Update _Dist/NeuralNetworks

Author: carefree0910

_Dist/NeuralNetworks/Base.py

@@ -37,7 +37,7 @@ def __init__(self, x, y, weights=None, name="Generator", shuffle=True):
                 self.n_class = 1
         self._name = name
         self._do_shuffle = shuffle
-        self._all_valid_data = np.hstack([self._x, self._y.reshape([-1, 1])])
+        self._all_valid_data = self._generate_all_valid_data()
         self._valid_indices = np.arange(len(self._all_valid_data))
         self._random_indices = self._valid_indices.copy()
         np.random.shuffle(self._random_indices)
@@ -72,6 +72,9 @@ def n_dim(self):
     def shape(self):
         return self.n_valid, self.n_dim
 
+    def _generate_all_valid_data(self):
+        return np.hstack([self._x, self._y.reshape([-1, 1])])
+
     def _cache_current_status(self):
         self._cache["_valid_indices"] = self._valid_indices
         self._cache["_random_indices"] = self._random_indices
@@ -149,6 +152,33 @@ def get_all_data(self):
         return self._get_data(self._valid_indices)
 
 
+class Generator3d(Generator):
+    @property
+    def n_time_step(self):
+        return self._x.shape[1]
+
+    @property
+    def shape(self):
+        return self.n_valid, self.n_time_step, self.n_dim
+
+    def _generate_all_valid_data(self):
+        return np.array([(x, y) for x, y in zip(self._x, self._y)])
+
+
+class Generator4d(Generator3d):
+    @property
+    def height(self):
+        return self._x.shape[1]
+
+    @property
+    def width(self):
+        return self._x.shape[2]
+
+    @property
+    def shape(self):
+        return self.n_valid, self.height, self.width, self.n_dim
+
+
 class Base:
     def __init__(self, name=None, model_param_settings=None, model_structure_settings=None):
         tf.reset_default_graph()
@@ -159,6 +189,7 @@ def __init__(self, name=None, model_param_settings=None, model_structure_setting
 
         self._settings_initialized = False
 
+        self._generator_base = Generator
         self._train_generator = self._test_generator = None
         self._sample_weights = self._tf_sample_weights = None
         self.n_dim = self.n_class = None
@@ -221,9 +252,9 @@ def init_from_data(self, x, y, x_test, y_test, sample_weights, names):
         else:
             self._tf_sample_weights = tf.placeholder(tf.float32, name="sample_weights")
 
-        self._train_generator = Generator(x, y, self._sample_weights, name="TrainGenerator")
+        self._train_generator = self._generator_base(x, y, self._sample_weights, name="TrainGenerator")
         if x_test is not None and y_test is not None:
-            self._test_generator = Generator(x_test, y_test, name="TestGenerator")
+            self._test_generator = self._generator_base(x_test, y_test, name="TestGenerator")
         else:
             self._test_generator = None
         self.n_random_train_subset = int(len(self._train_generator) * 0.1)
@@ -431,7 +462,8 @@ def _define_tf_collections(self):
     def add_tf_collections(self):
         for tensor in self.tf_collections:
             target = getattr(self, tensor)
-            tf.add_to_collection(tensor, target)
+            if target is not None:
+                tf.add_to_collection(tensor, target)
 
     def clear_tf_collections(self):
         for key in self.tf_collections:
@@ -450,6 +482,8 @@ def restore_collections(self, folder):
                 setattr(self, name, value)
         for tensor in self.tf_collections:
             target = tf.get_collection(tensor)
+            if target is None:
+                continue
             assert len(target) == 1, "{} available '{}' found".format(len(target), tensor)
             setattr(self, tensor, target[0])
         self.clear_tf_collections()

_Dist/NeuralNetworks/e_AdvancedNN/NN.py

@@ -37,11 +37,7 @@ def __init__(self, name=None, data_info=None, model_param_settings=None, model_s
         self._use_wide_network = self._dndf = self._pruner = None
 
         self._tf_p_keep = None
-        self._n_batch_placeholder = tf.placeholder(tf.int32, name="n_batch")
-
-    def init_all_settings(self):
-        super(Advanced, self).init_all_settings()
-        self.tf_collections.append("_n_batch_placeholder")
+        self._n_batch_placeholder = None
 
     def init_data_info(self):
         if self._data_info_initialized:
@@ -222,6 +218,7 @@ def _define_input_and_placeholder(self):
             self._is_training, lambda: self.dropout_keep_prob, lambda: 1.,
             name="p_keep"
         )
+        self._n_batch_placeholder = tf.placeholder(tf.int32, name="n_batch")
 
     def _define_py_collections(self):
         super(Advanced, self)._define_py_collections()
@@ -230,7 +227,7 @@ def _define_py_collections(self):
     def _define_tf_collections(self):
         super(Advanced, self)._define_tf_collections()
         self.tf_collections += [
-            "_deep_input", "_wide_input",
+            "_deep_input", "_wide_input", "_n_batch_placeholder",
             "_embedding", "_one_hot", "_embedding_with_one_hot",
             "_embedding_concat", "_one_hot_concat", "_embedding_with_one_hot_concat",
         ]
@@ -240,12 +237,15 @@ def add_tf_collections(self):
         super(Advanced, self).add_tf_collections()
         for tf_list in self.tf_list_collections:
             target_list = getattr(self, tf_list)
+            if target_list is None:
+                continue
             for tensor in target_list:
                 tf.add_to_collection(tf_list, tensor)
 
     def restore_collections(self, folder):
         for tf_list in self.tf_list_collections:
-            setattr(self, tf_list, tf.get_collection(tf_list))
+            if tf_list is not None:
+                setattr(self, tf_list, tf.get_collection(tf_list))
         super(Advanced, self).restore_collections(folder)
 
     def print_settings(self):

_Dist/NeuralNetworks/g_DistNN/NN.py

@@ -7,7 +7,6 @@
 import numpy as np
 import tensorflow as tf
 
-from _Dist.NeuralNetworks.Base import Generator
 from _Dist.NeuralNetworks.f_AutoNN.NN import Auto
 from _Dist.NeuralNetworks.NNUtil import Toolbox, PreProcessor
 
@@ -56,7 +55,7 @@ def rolling_fit(self, train_rate=0.8, cv_rate=0.1, sample_weights=None, **kwargs
                 if self._test_generator is None:
                     test_data, _ = self._train_generator.get_range(train_cursor, test_cursor)
                     x_test, y_test = test_data[..., :-1], test_data[..., -1]
-                    self._test_generator = Generator(x_test, y_test, name="TestGenerator")
+                    self._test_generator = self._generator_base(x_test, y_test, name="TestGenerator")
                 self._train_generator.set_range(cursor, train_cursor)
                 kwargs["print_settings"] = print_settings
                 self.fit(**kwargs)
@@ -96,9 +95,9 @@ def increment_fit(self, x=None, y=None, x_test=None, y_test=None, sample_weights
         self._handle_unbalance(y)
         self._handle_sparsity()
         if data is not None:
-            self._train_generator = Generator(x, y, self._sample_weights, name="Generator")
+            self._train_generator = self._generator_base(x, y, self._sample_weights, name="Generator")
             if x_test is not None and y_test is not None:
-                self._test_generator = Generator(x_test, y_test, name="TestGenerator")
+                self._test_generator = self._generator_base(x_test, y_test, name="TestGenerator")
         self.fit(**kwargs)
         x, y, _ = self._gen_batch(self._train_generator, self.n_random_train_subset, True)
         print("  -  Performance of increment fit", end=" | ")
@@ -170,7 +169,7 @@ def k_fold(self, k=10, data=None, test_rate=0., sample_weights=None, **kwargs):
         self._merge_preprocessors_from_k_series(names)
         self._sample_weights = sample_weights_store
         if x_test is not None and y_test is not None:
-            self._test_generator = Generator(x_test, y_test, name="TestGenerator")
+            self._test_generator = self._generator_base(x_test, y_test, name="TestGenerator")
         return self
 
     def k_random(self, k=3, data=None, cv_rate=0.1, test_rate=0., sample_weights=None, **kwargs):
@@ -201,7 +200,7 @@ def k_random(self, k=3, data=None, cv_rate=0.1, test_rate=0., sample_weights=Non
         self._merge_preprocessors_from_k_series(names)
         self._sample_weights = sample_weights_store
         if x_test is not None and y_test is not None:
-            self._test_generator = Generator(x_test, y_test, name="TestGenerator")
+            self._test_generator = self._generator_base(x_test, y_test, name="TestGenerator")
         return self