Fix broken workflow dependency updates

preprocessors/ner/Dockerfile

@@ -1,12 +1,11 @@
-FROM ubuntu:20.04
+FROM python:3.11-bookworm
 
-ENV JAVA_HOME /usr/lib/jvm/java-8-openjdk-amd64/
+ENV JAVA_HOME=/usr/lib/jvm/default-java
 
 RUN apt-get update && \
     DEBIAN_FRONTEND=noninteractive apt-get install -y \
-    openjdk-8-jdk \
-    python3-pip \
-    gcc \
+    build-essential \
+    default-jre-headless \
     git \
     curl \
     && apt-get clean && \

preprocessors/ner/clipscore.py

@@ -13,12 +13,9 @@
 import clip
 import torch
 from PIL import Image
-from sklearn.preprocessing import normalize
 from torchvision.transforms import Compose, Resize, CenterCrop, ToTensor, Normalize
-import torch
 import tqdm
 import numpy as np
-import sklearn.preprocessing
 import collections
 import os
 import pathlib
@@ -35,6 +32,12 @@
 from pycocoevalcap.spice.spice import Spice
 
 
+def normalize_rows(values: np.ndarray) -> np.ndarray:
+    """Normalize a 2D float array row-wise while leaving zero-norm rows as zeros."""
+    norms = np.linalg.norm(values, axis=1, keepdims=True)
+    return np.divide(values, norms, out=np.zeros_like(values), where=norms != 0)
+
+
 def get_all_metrics(refs, cands, return_per_cap=False):
     metrics = []
     names = []
@@ -212,14 +215,14 @@ def get_clip_score(model, images, candidates, device, w=2.5):
 
     #as of numpy 1.21, normalize doesn't work properly for float16
     if version.parse(np.__version__) < version.parse('1.21'):
-        images = sklearn.preprocessing.normalize(images, axis=1)
-        candidates = sklearn.preprocessing.normalize(candidates, axis=1)
+        images = normalize_rows(images)
+        candidates = normalize_rows(candidates)
     else:
         warnings.warn(
             'due to a numerical instability, new numpy normalization is slightly different than paper results. '
             'to exactly replicate paper results, please use numpy version less than 1.21, e.g., 1.20.3.')
-        images = images / np.sqrt(np.sum(images**2, axis=1, keepdims=True))
-        candidates = candidates / np.sqrt(np.sum(candidates**2, axis=1, keepdims=True))
+        images = normalize_rows(images)
+        candidates = normalize_rows(candidates)
 
     per = w*np.clip(np.sum(images * candidates, axis=1), 0, None)
     return np.mean(per), per, candidates
@@ -241,15 +244,15 @@ def get_refonlyclipscore(model, references, candidates, device):
     flattened_refs = extract_all_captions(flattened_refs, model, device)
 
     if version.parse(np.__version__) < version.parse('1.21'):
-        candidates = sklearn.preprocessing.normalize(candidates, axis=1)
-        flattened_refs = sklearn.preprocessing.normalize(flattened_refs, axis=1)
+        candidates = normalize_rows(candidates)
+        flattened_refs = normalize_rows(flattened_refs)
     else:
         warnings.warn(
             'due to a numerical instability, new numpy normalization is slightly different than paper results. '
             'to exactly replicate paper results, please use numpy version less than 1.21, e.g., 1.20.3.')
 
-        candidates = candidates / np.sqrt(np.sum(candidates**2, axis=1, keepdims=True))
-        flattened_refs = flattened_refs / np.sqrt(np.sum(flattened_refs**2, axis=1, keepdims=True))
+        candidates = normalize_rows(candidates)
+        flattened_refs = normalize_rows(flattened_refs)
 
     cand_idx2refs = collections.defaultdict(list)
     for ref_feats, cand_idx in zip(flattened_refs, flattened_refs_idxs):

preprocessors/ner/requirements.txt

@@ -2,14 +2,13 @@ bs4==0.0.1
 imgkit==1.2.2
 nltk==3.7
 gunicorn==23.0.0
-pillow==12.1.1
-torch==1.13.1
+pillow==12.2.0
+torch==2.6.0
 tqdm~=4.66.3
-torchvision==0.8.2
-clip-by-openai==1.1
-scikit-learn==1.1.2
+torchvision==0.21.0
 packaging==21.3
 jsonschema==4.23.0
 Flask==3.1.3
 pycocoevalcap==1.2
-git+https://github.com/openai/CLIP
+# Pin the tested CLIP commit used for the Python 3.11-compatible NER build.
+git+https://github.com/openai/CLIP@d05afc436d78f1c48dc0dbf8e5980a9d471f35f6

preprocessors/ner/test_clipscore.py

@@ -0,0 +1,23 @@
+import unittest
+
+import numpy as np
+
+from clipscore import normalize_rows
+
+
+class NormalizeRowsTest(unittest.TestCase):
+    def test_normalizes_rows(self):
+        values = np.array([[3.0, 4.0], [5.0, 12.0]])
+        normalized = normalize_rows(values)
+        expected = np.array([[0.6, 0.8], [5.0 / 13.0, 12.0 / 13.0]])
+        np.testing.assert_allclose(normalized, expected)
+
+    def test_preserves_zero_rows(self):
+        values = np.array([[0.0, 0.0], [0.0, 2.0]])
+        normalized = normalize_rows(values)
+        expected = np.array([[0.0, 0.0], [0.0, 1.0]])
+        np.testing.assert_allclose(normalized, expected)
+
+
+if __name__ == "__main__":
+    unittest.main()

services/espnet-tts-fr/requirements.txt

@@ -1,6 +1,6 @@
 espnet==202207
 espnet_model_zoo==0.1.7
-Flask==3.1.3
+Flask==2.2.5
 gunicorn==23.0.0
 jsonschema==4.4.0
 parallel_wavegan==0.5.5

services/espnet-tts/requirements.txt

@@ -1,6 +1,6 @@
 espnet==0.10.0
 espnet_model_zoo==0.1.7
-Flask==3.1.3
+Flask==2.2.5
 gunicorn==23.0.0
 jsonschema==4.4.0
 parallel_wavegan==0.5.4