imporve keras layer and kl algo

Signed-off-by: zehao-intel <[email protected]>

neural_compressor/tensorflow/keras/layers/conv2d.py

@@ -354,40 +354,15 @@ def get_config(self):
 def initialize_int8_conv2d(fp32_layer, q_config):
     kwargs = fp32_layer.get_config()
 
-    if "name" in kwargs:
-        del kwargs["name"]
-    if "filters" in kwargs:
-        del kwargs["filters"]
-    if "kernel_size" in kwargs:
-        del kwargs["kernel_size"]
-    if "strides" in kwargs:
-        del kwargs["strides"]
-    if "padding" in kwargs:
-        del kwargs["padding"]
-    if "data_format" in kwargs:
-        del kwargs["data_format"]
-    if "dilation_rate" in kwargs:
-        del kwargs["dilation_rate"]
-    if "groups" in kwargs:
-        del kwargs["groups"]
-    if "activation" in kwargs:
-        del kwargs["activation"]
-    if "use_bias" in kwargs:
-        del kwargs["use_bias"]
-    if "kernel_initializer" in kwargs:
-        del kwargs["kernel_initializer"]
-    if "bias_initializer" in kwargs:
-        del kwargs["bias_initializer"]
-    if "kernel_regularizer" in kwargs:
-        del kwargs["kernel_regularizer"]
-    if "activity_regularizer" in kwargs:
-        del kwargs["activity_regularizer"]
-    if "bias_regularizer" in kwargs:
-        del kwargs["bias_regularizer"]
-    if "kernel_constraint" in kwargs:
-        del kwargs["kernel_constraint"]
-    if "bias_constraint" in kwargs:
-        del kwargs["bias_constraint"]
+    param_list = ["name", "filters", "kernel_size", "strides", "padding",
+                  "data_format", "dilation_rate", "groups", "activation",
+                  "use_bias", "kernel_initializer", "bias_initializer",
+                  "kernel_regularizer", "activity_regularizer", "bias_regularizer",
+                  "kernel_constraint", "bias_constraint"
+                ]
+    for p in param_list:  # pragma: no cover
+        if p in kwargs:
+            del kwargs[p]
 
     return QConv2D(
         name=fp32_layer.name,

neural_compressor/tensorflow/keras/layers/dense.py

@@ -170,29 +170,14 @@ def get_config(self):
 def initialize_int8_dense(fp32_layer, q_config):
     kwargs = fp32_layer.get_config()
 
-    if "name" in kwargs:
-        del kwargs["name"]
-    if "units" in kwargs:
-        del kwargs["units"]
-    if "activation" in kwargs:
-        del kwargs["activation"]
-    if "use_bias" in kwargs:
-        del kwargs["use_bias"]
-    if "kernel_initializer" in kwargs:
-        del kwargs["kernel_initializer"]
-    if "bias_initializer" in kwargs:
-        del kwargs["bias_initializer"]
-    if "kernel_regularizer" in kwargs:
-        del kwargs["kernel_regularizer"]
-    if "activity_regularizer" in kwargs:
-        del kwargs["activity_regularizer"]
-    if "bias_regularizer" in kwargs:
-        del kwargs["bias_regularizer"]
-    if "kernel_constraint" in kwargs:
-        del kwargs["kernel_constraint"]
-    if "bias_constraint" in kwargs:
-        del kwargs["bias_constraint"]
-
+    param_list = ["name", "units", "activation", "use_bias", "kernel_initializer", 
+                  "bias_initializer", "kernel_regularizer", "activity_regularizer", 
+                  "bias_regularizer", "kernel_constraint", "bias_constraint"
+                ]
+    for p in param_list:  # pragma: no cover
+        if p in kwargs:
+            del kwargs[p]
+
     q_layer = QDense(
         name=fp32_layer.name,
         units=fp32_layer.units,

neural_compressor/tensorflow/keras/layers/depthwise_conv2d.py

@@ -372,42 +372,15 @@ def initialize_int8_depthwise_conv2d(fp32_layer, q_config):
     kwargs = fp32_layer.get_config()
     q_name = fp32_layer.name
 
-    if "name" in kwargs:
-        del kwargs["name"]
-    if "kernel_size" in kwargs:
-        del kwargs["kernel_size"]
-    if "strides" in kwargs:
-        del kwargs["strides"]
-    if "padding" in kwargs:
-        del kwargs["padding"]
-    if "depth_multiplier" in kwargs:
-        del kwargs["depth_multiplier"]
-    if "data_format" in kwargs:
-        del kwargs["data_format"]
-    if "dilation_rate" in kwargs:
-        del kwargs["dilation_rate"]
-    if "activation" in kwargs:
-        del kwargs["activation"]
-    if "use_bias" in kwargs:
-        del kwargs["use_bias"]
-    if "depthwise_initializer" in kwargs:
-        del kwargs["depthwise_initializer"]
-    if "bias_initializer" in kwargs:
-        del kwargs["bias_initializer"]
-    if "depthwise_regularizer" in kwargs:
-        del kwargs["depthwise_regularizer"]
-    if "activity_regularizer" in kwargs:
-        del kwargs["activity_regularizer"]
-    if "bias_regularizer" in kwargs:
-        del kwargs["bias_regularizer"]
-    if "depthwise_constraint" in kwargs:
-        del kwargs["depthwise_constraint"]
-    if "bias_constraint" in kwargs:
-        del kwargs["bias_constraint"]
-    if "min_value" in kwargs:
-        del kwargs["min_value"]
-    if "max_value" in kwargs:
-        del kwargs["max_value"]
+    param_list = ["name", "kernel_size", "strides", "padding", "depth_multiplier",
+                  "data_format", "dilation_rate", "activation",
+                  "use_bias", "depthwise_initializer", "bias_initializer",
+                  "depthwise_regularizer", "activity_regularizer", "bias_regularizer",
+                  "depthwise_constraint", "bias_constraint"
+                ]
+    for p in param_list:  # pragma: no cover
+        if p in kwargs:
+            del kwargs[p]
 
     return QDepthwiseConv2D(
         name=q_name,

neural_compressor/tensorflow/keras/layers/pool2d.py

@@ -215,16 +215,10 @@ def get_config(self):
 def initialize_int8_avgpool(fp32_layer, q_config):
     kwargs = fp32_layer.get_config()
 
-    if "name" in kwargs:
-        del kwargs["name"]
-    if "pool_size" in kwargs:
-        del kwargs["pool_size"]
-    if "strides" in kwargs:
-        del kwargs["strides"]
-    if "padding" in kwargs:
-        del kwargs["padding"]
-    if "data_format" in kwargs:
-        del kwargs["data_format"]
+    param_list = ["name", "pool_size", "strides", "padding", "data_format", ]
+    for p in param_list:  # pragma: no cover
+        if p in kwargs:
+            del kwargs[p]
 
     q_layer = QAvgPool2D(
         name=fp32_layer.name,
@@ -243,16 +237,10 @@ def initialize_int8_avgpool(fp32_layer, q_config):
 def initialize_int8_maxpool(fp32_layer, q_config):
     kwargs = fp32_layer.get_config()
 
-    if "name" in kwargs:
-        del kwargs["name"]
-    if "pool_size" in kwargs:
-        del kwargs["pool_size"]
-    if "strides" in kwargs:
-        del kwargs["strides"]
-    if "padding" in kwargs:
-        del kwargs["padding"]
-    if "data_format" in kwargs:
-        del kwargs["data_format"]
+    param_list = ["name", "pool_size", "strides", "padding", "data_format", ]
+    for p in param_list:  # pragma: no cover
+        if p in kwargs:
+            del kwargs[p]
 
     q_layer = QMaxPool2D(
         name=fp32_layer.name,

neural_compressor/tensorflow/keras/layers/separable_conv2d.py

@@ -364,50 +364,16 @@ def get_config(self):
 def initialize_int8_separable_conv2d(fp32_layer, q_config):
     kwargs = fp32_layer.get_config()
 
-    if "name" in kwargs:
-        del kwargs["name"]
-    if "filters" in kwargs:
-        del kwargs["filters"]
-    if "kernel_size" in kwargs:
-        del kwargs["kernel_size"]
-    if "strides" in kwargs:
-        del kwargs["strides"]
-    if "padding" in kwargs:
-        del kwargs["padding"]
-    if "data_format" in kwargs:
-        del kwargs["data_format"]
-    if "dilation_rate" in kwargs:
-        del kwargs["dilation_rate"]
-    if "depth_multiplier" in kwargs:
-        del kwargs["depth_multiplier"]
-    if "activation" in kwargs:
-        del kwargs["activation"]
-    if "use_bias" in kwargs:
-        del kwargs["use_bias"]
-    if "depthwise_initializer" in kwargs:
-        del kwargs["depthwise_initializer"]
-    if "pointwise_initializer" in kwargs:
-        del kwargs["pointwise_initializer"]
-    if "bias_initializer" in kwargs:
-        del kwargs["bias_initializer"]
-    if "depthwise_regularizer" in kwargs:
-        del kwargs["depthwise_regularizer"]
-    if "pointwise_regularizer" in kwargs:
-        del kwargs["pointwise_regularizer"]
-    if "activity_regularizer" in kwargs:
-        del kwargs["activity_regularizer"]
-    if "bias_regularizer" in kwargs:
-        del kwargs["bias_regularizer"]
-    if "depthwise_constraint" in kwargs:
-        del kwargs["depthwise_constraint"]
-    if "pointwise_constraint" in kwargs:
-        del kwargs["pointwise_constraint"]
-    if "bias_constraint" in kwargs:
-        del kwargs["bias_constraint"]
-    if "min_value" in kwargs:
-        del kwargs["min_value"]
-    if "max_value" in kwargs:
-        del kwargs["max_value"]
+    param_list = ["name", "filters", "kernel_size", "strides", "padding",
+                  "data_format", "dilation_rate", "depth_multiplier", "activation",
+                  "use_bias", "depthwise_initializer", "bias_initializer", "pointwise_initializer",
+                  "depthwise_regularizer", "activity_regularizer", "bias_regularizer",
+                  "pointwise_regularizer", "depthwise_constraint", "bias_constraint",
+                  "pointwise_constraint",
+                ]
+    for p in param_list:  # pragma: no cover
+        if p in kwargs:
+            del kwargs[p]
 
     return QSeparableConv2D(
         name=fp32_layer.name,

test/3x/tensorflow/keras/test_layers.py

@@ -93,7 +93,7 @@ def build_model2():
             keras.layers.InputLayer(input_shape=(28, 28)),
             keras.layers.Reshape(target_shape=(28, 28, 1)),
             keras.layers.SeparableConv2D(3, 3, activation="relu"),
-            keras.layers.MaxPooling2D(pool_size=(2, 2)),
+            keras.layers.AveragePooling2D(pool_size=(2, 2)),
             keras.layers.Flatten(),
             keras.layers.Dense(10, name="dense"),
         ]

test/3x/tensorflow/test_config.py

@@ -213,6 +213,35 @@ def test_static_quant_from_dict_advance(self):
 
         self.assertEqual(conv2d_quantized, False)
 
+    def test_static_quant_from_dict_advance2(self):
+        logger.info("test_static_quant_from_dict_advance2")
+        from neural_compressor.tensorflow import quantize_model
+        from neural_compressor.tensorflow.utils import DummyDataset
+
+        dataset = DummyDataset(shape=(100, 32, 32, 3), label=True)
+        calib_dataloader = MyDataLoader(dataset=dataset)
+        fp32_model = self.graph
+        quant_config = {
+            "static_quant": {
+                "global": {
+                    "weight_dtype": "int8",
+                    "weight_sym": True,
+                    "weight_granularity": "per_channel",
+                    "act_dtype": "int8",
+                    "act_sym": True,
+                    "act_granularity": "per_channel",
+                },
+                "local": {
+                    "conv1": {
+                        "weight_algorithm": "kl",
+                        "act_algorithm": "kl",
+                    }
+                },
+            }
+        }
+        qmodel = quantize_model(fp32_model, quant_config, calib_dataloader)
+        self.assertIsNotNone(qmodel)
+
     def test_static_quant_from_class_advance(self):
         logger.info("test_static_quant_from_class_advance")
         from neural_compressor.tensorflow import StaticQuantConfig, quantize_model