Update torch nightly and pin torchvision to fix CI (#2072)

ludwig-ai · May 28, 2022 · 59b2d1b · 59b2d1b
1 parent 8e1b329
commit 59b2d1b
Show file tree

Hide file tree

Showing 3 changed files with 7 additions and 6 deletions.
diff --git a/.github/workflows/pytest.yml b/.github/workflows/pytest.yml
@@ -87,8 +87,9 @@ jobs:
           if [ $PYTORCH == "nightly" ]; then
             cat requirements.txt | sed '/^torch[>=<]/d' > requirements-temp && mv requirements-temp requirements.txt
             extra_index_url=https://download.pytorch.org/whl/nightly/cpu
-            # TODO: temp fix for https://github.com/pytorch/vision/issues/5873
-            pip install --pre torch==1.12.0.dev20220421+cpu torchtext torchvision torchaudio --extra-index-url $extra_index_url
+            pip install --pre torch==1.13.0.dev20220527+cpu torchtext torchaudio --extra-index-url $extra_index_url
+            # TODO: temp fix for https://github.com/pytorch/vision/issues/6103
+            pip install https://download.pytorch.org/whl/nightly/cpu/torchvision-0.14.0.dev20220527%2Bcpu-cp39-cp39-linux_x86_64.whl
           else
             extra_index_url=https://download.pytorch.org/whl/cpu
             pip install torch==$PYTORCH torchtext torchvision torchaudio --extra-index-url $extra_index_url

diff --git a/ludwig/data/batcher/bucketed.py b/ludwig/data/batcher/bucketed.py
@@ -111,7 +111,7 @@ def set_epoch(self, epoch, batch_size):
         self.steps_per_epoch = self._compute_steps_per_epoch()
 
     def _compute_steps_per_epoch(self) -> int:
-        return int(np.asscalar(np.sum(np.ceil(self.bucket_sizes / self.batch_size))))
+        return int(np.sum(np.ceil(self.bucket_sizes / self.batch_size)).item())
 
 
 # dynamic_length_encoders = {

diff --git a/ludwig/visualize.py b/ludwig/visualize.py
@@ -1456,13 +1456,13 @@ def compare_classifiers_performance_from_prob(
         hits_at_k = 0
         for j in range(len(ground_truth)):
             hits_at_k += np.in1d(ground_truth[j], topk[j])
-        hits_at_ks.append(np.asscalar(hits_at_k) / len(ground_truth))
+        hits_at_ks.append(hits_at_k.item() / len(ground_truth))
 
         mrr = 0
         for j in range(len(ground_truth)):
             ground_truth_pos_in_probs = prob[j] == ground_truth[j]
             if np.any(ground_truth_pos_in_probs):
-                mrr += 1 / -(np.asscalar(np.argwhere(ground_truth_pos_in_probs)) - prob.shape[1])
+                mrr += 1 / -(np.argwhere(ground_truth_pos_in_probs).item() - prob.shape[1])
         mrrs.append(mrr / len(ground_truth))
 
     filename = None
@@ -1663,7 +1663,7 @@ def compare_classifiers_performance_subset(
         hits_at_k = 0
         for j in range(len(gt_subset)):
             hits_at_k += np.in1d(gt_subset[j], top3_subset[i, :])
-        hits_at_ks.append(np.asscalar(hits_at_k) / len(gt_subset))
+        hits_at_ks.append(hits_at_k.item() / len(gt_subset))
 
     title = None
     if subset == "ground_truth":