[fbsync] Keep 16bits png decoding private (#4732)

Reviewed By: NicolasHug

Differential Revision: D31916314

fbshipit-source-id: d06ace86997de7482b72edb922725e505cb24cad

Author: datumbox

test/test_image.py

@@ -22,6 +22,7 @@
     write_file,
     ImageReadMode,
     read_image,
+    _read_png_16,
 )
 
 IMAGE_ROOT = os.path.join(os.path.dirname(os.path.abspath(__file__)), "assets")
@@ -156,8 +157,21 @@ def test_decode_png(img_path, pil_mode, mode):
         img_pil = torch.from_numpy(np.array(img))
 
     img_pil = normalize_dimensions(img_pil)
-    data = read_file(img_path)
-    img_lpng = decode_image(data, mode=mode)
+
+    if "16" in img_path:
+        # 16 bits image decoding is supported, but only as a private API
+        # FIXME: see https://github.com/pytorch/vision/issues/4731 for potential solutions to making it public
+        with pytest.raises(RuntimeError, match="At most 8-bit PNG images are supported"):
+            data = read_file(img_path)
+            img_lpng = decode_image(data, mode=mode)
+
+        img_lpng = _read_png_16(img_path, mode=mode)
+        assert img_lpng.dtype == torch.int32
+        # PIL converts 16 bits pngs in uint8
+        img_lpng = torch.round(img_lpng / (2 ** 16 - 1) * 255).to(torch.uint8)
+    else:
+        data = read_file(img_path)
+        img_lpng = decode_image(data, mode=mode)
 
     tol = 0 if pil_mode is None else 1
 
@@ -168,11 +182,6 @@ def test_decode_png(img_path, pil_mode, mode):
         # TODO: remove once fix is released in PIL. Should be > 8.3.1.
         img_lpng, img_pil = img_lpng[0], img_pil[0]
 
-    if "16" in img_path:
-        # PIL converts 16 bits pngs in uint8
-        assert img_lpng.dtype == torch.int32
-        img_lpng = torch.round(img_lpng / (2 ** 16 - 1) * 255).to(torch.uint8)
-
     torch.testing.assert_close(img_lpng, img_pil, atol=tol, rtol=0)
 
 

torchvision/csrc/io/image/cpu/decode_png.cpp

@@ -5,7 +5,10 @@ namespace vision {
 namespace image {
 
 #if !PNG_FOUND
-torch::Tensor decode_png(const torch::Tensor& data, ImageReadMode mode) {
+torch::Tensor decode_png(
+    const torch::Tensor& data,
+    ImageReadMode mode,
+    bool allow_16_bits) {
   TORCH_CHECK(
       false, "decode_png: torchvision not compiled with libPNG support");
 }
@@ -16,7 +19,10 @@ bool is_little_endian() {
   return *(uint8_t*)&x;
 }
 
-torch::Tensor decode_png(const torch::Tensor& data, ImageReadMode mode) {
+torch::Tensor decode_png(
+    const torch::Tensor& data,
+    ImageReadMode mode,
+    bool allow_16_bits) {
   // Check that the input tensor dtype is uint8
   TORCH_CHECK(data.dtype() == torch::kU8, "Expected a torch.uint8 tensor");
   // Check that the input tensor is 1-dimensional
@@ -77,9 +83,12 @@ torch::Tensor decode_png(const torch::Tensor& data, ImageReadMode mode) {
     TORCH_CHECK(retval == 1, "Could read image metadata from content.")
   }
 
-  if (bit_depth > 16) {
+  auto max_bit_depth = allow_16_bits ? 16 : 8;
+  auto err_msg = "At most " + std::to_string(max_bit_depth) +
+      "-bit PNG images are supported currently.";
+  if (bit_depth > max_bit_depth) {
     png_destroy_read_struct(&png_ptr, &info_ptr, nullptr);
-    TORCH_CHECK(false, "At most 16-bit PNG images are supported currently.")
+    TORCH_CHECK(false, err_msg)
   }
 
   int channels = png_get_channels(png_ptr, info_ptr);

torchvision/csrc/io/image/cpu/decode_png.h

@@ -8,7 +8,8 @@ namespace image {
 
 C10_EXPORT torch::Tensor decode_png(
     const torch::Tensor& data,
-    ImageReadMode mode = IMAGE_READ_MODE_UNCHANGED);
+    ImageReadMode mode = IMAGE_READ_MODE_UNCHANGED,
+    bool allow_16_bits = false);
 
 } // namespace image
 } // namespace vision

torchvision/io/image.py

@@ -61,12 +61,7 @@ def decode_png(input: torch.Tensor, mode: ImageReadMode = ImageReadMode.UNCHANGE
     """
     Decodes a PNG image into a 3 dimensional RGB or grayscale Tensor.
     Optionally converts the image to the desired format.
-    The values of the output tensor are uint8 in [0, 255], except for
-    16-bits pngs which are int32 tensors in [0, 65535].
-
-    .. warning::
-        Should pytorch ever support the uint16 dtype natively, the dtype of the
-        output for 16-bits pngs will be updated from int32 to uint16.
+    The values of the output tensor are uint8 in [0, 255].
 
     Args:
         input (Tensor[1]): a one dimensional uint8 tensor containing
@@ -79,7 +74,7 @@ def decode_png(input: torch.Tensor, mode: ImageReadMode = ImageReadMode.UNCHANGE
     Returns:
         output (Tensor[image_channels, image_height, image_width])
     """
-    output = torch.ops.image.decode_png(input, mode.value)
+    output = torch.ops.image.decode_png(input, mode.value, False)
     return output
 
 
@@ -193,8 +188,7 @@ def decode_image(input: torch.Tensor, mode: ImageReadMode = ImageReadMode.UNCHAN
     operation to decode the image into a 3 dimensional RGB or grayscale Tensor.
 
     Optionally converts the image to the desired format.
-    The values of the output tensor are uint8 in [0, 255], except for
-    16-bits pngs which are int32 tensors in [0, 65535].
+    The values of the output tensor are uint8 in [0, 255].
 
     Args:
         input (Tensor): a one dimensional uint8 tensor containing the raw bytes of the
@@ -215,8 +209,7 @@ def read_image(path: str, mode: ImageReadMode = ImageReadMode.UNCHANGED) -> torc
     """
     Reads a JPEG or PNG image into a 3 dimensional RGB or grayscale Tensor.
     Optionally converts the image to the desired format.
-    The values of the output tensor are uint8 in [0, 255], except for
-    16-bits pngs which are int32 tensors in [0, 65535].
+    The values of the output tensor are uint8 in [0, 255].
 
     Args:
         path (str): path of the JPEG or PNG image.
@@ -230,3 +223,8 @@ def read_image(path: str, mode: ImageReadMode = ImageReadMode.UNCHANGED) -> torc
     """
     data = read_file(path)
     return decode_image(data, mode)
+
+
+def _read_png_16(path: str, mode: ImageReadMode = ImageReadMode.UNCHANGED) -> torch.Tensor:
+    data = read_file(path)
+    return torch.ops.image.decode_png(data, mode.value, True)