new class CNNfromConfig to be able to configure CNN in much more detail

mlair/model_modules/convolutional_networks.py

@@ -11,6 +11,124 @@ from mlair.model_modules.advanced_paddings import PadUtils, Padding2D, Symmetric
 import tensorflow.keras as keras
 
 
+class CNNfromConfig(AbstractModelClass):
+    _activation = {"relu": keras.layers.ReLU, "tanh": partial(keras.layers.Activation, "tanh"),
+                   "sigmoid": partial(keras.layers.Activation, "sigmoid"),
+                   "linear": partial(keras.layers.Activation, "linear"),
+                   "prelu": partial(keras.layers.PReLU, alpha_initializer=keras.initializers.constant(value=0.25)),
+                   "leakyrelu": partial(keras.layers.LeakyReLU)}
+    _initializer = {"tanh": "glorot_uniform", "sigmoid": "glorot_uniform", "linear": "glorot_uniform",
+                    "relu": keras.initializers.he_normal(), "selu": keras.initializers.lecun_normal(),
+                    "prelu": keras.initializers.he_normal()}
+    _optimizer = {"adam": keras.optimizers.Adam, "sgd": keras.optimizers.SGD}
+    _regularizer = {"l1": keras.regularizers.l1, "l2": keras.regularizers.l2, "l1_l2": keras.regularizers.l1_l2}
+    _requirements = ["lr", "beta_1", "beta_2", "epsilon", "decay", "amsgrad", "momentum", "nesterov", "l1", "l2"]
+
+    """
+    Use this class like the following. Note that all keys must match the corresponding tf/keras keys of the layer
+    
+    ```python
+        input_shape = [(65,1,9)]
+        output_shape = [(4, )]
+        layer_configuration=[
+            {"type": "Conv2D", "activation": "relu", "kernel_size": (1, 1), "filters": 8}, 
+            {"type": "Dropout", "rate": 0.2},
+            {"type": "Conv2D", "activation": "relu", "kernel_size": (5, 1), "filters": 16}, 
+            {"type": "Dropout", "rate": 0.2},
+            {"type": "MaxPooling2D", "pool_size": (8, 1), "strides": (1, 1)},
+            {"type": "Conv2D", "activation": "relu", "kernel_size": (1, 1), "filters": 16}, 
+            {"type": "Dropout", "rate": 0.2},
+            {"type": "Conv2D", "activation": "relu", "kernel_size": (5, 1), "filters": 32}, 
+            {"type": "Dropout", "rate": 0.2},
+            {"type": "MaxPooling2D", "pool_size": (8, 1), "strides": (1, 1)},
+            {"type": "Conv2D", "activation": "relu", "kernel_size": (1, 1), "filters": 32}, 
+            {"type": "Dropout", "rate": 0.2},
+            {"type": "Conv2D", "activation": "relu", "kernel_size": (5, 1), "filters": 64}, 
+            {"type": "Dropout", "rate": 0.2},
+            {"type": "MaxPooling2D", "pool_size": (8, 1), "strides": (1, 1)},
+            {"type": "Conv2D", "activation": "relu", "kernel_size": (1, 1), "filters": 64}, 
+            {"type": "Dropout", "rate": 0.2},
+            {"type": "Flatten"},
+            # {"type": "Dense", "units": 128, "activation": "relu"}
+        ]
+        model = CNNfromConfig(input_shape, output_shape, layer_configuration)
+    ```
+
+    """
+
+    def __init__(self, input_shape: list, output_shape: list, layer_configuration: list, **kwargs):
+
+        assert len(input_shape) == 1
+        assert len(output_shape) == 1
+        super().__init__(input_shape[0], output_shape[0])
+
+        self.conf = layer_configuration
+        activation_output = kwargs.pop("activation_output", "linear")
+        self.activation_output = self._activation.get(activation_output)
+        self.activation_output_name = activation_output
+        self.kwargs = kwargs
+
+        # apply to model
+
+        self.set_model()
+        self.set_compile_options()
+        self.set_custom_objects(loss=custom_loss([keras.losses.mean_squared_error, var_loss]), var_loss=var_loss)
+
+    def set_model(self):
+        x_input = keras.layers.Input(shape=self._input_shape)
+        x_in = x_input
+
+        for layer_opts in self.conf:
+            print(layer_opts)
+            layer, layer_kwargs, follow_up_layer = self._extract_layer_conf(layer_opts)
+            x_in = layer(**layer_kwargs)(x_in)
+            if follow_up_layer is not None:
+                x_in = follow_up_layer()(x_in)
+
+        x_in = keras.layers.Dense(self._output_shape)(x_in)
+        out = self.activation_output(name=f"{self.activation_output_name}_output")(x_in)
+        self.model = keras.Model(inputs=x_input, outputs=[out])
+        print(self.model.summary())
+
+    def _set_regularizer(self, regularizer, **kwargs):
+        if regularizer is None or (isinstance(regularizer, str) and regularizer.lower() == "none"):
+            return None
+        try:
+            reg_name = regularizer.lower()
+            reg = self._regularizer.get(reg_name)
+            reg_kwargs = {}
+            if reg_name in ["l1", "l2"]:
+                reg_kwargs = select_from_dict(kwargs, reg_name, remove_none=True)
+                if reg_name in reg_kwargs:
+                    reg_kwargs["l"] = reg_kwargs.pop(reg_name)
+            elif reg_name == "l1_l2":
+                reg_kwargs = select_from_dict(kwargs, ["l1", "l2"], remove_none=True)
+            return reg(**reg_kwargs)
+        except KeyError:
+            raise AttributeError(f"Given regularizer {regularizer} is not supported in this model class.")
+
+    def set_compile_options(self):
+        # self.compile_options = {"loss": [custom_loss([keras.losses.mean_squared_error, var_loss])],
+        #                         "metrics": ["mse", "mae", var_loss]}
+        self.compile_options = {"loss": [keras.losses.mean_squared_error],
+                                "metrics": ["mse", "mae", var_loss]}
+
+    def _extract_layer_conf(self, layer_opts):
+        follow_up_layer = None
+        layer_type = layer_opts.pop("type")
+        layer = getattr(keras.layers, layer_type, None)
+        activation_type = layer_opts.pop("activation", None)
+        if activation_type is not None:
+            activation = self._activation.get(activation_type)
+            kernel_initializer = self._initializer.get(activation_type, "glorot_uniform")
+            layer_opts["kernel_initializer"] = kernel_initializer
+            follow_up_layer = activation
+        regularizer_type = layer_opts.pop("kernel_regularizer", None)
+        if regularizer_type is not None:
+            layer_opts["kernel_regularizer"] = self._set_regularizer(regularizer_type, **self.kwargs)
+        return layer, layer_opts, follow_up_layer
+
+
 class CNN(AbstractModelClass):  # pragma: no cover
 
     _activation = {"relu": keras.layers.ReLU, "tanh": partial(keras.layers.Activation, "tanh"),