garf 2.7__tar.gz → 2.8__tar.gz

{garf-2.7 → garf-2.8}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: garf
-Version: 2.7
+Version: 2.8
 Summary: Utility tools for GATE ARF simulations
 Author-email: David Sarrut <david.sarrut@creatis.insa-lyon.fr>
 License:                                  Apache License
@@ -212,7 +212,7 @@ Requires-Python: >=3.8
 Description-Content-Type: text/markdown
 License-File: LICENSE
 Requires-Dist: itk
-Requires-Dist: numpy
+Requires-Dist: numpy<2.0.0
 Requires-Dist: pre-commit
 Requires-Dist: torch>=2.1.0
 Requires-Dist: tqdm

{garf-2.7 → garf-2.8}/garf/bin/garf_compare_image_profile.py

@@ -1,9 +1,8 @@
 #!/usr/bin/env python3
 # -*- coding: utf-8 -*-
 
-import sys
 import numpy as np
-import itk
+import SimpleITK as sitk
 import matplotlib.pyplot as plt
 import click
 
@@ -24,8 +23,8 @@ CONTEXT_SETTINGS = dict(help_option_names=["-h", "--help"])
 @click.option("--wslice", "-w", default=int(3), help="Slice width (to smooth)")
 def garf_compare_image_profile(image1_mhd, image2_mhd, islice, events, wslice):
     # Load image
-    img_ref = itk.imread(image1_mhd)
-    img = itk.imread(image2_mhd)
+    img_ref = sitk.ReadImage(image1_mhd)
+    img = sitk.ReadImage(image2_mhd)
     events = float(events)
     islice = int(islice)
     wslice = int(wslice)
@@ -34,8 +33,8 @@ def garf_compare_image_profile(image1_mhd, image2_mhd, islice, events, wslice):
     img = img * events
 
     # Get the pixels values as np array
-    data_ref = itk.GetArrayFromImage(img_ref).astype(float)
-    data = itk.GetArrayFromImage(img).astype(float)
+    data_ref = sitk.GetArrayFromImage(img_ref).astype(float)
+    data = sitk.GetArrayFromImage(img).astype(float)
 
     # Sometimes not same nb of slices -> crop the data_ref
     if len(data_ref) > len(data):

{garf-2.7 → garf-2.8}/garf/garf_detector.py

@@ -206,9 +206,7 @@ class GarfDetector:
         # FIXME : float instead of double because of w_pred is float
         # TODO : check float / double ?
         self.output_image = np.zeros(tuple(t_image_size), dtype=np.float32)
-        if self.current_gpu_mode == "mps":
-            self.output_image = self.output_image.astype(np.float32)
-        self.output_image = Tensor(torch.from_numpy(self.output_image))
+        self.output_image = torch.from_numpy(self.output_image)
         self.output_image = self.output_image.to(self.current_gpu_device)
 
     def initialize_planes_numpy(self, gaga_batch_size):
@@ -273,7 +271,7 @@ class GarfDetector:
             if len(projected_batch) > self.batch_size:
                 print(
                     f"Cannot use GARF, {len(projected_batch)} points while "
-                    f"batch size is {self.batch_size}"
+                    f"batch size is {self.batch_size}. Increase the garf batch_size."
                 )
                 exit(-1)
             projected_points[i] = projected_batch
@@ -410,7 +408,8 @@ class GarfDetector:
     def image_from_coordinates_add_torch(self, img, vu, w_pred):
         img_r = img.ravel()
         ind_r = vu[:, 1] * img.shape[0] + vu[:, 0]
-        img_r.index_add_(0, ind_r, w_pred)
+        # img_r.index_add_(dim=0, index=ind_r, source=w_pred) # <--- BUG on MPS
+        img_r = img_r.index_add(dim=0, index=ind_r, source=w_pred)
         img = img_r.reshape_as(img)
         return img
 
@@ -529,8 +528,12 @@ def normalize_logproba(x):
     exb_sum = torch.sum(exb, dim=1)
     # divide if not equal at zero
     p = torch.divide(exb.T, exb_sum, out=torch.zeros_like(exb.T)).T
+
+    # sometimes: nan.
+    p[torch.isnan(p)] = 0
+
     # check (should be equal to 1.0)
-    # check = np.sum(p, axis=1)
+    # check = torch.sum(p, dim=1)
     # print(check)
     return p
 
@@ -646,7 +649,11 @@ def normalize_proba_with_russian_roulette(w_pred, channel, rr):
 
     # normalize
     p_sum = torch.sum(w_pred, axis=1, keepdims=True)
-    w_pred = w_pred / p_sum
+    # w_pred = w_pred / p_sum
+    w_pred = torch.divide(w_pred, p_sum, out=torch.zeros_like(w_pred))
+
+    # remove nan
+    w_pred[torch.isnan(w_pred)] = 0
 
     # check
     # p_sum = torch.sum(w_pred, axis=1)
@@ -688,16 +695,21 @@ def image_from_coordinates_add_numpy(img, u, v, w_pred, hit_slice=False):
     # fill image using index broadcasting
     # Important: the >0 condition is to avoid outside elements.
     tiny = 0  # ??
-    for i in range(1, nb_ene):
-        # sum up values for pixel coordinates which occur multiple times
-        chx = np.bincount(uv32, weights=w_pred[:, i])
-        img[i, uv16Bins[chx > tiny, 0], uv16Bins[chx > tiny, 1]] += chx[chx > tiny]
-
-    # Consider the hit slice ?
     if hit_slice:
+        for i in range(1, nb_ene):
+            # sum up values for pixel coordinates which occur multiple times
+            chx = np.bincount(uv32, weights=w_pred[:, i])
+            img[i, uv16Bins[chx > tiny, 0], uv16Bins[chx > tiny, 1]] += chx[chx > tiny]
+        # Consider the hit slice
         chx = np.bincount(uv32)
         img[0, uv16Bins[chx > tiny, 0], uv16Bins[chx > tiny, 1]] += chx[chx > tiny]
 
+    else:
+        for i in range(0, nb_ene - 1):
+            # sum up values for pixel coordinates which occur multiple times
+            chx = np.bincount(uv32, weights=w_pred[:, i + 1])
+            img[i, uv16Bins[chx > tiny, 0], uv16Bins[chx > tiny, 1]] += chx[chx > tiny]
+
 
 def arf_plane_intersection(batch, plane, image_plane_size_mm):
     """

{garf-2.7 → garf-2.8}/garf/garf_train.py

@@ -220,28 +220,29 @@ def train_nn(x_train, y_train, params):
                 "{} epochs without improvement, early stop.".format(early_stopping)
             )
             stop = True
-            break
 
         # scheduler for learning rate
         scheduler.step(mean_loss)
 
         # FIXME WRONG
         # Check if need to print and store this epoch
-        if epoch % epoch_store_every == 0 or best_train_loss < previous_best:
-            tqdm.write(
-                "Epoch {} best is {:.5f} at epoch {:.0f}".format(
-                    epoch, best_loss, best_epoch
-                )
-            )
+        if best_train_loss < previous_best:
+            tqdm.write("Epoch {} loss is {:.5f}".format(epoch, best_loss))
+            previous_best = best_train_loss
+
+        if (
+            (epoch != 0 and epoch % epoch_store_every == 0)
+            or stop
+            or epoch >= epoch_max - 1
+        ):
             optim_data = dict()
+            print("Store weights", epoch)
             optim_data["epoch"] = epoch
             optim_data["train_loss"] = train_loss
             state = copy.deepcopy(model.state_dict())
             nn["optim"]["model_state"].append(state)
             nn["optim"]["data"].append(optim_data)
-            if best_train_loss < previous_best:
-                best_epoch_index = len(nn["optim"]["model_state"]) - 1
-                previous_best = best_train_loss
+            best_epoch_index = epoch
 
         # update progress bar
         pbar.update(1)

{garf-2.7 → garf-2.8}/garf.egg-info/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: garf
-Version: 2.7
+Version: 2.8
 Summary: Utility tools for GATE ARF simulations
 Author-email: David Sarrut <david.sarrut@creatis.insa-lyon.fr>
 License:                                  Apache License
@@ -212,7 +212,7 @@ Requires-Python: >=3.8
 Description-Content-Type: text/markdown
 License-File: LICENSE
 Requires-Dist: itk
-Requires-Dist: numpy
+Requires-Dist: numpy<2.0.0
 Requires-Dist: pre-commit
 Requires-Dist: torch>=2.1.0
 Requires-Dist: tqdm

{garf-2.7 → garf-2.8}/garf.egg-info/requires.txt

@@ -1,5 +1,5 @@
 itk
-numpy
+numpy<2.0.0
 pre-commit
 torch>=2.1.0
 tqdm

{garf-2.7 → garf-2.8}/pyproject.toml

@@ -17,7 +17,7 @@ find = { }  # Scanning implicit namespaces is active by default
 
 [project]
 name = "garf"
-version = "2.7"
+version = "2.8"
 dynamic = ["readme"]
 authors = [{ name = "David Sarrut", email = "david.sarrut@creatis.insa-lyon.fr" }]
 description = "Utility tools for GATE ARF simulations"
@@ -30,7 +30,7 @@ classifiers = [
 requires-python = ">=3.8"
 dependencies = [
         "itk",
-        "numpy",
+        "numpy<2.0.0",
         "pre-commit",
         "torch>=2.1.0",
         "tqdm",