From fbffafa41fa12ca20247ce14ef0379b6ff6e05d3 Mon Sep 17 00:00:00 2001
From: leufen1 <l.leufen@fz-juelich.de>
Date: Tue, 1 Mar 2022 16:50:40 +0100
Subject: [PATCH] improved CNN, first try
---
mlair/model_modules/convolutional_networks.py | 81 +++++++++++++++++--
1 file changed, 75 insertions(+), 6 deletions(-)
diff --git a/mlair/model_modules/convolutional_networks.py b/mlair/model_modules/convolutional_networks.py
index d8eb6eb3..5fd81133 100644
--- a/mlair/model_modules/convolutional_networks.py
+++ b/mlair/model_modules/convolutional_networks.py
@@ -17,7 +17,8 @@ class CNN(AbstractModelClass): # pragma: no cover
"sigmoid": partial(keras.layers.Activation, "sigmoid"),
"linear": partial(keras.layers.Activation, "linear"),
"selu": partial(keras.layers.Activation, "selu"),
- "prelu": partial(keras.layers.PReLU, alpha_initializer=keras.initializers.constant(value=0.25))}
+ "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()}
@@ -27,7 +28,9 @@ class CNN(AbstractModelClass): # pragma: no cover
_dropout = {"selu": keras.layers.AlphaDropout}
def __init__(self, input_shape: list, output_shape: list, activation="relu", activation_output="linear",
- optimizer="adam", regularizer=None, kernel_size=1, dropout=None, **kwargs):
+ optimizer="adam", regularizer=None, kernel_size=7, dropout=None, dropout_freq=None, pooling_freq=None,
+ n_layer=1, n_filter=10, layer_configuration=None, pooling_size=None,
+ dense_layer_configuration=None, **kwargs):
assert len(input_shape) == 1
assert len(output_shape) == 1
@@ -42,13 +45,24 @@ class CNN(AbstractModelClass): # pragma: no cover
self.kernel_regularizer = self._set_regularizer(regularizer, **kwargs)
self.kernel_size = kernel_size
self.optimizer = self._set_optimizer(optimizer, **kwargs)
+ self.layer_configuration = (n_layer, n_filter, self.kernel_size) if layer_configuration is None else layer_configuration
+ self.dense_layer_configuration = dense_layer_configuration or []
+ self.pooling_size = pooling_size
self.dropout, self.dropout_rate = self._set_dropout(activation, dropout)
+ self.dropout_freq = self._set_layer_freq(dropout_freq)
+ self.pooling_freq = self._set_layer_freq(pooling_freq)
# 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_layer_freq(self, param):
+ param = 0 if param is None else param
+ assert 0 <= param
+ assert isinstance(param, int)
+ return param
+
def _set_activation(self, activation):
try:
return self._activation.get(activation.lower())
@@ -91,6 +105,65 @@ class CNN(AbstractModelClass): # pragma: no cover
assert 0 <= dropout_rate < 1
return self._dropout.get(activation, keras.layers.Dropout), dropout_rate
+ def set_model(self):
+ """
+ Build the model.
+ """
+ if isinstance(self.layer_configuration, tuple) is True:
+ n_layer, n_hidden, kernel_size = self.layer_configuration
+ if isinstance(kernel_size, list):
+ assert len(kernel_size) == n_layer # use individual filter sizes for each layer
+ conf = [(n_hidden, kernel_size[i]) for i in range(n_layer)]
+ else:
+ assert isinstance(kernel_size, int) # use same filter size for all layers
+ conf = [(n_hidden, kernel_size) for _ in range(n_layer)]
+ else:
+ assert isinstance(self.layer_configuration, list) is True
+ if not isinstance(self.layer_configuration[0], tuple):
+ if isinstance(self.kernel_size, list):
+ assert len(self.kernel_size) == len(self.layer_configuration) # use individual filter sizes for each layer
+ conf = [(n_filter, self.kernel_size[i]) for i, n_filter in enumerate(self.layer_configuration)]
+ else:
+ assert isinstance(self.kernel_size, int) # use same filter size for all layers
+ conf = [(n_filter, self.kernel_size) for n_filter in self.layer_configuration]
+ else:
+ assert len(self.layer_configuration[0]) == 2
+ conf = self.layer_configuration
+
+ x_input = keras.layers.Input(shape=self._input_shape)
+ x_in = x_input
+ for layer, (n_filter, kernel_size) in enumerate(conf):
+ if self.pooling_size is not None and self.pooling_freq > 0 and layer % self.pooling_freq == 0 and layer > 0:
+ x_in = keras.layers.MaxPooling2D((self.pooling_size, 1), strides=(1, 1), padding='valid')(x_in)
+ x_in = keras.layers.Conv2D(filters=n_filter, kernel_size=(kernel_size, 1),
+ kernel_initializer=self.kernel_initializer,
+ kernel_regularizer=self.kernel_regularizer)(x_in)
+ x_in = self.activation()(x_in)
+ if self.dropout is not None and self.dropout_freq > 0 and layer % self.dropout_freq == 0:
+ x_in = self.dropout(self.dropout_rate)(x_in)
+
+ x_in = keras.layers.Flatten()(x_in)
+ for layer, n_hidden in enumerate(self.dense_layer_configuration):
+ if n_hidden < self._output_shape:
+ break
+ x_in = keras.layers.Dense(n_hidden, name=f"Dense_{len(conf) + layer + 1}",
+ kernel_initializer=self.kernel_initializer, )(x_in)
+ x_in = self.activation(name=f"{self.activation_name}_{len(conf) + layer + 1}")(x_in)
+ if self.dropout is not None:
+ x_in = self.dropout(self.dropout_rate)(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_compile_options(self):
+ self.compile_options = {"loss": [custom_loss([keras.losses.mean_squared_error, var_loss])],
+ "metrics": ["mse", "mae", var_loss]}
+
+
+class CNN_16_32_64(CNN):
+
def set_model(self):
"""
Build the model.
@@ -123,7 +196,3 @@ class CNN(AbstractModelClass): # pragma: no cover
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])
-
- def set_compile_options(self):
- self.compile_options = {"loss": [custom_loss([keras.losses.mean_squared_error, var_loss])],
- "metrics": ["mse", "mae", var_loss]}
--
GitLab