修改onnxexporter的导出逻辑 (#142)

OpenPPL · May 27, 2022 · 54c0e3f · 54c0e3f
1 parent 3b9d7ae
commit 54c0e3f
Show file tree

Hide file tree

Showing 4 changed files with 8 additions and 10 deletions.
diff --git a/ppq/IR/base/core.py b/ppq/IR/base/core.py
diff --git a/ppq/IR/morph.py b/ppq/IR/morph.py
@@ -671,7 +671,7 @@ def decompose_gemm(self):
             output_var = op.outputs[0]
 
             if op.num_of_input == 3:
-                bias_add  = graph.create_operation(op_type='Add')
+                bias_add  = graph.create_operation(op_type='Add', platform=op.platform)
                 bias_var  = op.inputs[-1]
 
                 graph.create_link_with_op(

diff --git a/ppq/parser/onnxruntime_exporter.py b/ppq/parser/onnxruntime_exporter.py
@@ -280,8 +280,9 @@ def convert_operation(self, graph: BaseGraph, op: QuantableOperation,
         for config, var in op.config_with_variable:
             meta = var.meta
             if var.is_parameter:
-                if not op.is_computing_op: continue
-
+                # we do not want to process clip value here.
+                if op.type in {'Clip'}: continue
+
                 assert len(var.dest_ops) == 1, (
                 f'Can not export variable {var.name}, cause it has more than 1 destination operations. '
                 'PPQ require all parameters to have only 1 destination operation.')
@@ -293,7 +294,7 @@ def convert_operation(self, graph: BaseGraph, op: QuantableOperation,
                 if config.state == QuantizationStates.PASSIVE_BAKED:
                     config.state = QuantizationStates.PASSIVE
 
-                if QuantizationStates.is_activated(config.state):
+                if QuantizationStates.can_export(config.state):
                     # if not quant parameter to int, all parameter should export as fp32.
                     # needs insert both quant and dequant op for them
                     if not quant_param_to_int:
@@ -309,7 +310,7 @@ def convert_operation(self, graph: BaseGraph, op: QuantableOperation,
             else:
                 if not process_activation: continue
 
-                if QuantizationStates.is_activated(config.state):
+                if QuantizationStates.can_export(config.state):
                     created = self.insert_quant_on_variable(
                         graph=graph, var=var, config=config, related_op=op, meta=meta)
                     var = created.outputs[0]

diff --git a/ppq/quantization/analyse/layerwise.py b/ppq/quantization/analyse/layerwise.py
@@ -106,17 +106,14 @@ def layerwise_error_analyse(
             fp_outputs = executor.forward(inputs=batch, output_names=interested_outputs)
 
             # manually override quantization state
-            for cfg, _ in operation.config_with_variable:
-                cfg.state = QuantizationStates.ACTIVATED
+            operation.restore_quantize_state()
             qt_outputs = executor.forward(inputs=batch, output_names=interested_outputs)
 
             for fp_output, qt_output in zip(fp_outputs, qt_outputs):
                 recorder.update(y_pred = qt_output, y_real = fp_output)
 
             # manually override quantization state
-            for cfg, _ in operation.config_with_variable:
-                cfg.state = QuantizationStates.DEQUANTIZED
-
+            operation.dequantize()
             if idx >= steps: break
 
     # restore quantization states