From d9fa688a6154f7df4fdacae445937f16027418f2 Mon Sep 17 00:00:00 2001
From: Guanghua Yu <742925032@qq.com>
Date: Mon, 17 Oct 2022 20:44:59 +0800
Subject: [PATCH] update dygraph new format QAT API (#1464)

---
 demo/dygraph/quant/train.py     |  5 ++--
 paddleslim/dygraph/quant/qat.py | 44 ++++++++++++++++++++-------------
 2 files changed, 30 insertions(+), 19 deletions(-)

diff --git a/demo/dygraph/quant/train.py b/demo/dygraph/quant/train.py
index 202b3eb4..31fc67be 100644
--- a/demo/dygraph/quant/train.py
+++ b/demo/dygraph/quant/train.py
@@ -165,6 +165,8 @@ def compress(args):
         'moving_rate': 0.9,
         # for dygraph quantization, layers of type in quantizable_layer_type will be quantized
         'quantizable_layer_type': ['Conv2D', 'Linear'],
+        # # Whether to export the quantized model with format of ONNX.
+        'onnx_format': args.onnx_format,
     }
 
     if args.use_pact:
@@ -360,8 +362,7 @@ def compress(args):
             input_spec=[
                 paddle.static.InputSpec(
                     shape=[None, 3, 224, 224], dtype='float32')
-            ],
-            onnx_format=args.onnx_format)
+            ])
 
 
 def main():
diff --git a/paddleslim/dygraph/quant/qat.py b/paddleslim/dygraph/quant/qat.py
index 34b1ae7b..05b1b40d 100644
--- a/paddleslim/dygraph/quant/qat.py
+++ b/paddleslim/dygraph/quant/qat.py
@@ -58,6 +58,8 @@ _quant_config_default = {
     'quantizable_layer_type': ['Conv2D', 'Linear'],
     # whether fuse conv and bn before QAT
     'fuse_conv_bn': False,
+    # Whether to export the quantized model with format of ONNX. Default is False.
+    'onnx_format': False,
 }
 
 
@@ -215,7 +217,9 @@ class QAT(object):
                 weight_preprocess_layer=self.weight_preprocess,
                 act_preprocess_layer=self.act_preprocess,
                 weight_quantize_layer=self.weight_quantize,
-                act_quantize_layer=self.act_quantize)
+                act_quantize_layer=self.act_quantize,
+                onnx_format=self.config['onnx_format'],  # support Paddle >= 2.4
+            )
         except:
             self.imperative_qat = ImperativeQuantAware(
                 weight_bits=self.config['weight_bits'],
@@ -257,11 +261,7 @@ class QAT(object):
 
         return quant_model
 
-    def save_quantized_model(self,
-                             model,
-                             path,
-                             input_spec=None,
-                             onnx_format=False):
+    def save_quantized_model(self, model, path, input_spec=None):
         """
         Save the quantized inference model.
 
@@ -287,20 +287,30 @@ class QAT(object):
                 model.eval()
 
         self.imperative_qat.save_quantized_model(
-            layer=model,
-            path=path,
-            input_spec=input_spec,
-            onnx_format=onnx_format)
+            layer=model, path=path, input_spec=input_spec)
 
     def _remove_preprocess(self, model):
         state_dict = model.state_dict()
-        self.imperative_qat = ImperativeQuantAware(
-            weight_bits=self.config['weight_bits'],
-            activation_bits=self.config['activation_bits'],
-            weight_quantize_type=self.config['weight_quantize_type'],
-            activation_quantize_type=self.config['activation_quantize_type'],
-            moving_rate=self.config['moving_rate'],
-            quantizable_layer_type=self.config['quantizable_layer_type'])
+        try:
+            self.imperative_qat = ImperativeQuantAware(
+                weight_bits=self.config['weight_bits'],
+                activation_bits=self.config['activation_bits'],
+                weight_quantize_type=self.config['weight_quantize_type'],
+                activation_quantize_type=self.config[
+                    'activation_quantize_type'],
+                moving_rate=self.config['moving_rate'],
+                quantizable_layer_type=self.config['quantizable_layer_type'],
+                onnx_format=self.config['onnx_format'],  # support Paddle >= 2.4
+            )
+        except:
+            self.imperative_qat = ImperativeQuantAware(
+                weight_bits=self.config['weight_bits'],
+                activation_bits=self.config['activation_bits'],
+                weight_quantize_type=self.config['weight_quantize_type'],
+                activation_quantize_type=self.config[
+                    'activation_quantize_type'],
+                moving_rate=self.config['moving_rate'],
+                quantizable_layer_type=self.config['quantizable_layer_type'])
 
         with paddle.utils.unique_name.guard():
             if hasattr(model, "_layers"):
-- 
GitLab