Add wavelet transformer

ecg/feature.py

@@ -1,12 +1,39 @@
 from sklearn import preprocessing
+import pywt
+import numpy as np
+from builtins import zip
 
 
 class Normalizer(object):
     def __init__(self):
         self.scaler = None
 
+    def _dim_fix(self, x):
+        if (len(x.shape) == 2):
+            x = np.expand_dims(x, axis=-1)
+        assert(len(x.shape) == 3)
+        return x
+
     def fit(self, x):
+        x = self._dim_fix(x)
+        x = x.reshape((x.shape[0]*x.shape[1], x.shape[2]))
         self.scaler = preprocessing.StandardScaler().fit(x)
 
     def transform(self, x):
-        return self.scaler.transform(x)
+        x = self._dim_fix(x)
+        original_shape = x.shape
+        new_shape = (x.shape[0]*x.shape[1], x.shape[2])
+        return self.scaler.transform(
+            x.reshape(new_shape)).reshape(original_shape)
+
+
+class WaveletTransformer(object):
+    def __init__(self, wavelet_type='db1'):
+        self.wavelet_type = wavelet_type
+
+    def fit(self, x):
+        pass
+
+    def transform(self, x):
+        x_new = np.array([np.array(pywt.dwt(x_indiv, self.wavelet_type)).T for x_indiv in x])
+        return x_new

ecg/keras_models/model.py

@@ -27,7 +27,7 @@ def add_recurrent_layers(model, **params):
             Recurrent = LSTM
         rec_layer = Recurrent(
                     params["recurrent_hidden"],
-                    consume_less="gpu",
+                    consume_less="mem",
                     dropout_W=params["recurrent_dropout"],
                     dropout_U=params["recurrent_dropout"],
                     return_sequences=True)

ecg/loader.py

@@ -51,10 +51,11 @@ def _postprocess(self, use_one_hot):
         self.x_train = np.array(self.x_train)
         self.x_test = np.array(self.x_test)
 
-        transformer = feature.Normalizer()
-        transformer.fit(self.x_train)
-        self.x_train = transformer.transform(self.x_train)
-        self.x_test = transformer.transform(self.x_test)
+        for transformer_fn in [feature.WaveletTransformer, feature.Normalizer]:
+            transformer = transformer_fn()
+            transformer.fit(self.x_train)
+            self.x_train = transformer.transform(self.x_train)
+            self.x_test = transformer.transform(self.x_test)
 
         label_counter = collections.Counter(l for labels in self.y_train
                                             for l in labels)
@@ -158,5 +159,4 @@ def output_dim(self):
         count += 1
         assert len(ecgs) == len(labels) == batch_size, \
             "Invalid number of examples."
-        assert len(ecgs[0].shape) == 1, "ECG array should be 1D"
     assert count == len(ldr.x_train) // batch_size, "Wrong number of batches."

ecg/train-keras.py

@@ -64,11 +64,11 @@ def save_params(params, start_time, net_type):
         seed=2016,
         use_cached_if_available=not args.refresh)
 
-    x_train = dl.x_train[:, :, np.newaxis]
+    x_train = dl.x_train
     y_train = dl.y_train
     print("Training size: " + str(len(x_train)) + " examples.")
 
-    x_val = dl.x_test[:, :, np.newaxis]
+    x_val = dl.x_test
     y_val = dl.y_test
     print("Validation size: " + str(len(x_val)) + " examples.")
 

requirements.txt

@@ -15,6 +15,7 @@ pydot-ng==1.0.0
 pyparsing==2.1.10
 python-dateutil==2.6.0
 pytz==2016.10
+PyWavelets==0.5.1
 PyYAML==3.12
 scikit-learn==0.18.1
 scipy==0.18.1