Merge pull request #62 from Kevin2/same_img_size

Use the same image size in filter_deep

Author: FeGeyer

.gitignore

@@ -19,6 +19,7 @@ __pycache__/
 *.jpg
 *.pdf
 *.gif
+*.png
 
 # make
 *_done

.travis.yml

@@ -2,9 +2,15 @@ language: python
 python:
   - "3.7"
 
+# directories:
+#  - "/tmp/texlive"
+#  - "$HOME/.texlive"
+
 before_install:
   - sudo apt-get -y install libgeos-dev
   - sudo apt-get -y install libproj-dev
+  # - travis_wait 45 bash ./utilities/travis_setup.sh
+  # - export PATH="/tmp/texlive/bin/x86_64-linux:$PATH"
 
 install:
   - pip install .

dl_framework/architectures.py

@@ -557,10 +557,10 @@ def __init__(self, img_size):
         )
 
         self.conv4_amp = nn.Sequential(
-             *conv_amp(1, 4, (5, 5), 1, 3, 2)
+             *conv_amp(1, 4, (5, 5), 1, 4, 2)
         )
         self.conv4_phase = nn.Sequential(
-             *conv_phase(1, 4, (5, 5), 1, 3, 2, add=1-pi)
+             *conv_phase(1, 4, (5, 5), 1, 4, 2, add=1-pi)
         )
         self.conv5_amp = nn.Sequential(
              *conv_amp(4, 8, (5, 5), 1, 2, 1)
@@ -569,10 +569,10 @@ def __init__(self, img_size):
              *conv_phase(4, 8, (5, 5), 1, 2, 1, add=1-pi)
         )
         self.conv6_amp = nn.Sequential(
-             *conv_amp(8, 12, (3, 3), 1, 3, 2)
+             *conv_amp(8, 12, (3, 3), 1, 2, 2)
         )
         self.conv6_phase = nn.Sequential(
-             *conv_phase(8, 12, (3, 3), 1, 3, 2, add=1-pi)
+             *conv_phase(8, 12, (3, 3), 1, 2, 2, add=1-pi)
         )
         self.conv7_amp = nn.Sequential(
              *conv_amp(12, 16, (3, 3), 1, 1, 1)
@@ -710,13 +710,13 @@ def __init__(self, img_size):
         )
 
         self.conv4_amp = nn.Sequential(
-             *conv_amp(1, 4, (5, 5), 1, 3, 2)
+             *conv_amp(1, 4, (5, 5), 1, 4, 2)
         )
         self.conv5_amp = nn.Sequential(
              *conv_amp(4, 8, (5, 5), 1, 2, 1)
         )
         self.conv6_amp = nn.Sequential(
-             *conv_amp(8, 12, (3, 3), 1, 3, 2)
+             *conv_amp(8, 12, (3, 3), 1, 2, 2)
         )
         self.conv7_amp = nn.Sequential(
              *conv_amp(12, 16, (3, 3), 1, 1, 1)
@@ -787,51 +787,51 @@ class filter_deep_phase(nn.Module):
     def __init__(self, img_size):
         super().__init__()
         self.conv1_phase = nn.Sequential(
-            *conv_phase(1, 4, (23, 23), 1, 11, 1, add=-2.1415)
+            *conv_phase(1, 4, (23, 23), 1, 11, 1, add=1-pi)
         )
         self.conv2_phase = nn.Sequential(
-             *conv_phase(4, 8, (21, 21), 1, 10, 1, add=-2.1415)
+             *conv_phase(4, 8, (21, 21), 1, 10, 1, add=1-pi)
         )
         self.conv3_phase = nn.Sequential(
-             *conv_phase(8, 12, (17, 17), 1, 8, 1, add=-2.1415)
+             *conv_phase(8, 12, (17, 17), 1, 8, 1, add=1-pi)
         )
         self.conv_con1_phase = nn.Sequential(
             LocallyConnected2d(12, 1, img_size, 1, stride=1, bias=False),
             nn.BatchNorm2d(1),
-            GeneralELU(-2.1415),
+            GeneralELU(1-pi),
         )
 
         self.conv4_phase = nn.Sequential(
-             *conv_phase(1, 4, (5, 5), 1, 3, 2, add=-2.1415)
+             *conv_phase(1, 4, (5, 5), 1, 4, 2, add=1-pi)
         )
         self.conv5_phase = nn.Sequential(
-             *conv_phase(4, 8, (5, 5), 1, 2, 1, add=-2.1415)
+             *conv_phase(4, 8, (5, 5), 1, 2, 1, add=1-pi)
         )
         self.conv6_phase = nn.Sequential(
-             *conv_phase(8, 12, (3, 3), 1, 3, 2, add=-2.1415)
+             *conv_phase(8, 12, (3, 3), 1, 2, 2, add=1-pi)
         )
         self.conv7_phase = nn.Sequential(
-             *conv_phase(12, 16, (3, 3), 1, 1, 1, add=-2.1415)
+             *conv_phase(12, 16, (3, 3), 1, 1, 1, add=1-pi)
         )
         self.conv_con2_phase = nn.Sequential(
             LocallyConnected2d(16, 1, img_size, 1, stride=1, bias=False),
             nn.BatchNorm2d(1),
-            GeneralELU(-2.1415),
+            GeneralELU(1-pi),
         )
 
         self.conv8_phase = nn.Sequential(
-             *conv_phase(1, 4, (3, 3), 1, 1, 1, add=-2.1415)
+             *conv_phase(1, 4, (3, 3), 1, 1, 1, add=1-pi)
         )
         self.conv9_phase = nn.Sequential(
-             *conv_phase(4, 8, (3, 3), 1, 1, 1, add=-2.1415)
+             *conv_phase(4, 8, (3, 3), 1, 1, 1, add=1-pi)
         )
         self.conv10_phase = nn.Sequential(
-             *conv_phase(8, 12, (3, 3), 1, 2, 2, add=-2.1415)
+             *conv_phase(8, 12, (3, 3), 1, 2, 2, add=1-pi)
         )
         self.conv_con3_phase = nn.Sequential(
             LocallyConnected2d(12, 1, img_size, 1, stride=1, bias=False),
             nn.BatchNorm2d(1),
-            GeneralELU(-2.1415),
+            GeneralELU(1-pi),
         )
         self.symmetry_imag = Lambda(partial(symmetry, mode='imag'))
 

dl_framework/callbacks.py

@@ -135,8 +135,22 @@ def plot_lr(self):
         plt.tight_layout()
 
     def plot_loss(self, log=True):
-        plt.plot(self.train_losses, label="train loss")
-        plt.plot(self.valid_losses, label="valid loss")
+        import matplotlib as mpl
+
+        # make nice Latex friendly plots
+        # mpl.use("pgf")
+        # mpl.rcParams.update(
+        #     {
+        #         "font.size": 12,
+        #         "font.family": "sans-serif",
+        #         "text.usetex": True,
+        #         "pgf.rcfonts": False,
+        #         "pgf.texsystem": "lualatex",
+        #     }
+        # )
+
+        plt.plot(self.train_losses, label="training loss")
+        plt.plot(self.valid_losses, label="validation loss")
         if log:
             plt.yscale("log")
         plt.xlabel(r"Number of Epochs")

dl_framework/data.py

@@ -108,6 +108,8 @@ def open_image(self, var, i):
 
                 data_channel = torch.cat([data_amp, data_phase], dim=1)
         else:
+            if data.shape[1] == 2:
+                raise ValueError("Two channeled data is used despite Fourier being False. Set Fourier to True!")
             if len(i) == 1:
                 data_channel = data.reshape(data.shape[-1] ** 2)
             else:

dl_framework/inspection.py

@@ -7,6 +7,18 @@
 import dl_framework.architectures as architecture
 from dl_framework.model import load_pre_model
 
+# make nice Latex friendly plots
+# mpl.use("pgf")
+# mpl.rcParams.update(
+#     {
+#         "font.size": 12,
+#         "font.family": "sans-serif",
+#         "text.usetex": True,
+#         "pgf.rcfonts": False,
+#         "pgf.texsystem": "lualatex",
+#     }
+# )
+
 
 def load_pretrained_model(arch_name, model_path):
     """
@@ -121,7 +133,7 @@ def plot_loss(learn, model_path):
     save_path = model_path.split(".model")[0]
     print("\nPlotting Loss for: {}\n".format(name_model))
     learn.recorder.plot_loss()
-    plt.title(r"{}".format(name_model))
+    plt.title(r"{}".format(name_model.replace("_", " ")))
     plt.savefig("{}_loss.pdf".format(save_path), bbox_inches="tight", pad_inches=0.01)
     plt.clf()
     mpl.rcParams.update(mpl.rcParamsDefault)

dl_framework/learner.py

@@ -7,7 +7,16 @@
 from tqdm import tqdm
 import sys
 from functools import partial
-from dl_framework.loss_functions import init_feature_loss, loss_amp, loss_phase
+from dl_framework.loss_functions import (
+    init_feature_loss,
+    loss_amp,
+    loss_phase,
+    loss_msssim,
+    loss_mse_msssim,
+    loss_mse_msssim_phase,
+    loss_mse_msssim_amp,
+    loss_msssim_amp,
+)
 from dl_framework.callbacks import (
     AvgStatsCallback,
     BatchTransformXCallback,
@@ -188,6 +197,7 @@ def define_learner(
     opt_func=torch.optim.Adam,
 ):
     cbfs.extend([
+        # commented out because of normed and limited input values
         # partial(BatchTransformXCallback, norm),
     ])
     if not test:
@@ -202,7 +212,7 @@ def define_learner(
         ])
     if not test and not lr_find:
         cbfs.extend([
-            partial(LoggerCallback, model_name=model_name), 
+            partial(LoggerCallback, model_name=model_name),
             data_aug,
         ])
 
@@ -216,8 +226,18 @@ def define_learner(
         loss_func = loss_amp
     elif loss_func == "loss_phase":
         loss_func = loss_phase
+    elif loss_func == "msssim":
+        loss_func = loss_msssim
+    elif loss_func == "mse_msssim":
+        loss_func = loss_mse_msssim
+    elif loss_func == "mse_msssim_phase":
+        loss_func = loss_mse_msssim_phase
+    elif loss_func == "mse_msssim_amp":
+        loss_func = loss_mse_msssim_amp
+    elif loss_func == "msssim_amp":
+        loss_func = loss_msssim_amp
     else:
-        print("\n No matching loss function! Exiting. \n")
+        print("\n No matching loss function or architecture! Exiting. \n")
         sys.exit(1)
 
     # Combine model and data in learner