fix ipex prepare model unsupport deepcopy issue and add ut (#1174)

Author: changwangss

neural_compressor/adaptor/pytorch.py

@@ -2362,8 +2362,8 @@ def _get_quantizable_ops_recursively(self, model, prefix, quantizable_ops):
             assert isinstance(model, torch.nn.Module), \
                     "The model passed in is not the instance of torch.nn.Module"
 
-            model_ = copy.deepcopy(model)
             if not IPEX_110 and not IPEX_112:
+                model_ = copy.deepcopy(model)
                 model_.eval().to(ipex.DEVICE)
                 try:
                     init_model = torch.jit.script(model_)
@@ -2377,10 +2377,15 @@ def _get_quantizable_ops_recursively(self, model, prefix, quantizable_ops):
                             "Fail to convert this model to PyTorch Script model"
                         )
                         init_model = model_
+            elif IPEX_110:
+                init_model = copy.deepcopy(model)
+                init_model.eval()
             else:
-                model_.eval()
-                init_model = model_
-
+                if hasattr(model,'save_qconf_summary'):
+                    init_model = ipex.quantization._quantize_utils.copy_prepared_model(model)
+                else:
+                    init_model = copy.deepcopy(model)
+                    init_model.eval()
             # create a quantization config file for intel pytorch extension model
             os.makedirs(os.path.dirname(self.ipex_config_path), exist_ok=True)
             if not IPEX_110 and not IPEX_112: 

test/ipex/test_adaptor_ipex.py

@@ -1,14 +1,9 @@
 import torch
 import unittest
 import os
-from neural_compressor.adaptor import FRAMEWORKS
-from neural_compressor.model import MODELS
 from neural_compressor.adaptor.pytorch import PyTorchVersionMode
 import neural_compressor.adaptor.pytorch as nc_torch
 from neural_compressor.experimental import Quantization, common
-from neural_compressor.conf.config import QuantConf
-from neural_compressor.utils.pytorch import load
-from neural_compressor.utils.utility import recover
 import shutil
 import copy
 import numpy as np
@@ -20,6 +15,7 @@
 except:
     TEST_IPEX = False
 
+torch.manual_seed(9527)
 assert TEST_IPEX, "Please install intel extension for pytorch"
 # get torch and IPEX version
 PT_VERSION = nc_torch.get_torch_version()
@@ -89,7 +85,7 @@ def test_tuning_ipex(self):
         ipex_conf = ipex.quantization.QuantConf(
                     configure_file="./saved/best_configure.json",
                 )
-        q_model = ipex.quantization.convert(model, ipex_conf, torch.randn(1, 3, 224, 224))
+        q_model = ipex.quantization.convert(model, ipex_conf, torch.ones(1, 3, 224, 224))
         from neural_compressor.experimental import Benchmark
         evaluator = Benchmark('ipex_yaml.yaml')
         evaluator.model = q_model
@@ -121,7 +117,7 @@ def test_tuning_ipex(self):
         nc_model = quantizer.fit()
         nc_model.save('./saved')
         qconfig = ipex.quantization.default_static_qconfig
-        prepared_model = ipex.quantization.prepare(model, qconfig, example_inputs=torch.randn(1, 3, 224, 224), inplace=False)
+        prepared_model = ipex.quantization.prepare(model, qconfig, example_inputs=torch.ones(1, 3, 224, 224), inplace=False)
         prepared_model.load_qconf_summary(qconf_summary = "./saved/best_configure.json")
         convert_model = ipex.quantization.convert(prepared_model)
         from neural_compressor.experimental import Benchmark
@@ -130,5 +126,27 @@ def test_tuning_ipex(self):
         evaluator.b_dataloader = common.DataLoader(dataset)
         evaluator.fit('accuracy')
 
+    def test_tuning_ipex_for_ipex_autotune_func(self):
+        from neural_compressor.experimental import Quantization
+        model = M()
+        qconfig = ipex.quantization.default_static_qconfig
+        prepared_model = ipex.quantization.prepare(model, qconfig, example_inputs=torch.ones(1, 3, 224, 224), inplace=False)
+        quantizer = Quantization('ipex_yaml.yaml')
+        quantizer.conf.usr_cfg.tuning.exit_policy['performance_only'] = True
+        dataset = quantizer.dataset('dummy', (100, 3, 224, 224), label=True)
+        quantizer.model = prepared_model
+        quantizer.calib_dataloader = common.DataLoader(dataset)
+        quantizer.eval_dataloader = common.DataLoader(dataset)
+        nc_model = quantizer.fit()
+        nc_model.save('./saved')
+        qconfig = ipex.quantization.default_static_qconfig
+        prepared_model = ipex.quantization.prepare(model, qconfig, example_inputs=torch.ones(1, 3, 224, 224), inplace=False)
+        prepared_model.load_qconf_summary(qconf_summary = "./saved/best_configure.json")
+        convert_model = ipex.quantization.convert(prepared_model)
+        from neural_compressor.experimental import Benchmark
+        evaluator = Benchmark('ipex_yaml.yaml')
+        evaluator.model = convert_model
+        evaluator.b_dataloader = common.DataLoader(dataset)
+        evaluator.fit('accuracy')
 if __name__ == "__main__":
     unittest.main()