Upgrade antspy version (#25)

* Upgrade antspy version

* Extend to other python versions

* Modigy github.workflow

* Try to make github actions work

* Remove 3.8 and 3.9 because tf version

* Try to remove smaller constrain on tensorflow

* Try python version 3.8

* Try version 3.8 (2)

* Remove tensorflow constrains and add all python versions

* Try to add colors to CI

* Remove Keras version

* Focus on 3.8

* Upgrade tensorflow

* Fix tf seed

* Update to KerasTensor and compat.v1

* Remove Keras dependency

* Remove allensdk dependency

* Remove SimpleITK dependency

* Fix issues

* Add macos in github actions

* Fix typo

Author: EmilieDel

.github/workflows/run-tests.yml

@@ -11,11 +11,24 @@ jobs:
 
   run_test:
 
-    runs-on: ubuntu-latest
+    runs-on: ${{ matrix.os }}
 
     strategy:
+
       matrix:
-        python-version: [3.7]  # currently ANTsPy wheels are only avail for py37
+        os: [ubuntu-latest] # macos 11 is currently in preview, macos-latest == 1.10.15
+        python-version: [
+          3.7,
+          3.8,
+          3.9,
+        ]
+        include:
+          - python-version: 3.7
+            tox-env: py37
+          - python-version: 3.8
+            tox-env: py38
+          - python-version: 3.9
+            tox-env: py39
 
     steps:
 
@@ -30,13 +43,13 @@ jobs:
       - name: install python dependencies
         run: |
           python -m pip install --upgrade pip
-          pip install tox
+          pip install tox tox-gh-actions
 
       - name: linting and code style
         run: tox -vv -e lint
 
       - name: tests and coverage
-        run: tox -vv -e py37
+        run: tox -vv -e ${{ matrix.tox-env }} -- --color=yes
 
       - name: docs
         run: tox -vv -e docs

atlalign/allen/utils.py

@@ -27,16 +27,16 @@
 """
 
 import os
+import urllib
 
 import matplotlib.pyplot as plt
 import numpy as np
 import requests
-from allensdk.api.queries.image_download_api import ImageDownloadApi
 
 CACHE_FOLDER = os.path.expanduser("~/.atlalign/")
 
 
-def get_image(image_id, folder=None, **kwargs):
+def get_image(image_id, folder=None, expression=False):
     """Get any image from Allen's database just by its id.
 
     Notes
@@ -52,8 +52,9 @@ def get_image(image_id, folder=None, **kwargs):
     folder : str or LocalPath or None
         Local folder where image saved. If None then automatically defaults to `CACHE_FOLDER`.
 
-    **kwargs
-        Additional parameters to be passed onto the `download_image` method of ``ImageDownloadApi``. See
+    expression : bool
+        If True, retrieve the specified SectionImage's expression mask image.
+        Otherwise, retrieve the specified SectionImage.
         See references for details.
 
     Returns
@@ -76,11 +77,7 @@ def get_image(image_id, folder=None, **kwargs):
         os.makedirs(folder)
 
     # Create full path
-    additional_speficier = "_".join(
-        sorted(["{}_{}".format(k, v) for k, v in kwargs.items()])
-    )
-    if additional_speficier:
-        additional_speficier = "_{}".format(additional_speficier)
+    additional_speficier = "_expression" if expression else ""
     path = "{}{}{}.jpg".format(folder, image_id, additional_speficier)
 
     # Check image exists
@@ -93,10 +90,13 @@ def get_image(image_id, folder=None, **kwargs):
         return img
 
     else:
-
-        img_api = ImageDownloadApi()
-        img_api.download_image(image_id, file_path=path, **kwargs)
-        return get_image(image_id, **kwargs)
+        image_url = (
+            f"http://api.brain-map.org/api/v2/section_image_download/{str(image_id)}"
+        )
+        if expression:
+            image_url += "?view=expression"
+        urllib.request.urlretrieve(image_url, path)
+        return get_image(image_id, expression=expression)
 
 
 def get_2d(image_id, ref2inp=False, add_last=False):

atlalign/base.py

@@ -37,12 +37,6 @@
     SmoothBivariateSpline,
     griddata,
 )
-
-try:
-    import SimpleITK as sitk
-except ImportError as e:
-    print(e)
-
 from skimage.transform import resize
 
 from atlalign.utils import griddata_custom
@@ -990,24 +984,23 @@ def pseudo_inverse(
         """
         interpolator_kwargs = interpolator_kwargs or {}
 
-        if interpolation_method != "itk":
-            x, y = np.meshgrid(list(range(self.shape[1])), list(range(self.shape[0])))
-            xi = (y, x)
-            x_r, y_r = x.ravel(), y.ravel()
+        x, y = np.meshgrid(list(range(self.shape[1])), list(range(self.shape[0])))
+        xi = (y, x)
+        x_r, y_r = x.ravel(), y.ravel()
 
-            points = np.hstack(
-                (
-                    (y_r + self.delta_y.ravel()).reshape(-1, 1),
-                    (x_r + self.delta_x.ravel()).reshape(-1, 1),
-                )
+        points = np.hstack(
+            (
+                (y_r + self.delta_y.ravel()).reshape(-1, 1),
+                (x_r + self.delta_x.ravel()).reshape(-1, 1),
             )
+        )
 
-            # Downsampling
-            points = points[::ds_f]
-            x_r_ds = x_r[::ds_f]
-            y_r_ds = y_r[::ds_f]
+        # Downsampling
+        points = points[::ds_f]
+        x_r_ds = x_r[::ds_f]
+        y_r_ds = y_r[::ds_f]
 
-            x_, y_ = points[:, 1], points[:, 0]
+        x_, y_ = points[:, 1], points[:, 0]
 
         if interpolation_method == "griddata":
             values_grid_x = griddata(points=points, values=x_r_ds, xi=xi)
@@ -1023,35 +1016,6 @@ def pseudo_inverse(
             delta_x = values_grid_x.reshape(self.shape) - x
             delta_y = values_grid_y.reshape(self.shape) - y
 
-        elif interpolation_method == "itk":
-            # ~ 30 ms per image
-            df_sitk = sitk.GetImageFromArray(
-                np.concatenate(
-                    (self.delta_x[..., np.newaxis], self.delta_y[..., np.newaxis]),
-                    axis=2,
-                ),
-                isVector=True,
-            )
-
-            invertor = sitk.InvertDisplacementFieldImageFilter()
-
-            # Set behaviour
-            user_spec = {
-                "n_iter": interpolator_kwargs.get("n_iter", 20),
-                "tol": interpolator_kwargs.get("tol", 1e-3),
-            }
-
-            # invertor.EnforceBoundaryConditionOn()
-            invertor.SetMeanErrorToleranceThreshold(user_spec["tol"])  # big effect
-            invertor.SetMaximumNumberOfIterations(user_spec["n_iter"])  # big effect
-
-            # Run
-            df_sitk_inv = invertor.Execute(df_sitk)
-
-            delta_xy = sitk.GetArrayFromImage(df_sitk_inv)
-
-            delta_x, delta_y = delta_xy[..., 0], delta_xy[..., 1]
-
         elif interpolation_method == "noop":
             # for benchmarking purposes
             delta_x, delta_y = np.zeros(self.shape), np.zeros(self.shape)

atlalign/metrics.py

@@ -1000,13 +1000,14 @@ def perceptual_loss_img(y_true, y_pred, model="net-lin", net="vgg"):
     y_true = np.stack((y_true,) * 3, axis=-1)
     y_pred = np.stack((y_pred,) * 3, axis=-1)
 
-    image0_ph = tf.placeholder(tf.float32)
-    image1_ph = tf.placeholder(tf.float32)
+    image0_ph = tf.Variable(tf.float32)  # noqa: F841
+    image1_ph = tf.Variable(tf.float32)  # noqa: F841
 
-    distance_t = lpips_tf.lpips(image0_ph, image1_ph, model=model, net=net)
+    @tf.function
+    def lpips(image0_ph, image1_ph):
+        return lpips_tf.lpips(image0_ph, image1_ph, model=model, net=net)
 
-    with tf.Session() as session:
-        pl = session.run(distance_t, feed_dict={image0_ph: y_true, image1_ph: y_pred})
+    pl = lpips(y_true, y_pred)
 
     tf.reset_default_graph()
 

atlalign/ml_utils/callbacks.py

@@ -22,9 +22,9 @@
 import pathlib
 
 import h5py
-import keras
 import mlflow
 import pandas as pd
+from tensorflow import keras
 
 from atlalign.data import annotation_volume, segmentation_collapsing_labels
 from atlalign.metrics import evaluate_single

atlalign/ml_utils/io.py

@@ -24,9 +24,9 @@
 import pathlib
 
 import h5py
-import keras
 import mlflow
 import numpy as np
+from tensorflow import keras
 
 from atlalign.base import DisplacementField
 from atlalign.data import nissl_volume

atlalign/ml_utils/layers.py

@@ -21,18 +21,19 @@
 
 import numpy as np
 import tensorflow as tf
-from keras import backend as K
-from keras.engine.topology import Layer
-from keras.layers import (
+from tensorflow.keras import backend as K
+from tensorflow.keras.layers import (
     BatchNormalization,
     Conv2D,
     Dense,
     Flatten,
     Lambda,
+    Layer,
     MaxPool2D,
     ReLU,
     Reshape,
 )
+from tensorflow_addons.image import resampler
 
 
 def K_meshgrid(x, y):
@@ -349,7 +350,7 @@ def call(self, tensors, mask=None):
 
         f_x_f_y = grid + dvfs
 
-        output = tf.contrib.resampler.resampler(imgs, f_x_f_y)
+        output = resampler(imgs, f_x_f_y)
 
         return output
 

atlalign/ml_utils/losses.py

@@ -19,8 +19,8 @@
     along with this program.  If not, see <https://www.gnu.org/licenses/>.
 """
 
-import keras
-import keras.backend as K
+import tensorflow.keras.backend as K
+from tensorflow import keras
 
 try:
     import lpips_tf
@@ -141,7 +141,7 @@ def ncc(self, I, J):  # noqa
         cc = cross * cross / (I_var * J_var + self.eps)
 
         # return negative cc.
-        return tf.reduce_mean(cc)
+        return tf.reduce_mean(input_tensor=cc)
 
     def loss(self, I, J):  # noqa
         """Compute loss."""
@@ -269,12 +269,12 @@ def _diffs(self, y):
     def loss(self, _, y_pred):
         """Compute loss."""
         if self.penalty == "l1":
-            df = [tf.reduce_mean(tf.abs(f)) for f in self._diffs(y_pred)]
+            df = [tf.reduce_mean(input_tensor=tf.abs(f)) for f in self._diffs(y_pred)]
         else:
             assert self.penalty == "l2", (
                 "penalty can only be l1 or l2. Got: %s" % self.penalty
             )
-            df = [tf.reduce_mean(f * f) for f in self._diffs(y_pred)]
+            df = [tf.reduce_mean(input_tensor=f * f) for f in self._diffs(y_pred)]
         return tf.add_n(df) / len(df)
 
 
@@ -310,7 +310,8 @@ def jacobian(_, y_pred):
 
     n_pixels = tf.constant(np.prod(keras.backend.int_shape(det)[1:]), dtype=tf.float32)
     count_artifacts = tf.cast(
-        tf.count_nonzero(tf.greater_equal(-det, 0.0), axis=(1, 2)), dtype=tf.float32
+        tf.math.count_nonzero(tf.greater_equal(-det, 0.0), axis=(1, 2)),
+        dtype=tf.float32,
     )
     perc_artifacts = count_artifacts / n_pixels
 
@@ -394,10 +395,12 @@ def vector_distance(y_true, y_pred):
     diff = y_pred - y_true
 
     vector_distance_per_output = tf.reduce_mean(
-        tf.sqrt(tf.abs(tf.square(diff[..., 0]) + tf.square(diff[..., 1]) + 0.001)),
+        input_tensor=tf.sqrt(
+            tf.abs(tf.square(diff[..., 0]) + tf.square(diff[..., 1]) + 0.001)
+        ),
         axis=0,
     )
-    vector_distance_average = tf.reduce_mean(vector_distance_per_output)
+    vector_distance_average = tf.reduce_mean(input_tensor=vector_distance_per_output)
 
     return vector_distance_average
 
@@ -421,8 +424,8 @@ def mse_po(y_true, y_pred):
     diff = y_pred - y_true
 
     vector_distance_per_output = tf.reduce_mean(
-        tf.square(diff[..., 0]) + tf.square(diff[..., 1]), axis=0
+        input_tensor=tf.square(diff[..., 0]) + tf.square(diff[..., 1]), axis=0
     )
-    vector_distance_average = tf.reduce_mean(vector_distance_per_output)
+    vector_distance_average = tf.reduce_mean(input_tensor=vector_distance_per_output)
 
     return vector_distance_average

atlalign/ml_utils/models.py

@@ -22,10 +22,10 @@
 import pathlib
 from copy import deepcopy
 
-from keras.layers import Lambda, concatenate
-from keras.models import Model
-from keras.models import load_model as load_model_keras
-from keras.models import model_from_json
+from tensorflow.keras.layers import Lambda, concatenate
+from tensorflow.keras.models import Model
+from tensorflow.keras.models import load_model as load_model_keras
+from tensorflow.keras.models import model_from_json
 
 from atlalign.ml_utils import (
     Affine2DVF,
@@ -120,7 +120,7 @@ def save_model(model, path, separate=True, overwrite=True):
         raise ValueError("Please specify a path without extension (folder).")
 
     if not separate:
-        model.save(str(path) + ".h5", overwrite=overwrite)
+        model.save(str(path) + ".h5", overwrite=overwrite, save_format="h5")
 
     else:
         path_architecture = path / (path.stem + ".json")

atlalign/nn.py

@@ -19,10 +19,10 @@
     along with this program.  If not, see <https://www.gnu.org/licenses/>.
 """
 
-import keras
 import mlflow
 import numpy as np
-from keras.layers import (
+from tensorflow import keras
+from tensorflow.keras.layers import (
     Conv2D,
     Cropping2D,
     Dense,
@@ -36,7 +36,7 @@
     ZeroPadding2D,
     concatenate,
 )
-from keras.models import Model
+from tensorflow.keras.models import Model
 
 from atlalign.ml_utils import (
     ALL_DVF_LOSSES,