Merge branch 'master' of gitext.gfz-potsdam.de:sec2.1_ml/deeplearning

pinky/README.md

@@ -18,3 +18,21 @@ You can dump your examples to TFRecordDatasets to accelerate io operations:
 
 and use the newly created config file to run `--train`
 
+
+Tests
+-----
+
+ - basic learning (synthetics, real data)
+   # TODO: split data -> training, evaluation
+ - synthetics, layered cake, train with top and bottom layer containing events.
+        Validate with events within middle layer
+ - evaluation using 'unknown' velocity model
+ - test extrapolation with fault plane geometry
+ - synthetic pretraining
+ - hyperparameter optimization using skopt
+
+Outlook
+-------
+
+ - increase z error weight -> improve z estimates
+

pinky/src/model.py

@@ -68,6 +68,8 @@ class Model(Object):
         :param kernel_width: convolution kernel size accross time axis
 
         (Needs some debugging and checking)
+        TODO:
+         - cross_channel_kernel
         '''
         _, n_channels, n_samples, _ = input.shape
 
@@ -113,7 +115,7 @@ class Model(Object):
 
         # tf.summary.image('input', features)
         for ilayer in range(params.get('n_layers', 3)):
-            input = self.time_axis_cnn(input, n_filters, 1,
+            input = self.time_axis_cnn(input, n_filters * (1 + ilayer), 1,
                     kernel_width=int(kernel_width + ilayer*kernel_width_factor),
                     name='conv_%s' % ilayer, training=training)
         # conv = self.time_axis_cnn(conv, n_filters*2, 1, kernel_width=2, name='conv2',

pinky/src/optimize.py

+import matplotlib as mpl
+mpl.use('PDF')
+
 import os
 from .util import delete_if_exists
 from skopt import gp_minimize
 from skopt import dump as dump_result
 from skopt import load as load_result
 from skopt.space import Real, Categorical, Integer
-from skopt.plots import plot_convergence, plot_objective_2D
+# from skopt.plots import plot_convergence, plot_objective_2D
+from skopt.plots import plot_convergence, plot_objective
 from skopt.plots import plot_objective, plot_evaluations
+import matplotlib.pyplot as plt
+import logging
+from pyrocko.guts import Object, Int, Float, List, Tuple, String
+
 try:
     from skopt.plots import plot_histogram
     _plot_histogram_error = False
 except ImportError as e:
     _plot_histogram_error = e
-    logger.debug(e)
-
-from pyrocko.guts import Object, Int, Float, List, Tuple, String
-import logging
+    logging.debug(e)
 
 logger = logging.getLogger()
 
@@ -30,6 +35,12 @@ class Optimizer(Object):
     path_out = String.T(default='optimizer-results', help='base path where to store results, plots and logs')
 
     def __init__(self, **kwargs):
+
+        '''
+        TODO:
+         - optimize kernel heigth (cross_channel_kernel)
+
+        '''
         super().__init__(**kwargs)
         self.model = None
         self.result = None
@@ -101,17 +112,17 @@ class Optimizer(Object):
                 n_calls=self.n_calls,
                 x0=self.optimizer_values)
 
-        dump_result(self.result, self.fn_result)
+        # dump_result(self.result, self.fn_result)
         self.evaluate_result()
         self.plot_results()
 
     def ensure_result(self):
         ''' Load and set minimizer result.'''
         if self.result is None:
+            if self.fn_result is None:
+                logger.warn(
+                    'Cannot load results from filename: %s' % self.fn_result)
             self.result = load_result(self.fn_result)
-        else:
-            logger.warn(
-                'Cannot load results from filename: %s' % self.fn_result)
 
     def extend_path(self, *path):
         '''Prepend `self.path_out` to `path`.'''
@@ -120,33 +131,41 @@ class Optimizer(Object):
     def evaluate_result(self):
         self.ensure_result()
 
-        best = self.result.space.point_to_dict(self.result.x)
+        # best = self.result.space.point_to_dict(self.result.x)
+        best = self.result.x
         logger.info('Best parameter set:')
         logger.info(best)
 
         logger.info('Best parameter loss:')
         logger.info(self.result.fun)
 
+    def ensure_directory(self, directory):
+        if not os.path.exists(directory):
+            os.makedirs(directory)
+
     def plot_results(self):
         '''Produce and save result plots. '''
         self.ensure_result()
+        self.ensure_directory(self.extend_path('plots'))
 
         if _plot_histogram_error:
             logger.warn(_plot_histogram_error)
         else:
             for dim_name in self.optimizer_keys:
-                fig, ax = plot_histogram(result=self.result, dimension_name=dim_name)
-                fig.savefig(extend_path('plots/histogram_%s.pdf' % dim_name))
-
-        fig, ax = plot_objective(
-            result=self.result,
-            dimension_names=self.non_categorical_dimensions)
-        fig.savefig(extend_path('plots/objectives.pdf'))
-
-        fig, ax = plot_evaluations(
-            result=self.result,
-            dimension_names=self.non_categorical_dimensions)
-        fig.savefig(extend_path('plots/evaluations.pdf'))
+                fig, ax = plot_histogram(result=self.result) #, dimension_name=dim_name)
+                fig.savefig(self.extend_path('plots/histogram_%s.pdf' % dim_name))
+
+        ax = plot_objective(
+            result=self.result,)
+            # dimension_names=self.non_categorical_dimensions)
+        fig = plt.gcf()
+        fig.savefig(self.extend_path('plots/objectives.pdf'))
+
+        ax = plot_evaluations(
+            result=self.result,)
+            # dimension_names=self.non_categorical_dimensions)
+        fig = plt.gcf()
+        fig.savefig(self.extend_path('plots/evaluations.pdf'))
 
     def log_dir_name(self, params):
         '''Helper function to transform `params` into a logging directory