docs(server): docs document tensor (#679)

docs/user-guides/client.md

@@ -96,7 +96,9 @@ This feature uses [DocArray](https://docarray.jina.ai), which is installed toget
 
 If auto-detection on a list of raw string is too "sci-fi" to you, then you may use `docarray.Document` to make the input more explicit and organized. `Document` can be used as a container to easily represent a sentence or an image.
 
-- Input: each Document must be filled with `.text` or `.uri` or `.blob` attribute. Document filled with `.text` is considered as sentence, Document filled with `.uri` or `.blob` is considered as image.
+- Input: each Document must be filled with `.text` or `.uri` or `.blob` or `.tensor` attribute. 
+  - Document filled with `.text` is considered as sentence;
+  - Document filled with `.uri` or `.blob` or `.tensor` is considered as image. If `.tensor` is filled, then its shape must be in `[H, W, C]` format.
 - Output: a `DocumentArray` of the same input length. Each Document in it is now filled with `.embedding` attribute.
 
 The explicit comes from now you have to put the string into the Document attributes. For example, we can rewrite the above example as below:

tests/test_server.py

@@ -3,6 +3,7 @@
 import pytest
 from clip_server.model.clip import _transform_ndarray, _transform_blob
 from docarray import Document
+import numpy as np
 
 
 @pytest.mark.parametrize(
@@ -22,3 +23,16 @@ def test_server_preprocess_ndarray_image(image_uri, size):
     t1 = _transform_blob(size)(d1.blob).numpy()
     t2 = _transform_ndarray(size)(d2.tensor).numpy()
     assert t1.shape == t2.shape
+
+
+@pytest.mark.parametrize(
+    'tensor',
+    [
+        np.random.random([100, 100, 3]),
+        np.random.random([1, 1, 3]),
+        np.random.random([5, 50, 3]),
+    ],
+)
+def test_transform_arbitrary_tensor(tensor):
+    d = Document(tensor=tensor)
+    assert _transform_ndarray(224)(d.tensor).numpy().shape == (3, 224, 224)