remove native conv cpp impl in cpp operation/; rename github workflow…

… stages

.asf.yaml

@@ -24,3 +24,5 @@ github:
     wiki: true
     # Enable issues on github
     issues: true
+    # Enable settings on github
+    settings: true

.github/workflows/conda.yaml

@@ -27,7 +27,7 @@ on:
 
 # A workflow run is made up of one or more jobs that can run sequentially or in parallel
 jobs:
-  build-ubuntu-conda:
+  build-pytest-on-cpu:
     runs-on: ubuntu-latest
 
     steps:
@@ -36,5 +36,5 @@ jobs:
         run: conda install conda-build
       - name: conda-config
         run: conda config --add channels conda-forge && conda config --add channels nusdbsystem
-      - name: build-singa-conda
+      - name: build-pytest
         run: conda build tool/conda/singa

.github/workflows/macOS.yaml

@@ -15,14 +15,14 @@
 # limitations under the License.
 #
 
-name: macOS
+name: Native-MacOS
 
 on:
   push:
   pull_request:
 
 jobs:
-  build-macOS-cpu:
+  build-cpptest-cpu:
     runs-on: macos-latest
 
     steps:

.github/workflows/cpp.yaml → .github/workflows/ubuntu.yaml

@@ -17,7 +17,7 @@
 
 # This is a basic workflow to help you get started with Actions
 
-name: C++
+name: Native-Ubuntu
 
 # Controls when the action will run. Triggers the workflow on push or pull request
 # events but only for the master branch
@@ -41,7 +41,7 @@ jobs:
   #    - name: C++ test
   #      run: build/bin/test_singa
 
-  build-ubuntu-oneDNN:
+  build-cpptest-on-cpu:
     runs-on: ubuntu-latest
 
     steps:

examples/onnx/utils.py

@@ -21,7 +21,6 @@
 import tarfile
 import glob
 import onnx
-from onnx import numpy_helper
 import logging
 logging.basicConfig(level=logging.INFO, format='%(asctime)-15s %(message)s')
 
@@ -41,7 +40,7 @@ def load_dataset(test_data_dir):
         onnx_tensor = onnx.TensorProto()
         with open(input_file, 'rb') as f:
             onnx_tensor.ParseFromString(f.read())
-        inputs.append(numpy_helper.to_array(onnx_tensor))
+        inputs.append(onnx.numpy_helper.to_array(onnx_tensor))
 
     # load reference outputs
     ref_outputs = []
@@ -51,7 +50,7 @@ def load_dataset(test_data_dir):
         onnx_tensor = onnx.TensorProto()
         with open(output_file, 'rb') as f:
             onnx_tensor.ParseFromString(f.read())
-        ref_outputs.append(numpy_helper.to_array(onnx_tensor))
+        ref_outputs.append(onnx.numpy_helper.to_array(onnx_tensor))
     return inputs, ref_outputs
 
 
@@ -63,4 +62,3 @@ def check_exist_or_download(url):
         logging.info("Downloading %s" % url)
         urllib.request.urlretrieve(url, filename)
     return filename
-

python/singa/sonnx.py

@@ -21,18 +21,21 @@
 
 import numpy as np
 import onnx.utils
-import onnx
+import onnx.save 
+import onnx.load
 from onnx.backend.base import Backend, BackendRep
 from onnx import (checker, helper, numpy_helper, GraphProto, NodeProto,
-                  TensorProto, OperatorSetIdProto, optimizer, mapping)
+                  TensorProto, OperatorSetIdProto, optimizer, mapping,
+                  shape_inference)
 import warnings
 
-from singa import device
-from . import singa_wrap as singa
-from . import autograd, layer
-from . import tensor
-from . import model
-from singa import utils
+import device
+import autograd
+import layer
+import tensor
+import model
+import utils
+from singa_wrap as singa
 
 import collections
 OrderedDict = collections.OrderedDict
@@ -390,10 +393,9 @@ def _create_constantOfShape(cls, op, op_t):
             the onnx node
         """
         node = cls._common_singa_tensor_to_onnx_node(op, op_t)
-        tensor_type = onnx.TensorProto.FLOAT if isinstance(
-            op.value, float) else onnx.TensorProto.INT32
-        tensor_value = onnx.helper.make_tensor("value", tensor_type, [1],
-                                               [op.value])
+        tensor_type = TensorProto.FLOAT if isinstance(
+            op.value, float) else TensorProto.INT32
+        tensor_value = helper.make_tensor("value", tensor_type, [1], [op.value])
         node.attribute.extend([
             helper.make_attribute('value', tensor_value),
         ])
@@ -1895,7 +1897,7 @@ def prepare(cls, model, device='CPU', **kwargs):
         try:
             model = onnx.utils.polish_model(model)
         except IndexError as err:
-            model = onnx.shape_inference.infer_shapes(model)
+            model = shape_inference.infer_shapes(model)
 
         # check the opset version and ir version
         # SINGA supports opset version(11), ir version(1.6.0 -> 6)
@@ -2199,4 +2201,4 @@ def forward(self, *input, aux_output=(), **kwargs):
 get_op = SingaBackend._onnx_node_to_singa_op
 to_onnx = SingaFrontend.singa_to_onnx_model
 save = onnx.save
-load = onnx.load
+load = onnx.load

setup.py

@@ -55,8 +55,6 @@
 import io
 import sys
 import subprocess
-import textwrap
-import traceback
 import shutil
 import shlex
 from pathlib import Path
@@ -192,10 +190,8 @@ def generate_singa_config(with_cuda, with_nccl):
 
 
 def get_cpp_flags():
-    last_err = None
     default_flags = ['-std=c++11', '-fPIC', '-g', '-O2', '-Wall', '-pthread']
     # avx_flags = [ '-mavx'] #'-mf16c',
-    flags_to_try = []
     if sys.platform == 'darwin':
         # Darwin most likely will have Clang, which has libc++.
         return default_flags + ['-stdlib=libc++']

src/model/operation/convolution.cc

@@ -20,10 +20,10 @@
  ************************************************************/
 // #include "../layer/convolution.h"
 
-#include <cctype>
-
 #include "convolution.h"
 
+#include <cctype>
+
 namespace singa {
 
 ConvHandle::ConvHandle(const Tensor &input,
@@ -188,7 +188,8 @@ Tensor CpuConvForward(const Tensor &x, Tensor &W, Tensor &b,
       {x.block(), W.block(), b.block()}, {output.block()}, "CpuConvForward");
 
   return output;
-#else   // cpp naive
+#else   // cpp naive, error due to Im2col importing
+/*
   Shape w_shape = W.shape();
   Shape b_shape;
   if (ch.bias_term) b_shape = b.shape();
@@ -219,6 +220,7 @@ Tensor CpuConvForward(const Tensor &x, Tensor &W, Tensor &b,
   W.Reshape(w_shape);
   if (ch.bias_term) b.Reshape(b_shape);
   return output;
+*/
 #endif  // USE_DNNL
 }
 
@@ -284,6 +286,7 @@ Tensor CpuConvBackwardx(const Tensor &dy, Tensor &W, const Tensor &x,
   return dx;
 
 #else   // NOT USE_DNNL
+/*  // error due to importing Col2im
   Shape w_shape = W.shape();
   W.Reshape(Shape{ch.num_filters, ch.col_height});
 
@@ -303,6 +306,7 @@ Tensor CpuConvBackwardx(const Tensor &dy, Tensor &W, const Tensor &x,
   }
   W.Reshape(w_shape);
   return dx;
+*/
 #endif  // USE_DNNL
 }
 
@@ -372,10 +376,12 @@ Tensor CpuConvBackwardW(const Tensor &dy, const Tensor &x, const Tensor &W,
                       {DNNL_ARG_DIFF_BIAS, conv_diff_bias_memory}});
         ctx->dnnl_stream.wait();
       },
-      {x.block(), dy.block(), W.block()}, {dW.block(), ch.db->block()}, "CpuConvBackwardW");
+      {x.block(), dy.block(), W.block()}, {dW.block(), ch.db->block()},
+      "CpuConvBackwardW");
 
   return dW;
 #else   // native cpp
+/* // error due to importing Im2col
   Tensor dW;
   dW.ResetLike(W);
   dW.SetValue(0.0f);
@@ -398,6 +404,7 @@ Tensor CpuConvBackwardW(const Tensor &dy, const Tensor &x, const Tensor &W,
   }
   dW.Reshape(w_shape);
   return dW;
+*/
 #endif  // USE_DNNL
 }
 
@@ -598,7 +605,8 @@ Tensor GpuConvForward(const Tensor &x, const Tensor &W, const Tensor &b,
                                 cch.workspace_count * sizeof(float), &beta,
                                 cch.y_desc, outblock->mutable_data());
       },
-      {x.block(), W.block()}, {output.block(), cch.workspace.block()}, "cudnnConvForward");
+      {x.block(), W.block()}, {output.block(), cch.workspace.block()},
+      "cudnnConvForward");
 
   if (cch.bias_term) {
     Tensor outputFake(output);
@@ -634,7 +642,8 @@ Tensor GpuConvBackwardx(const Tensor &dy, const Tensor &W, const Tensor &x,
             cch.workspace_count * sizeof(float), &beta, cch.x_desc,
             dxblock->mutable_data());
       },
-      {dy.block(), W.block()}, {dx.block(), cch.workspace.block()}, "cudnnConvolutionBackwardData");
+      {dy.block(), W.block()}, {dx.block(), cch.workspace.block()},
+      "cudnnConvolutionBackwardData");
 
   return dx;
 }
@@ -658,7 +667,8 @@ Tensor GpuConvBackwardW(const Tensor &dy, const Tensor &x, const Tensor &W,
             cch.workspace_count * sizeof(float), &beta, cch.filter_desc,
             dwblock->mutable_data());
       },
-      {dy.block(), x.block()}, {dW.block(), cch.workspace.block()}, "cudnnConvolutionBackwardFilter");
+      {dy.block(), x.block()}, {dW.block(), cch.workspace.block()},
+      "cudnnConvolutionBackwardFilter");
 
   return dW;
 }