Support subword and word language models

Closes #199 (closed)

Ref https://redmine.teklia.com/issues/4941

dan/datasets/extract/utils.py

@@ -57,10 +57,18 @@ def download_image(url):
     Download an image and open it with Pillow
     """
     assert url.startswith("http"), "Image URL must be HTTP(S)"
+
     # Download the image
     # Cannot use stream=True as urllib's responses do not support the seek(int) method,
     # which is explicitly required by Image.open on file-like objects
-    resp = _retried_request(url)
+    try:
+        resp = _retried_request(url)
+    except requests.HTTPError as e:
+        if "/full/" in url and 400 <= e.response.status_code < 500:
+            # Retry with max instead of full as IIIF size
+            resp = _retried_request(url.replace("/full/", "/max/"))
+        else:
+            raise e
 
     # Preprocess the image and prepare it for classification
     image = Image.open(BytesIO(resp.content)).convert("RGB")