Auto regressive timeseries sampling method

Author: Francisco Santos

examples/timeseries/tscwgan_example.py

@@ -1,4 +1,4 @@
-from numpy import squeeze
+from numpy import reshape
 
 from ydata_synthetic.preprocessing.timeseries import processed_stock
 from ydata_synthetic.synthesizers.timeseries import TSCWGAN
@@ -51,9 +51,9 @@
 #Sampling the data
 #Note that the data returned is not inverse processed.
 cond_index = 100  # Arbitrary sequence for conditioning
-cond_array = squeeze(processed_data[cond_index][:cond_dim], axis=1)
+cond_array = reshape(processed_data[cond_index][:cond_dim], (1,-1))
 
-data_sample = synth.sample(cond_array, 1000)
+data_sample = synth.sample(cond_array, 1000, 100)
 
 # Inverting the scaling of the synthetic samples
 data_sample = inverse_transform(data_sample, scaler)

src/ydata_synthetic/synthesizers/timeseries/tscwgan/model.py

@@ -4,11 +4,11 @@
 And on: https://github.com/CasperHogenboom/WGAN_financial_time-series
 """
 from tqdm import trange
-from numpy import array, vstack
+from numpy import array, vstack, hstack
 from numpy.random import normal
 from typing import List
 
-from tensorflow import concat, float32, convert_to_tensor, reshape, GradientTape, reduce_mean, make_ndarray, make_tensor_proto, tile, constant
+from tensorflow import concat, float32, convert_to_tensor, reshape, GradientTape, reduce_mean, tile
 from tensorflow import data as tfdata
 from tensorflow.keras import Model, Sequential
 from tensorflow.keras.optimizers import Adam
@@ -17,7 +17,6 @@
 from ydata_synthetic.synthesizers.gan import BaseModel
 from ydata_synthetic.synthesizers import TrainParameters
 from ydata_synthetic.synthesizers.loss import Mode, gradient_penalty
-from ydata_synthetic.synthesizers.timeseries import TimeSeriesDataProcessor
 
 class TSCWGAN(BaseModel):
 
@@ -148,31 +147,35 @@ def get_batch_data(self, data, n_windows= None):
                                 .shuffle(buffer_size=n_windows)
                                 .batch(self.batch_size).repeat())
 
-    def sample(self, cond_array: array, n_samples: int, inverse_transform: bool = True):
-        """Provided that cond_array is passed, produce n_samples for each condition vector in cond_array.
-        The returned samples per condition will always be a multiple of batch_size and equal or bigger than n_samples.
-
-        Arguments:
-            cond_array (numpy array): Array with the set of conditions for the sampling process.
-            n_samples (int): Number of samples to be taken for each condition in cond_array.
-            inverse_transform (bool): """
-        assert len(cond_array.shape) == 2, "Condition array should be two-dimensional. N_conditions x cond_dim"
-        assert cond_array.shape[1] == self.cond_dim, \
-            f"The condition sequences should have a {self.cond_dim} length."
-        steps = n_samples // self.batch_size + 1
+    def sample(self, condition: array, n_samples: int = 100, seq_len: int = 24):
+        """For a given condition, produce n_samples of length seq_len.
+        The samples are returned in the original data format (any preprocessing transformation is inverted).
+
+        Args:
+            condition (numpy.array): Condition for the generated samples, must have the same length .
+            n_samples (int): Minimum number of generated samples (returns always a multiple of batch_size).
+            seq_len (int): Length of the generated samples.
+
+        Returns:
+            data (numpy.array): An array of data of shape [n_samples, seq_len]"""
+        assert len(condition.shape) == 2, "Condition array should be two-dimensional."
+        assert condition.shape[1] == self.cond_dim, \
+            f"The condition sequence should have {self.cond_dim} length."
+        batches = n_samples // self.batch_size + 1
+        ar_steps = seq_len // self.data_dim + 1
         data = []
         z_dist = self.get_batch_noise()
-        for condition in cond_array:
+        for batch in trange(batches, desc=f'Synthetic data generation'):
+            data_ = []
             cond_seq = convert_to_tensor(condition, float32)
-            cond_seq = tile(cond_seq, multiples=[self.batch_size, 1])
-            for step in trange(steps, desc=f'Synthetic data generation'):
-                gen_input = concat([cond_seq, next(z_dist)], axis=1)
+            gen_input = concat([tile(cond_seq, multiples=[self.batch_size, 1]), next(z_dist)], axis=1)
+            for step in range(ar_steps):
                 records = self.generator(gen_input, training=False)
-                data.append(records)
-        data = array(vstack(data))
-        if inverse_transform:
-            return self.processor.inverse_transform(data)
-        return data
+                gen_input = concat([records[:, -self.cond_dim:], next(z_dist)], axis=1)
+                data_.append(records)
+            data_ = hstack(data_)[:, :seq_len]
+            data.append(data_)
+        return self.processor.inverse_transform(array(vstack(data)))
 
 
 class Generator(Model):