[hackathon 7th] PaddleSpeech line: fix when shape=() can not load (Pa…

…ddlePaddle#69907)

python/paddle/static/quantization/utils.py

@@ -116,7 +116,11 @@ def load_variable_data(scope, var_name):
     '''
     var_node = scope.find_var(var_name)
     assert var_node is not None, "Cannot find " + var_name + " in scope."
-    return np.array(var_node.get_tensor())
+    tensor = np.array(var_node.get_tensor())
+    if tensor.shape == ():
+        return tensor.reshape(1)
+    else:
+        return tensor
 
 
 def set_variable_data(scope, place, var_name, np_value):

test/quantization/test_post_training_quantization_mobilenetv1.py

@@ -521,6 +521,152 @@ def test_post_training_avg_mobilenetv1(self):
         )
 
 
+class TestPostTrainingavgForPwganCsmsc(TestPostTrainingQuantization):
+    def setUp(self):
+        self.int8_download = 'int8/download'
+        self.cache_folder = os.path.expanduser(
+            '~/.cache/paddle/dataset/' + self.int8_download
+        )
+        self.data_cache_folder = ''
+        data_urls = []
+        data_md5s = []
+        if os.environ.get('DATASET') == 'full':
+            data_urls.append(
+                'https://paddlespeech.bj.bcebos.com/tmp/csmsc_voc1.npy'
+            )
+            data_md5s.append('47950146167ca8d885a78d71e74f1a2b')
+            self.data_cache_folder = self.download_data(
+                data_urls, data_md5s, "full_data", False
+            )
+        else:
+            data_urls.append(
+                'https://paddlespeech.bj.bcebos.com/tmp/csmsc_voc1.npy'
+            )
+            data_md5s.append('47950146167ca8d885a78d71e74f1a2b')
+            self.data_cache_folder = self.download_data(
+                data_urls, data_md5s, "small_data", False
+            )
+
+        # reader/decorator.py requires the relative path to the data folder
+        if not os.path.exists("./data/BZNSYP"):
+            cmd = 'rm -rf {0} && ln -s {1} {0}'.format(
+                "data", self.data_cache_folder
+            )
+            os.system(cmd)
+
+        self.batch_size = 1 if os.environ.get('DATASET') == 'full' else 50
+        self.infer_iterations = (
+            50000 if os.environ.get('DATASET') == 'full' else 2
+        )
+
+        self.root_path = tempfile.TemporaryDirectory()
+        self.int8_model = os.path.join(
+            self.root_path.name, "post_training_quantization"
+        )
+
+    def cache_unzipping(self, target_folder, zip_path):
+        if not os.path.exists(target_folder):
+            if zip_path.endswith('.tar.gz'):
+                cmd = f'mkdir {target_folder} && tar xf {zip_path} -C {target_folder}'
+            elif zip_path.endswith('.zip'):
+                cmd = f'mkdir {target_folder} && unzip -o {zip_path} -d {target_folder}'
+            else:
+                cmd = f'mkdir {target_folder}'
+            os.system(cmd)
+
+    def generate_quantized_model(
+        self,
+        model_path,
+        model_filename,
+        params_filename,
+        quantizable_op_type,
+        batch_size,
+        algo="avg",
+        round_type="round",
+        is_full_quantize=False,
+        is_use_cache_file=False,
+        is_optimize_model=False,
+        batch_nums=1,
+        onnx_format=False,
+        deploy_backend=None,
+        feed_name="inputs",
+    ):
+        try:
+            os.system("mkdir " + self.int8_model)
+        except Exception as e:
+            _logger.info(f"Failed to create {self.int8_model} due to {e}")
+            sys.exit(-1)
+
+        place = paddle.CPUPlace()
+        exe = paddle.static.Executor(place)
+        val_dataset = "~/.cache/paddle/dataset/int8/download/csmsc_voc1.npy"
+        data_loader = paddle.io.DataLoader(
+            val_dataset,
+            places=place,
+            drop_last=False,
+            batch_size=2,
+        )
+        ptq = PostTrainingQuantization(
+            executor=exe,
+            data_loader=data_loader,
+            model_dir=model_path,
+            model_filename=model_filename,
+            params_filename=params_filename,
+            batch_size=batch_size,
+            batch_nums=batch_nums,
+            algo=algo,
+            quantizable_op_type=quantizable_op_type,
+            round_type=round_type,
+            is_full_quantize=is_full_quantize,
+            optimize_model=is_optimize_model,
+            onnx_format=onnx_format,
+            is_use_cache_file=is_use_cache_file,
+            deploy_backend=deploy_backend,
+        )
+        ptq.quantize()
+        ptq.save_quantized_model(
+            self.int8_model,
+            model_filename=model_filename,
+            params_filename=params_filename,
+        )
+
+
+class TestPostTraininghistForNoneShape(TestPostTrainingavgForPwganCsmsc):
+    def test_post_training_avg_pwgancsmsc(self):
+        model = "pwg_baker_static_0.4"
+        algo = "avg"
+        round_type = "round"
+        data_urls = [
+            'https://paddlespeech.bj.bcebos.com/Parakeet/released_models/pwgan/pwg_baker_static_0.4.zip'
+        ]
+        data_md5s = ['e3504aed9c5a290be12d1347836d2742']
+        quantizable_op_type = [
+            "conv2d",
+            "depthwise_conv2d",
+            "mul",
+        ]
+        is_full_quantize = False
+        is_use_cache_file = False
+        is_optimize_model = True
+        diff_threshold = 0.05
+        self.run_test(
+            model,
+            'pwgan_csmsc.pdmodel',
+            'pwgan_csmsc.pdiparams',
+            algo,
+            round_type,
+            data_urls,
+            data_md5s,
+            "pwg_baker_static_0.4",
+            quantizable_op_type,
+            is_full_quantize,
+            is_use_cache_file,
+            is_optimize_model,
+            diff_threshold,
+            batch_nums=2,
+        )
+
+
 class TestPostTraininghistForMobilenetv1(TestPostTrainingQuantization):
     def test_post_training_hist_mobilenetv1(self):
         model = "MobileNet-V1"