kwg_size / local_ws[0]);
+    local_ws[2] = std::min<uint32_t>(height * batch, kwg_size / (local_ws[0] * local_ws[1]));
     return {{8, 128, 1}, //SNPE size
+            {local_ws[0], local_ws[1], local_ws[2]},
             {kwg_size / 16, 4, 4},
             {kwg_size / 32, 4, 8},
             {kwg_size / 32, 8, 4},

mace/kernels/opencl/concat.cc

@@ -6,6 +6,7 @@
 #include "mace/core/runtime/opencl/opencl_runtime.h"
 #include "mace/kernels/opencl/helper.h"
 #include "mace/utils/utils.h"
+#include "mace/utils/tuner.h"
 
 namespace mace {
 namespace kernels {
@@ -41,21 +42,57 @@ static void Concat2(const Tensor *input0,
   concat_kernel.setArg(idx++, static_cast<int32_t>(input0->dim(3)));
   concat_kernel.setArg(idx++, *(static_cast<cl::Image2D *>(output->buffer())));
 
+  const uint32_t gws[3] = {
+      static_cast<uint32_t>(channel_blk),
+      static_cast<uint32_t>(width),
+      static_cast<uint32_t>(batch * height),
+  };
   const uint32_t kwg_size = runtime->GetKernelMaxWorkGroupSize(concat_kernel);
-
-  uint32_t lws[3] = {8, 16, 8};
-//  lws[0] = std::min<uint32_t>(channel_blk, kwg_size);
-//  lws[1] = std::min<uint32_t>(width, kwg_size / lws[0]);
-//  lws[2] = std::min<uint32_t>(height * batch, kwg_size / (lws[0] * lws[1]));
-
+  std::vector<uint32_t> lws = {8, 16, 8};
+  auto params_generator = [&]() -> std::vector<std::vector<uint32_t>> {
+    std::vector<uint32_t> local_ws(3, 0);
+    local_ws[0] = std::min<uint32_t>(channel_blk, kwg_size);
+    local_ws[1] = std::min<uint32_t>(width, kwg_size / local_ws[0]);
+    local_ws[2] = std::min<uint32_t>(height * batch, kwg_size / (local_ws[0] * local_ws[1]));
+    return {{4, 15, 8}, //SNPE size
+            {local_ws[0], local_ws[1], local_ws[2]},
+            {kwg_size / 16, 4, 4},
+            {kwg_size / 32, 4, 8},
+            {kwg_size / 32, 8, 4},
+            {kwg_size / 64, 8, 8},
+            {kwg_size / 64, 16, 4},
+            {kwg_size / 128, 8, 16},
+            {kwg_size / 128, 16, 8},
+            {kwg_size / 128, 32, 4},
+            {1, kwg_size / 32, 32},
+            {1, kwg_size / 64, 64},
+            {1, kwg_size / 128, 128},
+            {3, 15, 9},
+            {7, 15, 9},
+            {9, 7, 15},
+            {15, 7, 9},
+            {1, kwg_size, 1}};
+  };
+  auto func = [&](const std::vector<uint32_t> &params) -> cl_int {
     cl_int error = runtime->command_queue().enqueueNDRangeKernel(
         concat_kernel, cl::NullRange,
-      cl::NDRange(static_cast<uint32_t>(channel_blk),
-                  static_cast<uint32_t>(width),
-                  static_cast<uint32_t>(height * batch)),
-      cl::NDRange(lws[0], lws[1], lws[2]),
+        cl::NDRange(gws[0], gws[1], gws[2]),
+        cl::NDRange(params[0], params[1], params[2]),
         NULL, OpenCLRuntime::Get()->GetDefaultEvent());
-  MACE_CHECK(error == CL_SUCCESS);
+
+    MACE_CHECK(error == CL_SUCCESS) << "Error code: " << error;
+    return error;
+  };
+  std::stringstream ss;
+  ss << "concat_opencl_kernel_"
+     << output->dim(0) << "_"
+     << output->dim(1) << "_"
+     << output->dim(2) << "_"
+     << output->dim(3);
+  Tuner<uint32_t>::Get()->template TuneOrRun<cl_int>(ss.str(),
+                                                     lws,
+                                                     params_generator,
+                                                     func);
 }
 
 template<typename T>

mace/kernels/opencl/conv_2d_opencl_1x1.cc

@@ -68,8 +68,13 @@ void Conv1x1(const Tensor *input,
                            static_cast<uint32_t>(height * batch)};
   const std::vector<uint32_t> lws = {8, 15, 8};
   const uint32_t kwg_size = runtime->GetKernelMaxWorkGroupSize(conv_2d_kernel);
-  auto params_generator = [&kwg_size]()->std::vector<std::vector<uint32_t>> {
+  auto params_generator = [&]()->std::vector<std::vector<uint32_t>> {
+    std::vector<uint32_t> local_ws(3, 0);
+    local_ws[0] = std::min<uint32_t>(channel_blocks, kwg_size);
+    local_ws[1] = std::min<uint32_t>(width_blocks, kwg_size / local_ws[0]);
+    local_ws[2] = std::min<uint32_t>(height * batch, kwg_size / (local_ws[0] * local_ws[1]));
     return {{4, 15, 8}, //SNPE size
+            {local_ws[0], local_ws[1], local_ws[2]},
             {kwg_size/16, 4, 4},
             {kwg_size/32, 4, 8},
             {kwg_size/32, 8, 4},

mace/kernels/opencl/conv_2d_opencl_3x3.cc

@@ -60,8 +60,13 @@ static void Conv2d3x3S12(const Tensor *input, const Tensor *filter,
                            static_cast<uint32_t>(height * batch)};
   const std::vector<uint32_t> lws = {4, 15, 8};
   const uint32_t kwg_size = runtime->GetKernelMaxWorkGroupSize(conv_2d_kernel);
-  auto params_generator = [&kwg_size]() -> std::vector<std::vector<uint32_t>> {
+  auto params_generator = [&]() -> std::vector<std::vector<uint32_t>> {
+    std::vector<uint32_t> local_ws(3, 0);
+    local_ws[0] = std::min<uint32_t>(channel_blocks, kwg_size);
+    local_ws[1] = std::min<uint32_t>(width_blocks, kwg_size / local_ws[0]);
+    local_ws[2] = std::min<uint32_t>(height * batch, kwg_size / (local_ws[0] * local_ws[1]));
     return {{4, 15, 8}, //SNPE size
+            {local_ws[0], local_ws[1], local_ws[2]},
             {kwg_size / 16, 4, 4},
             {kwg_size / 32, 4, 8},
             {kwg_size / 32, 8, 4},

mace/kernels/opencl/conv_2d_opencl_general.cc

@@ -62,8 +62,13 @@ void Conv2dOpencl(const Tensor *input, const Tensor *filter,
                            static_cast<uint32_t>(height * batch)};
   const std::vector<uint32_t> lws = {8, 16, 8};
   const uint32_t kwg_size = runtime->GetKernelMaxWorkGroupSize(conv_2d_kernel);
-  auto params_generator = [&kwg_size]() -> std::vector<std::vector<uint32_t>> {
+  auto params_generator = [&]() -> std::vector<std::vector<uint32_t>> {
+    std::vector<uint32_t> local_ws(3, 0);
+    local_ws[0] = std::min<uint32_t>(channel_blocks, kwg_size);
+    local_ws[1] = std::min<uint32_t>(width_blocks, kwg_size / local_ws[0]);
+    local_ws[2] = std::min<uint32_t>(height * batch, kwg_size / (local_ws[0] * local_ws[1]));
     return {{4, 15, 8}, //SNPE size
+            {local_ws[0], local_ws[1], local_ws[2]},
             {kwg_size / 16, 4, 4},
             {kwg_size / 32, 4, 8},
             {kwg_size / 32, 8, 4},

mace/kernels/opencl/pooling_opencl.cc

@@ -6,6 +6,7 @@
 #include "mace/core/runtime/opencl/cl2_header.h"
 #include "mace/core/runtime/opencl/opencl_runtime.h"
 #include "mace/kernels/opencl/helper.h"
+#include "mace/utils/tuner.h"
 
 namespace mace {
 namespace kernels {
@@ -23,11 +24,6 @@ static void Pooling(const Tensor *input,
   index_t channels = output->dim(3);
 
   index_t channel_blocks = (channels + 3) / 4;
-  const uint32_t gws[3] = {
-      static_cast<uint32_t>(channel_blocks),
-      static_cast<uint32_t>(out_width),
-      static_cast<uint32_t>(batch * out_height),
-  };
 
   auto runtime = OpenCLRuntime::Get();
   std::set<std::string> built_options;
@@ -44,13 +40,6 @@ static void Pooling(const Tensor *input,
   }
   auto pooling_kernel = runtime->BuildKernel("pooling", "pooling", built_options);
 
-  const uint32_t kwg_size = runtime->GetKernelMaxWorkGroupSize(pooling_kernel);
-
-  uint32_t lws[3];
-  lws[0] = std::min<uint32_t>(channel_blocks, kwg_size);
-  lws[1] = std::min<uint32_t>(out_width, kwg_size / lws[0]);
-  lws[2] = std::min<uint32_t>(out_height * batch, kwg_size / (lws[0] * lws[1]));
-
   uint32_t idx = 0;
   pooling_kernel.setArg(idx++, *(static_cast<const cl::Image2D *>(input->buffer())));
   pooling_kernel.setArg(idx++, static_cast<int32_t>(input->dim(1)));
@@ -62,12 +51,60 @@ static void Pooling(const Tensor *input,
   pooling_kernel.setArg(idx++, pooling_size);
   pooling_kernel.setArg(idx++, *(static_cast<cl::Image2D *>(output->buffer())));
 
+  const uint32_t gws[3] = {
+      static_cast<uint32_t>(channel_blocks),
+      static_cast<uint32_t>(out_width),
+      static_cast<uint32_t>(batch * out_height),
+  };
+  const uint32_t kwg_size = runtime->GetKernelMaxWorkGroupSize(pooling_kernel);
+  std::vector<uint32_t> lws(3, 0);
+  lws[0] = std::min<uint32_t>(channel_blocks, kwg_size);
+  lws[1] = std::min<uint32_t>(out_width, kwg_size / lws[0]);
+  lws[2] = std::min<uint32_t>(out_height * batch, kwg_size / (lws[0] * lws[1]));
+  auto params_generator = [&]() -> std::vector<std::vector<uint32_t>> {
+    std::vector<uint32_t> local_ws(3, 0);
+    local_ws[0] = std::min<uint32_t>(channel_blocks, kwg_size);
+    local_ws[1] = std::min<uint32_t>(out_width, kwg_size / local_ws[0]);
+    local_ws[2] = std::min<uint32_t>(out_height * batch, kwg_size / (local_ws[0] * local_ws[1]));
+    return {{4, 15, 8}, //SNPE size
+            {local_ws[0], local_ws[1], local_ws[2]},
+            {kwg_size / 16, 4, 4},
+            {kwg_size / 32, 4, 8},
+            {kwg_size / 32, 8, 4},
+            {kwg_size / 64, 8, 8},
+            {kwg_size / 64, 16, 4},
+            {kwg_size / 128, 8, 16},
+            {kwg_size / 128, 16, 8},
+            {kwg_size / 128, 32, 4},
+            {1, kwg_size / 32, 32},
+            {1, kwg_size / 64, 64},
+            {1, kwg_size / 128, 128},
+            {3, 15, 9},
+            {7, 15, 9},
+            {9, 7, 15},
+            {15, 7, 9},
+            {1, kwg_size, 1}};
+  };
+  auto func = [&](const std::vector<uint32_t> &params) -> cl_int {
     cl_int error = runtime->command_queue().enqueueNDRangeKernel(
         pooling_kernel, cl::NullRange,
         cl::NDRange(gws[0], gws[1], gws[2]),
-      cl::NDRange(lws[0], lws[1], lws[2]),
+        cl::NDRange(params[0], params[1], params[2]),
         NULL, OpenCLRuntime::Get()->GetDefaultEvent());
-  MACE_CHECK(error == CL_SUCCESS) << error;
+
+    MACE_CHECK(error == CL_SUCCESS) << "Error code: " << error;
+    return error;
+  };
+  std::stringstream ss;
+  ss << "pooling_opencl_kernel_"
+     << output->dim(0) << "_"
+     << output->dim(1) << "_"
+     << output->dim(2) << "_"
+     << output->dim(3);
+  Tuner<uint32_t>::Get()->template TuneOrRun<cl_int>(ss.str(),
+                                                     lws,
+                                                     params_generator,
+                                                     func);
 }
 
 template<typename T>

mace/kernels/opencl/relu_opencl.cc

@@ -50,8 +50,13 @@ void ReluFunctor<DeviceType::OPENCL, T>::operator()(const Tensor *input,
                            static_cast<uint32_t>(height * batch)};
   const std::vector<uint32_t> lws = {8, 16, 8};
   const uint32_t kwg_size = runtime->GetKernelMaxWorkGroupSize(relu_kernel);
-  auto params_generator = [&kwg_size]() -> std::vector<std::vector<uint32_t>> {
+  auto params_generator = [&]() -> std::vector<std::vector<uint32_t>> {
+    std::vector<uint32_t> local_ws(3, 0);
+    local_ws[0] = std::min<uint32_t>(channel_blocks, kwg_size);
+    local_ws[1] = std::min<uint32_t>(width, kwg_size / local_ws[0]);
+    local_ws[2] = std::min<uint32_t>(height * batch, kwg_size / (local_ws[0] * local_ws[1]));
     return {{4, 15, 8}, //SNPE size
+            {local_ws[0], local_ws[1], local_ws[2]},
             {kwg_size / 16, 4, 4},
             {kwg_size / 32, 4, 8},
             {kwg_size / 32, 8, 4},

mace/kernels/opencl/resize_bilinear_opencl.cc

@@ -7,6 +7,7 @@
 #include "mace/kernels/resize_bilinear.h"
 #include "mace/kernels/opencl/helper.h"
 #include "mace/utils/utils.h"
+#include "mace/utils/tuner.h"
 
 namespace mace {
 namespace kernels {
@@ -44,8 +45,6 @@ void ResizeBilinearFunctor<DeviceType::OPENCL, T>::operator()(
   built_options.emplace("-DCMD_DATA_TYPE=" + DtToUpstreamCLCMDDt(dt));
   auto rb_kernel  = runtime->BuildKernel("resize_bilinear", "resize_bilinear_nocache", built_options);
 
-  const uint32_t kwg_size = runtime->GetKernelMaxWorkGroupSize(rb_kernel);
-
   uint32_t idx = 0;
   rb_kernel.setArg(idx++, *(static_cast<const cl::Image2D *>(input->buffer())));
   rb_kernel.setArg(idx++, *(static_cast<cl::Image2D *>(output->buffer())));
@@ -55,17 +54,52 @@ void ResizeBilinearFunctor<DeviceType::OPENCL, T>::operator()(
   rb_kernel.setArg(idx++, static_cast<int32_t>(in_width));
   rb_kernel.setArg(idx++, static_cast<int32_t>(out_height));
 
-  auto command_queue = runtime->command_queue();
-
-  cl_int error = command_queue.enqueueNDRangeKernel(
+  const uint32_t gws[3] = {static_cast<uint32_t>(channel_blocks),
+                           static_cast<uint32_t>(out_width),
+                           static_cast<uint32_t>(out_height * batch)};
+  const std::vector<uint32_t> lws = {8, 16, 8};
+  const uint32_t kwg_size = runtime->GetKernelMaxWorkGroupSize(rb_kernel);
+  auto params_generator = [&]() -> std::vector<std::vector<uint32_t>> {
+    std::vector<uint32_t> local_ws(3, 0);
+    local_ws[0] = std::min<uint32_t>(channel_blocks, kwg_size);
+    local_ws[1] = std::min<uint32_t>(out_width, kwg_size / local_ws[0]);
+    local_ws[2] = std::min<uint32_t>(out_height * batch, kwg_size / (local_ws[0] * local_ws[1]));
+    return {{4, 15, 8}, //SNPE size
+            {local_ws[0], local_ws[1], local_ws[2]},
+            {kwg_size / 16, 4, 4},
+            {kwg_size / 32, 4, 8},
+            {kwg_size / 32, 8, 4},
+            {kwg_size / 64, 8, 8},
+            {kwg_size / 64, 16, 4},
+            {kwg_size / 128, 8, 16},
+            {kwg_size / 128, 16, 8},
+            {kwg_size / 128, 32, 4},
+            {1, kwg_size / 32, 32},
+            {1, kwg_size / 64, 64},
+            {1, kwg_size / 128, 128},
+            {1, kwg_size, 1}};
+  };
+  auto func = [&](const std::vector<uint32_t> &params) -> cl_int {
+    cl_int error = runtime->command_queue().enqueueNDRangeKernel(
         rb_kernel, cl::NullRange,
-      cl::NDRange(static_cast<int32_t>(channel_blocks),
-                  static_cast<int32_t>(out_width),
-                  static_cast<int32_t>(out_height * batch)),
-      // TODO tuning
-      cl::NDRange(1, static_cast<int32_t>(out_width > kwg_size ? kwg_size : out_width), 1),
-      nullptr, OpenCLRuntime::Get()->GetDefaultEvent());
-  MACE_CHECK(error == CL_SUCCESS, error);
+        cl::NDRange(gws[0], gws[1], gws[2]),
+        cl::NDRange(params[0], params[1], params[2]),
+        NULL, OpenCLRuntime::Get()->GetDefaultEvent());
+
+    MACE_CHECK(error == CL_SUCCESS) << "Error code: " << error;
+    return error;
+  };
+  std::stringstream ss;
+  ss << "resize_bilinear_opencl_kernel_"
+     << output->dim(0) << "_"
+     << output->dim(1) << "_"
+     << output->dim(2) << "_"
+     << output->dim(3);
+  Tuner<uint32_t>::Get()->template TuneOrRun<cl_int>(ss.str(),
+                                                     lws,
+                                                     params_generator,
+                                                     func);
+
 }
 
 template struct ResizeBilinearFunctor<DeviceType::OPENCL, float>;

mace/python/tools/BUILD

@@ -8,6 +8,7 @@ py_library(
     ],
     srcs_version = "PY2AND3",
     deps = [
+        ":memory_optimizer",
         "//mace/proto:mace_py",
     ],
 )

mace/python/tools/memory_optimizer.py

@@ -65,7 +65,7 @@ class MemoryOptimizer(object):
             raise Exception('ref count is less than 0')
 
     for mem in self.mem_block:
-      arena = net_def.mem_arena
+      arena = self.net_def.mem_arena
       block = arena.mem_block.add()
       block.mem_id = mem
       block.x = self.mem_block[mem][0]
@@ -83,20 +83,7 @@ class MemoryOptimizer(object):
 
     print('origin mem: %d, optimized mem: %d', origin_mem_size, optimized_mem_size)
 
-if __name__ == '__main__':
-  model_file = sys.argv[1]
-  opt_model_file = sys.argv[2]
-  with open(model_file, "rb") as f:
-    net_def = mace_pb2.NetDef()
-    net_def.ParseFromString(f.read())
-    optimizer = MemoryOptimizer(net_def)
-    optimizer.optimize()
-
-    with open(opt_model_file, "wb") as f:
-      f.write(net_def.SerializeToString())
-    with open(opt_model_file + '_txt', "wb") as f:
-      net_def.ClearField('tensors')
-      f.write(str(net_def))
-
-
 
+def optimize_memory(net_def):
+  mem_optimizer = MemoryOptimizer(net_def)
+  mem_optimizer.optimize()
\ No newline at end of file

mace/python/tools/tf_converter_lib.py

 from mace.proto import mace_pb2
 import tensorflow as tf
 import numpy as np
-from mace.python.tools.convert_util import tf_dtype_2_mace_dtype
+from mace.python.tools import memory_optimizer
 
 # TODO: support NCHW formt, now only support NHWC.
 padding_mode = {
@@ -25,22 +25,10 @@ data_type_map = {
   'DT_FLOAT': mace_pb2.DT_FLOAT
 }
 
-def convert_tensor(op, tensor):
-  tf_tensor = op.outputs[0].eval()
-  tensor.name = op.outputs[0].name
+BATCH_NORM_ORDER = ["Add", "Rsqrt", "Mul", "Mul", "Mul", "Sub", "Add"]
 
-  shape = list(tf_tensor.shape)
-  tensor.dims.extend(shape)
-
-  tf_dt = op.get_attr('dtype')
-  if tf_dt == tf.float32:
-    tensor.data_type = mace_pb2.DT_FLOAT
-    tensor.float_data.extend(tf_tensor.astype(float).flat)
-  elif tf_dt == tf.int32:
-    tensor.data_type = mace_pb2.DT_INT32
-    tensor.int32_data.extend(tf_tensor.astype(np.int32).flat)
-  else:
-    raise Exception("Not supported tensor type: " + tf_dt.name)
+MACE_INPUT_NODE_NAME = "mace_input_node"
+MACE_OUTPUT_NODE_NAME = "mace_output_node"
 
 def get_input_tensor(op, index):
   input_tensor = op.inputs[index]
@@ -48,9 +36,26 @@ def get_input_tensor(op, index):
     input_tensor = get_input_tensor(input_tensor.op, 0)
   return input_tensor
 
-def add_buffer_to_image(input_name, input_type, dt, net_def):
+class TFConverter(object):
+  def __init__(self, tf_ops, net_def, dt, device):
+    self.net_def = net_def
+    self.tf_ops = tf_ops
+    self.dt = dt
+    self.device = device
+    self.tf_graph = {}
+    self.resolved_ops = {}
+
+    for op in tf_ops:
+      self.resolved_ops[op.name] = 0
+      for input in op.inputs:
+        input_name = input.name[:-2]
+        if input_name not in self.tf_graph:
+          self.tf_graph[input_name] = []
+        self.tf_graph[input_name].append(op)
+
+  def add_buffer_to_image(self, input_name, input_type):
     output_name = input_name[:-2] + "_b2i" + input_name[-2:]
-  op_def = net_def.op.add()
+    op_def = self.net_def.op.add()
     op_def.name = output_name[:-2]
     op_def.type = 'BufferToImage'
     op_def.input.extend([input_name])
@@ -64,28 +69,12 @@ def add_buffer_to_image(input_name, input_type, dt, net_def):
     arg.i = 0
     arg = op_def.arg.add()
     arg.name = 'T'
-  arg.i = dt
+    arg.i = self.dt
     return output_name
 
-def add_image_to_buffer(input_name, input_type, dt, net_def):
-  output_name = input_name[:-2] + "_i2b" + input_name[-2:]
-  op_def = net_def.op.add()
-  op_def.name = output_name[:-2]
-  op_def.type = 'ImageToBuffer'
-  op_def.input.extend([input_name])
-  op_def.output.extend([output_name])
-
-  arg = op_def.arg.add()
-  arg.name = 'buffer_type'
-  arg.i = buffer_type_map[input_type]
-  arg = op_def.arg.add()
-  arg.name = 'T'
-  arg.i = dt
-  return output_name
-
-def add_input_transform(name, dt, net_def):
-  new_input_name = "mace_input_node:0"
-  op_def = net_def.op.add()
+  def add_input_transform(self, name):
+    new_input_name = MACE_INPUT_NODE_NAME + ":0"
+    op_def = self.net_def.op.add()
     op_def.name = name
     op_def.type = 'BufferToImage'
     op_def.input.extend([new_input_name])
@@ -97,11 +86,11 @@ def add_input_transform(name, dt, net_def):
 
     arg = op_def.arg.add()
     arg.name = 'T'
-  arg.i = dt
+    arg.i = self.dt
 
-def add_output_transform(name, net_def):
-  output_name = "mace_output_node:0"
-  op_def = net_def.op.add()
+  def add_output_transform(self, name):
+    output_name = MACE_OUTPUT_NODE_NAME + ":0"
+    op_def = self.net_def.op.add()
     op_def.name = output_name[:-2]
     op_def.type = 'ImageToBuffer'
     op_def.input.extend([name+':0'])
@@ -111,197 +100,322 @@ def add_output_transform(name, net_def):
     epsilon_arg.name = 'buffer_type'
     epsilon_arg.i = buffer_type_map['IN_OUT']
 
-
-def convert_op_outputs(mace_op_def, tf_op):
-  mace_op_def.output.extend([output.name for output in tf_op.outputs])
-  mace_op_def.output_type.extend([tf_dtype_2_mace_dtype(output.dtype)
-                                  for output in tf_op.outputs])
+  @staticmethod
+  def add_output_shape(outputs, op):
     output_shapes = []
-  for output in tf_op.outputs:
+    for output in outputs:
+      if output.shape.num_elements() is not None:
         output_shape = mace_pb2.OutputShape()
         output_shape.dims.extend(output.shape.as_list())
         output_shapes.append(output_shape)
-  mace_op_def.output_shape.extend(output_shapes)
+    op.output_shape.extend(output_shapes)
 
+  def convert_tensor(self, op):
+    tensor = self.net_def.tensors.add()
+    tf_tensor = op.outputs[0].eval()
+    tensor.name = op.outputs[0].name
 
-def convert_ops(unresolved_ops, dt, net_def, device):
-  ops_count = len(unresolved_ops)
-  resolved_count = 1
-
-  first_op = unresolved_ops[0]
+    shape = list(tf_tensor.shape)
+    tensor.dims.extend(shape)
 
-  if first_op.type in ['Placeholder', 'Reshape', 'Identity']:
-    pass
-  elif first_op.type == 'Const':
-    tensor = net_def.tensors.add()
-    convert_tensor(first_op, tensor)
+    tf_dt = op.get_attr('dtype')
+    if tf_dt == tf.float32:
+      tensor.data_type = mace_pb2.DT_FLOAT
+      tensor.float_data.extend(tf_tensor.astype(np.float32).flat)
+    elif tf_dt == tf.int32:
+      tensor.data_type = mace_pb2.DT_INT32
+      tensor.int32_data.extend(tf_tensor.astype(np.int32).flat)
     else:
-    op_def = net_def.op.add()
+      raise Exception("Not supported tensor type: " + tf_dt.name)
+    self.resolved_ops[op.name] = 1
+
+  def convert_conv2d(self, op):
+    op_def = mace_pb2.OperatorDef()
     arg = op_def.arg.add()
     arg.name = 'T'
-    arg.i = dt
-
-    if first_op.type == 'Conv2D' or first_op.type == 'DepthwiseConv2dNative':
-      op_def.name = first_op.name
-      if first_op.type == 'DepthwiseConv2dNative':
+    arg.i = self.dt
+    op_def.name = op.name
+    if op.type == 'DepthwiseConv2dNative':
       op_def.type = 'DepthwiseConv2d'
     else:
-        op_def.type = first_op.type
-      if device == 'gpu':
-        op_def.input.extend([first_op.inputs[0].name])
-        output_name = add_buffer_to_image(first_op.inputs[1].name, "FILTER", dt, net_def)
+      op_def.type = op.type
+    if self.device == 'gpu':
+      op_def.input.extend([op.inputs[0].name])
+      output_name = self.add_buffer_to_image(op.inputs[1].name, "FILTER")
       op_def.input.extend([output_name])
     else:
-        op_def.input.extend([input.name for input in first_op.inputs])
+      op_def.input.extend([input.name for input in op.inputs])
 
     padding_arg = op_def.arg.add()
     padding_arg.name = 'padding'
-      padding_arg.i = padding_mode[first_op.get_attr('padding')]
+    padding_arg.i = padding_mode[op.get_attr('padding')]
     strides_arg = op_def.arg.add()
     strides_arg.name = 'strides'
-      strides_arg.ints.extend(first_op.get_attr('strides')[1:3])
+    strides_arg.ints.extend(op.get_attr('strides')[1:3])
     data_format_arg = op_def.arg.add()
     data_format_arg.name = 'data_format'
     data_format_arg.s = 'NHWC'
-      final_op = first_op
-
-      if ops_count >= 3 and unresolved_ops[1].type == 'Const' and unresolved_ops[2].type == 'BiasAdd' :
-        bias_tensor = unresolved_ops[1]
-        tensor = net_def.tensors.add()
-        convert_tensor(bias_tensor, tensor)
+    final_op = op
+    self.resolved_ops[op.name] = 1
 
-        bias_add_op = unresolved_ops[2]
-        if device == 'gpu':
-          output_name = add_buffer_to_image(bias_add_op.inputs[1].name, "ARGUMENT", dt, net_def)
+    if len(self.tf_graph[op.name]) == 1 and self.tf_graph[op.name][0].type == 'BiasAdd' :
+      bias_add_op = self.tf_graph[op.name][0]
+      if self.device == 'gpu':
+        output_name = self.add_buffer_to_image(bias_add_op.inputs[1].name, "ARGUMENT")
         op_def.input.extend([output_name])
       else:
         op_def.input.extend([bias_add_op.inputs[1].name])
       final_op = bias_add_op
-        resolved_count = 3
+      self.resolved_ops[bias_add_op.name] = 1
 
-      if ops_count >= 4 and unresolved_ops[3].type == 'Relu':
-        relu_op = unresolved_ops[3];
+    if len(self.tf_graph[final_op.name]) == 1 \
+        and self.tf_graph[final_op.name][0].type == 'Relu':
+      relu_op = self.tf_graph[final_op.name][0]
       op_def.type = "FusedConv2D"
       final_op = relu_op
-        resolved_count = 4
+      self.resolved_ops[relu_op.name] = 1
 
-      convert_op_outputs(op_def, final_op)
+    op_def.output.extend([output.name for output in final_op.outputs])
+    self.add_output_shape(final_op.outputs, op_def)
+    self.net_def.op.extend([op_def])
 
-    elif first_op.type == 'FusedBatchNorm':
-      op_def.name = first_op.name
+  def convert_fused_batchnorm(self, op):
+    op_def = mace_pb2.OperatorDef()
+    arg = op_def.arg.add()
+    arg.name = 'T'
+    arg.i = self.dt
+    op_def.name = op.name
     op_def.type = 'BatchNorm'
-      if device == 'gpu':
-        op_def.input.extend([first_op.inputs[0].name])
-        for i in range(1, len(first_op.inputs)):
-          output_name = add_buffer_to_image(first_op.inputs[i].name, "ARGUMENT", dt, net_def)
+    if self.device == 'gpu':
+      op_def.input.extend([op.inputs[0].name])
+      for i in range(1, len(op.inputs)):
+        output_name = self.add_buffer_to_image(op.inputs[i].name, "ARGUMENT")
         op_def.input.extend([output_name])
     else:
-        op_def.input.extend([input.name for input in first_op.inputs])
-      op_def.output.extend([first_op.outputs[0].name])
+      op_def.input.extend([input.name for input in op.inputs])
+    op_def.output.extend([op.outputs[0].name])
 
-      output_shape = mace_pb2.OutputShape()
-      output_shape.dims.extend(first_op.outputs[0].shape.as_list())
-      op_def.output_shape.extend([output_shape])
+    self.add_output_shape(op.outputs, op_def)
 
     epsilon_arg = op_def.arg.add()
     epsilon_arg.name = 'epsilon'
-      epsilon_arg.f = first_op.get_attr('epsilon')
+    epsilon_arg.f = op.get_attr('epsilon')
     data_format_arg = op_def.arg.add()
     data_format_arg.name = 'data_format'
     data_format_arg.s = 'NHWC'
-    elif first_op.type == 'Add' and first_op.name.endswith(
-        'batchnorm/add') and ops_count > 7:
-      add_op = first_op
-      mul_op = unresolved_ops[2]
-      mul_1_op = unresolved_ops[3]
-      mul_2_op = unresolved_ops[4]
-      sub_op = unresolved_ops[5]
-      add_1_op = unresolved_ops[6]
-      # print (mul_op.type, mul_2_op.type, mul_1_op.type, sub_op.type)
-      if mul_op.type != 'Mul' or mul_2_op.type != 'Mul' or \
-                      mul_1_op.type != 'Mul' or sub_op.type != 'Sub' or add_1_op.type != 'Add':
+    self.resolved_ops[op.name] = 1
+    self.net_def.op.extend([op_def])
+
+  def convert_batchnorm(self, op):
+    bn_ops = []
+    bn_ops.append(op)
+    for i in range(1, 3):
+      if len(self.tf_graph[bn_ops[i-1].name]) == 1 \
+          and self.tf_graph[bn_ops[i-1].name][0].type == BATCH_NORM_ORDER[i]:
+        bn_ops.append(self.tf_graph[bn_ops[i-1].name][0])
+      else:
+        raise Exception('Invalid BatchNorm Op')
+    if len(self.tf_graph[bn_ops[2].name]) == 2 \
+        and self.tf_graph[bn_ops[2].name][0].type == BATCH_NORM_ORDER[3] \
+        and self.tf_graph[bn_ops[2].name][1].type == BATCH_NORM_ORDER[4]:
+      bn_ops.append(self.tf_graph[bn_ops[2].name][0])
+      bn_ops.append(self.tf_graph[bn_ops[2].name][1])
+    else:
       raise Exception('Invalid BatchNorm Op')
+    bn_ops.append(self.tf_graph[bn_ops[4].name][0])
+    bn_ops.append(self.tf_graph[bn_ops[3].name][0])
+
+    op_def = mace_pb2.OperatorDef()
+    arg = op_def.arg.add()
+    arg.name = 'T'
+    arg.i = self.dt
 
-      get_input_tensor(mul_1_op, 0)
-      input_name = get_input_tensor(mul_1_op, 0).name
-      gamma = get_input_tensor(mul_op, 1).name
-      beta = get_input_tensor(sub_op, 0).name
-      mean = get_input_tensor(mul_2_op, 0).name
-      variance = get_input_tensor(add_op, 0).name
-      epsilon = get_input_tensor(add_op, 1).name
+    input_name = get_input_tensor(bn_ops[3], 0).name
+    gamma = get_input_tensor(bn_ops[2], 1).name
+    beta = get_input_tensor(bn_ops[5], 0).name
+    mean = get_input_tensor(bn_ops[4], 0).name
+    variance = get_input_tensor(bn_ops[0], 0).name
 
-      op_def.name = first_op.name[:-4]  # remove /add
+    op_def.name = op.name[:-4]  # remove /add
     op_def.type = 'BatchNorm'
-      op_def.input.extend([input_name, gamma, beta, mean, variance, epsilon])
-      convert_op_outputs(op_def, add_1_op)
+    if self.device == 'gpu':
+      op_def.input.extend([input_name])
+      for tensor_name in [gamma, beta, mean, variance]:
+        output_name = self.add_buffer_to_image(tensor_name, "ARGUMENT")
+        op_def.input.extend([output_name])
+    else:
+      op_def.input.extend([input_name, gamma, beta, mean, variance])
+    op_def.output.extend([output.name for output in bn_ops[6].outputs])
+    self.add_output_shape(bn_ops[6].outputs, op_def)
+    epsilon_arg = op_def.arg.add()
+    epsilon_arg.name = 'epsilon'
+    epsilon_arg.f = get_input_tensor(op, 1).eval().astype(np.float)
+    data_format_arg = op_def.arg.add()
+    data_format_arg.name = 'data_format'
+    data_format_arg.s = 'NHWC'
 
-      resolved_count = 7
-    elif first_op.type == 'Relu6':
-      op_def.name = first_op.name
-      op_def.type = 'Relu'
-      op_def.input.extend([input.name for input in first_op.inputs])
-      convert_op_outputs(op_def, first_op)
+    self.net_def.op.extend([op_def])
+    for i in range(0, 7):
+      self.resolved_ops[bn_ops[i].name] = 1
 
-      max_limit_arg = op_def.arg.add()
-      max_limit_arg.name = 'max_limit'
-      max_limit_arg.f = 6
-    elif first_op.type == 'AvgPool' or first_op.type == 'MaxPool':
-      op_def.name = first_op.name
+  def convert_pooling(self, op):
+    op_def = self.net_def.op.add()
+    arg = op_def.arg.add()
+    arg.name = 'T'
+    arg.i = self.dt
+    op_def.name = op.name
     op_def.type = 'Pooling'
-      op_def.input.extend([input.name for input in first_op.inputs])
-      convert_op_outputs(op_def, first_op)
-
+    op_def.input.extend([input.name for input in op.inputs])
+    op_def.output.extend([output.name for output in op.outputs])
+    self.add_output_shape(op.outputs, op_def)
     pooling_type_arg = op_def.arg.add()
     pooling_type_arg.name = 'pooling_type'
-      pooling_type_arg.i = pooling_type_mode[first_op.type]
+    pooling_type_arg.i = pooling_type_mode[op.type]
     padding_arg = op_def.arg.add()
     padding_arg.name = 'padding'
-      padding_arg.i = padding_mode[first_op.get_attr('padding')]
+    padding_arg.i = padding_mode[op.get_attr('padding')]
     strides_arg = op_def.arg.add()
     strides_arg.name = 'strides'
-      strides_arg.ints.extend(first_op.get_attr('strides')[1:3])
+    strides_arg.ints.extend(op.get_attr('strides')[1:3])
     kernels_arg = op_def.arg.add()
     kernels_arg.name = 'kernels'
-      kernels_arg.ints.extend(first_op.get_attr('ksize')[1:3])
+    kernels_arg.ints.extend(op.get_attr('ksize')[1:3])
     data_format_arg = op_def.arg.add()
     data_format_arg.name = 'data_format'
     data_format_arg.s = 'NHWC'
-    elif first_op.type == 'Add':
-      op_def.name = first_op.name
+    self.resolved_ops[op.name] = 1
+
+  def convert_relu6(self, op):
+    op_def = self.net_def.op.add()
+    arg = op_def.arg.add()
+    arg.name = 'T'
+    arg.i = self.dt
+    op_def.name = op.name
+    op_def.type = 'Relu'
+    op_def.input.extend([input.name for input in op.inputs])
+    op_def.output.extend([output.name for output in op.outputs])
+    self.add_output_shape(op.outputs, op_def)
+    max_limit_arg = op_def.arg.add()
+    max_limit_arg.name = 'max_limit'
+    max_limit_arg.f = 6
+    self.resolved_ops[op.name] = 1
+
+  def convert_add(self, op):
+    op_def = self.net_def.op.add()
+    arg = op_def.arg.add()
+    arg.name = 'T'
+    arg.i = self.dt
+    op_def.name = op.name
     op_def.type = "AddN"
-      op_def.input.extend([input.name for input in first_op.inputs])
-      convert_op_outputs(op_def, first_op)
-    elif first_op.type == 'ConcatV2':
-      op_def.name = first_op.name
+    op_def.input.extend([input.name for input in op.inputs])
+    op_def.output.extend([output.name for output in op.outputs])
+    self.add_output_shape(op.outputs, op_def)
+    self.resolved_ops[op.name] = 1
+
+  def convert_concat(self, op):
+    op_def = self.net_def.op.add()
+    arg = op_def.arg.add()
+    arg.name = 'T'
+    arg.i = self.dt
+    op_def.name = op.name
     op_def.type = "Concat"
-      op_def.input.extend([first_op.inputs[i].name for i in xrange(2)])
+    op_def.input.extend([op.inputs[i].name for i in xrange(2)])
+    op_def.output.extend([output.name for output in op.outputs])
     axis_arg = op_def.arg.add()
     axis_arg.name = 'axis'
-      axis_arg.i = get_input_tensor(first_op, 2).eval().astype(np.int32)
-      convert_op_outputs(op_def, first_op)
-    elif first_op.type == 'ResizeBilinear':
-      op_def.name = first_op.name
+    axis_arg.i = get_input_tensor(op, 2).eval().astype(np.int32)
+    self.add_output_shape(op.outputs, op_def)
+    self.resolved_ops[op.name] = 1
+
+  def convert_resize_bilinear(self, op):
+    op_def = self.net_def.op.add()
+    arg = op_def.arg.add()
+    arg.name = 'T'
+    arg.i = self.dt
+    op_def.name = op.name
     op_def.type = "ResizeBilinear"
-      op_def.input.extend([first_op.inputs[0].name])
+    op_def.input.extend([op.inputs[0].name])
+    op_def.output.extend([output.name for output in op.outputs])
     size_arg = op_def.arg.add()
     size_arg.name = 'size'
-      size_arg.ints.extend(get_input_tensor(first_op, 1).eval().astype(np.int32).flat)
+    size_arg.ints.extend(get_input_tensor(op, 1).eval().astype(np.int32).flat)
     size_arg = op_def.arg.add()
     size_arg.name = 'align_corners'
-      size_arg.i = first_op.get_attr('align_corners')
-      convert_op_outputs(op_def, first_op)
-    elif first_op.type in ['Relu', 'SpaceToBatchND', 'BatchToSpaceND', 'BiasAdd']:
-      op_def.name = first_op.name
-      op_def.type = first_op.type
-      op_def.input.extend([input.name for input in first_op.inputs])
-      convert_op_outputs(op_def, first_op)
+    size_arg.i = op.get_attr('align_corners')
+    self.add_output_shape(op.outputs, op_def)
+    self.resolved_ops[op.name] = 1
+
+  def convert_bias_add(self, op):
+    op_def = mace_pb2.OperatorDef()
+    arg = op_def.arg.add()
+    arg.name = 'T'
+    arg.i = self.dt
+    op_def.name = op.name
+    op_def.type = "BiasAdd"
+    op_def.input.extend([op.inputs[0].name])
+    if self.device == 'gpu':
+      output_name = self.add_buffer_to_image(op.inputs[1].name, "ARGUMENT")
+      op_def.input.extend([output_name])
     else:
-      raise Exception('Unknown Op: %s, type: %s' % (first_op.name, first_op.type))
+      op_def.input.extend([op.inputs[1].name])
+    op_def.output.extend([output.name for output in op.outputs])
+    self.add_output_shape(op.outputs, op_def)
+    self.net_def.op.extend([op_def])
+    self.resolved_ops[op.name] = 1
+
+  def convert_normal_op(self, op):
+    op_def = self.net_def.op.add()
+    arg = op_def.arg.add()
+    arg.name = 'T'
+    arg.i = self.dt
+    op_def.name = op.name
+    op_def.type = op.type
+    op_def.input.extend([input.name for input in op.inputs])
+    op_def.output.extend([output.name for output in op.outputs])
+    self.add_output_shape(op.outputs, op_def)
+    self.resolved_ops[op.name] = 1
+
+  def convert(self, input_node, output_node):
+    if self.device == 'gpu':
+      self.add_input_transform(input_node)
+
+    for op in self.tf_ops:
+      if self.resolved_ops[op.name] == 1:
+        continue
+      if op.type in ['Placeholder', 'Reshape', 'Identity']:
+        self.resolved_ops[op.name] = 1
         pass
+      elif op.type == 'Const':
+        self.convert_tensor(op)
+      elif op.type == 'Conv2D' or op.type == 'DepthwiseConv2dNative':
+        self.convert_conv2d(op)
+      elif op.type == 'FusedBatchNorm':
+        self.convert_fused_batchnorm(op)
+      elif op.type == 'Add' and op.name.endswith('batchnorm/add'):
+        self.convert_batchnorm(op)
+      elif op.type == 'AvgPool' or op.type == 'MaxPool':
+        self.convert_pooling(op)
+      elif op.type == 'Relu6':
+        self.convert_relu6(op)
+      elif op.type == 'Add':
+        self.convert_add(op)
+      elif op.type == 'ConcatV2':
+        self.convert_concat(op)
+      elif op.type == 'ResizeBilinear':
+        self.convert_resize_bilinear(op)
+      elif op.type == 'BiasAdd':
+        self.convert_bias_add(op)
+      elif op.type in ['Relu', 'SpaceToBatchND', 'BatchToSpaceND']:
+        self.convert_normal_op(op)
+      else:
+        raise Exception('Unknown Op: %s, type: %s' % (op.name, op.type))
 
-  for i in range(resolved_count):
-    del unresolved_ops[0]
+    if self.device == 'gpu':
+      self.add_output_transform(output_node)
 
+    for key in self.resolved_ops:
+      if self.resolved_ops[key] != 1:
+        print 'Unresolve Op: %s' % key
 
 def convert_to_mace_pb(input_graph_def, input_node, output_node, data_type, device):
   net_def = mace_pb2.NetDef()
@@ -311,14 +425,11 @@ def convert_to_mace_pb(input_graph_def, input_node, output_node, data_type, devi
     with session.graph.as_default() as graph:
       tf.import_graph_def(input_graph_def, name="")
       ops = graph.get_operations()
-      unresolved_ops = ops
-      if device == 'gpu':
-        add_input_transform(input_node, dt, net_def)
-      while len(unresolved_ops) > 0:
-        convert_ops(unresolved_ops, dt, net_def, device)
-      if device == 'gpu':
-        add_output_transform(output_node, net_def)
-
-  print "PB Parsed."
+      converter = TFConverter(ops, net_def, dt, device)
+      converter.convert(input_node, output_node)
+      print "PB Converted, start optimize memory."
+      mem_optimizer = memory_optimizer.MemoryOptimizer(net_def)
+      mem_optimizer.optimize()
+      print "Memory optimization done."
 
   return net_def

mace/python/tools/tf_dsp_converter_lib.py

@@ -149,6 +149,7 @@ def convert_ops(unresolved_ops, resolved_ops, net_def, output_node, dsp_ops):
     elif is_node_flatten_reshape(first_op):
       op_def.type = 'Flatten'
       op_def.input.extend([t.name for t in first_op.inputs])
+      op_def.out_max_byte_size.extend([max_elem_size(out) for out in first_op.outputs])
       convert_op_outputs(op_def, first_op)
     elif dsp_ops.has_op(first_op.type):
       op_def.input.extend([t.name for t in first_op.inputs])

tools/validate.py

@@ -4,6 +4,7 @@ import os
 import os.path
 import tensorflow as tf
 import numpy as np
+from scipy import spatial
 
 from tensorflow import gfile
 
@@ -34,9 +35,12 @@ def load_data(file):
 def valid_output(out_shape, mace_out_file, tf_out_value):
   mace_out_value = load_data(mace_out_file)
   if mace_out_value.size != 0:
+    similarity = (1 - spatial.distance.cosine(tf_out_value.flat, mace_out_value))
+    print 'MACE VS TF similarity: ', similarity
+    if similarity > 0.999:
+      print '=======================Passed! Haha======================'
     mace_out_value = mace_out_value.reshape(out_shape)
     np.testing.assert_allclose(mace_out_value, tf_out_value, rtol=0.05)
-    print '=======================Passed! Haha======================'
   else:
     print '=======================Skip empty node==================='
 
@@ -62,7 +66,7 @@ def run_model(input_shape):
         input_value = input_value.reshape(input_shape)
         
         output_value = session.run(output_node, feed_dict={input_node: [input_value]})
-        # output_value.astype(np.float32).tofile( os.path.dirname(FLAGS.input_file) + '/tf_weight')
+        output_value.astype(np.float32).tofile( os.path.dirname(FLAGS.input_file) + '/tf_out')
         return output_value
 
 def main(unused_args):

tools/validate_gcn.sh

@@ -2,10 +2,10 @@
 # Must run at root dir of mace project.
 set +x
 Usage() {
-  echo 'Usage: bash tools/validate_gcn.sh tf_model_file'
+  echo 'Usage: bash tools/validate_gcn.sh tf_model_path image_size'
 }
 
-if [ $# != 1 ];then
+if [ $# != 2 ];then
   Usage
   exit -1
 fi
@@ -13,18 +13,18 @@ fi
 TF_MODEL_FILE_PATH=$1
 MODEL_DIR=$(dirname ${TF_MODEL_FILE_PATH})
 MACE_MODEL_NAME='mace_model.pb'
-MACE_OPT_MODEL_NAME='mace_opt_model.pb'
 INPUT_FILE_NAME='model_input'
 OUTPUT_FILE_NAME='gcn.out'
 OUTPUT_LIST_FILE='gcn.list'
 PHONE_DATA_DIR="/data/local/tmp/${MACE_MODEL_NAME}"
 KERNEL_DIR="${PHONE_DATA_DIR}/cl/"
+IMAGE_SIZE=$2
 
 # Step 1: Generate input data
 echo "Step 1: Generate input data"
 python tools/validate.py --generate_data true --random_seed 1 \
  --input_file=${MODEL_DIR}/${INPUT_FILE_NAME} \
- --input_shape=512,512,3
+ --input_shape="${IMAGE_SIZE},${IMAGE_SIZE},3"
 
 # Step 2: convert tf model to mace model
 echo "Step 2: convert tf model to mace model and optimize memory"
@@ -35,10 +35,6 @@ bazel-bin/mace/python/tools/tf_converter --input=${TF_MODEL_FILE_PATH} \
                                          --output_node=GCN/br_result_2/fcn_br \
                                          --data_type=DT_HALF \
                                          --runtime=gpu
-bazel build mace/python/tools:memory_optimizer
-bazel-bin/mace/python/tools/memory_optimizer ${MODEL_DIR}/${MACE_MODEL_NAME} \
-                                             ${MODEL_DIR}/${MACE_OPT_MODEL_NAME}
-
 
 # Step 3: Run model on the phone
 echo "Step 3: Run model on the phone"
@@ -49,21 +45,22 @@ bazel build -c opt --strip always mace/examples:mace_run  \
 
 adb shell "mkdir -p ${PHONE_DATA_DIR}"
 adb shell "mkdir -p ${KERNEL_DIR}"
-adb push mace/kernels/opencl/cl/ ${KERNEL_DIR}
-adb push ${MODEL_DIR}/${MACE_OPT_MODEL_NAME} ${PHONE_DATA_DIR}
+adb push mace/kernels/opencl/cl/* ${KERNEL_DIR}
+adb push ${MODEL_DIR}/${MACE_MODEL_NAME} ${PHONE_DATA_DIR}
 adb push ${MODEL_DIR}/${INPUT_FILE_NAME} ${PHONE_DATA_DIR}
 adb push bazel-bin/mace/examples/mace_run ${PHONE_DATA_DIR}
 
 num_threads=${1:-4}
 
-adb </dev/null shell MACE_RUN_PARAMETER_PATH=${PHONE_DATA_DIR}/mace_run.config \
+adb </dev/null shell MACE_CPP_MIN_VLOG_LEVEL=0 \
+        MACE_RUN_PARAMETER_PATH=${PHONE_DATA_DIR}/mace_run.config \
         MACE_KERNEL_PATH=$KERNEL_DIR \
         OMP_NUM_THREADS=$num_threads \
         ${PHONE_DATA_DIR}/mace_run \
-          --model=${PHONE_DATA_DIR}/${MACE_OPT_MODEL_NAME} \
+          --model=${PHONE_DATA_DIR}/${MACE_MODEL_NAME} \
           --input=mace_input_node \
           --output=mace_output_node \
-          --input_shape=1,512,512,3\
+          --input_shape="1,${IMAGE_SIZE},${IMAGE_SIZE},3"\
           --input_file=${PHONE_DATA_DIR}/${INPUT_FILE_NAME} \
           --output_file=${PHONE_DATA_DIR}/${OUTPUT_FILE_NAME} \
           --device=OPENCL   \
@@ -81,4 +78,5 @@ python tools/validate.py --model_file ${TF_MODEL_FILE_PATH} \
     --mace_out_file ${MODEL_DIR}/${OUTPUT_FILE_NAME} \
     --input_node input \
     --output_node GCN/br_result_2/fcn_br\
-    --output_shape 1,512,512,2
+    --input_shape "${IMAGE_SIZE},${IMAGE_SIZE},3" \
+    --output_shape "1,${IMAGE_SIZE},${IMAGE_SIZE},2"