Merge branch 'lukas_issue284_feat_create-fcn-model-class' into 'develop'

new var loss and custom loss methods

See merge request toar/mlair!265

mlair/model_modules/fully_connected_networks.py

@@ -5,6 +5,7 @@ from functools import reduce, partial
 
 from mlair.model_modules import AbstractModelClass
 from mlair.helpers import select_from_dict
+from mlair.model_modules.loss import var_loss, custom_loss
 
 import keras
 
@@ -64,7 +65,9 @@ class FCN(AbstractModelClass):
 
     _activation = {"relu": keras.layers.ReLU, "tanh": partial(keras.layers.Activation, "tanh"),
                    "sigmoid": partial(keras.layers.Activation, "sigmoid"),
-                   "linear": partial(keras.layers.Activation, "linear")}
+                   "linear": partial(keras.layers.Activation, "linear"),
+                   "selu": partial(keras.layers.Activation, "selu")}
+    _initializer = {"selu": keras.initializers.lecun_normal()}
     _optimizer = {"adam": keras.optimizers.adam, "sgd": keras.optimizers.SGD}
     _requirements = ["lr", "beta_1", "beta_2", "epsilon", "decay", "amsgrad", "momentum", "nesterov"]
 
@@ -87,6 +90,7 @@ class FCN(AbstractModelClass):
         self.optimizer = self._set_optimizer(optimizer, **kwargs)
         self.layer_configuration = (n_layer, n_hidden)
         self._update_model_name()
+        self.kernel_initializer = self._initializer.get(activation, "glorot_uniform")
 
         # apply to model
         self.set_model()
@@ -126,11 +130,12 @@ class FCN(AbstractModelClass):
         x_in = keras.layers.Flatten()(x_input)
         n_layer, n_hidden = self.layer_configuration
         for layer in range(n_layer):
-            x_in = keras.layers.Dense(n_hidden)(x_in)
+            x_in = keras.layers.Dense(n_hidden, kernel_initializer=self.kernel_initializer)(x_in)
             x_in = self.activation()(x_in)
         x_in = keras.layers.Dense(self._output_shape)(x_in)
         out = self.activation_output()(x_in)
         self.model = keras.Model(inputs=x_input, outputs=[out])
 
     def set_compile_options(self):
-        self.compile_options = {"loss": [keras.losses.mean_squared_error], "metrics": ["mse", "mae"]}
+        self.compile_options = {"loss": [custom_loss([keras.losses.mean_squared_error, var_loss])],
+                                "metrics": ["mse", "mae", var_loss]}

mlair/model_modules/loss.py

@@ -20,3 +20,21 @@ def l_p_loss(power: int) -> Callable:
         return K.mean(K.pow(K.abs(y_pred - y_true), power), axis=-1)
 
     return loss
+
+
+def var_loss(y_true, y_pred) -> Callable:
+    return K.mean(K.square(K.var(y_true) - K.var(y_pred)))
+
+
+def custom_loss(loss_list, loss_weights=None) -> Callable:
+    n = len(loss_list)
+    if loss_weights is None:
+        loss_weights = [1. / n for _ in range(n)]
+    else:
+        assert len(loss_weights) == n
+        loss_weights = [w / sum(loss_weights) for w in loss_weights]
+
+    def loss(y_true, y_pred):
+        return sum([loss_weights[i] * loss_list[i](y_true, y_pred) for i in range(n)])
+
+    return loss

test/test_model_modules/test_loss.py

 import keras
 import numpy as np
 
-from mlair.model_modules.loss import l_p_loss
+from mlair.model_modules.loss import l_p_loss, var_loss, custom_loss
 
+import pytest
 
-class TestLoss:
+
+class TestLPLoss:
 
     def test_l_p_loss(self):
         model = keras.Sequential()
@@ -15,3 +17,41 @@ class TestLoss:
         model.compile(optimizer=keras.optimizers.Adam(), loss=l_p_loss(3))
         hist = model.fit(np.array([1, 0, -2, 0.5]), np.array([1, 1, 0, 0.5]), epochs=1)
         assert hist.history['loss'][0] == 2.25
+
+
+class TestVarLoss:
+
+    def test_var_loss(self):
+        model = keras.Sequential()
+        model.add(keras.layers.Lambda(lambda x: x, input_shape=(None,)))
+        model.compile(optimizer=keras.optimizers.Adam(), loss=var_loss)
+        hist = model.fit(np.array([1, 0, 2, 0.5]), np.array([1, 1, 0, 0.5]), epochs=1)
+        assert hist.history['loss'][0] == 0.140625
+
+
+class TestCustomLoss:
+
+    def test_custom_loss_no_weights(self):
+        cust_loss = custom_loss([l_p_loss(2), var_loss])
+        model = keras.Sequential()
+        model.add(keras.layers.Lambda(lambda x: x, input_shape=(None,)))
+        model.compile(optimizer=keras.optimizers.Adam(), loss=cust_loss)
+        hist = model.fit(np.array([1, 0, 2, 0.5]), np.array([1, 1, 0, 0.5]), epochs=1)
+        assert hist.history['loss'][0] == (0.5 * 0.140625 + 0.5 * 1.25)
+
+    @pytest.mark.parametrize("weights", [[0.3, 0.7], [0.5, 0.5], [1, 1], [4, 1]])
+    def test_custom_loss_with_weights(self, weights):
+        cust_loss = custom_loss([l_p_loss(2), var_loss], weights)
+        model = keras.Sequential()
+        model.add(keras.layers.Lambda(lambda x: x, input_shape=(None,)))
+        model.compile(optimizer=keras.optimizers.Adam(), loss=cust_loss)
+        hist = model.fit(np.array([1, 0, 2, 0.5]), np.array([1, 1, 0, 0.5]), epochs=1)
+        weights_adjusted = list(map(lambda x: x / sum(weights), weights))
+        expected = (weights_adjusted[0] * 1.25 + weights_adjusted[1] * 0.140625)
+        assert np.testing.assert_almost_equal(hist.history['loss'][0], expected, decimal=6) is None
+
+    def test_custom_loss_invalid_weights(self):
+        with pytest.raises(AssertionError):
+            custom_loss([l_p_loss(2), var_loss], [0.3])
+        with pytest.raises(AssertionError):
+            custom_loss([l_p_loss(2), var_loss], [0.4, 3, 1])