Bugfixing

Author: bcebere

fflows/filters/spectral.py

@@ -50,17 +50,9 @@ def __init__(self, d, k, FFT, hidden, flip=False, RNN=False):
 
         self.d, self.k = d, k
 
-        if FFT:
-
-            self.out_size = self.d - self.k + 1
-            self.pz_size = self.d + 1
-            self.in_size = self.k
-
-        else:
-
-            self.out_size = self.d - self.k
-            self.pz_size = self.d
-            self.in_size = self.k
+        self.out_size = self.d - self.k
+        self.pz_size = self.d
+        self.in_size = self.k
 
         if flip:
 

fflows/fourier/transforms.py

@@ -7,9 +7,6 @@
 Used as the "Fourier Transform" step in a Fourier Flow, can be preceded by other torch modules.
 
 """
-
-from __future__ import absolute_import, division, print_function
-
 import sys
 import warnings
 
@@ -60,10 +57,10 @@ def reconstruct_DFT(x, component="real"):
     """
 
     if component == "real":
-        x_rec = torch.cat([x[0, :], flip(x[0, :], dim=0)[1:]], dim=0)
+        x_rec = torch.cat([x[0, :], flip(x[0, :], dim=0)], dim=0)
 
     elif component == "imag":
-        x_rec = torch.cat([x[1, :], -1 * flip(x[1, :], dim=0)[1:]], dim=0)
+        x_rec = torch.cat([x[1, :], -1 * flip(x[1, :], dim=0)], dim=0)
 
     return x_rec
 
@@ -93,7 +90,7 @@ def __init__(self, N_fft=100):
         super(DFT, self).__init__()
 
         self.N_fft = N_fft
-        self.crop_size = int(self.N_fft / 2) + 1
+        self.crop_size = int(np.ceil(self.N_fft / 2))
         base_mu, base_cov = torch.zeros(self.crop_size * 2), torch.eye(
             self.crop_size * 2
         )

fflows/sequential_flows.py

@@ -26,7 +26,7 @@ def __init__(self, hidden, n_flows, FFT=True, flip=True, normalize=False):
 
         self.FFT = FFT
         self.normalize = normalize
-        self.n_flows =n_flows
+        self.n_flows = n_flows
         self.hidden = hidden
 
         if flip:
@@ -42,7 +42,7 @@ def forward(self, x):
             if self.normalize:
                 x = (x - self.fft_mean) / (self.fft_std + 1e-8)
 
-            x = x.view(-1, self.d + 1)
+            x = x.view(-1, self.d)
 
         log_jacobs = []
 
@@ -55,17 +55,12 @@ def forward(self, x):
         return x, log_pz, sum(log_jacobs)
 
     def inverse(self, z):
-
         for bijector, f in zip(reversed(self.bijectors), reversed(self.flips)):
-
             z = bijector.inverse(z, flip=f)
 
         if self.FFT:
-
             if self.normalize:
-                z = z * self.fft_std.view(-1, self.d + 1) + self.fft_mean.view(
-                    -1, self.d + 1
-                )
+                z = z * self.fft_std.view(-1, self.d) + self.fft_mean.view(-1, self.d)
 
             z = self.FourierTransform.inverse(z)
 
@@ -74,10 +69,18 @@ def inverse(self, z):
     def fit(self, X, epochs=500, batch_size=128, learning_rate=1e-3, display_step=100):
         X_train = torch.from_numpy(np.array(X)).float()
 
+        self.carry_flag = False
+        if np.prod(X_train.shape[1:]) % 2 == 1:
+            repeat_last = X_train[:, :, -1:]
+            X_train = torch.concat([X_train, repeat_last], dim=2)
+            self.carry_flag = True
+
         self.individual_shape = X_train.shape[1:]
 
         self.d = np.prod(self.individual_shape)
-        self.k = int(np.floor(self.d / 2))
+        self.k = int(np.ceil(self.d / 2))
+
+        assert self.d % 2 == 0
 
         # Prepare models
         self.bijectors = nn.ModuleList(
@@ -95,6 +98,8 @@ def fit(self, X, epochs=500, batch_size=128, learning_rate=1e-3, display_step=10
 
         # for normalizing the spectral transforms
         X_train_spectral = self.FourierTransform(X_train)[0]
+        assert X_train_spectral.shape[-1] == self.k
+
         self.fft_mean = torch.mean(X_train_spectral, dim=0)
         self.fft_std = torch.std(X_train_spectral, dim=0)
         optim = torch.optim.Adam(self.parameters(), lr=learning_rate)
@@ -132,21 +137,18 @@ def fit(self, X, epochs=500, batch_size=128, learning_rate=1e-3, display_step=10
         return losses
 
     def sample(self, n_samples):
-
-        if self.FFT:
-
-            mu, cov = torch.zeros(self.d + 1), torch.eye(self.d + 1)
-
-        else:
-
-            mu, cov = torch.zeros(self.d), torch.eye(self.d)
+        mu, cov = torch.zeros(self.d), torch.eye(self.d)
 
         p_Z = MultivariateNormal(mu, cov)
         z = p_Z.rsample(sample_shape=(n_samples,))
 
         X_sample = self.inverse(z)
+        X_sample = X_sample.reshape(-1, *self.individual_shape)
 
-        return X_sample.reshape(-1, *self.individual_shape)
+        if self.carry_flag:
+            X_sample = X_sample[:, :, :-1]
+
+        return X_sample
 
 
 class RealNVP(nn.Module):
@@ -173,7 +175,7 @@ def forward(self, x):
             if self.normalize:
                 x = (x - self.fft_mean) / (self.fft_std + 1e-8)
 
-            x = x.view(-1, self.d + 1)
+            x = x.view(-1, self.d)
 
         log_jacobs = []
 
@@ -194,9 +196,7 @@ def inverse(self, z):
         if self.FFT:
 
             if self.normalize:
-                z = z * self.fft_std.view(-1, self.d + 1) + self.fft_mean.view(
-                    -1, self.d + 1
-                )
+                z = z * self.fft_std.view(-1, self.d) + self.fft_mean.view(-1, self.d1)
 
             z = self.FourierTransform.inverse(z)
 
@@ -259,7 +259,7 @@ def sample(self, n_samples):
 
         if self.FFT:
 
-            mu, cov = torch.zeros(self.d + 1), torch.eye(self.d + 1)
+            mu, cov = torch.zeros(self.d), torch.eye(self.d)
 
         else:
 
@@ -309,7 +309,7 @@ def forward(self, x):
             if self.normalize:
                 x = (x - self.fft_mean) / self.fft_std
 
-            x = x.view(-1, self.d + 1)
+            x = x.view(-1, self.d)
 
         log_jacobs = []
 
@@ -330,9 +330,7 @@ def inverse(self, z):
         if self.FFT:
 
             if self.normalize:
-                z = z * self.fft_std.view(-1, self.d + 1) + self.fft_mean.view(
-                    -1, self.d + 1
-                )
+                z = z * self.fft_std.view(-1, self.d) + self.fft_mean.view(-1, self.d)
 
             z = self.FourierTransform.inverse(z)
 
@@ -384,7 +382,7 @@ def sample(self, n_samples):
 
         if self.FFT:
 
-            mu, cov = torch.zeros(self.d + 1), torch.eye(self.d + 1)
+            mu, cov = torch.zeros(self.d), torch.eye(self.d)
 
         else:
 

fflows/utils/data_padding.py

@@ -1,8 +1,5 @@
 # Copyright (c) 2020, Ahmed M. Alaa
 # Licensed under the BSD 3-clause license (see LICENSE.txt)
-
-from __future__ import absolute_import, division, print_function
-
 import sys
 
 import numpy as np

fflows/utils/make_data.py

@@ -1,8 +1,5 @@
 # Copyright (c) 2020, Ahmed M. Alaa
 # Licensed under the BSD 3-clause license (see LICENSE.txt)
-
-from __future__ import absolute_import, division, print_function
-
 import sys
 
 import numpy as np
@@ -38,7 +35,7 @@ def create_autoregressive_data(
     # Create the input features
 
     X = [np.random.normal(X_m, X_v, (seq_len, n_features)) for k in range(n_samples)]
-    w = np.array([memory_factor ** k for k in range(seq_len)])
+    w = np.array([memory_factor**k for k in range(seq_len)])
 
     if mode == "noise-sweep":
 

fflows/utils/spectral.py

@@ -1,9 +1,5 @@
 # Copyright (c) 2021, Ahmed M. Alaa
 # Licensed under the BSD 3-clause license (see LICENSE.txt)
-
-
-from __future__ import absolute_import, division, print_function
-
 import numpy as np
 
 

fflows/version.py

@@ -1 +1 @@
-__version__ = "0.0.2"
+__version__ = "0.0.3"

tests/test_fourier_flows.py

@@ -4,18 +4,21 @@
 from data import sine_data_generation
 
 
-@pytest.mark.parametrize("normalize", [True, False])
+@pytest.mark.parametrize("normalize", [False, True])
 @pytest.mark.parametrize("FFT", [True, False])
 @pytest.mark.parametrize("flip", [True, False])
-def test_training(normalize: bool, FFT: bool, flip: bool) -> None:
-    T = 11
+@pytest.mark.parametrize("T", [12, 11])
+@pytest.mark.parametrize("dims", [4, 6, 3])
+@pytest.mark.parametrize("n_flows", [5, 10, 15])
+def test_training(
+    normalize: bool, FFT: bool, flip: bool, T: int, dims: int, n_flows: int
+) -> None:
     n_samples = 100
-    dims = 3
-    X = sine_data_generation(no=n_samples, seq_len=T, dim=3)
+    X = sine_data_generation(no=n_samples, seq_len=T, dim=dims)
 
     ff_params = {
         "hidden": 11,
-        "n_flows": 11,
+        "n_flows": n_flows,
         "normalize": normalize,
         "FFT": FFT,
         "flip": flip,
@@ -31,10 +34,10 @@ def test_training(normalize: bool, FFT: bool, flip: bool) -> None:
     _ = model.fit(X, **train_params)
 
 
-def test_generation() -> None:
-    T = 11
+@pytest.mark.parametrize("T", [12, 11])
+@pytest.mark.parametrize("dims", [4, 3])
+def test_generation(T: int, dims: int) -> None:
     n_samples = 100
-    dims = 3
     X = sine_data_generation(no=n_samples, seq_len=T, dim=dims)
 
     ff_params = {"hidden": 11, "n_flows": 11, "normalize": False}

tests/test_real_nvp.py

@@ -5,11 +5,11 @@
 
 
 @pytest.mark.parametrize("normalize", [True, False])
-def test_training(normalize: bool) -> None:
-    T = 11
+@pytest.mark.parametrize("T", [11, 12])
+@pytest.mark.parametrize("dims", [3, 4, 6])
+def test_training(normalize: bool, T: int, dims: int) -> None:
     n_samples = 100
-    dims = 3
-    X = sine_data_generation(no=n_samples, seq_len=T, dim=3)
+    X = sine_data_generation(no=n_samples, seq_len=T, dim=dims)
 
     model_params = {"hidden": 11, "n_flows": 11, "normalize": normalize}
     train_params = {