[CI] Fix lint test python version. (#225)

* [CI] Fix lint test python version.

Co-authored-by: zhangqi3 <zhangqi3@sensetime.com>

.github/workflows/lint-and-test.yml

@@ -4,7 +4,7 @@ on: [push]
 
 jobs:
   Lint-and-test:
-    runs-on: ubuntu-latest
+    runs-on: ubuntu-18.04
     strategy:
       max-parallel: 5
 
@@ -13,28 +13,21 @@ jobs:
     - name: Set up Python 3.6
       uses: actions/setup-python@v2
       with:
-        python-version: 3.6
-    - name: Add conda to system path
-      run: |
-        # $CONDA is an environment variable pointing to the root of the miniconda directory
-        echo $CONDA/bin >> $GITHUB_PATH
+        python-version: 3.7
     - name: Lint with flake8
       run: |
-        conda install flake8
+        pip install flake8
         flake8 .
-    - name: Install onnxruntime and onnxsim
-      run:
-        pip install onnxruntime onnx-simplifier
-    - name: Install prettytable
-      run:
-        pip install prettytable
     - name: Install Protobuf
       run:
-        conda install protobuf=3.20.1
+        pip install protobuf==3.19.0
+    - name: Install onnx onnxruntime and onnxsim
+      run:
+        pip install onnx==1.7.0 onnxruntime onnx-simplifier
     - name: Install MQBench
       run: |
         python setup.py develop
     - name: Test with pytest
       run: |
-        conda install pytest
+        pip install pytest
         pytest test --junitxml=report.xml

mqbench/custom_quantizer/onnx_qnn_quantizer.py

@@ -56,9 +56,9 @@ def _qat_swap_modules(self, root: GraphModule, additional_qat_module_mapping: Di
         all_mappings = get_combined_dict(
             get_default_qat_module_mappings(), additional_qat_module_mapping)
         # There is no QLinearFC in ONNX for now.
-        del(all_mappings[torch.nn.modules.linear.Linear])
-        del(all_mappings[torch.nn.intrinsic.modules.fused.LinearReLU])
-        del(all_mappings[qnni.modules.fused.LinearBn1d])
+        del all_mappings[torch.nn.modules.linear.Linear]
+        del all_mappings[torch.nn.intrinsic.modules.fused.LinearReLU]
+        del all_mappings[qnni.modules.fused.LinearBn1d]
         root = self._convert(root, all_mappings, inplace=True)
         return root
 

mqbench/deploy/common.py

@@ -93,7 +93,7 @@ def insert_node_purely(self, node, idx=0):
 
     def del_initializer(self, initializer_name):
         if initializer_name in self.initializer:
-            del(self.initializer[initializer_name])
+            del self.initializer[initializer_name]
 
     def optimize_model(self):
         # Delete redundant nodes.

mqbench/deploy/deploy_stpu.py

@@ -3,7 +3,6 @@
 from collections import OrderedDict
 
 import onnx
-from onnx import numpy_helper
 
 from mqbench.deploy.common import (get_constant_inputs, prepare_data,
                                    prepare_initializer,
@@ -15,10 +14,6 @@
 
 class STPU_process(LinearQuantizer_process):
 
-    @staticmethod
-    def get_constant(node: onnx.NodeProto):
-        return numpy_helper.to_array(node.attribute[0].t).tolist()
-
     def remove_fakequantize_and_collect_params(self, onnx_path, model_name):
         model = onnx.load(onnx_path)
         graph = model.graph

mqbench/fuser_method_mappings.py

@@ -47,7 +47,7 @@ def fuse_linear_bn(linear, bn):
         >>> b1 = nn.BatchNorm1d(20)
         >>> m2 = fuse_linear_bn(m1, b1)
     """
-    assert(linear.training == bn.training),\
+    assert linear.training == bn.training, \
         "Linear and BN both must be in the same mode (train or eval)."
 
     if linear.training:
@@ -59,7 +59,7 @@ def fuse_linear_bn(linear, bn):
 
 
 def fuse_deconv_bn(deconv, bn):
-    assert(deconv.training == bn.training),\
+    assert deconv.training == bn.training, \
         'DeConv and BN must be in the same mode (train or eval)'
 
     if deconv.training:
@@ -72,7 +72,7 @@ def fuse_deconv_bn(deconv, bn):
 
 
 def fuse_deconv_bn_relu(deconv, bn, relu):
-    assert(deconv.training == bn.training == relu.training),\
+    assert deconv.training == bn.training == relu.training, \
         "DeConv and BN both must be in the same mode (train or eval)."
 
     if deconv.training:
@@ -85,7 +85,7 @@ def fuse_deconv_bn_relu(deconv, bn, relu):
 
 
 def fuse_conv_freezebn(conv, bn):
-    assert(bn.training is False), "Freezebn must be eval."
+    assert bn.training is False, "Freezebn must be eval."
 
     fused_module_class_map = {
         nn.Conv2d: qnni.ConvFreezebn2d,
@@ -102,7 +102,7 @@ def fuse_conv_freezebn(conv, bn):
 
 
 def fuse_conv_freezebn_relu(conv, bn, relu):
-    assert(conv.training == relu.training and bn.training is False), "Conv and relu both must be in the same mode (train or eval) and bn must be eval."
+    assert conv.training == relu.training and bn.training is False, "Conv and relu both must be in the same mode (train or eval) and bn must be eval."
     fused_module : Optional[Type[nn.Sequential]] = None
     if conv.training:
         map_to_fused_module_train = {
@@ -123,7 +123,7 @@ def fuse_conv_freezebn_relu(conv, bn, relu):
 
 
 def fuse_deconv_freezebn(deconv, bn):
-    assert(bn.training is False), "Freezebn must be eval."
+    assert bn.training is False, "Freezebn must be eval."
 
     if deconv.training:
         assert bn.num_features == deconv.out_channels, 'Output channel of ConvTranspose2d must match num_features of BatchNorm2d'
@@ -135,7 +135,7 @@ def fuse_deconv_freezebn(deconv, bn):
 
 
 def fuse_deconv_freezebn_relu(deconv, bn, relu):
-    assert(deconv.training == relu.training and bn.training is False), "Conv and relu both must be in the same mode (train or eval) and bn must be eval."
+    assert deconv.training == relu.training and bn.training is False, "Conv and relu both must be in the same mode (train or eval) and bn must be eval."
 
     if deconv.training:
         assert bn.num_features == deconv.out_channels, 'Output channel of ConvTranspose2d must match num_features of BatchNorm2d'

mqbench/mix_precision/mix_precision.py

@@ -163,7 +163,7 @@ def hawq(model: Module, data: Tuple, criterion, type='trace'):
     elif type == 'trace':
         return hessian_comp.layer_trace()
     else:
-        raise(NotImplementedError, "{} is not supported, only trace and eigenvalues.".format(type))
+        raise NotImplementedError("{} is not supported, only trace and eigenvalues.".format(type))
 
 
 def mixprecision_bit_selection(bitwidth_list, sensetive_dict, layer_parameters_dict, model_size_constraints, latency_constraints):

requirements.txt

@@ -1,4 +1,4 @@
 torch==1.10.0
 torchvision==0.11.1
-onnx
-prettytable
+onnx==1.7.0
+prettytable