pixel_unshuffle function added for Paddle Frontend (ivy-llc#21754)

ivy/functional/frontends/paddle/nn/functional/vision.py

@@ -8,6 +8,45 @@
 from ivy.utils.assertions import check_equal
 
 
+@to_ivy_arrays_and_back
+def pixel_unshuffle(x, downscale_factor, data_format="NCHW"):
+    input_shape = ivy.shape(x)
+
+    if data_format == "NCHW":
+        b, c, h, w = input_shape
+    else:
+        b, h, w, c = input_shape
+
+    check_equal(
+        c % (downscale_factor ** 2),
+        0,
+        message=(
+            "pixel unshuffle expects input channel to be divisible by square of downscale"
+            " factor, but got input with size {}, downscale factor={}, and"
+            " self.size(1)={}, is not divisible by {}".format(
+                input_shape, downscale_factor, c, downscale_factor ** 2
+            )
+        ),
+        as_array=False,
+    )
+
+    oc = c // (downscale_factor ** 2)
+    oh = h // downscale_factor
+    ow = w // downscale_factor
+
+    if data_format == "NCHW":
+        x_reshaped = ivy.reshape(x, (b, oc, downscale_factor, downscale_factor, oh, ow))
+    else:
+        x_reshaped = ivy.reshape(x, (b, oh, ow, downscale_factor, downscale_factor, oc))
+
+    if data_format == "NCHW":
+        return ivy.reshape(
+            ivy.permute_dims(x_reshaped, (0, 1, 4, 2, 5, 3)), (b, oc, oh, ow)
+        )
+    return ivy.reshape(
+        ivy.permute_dims(x_reshaped, (0, 4, 1, 5, 2, 3)), (b, oh, ow, oc)
+    )
+
 @to_ivy_arrays_and_back
 @with_unsupported_dtypes({"2.5.1 and below": ("float16", "bfloat16")}, "paddle")
 def affine_grid(theta, out_shape, align_corners=True):

ivy_tests/test_ivy/test_frontends/test_paddle/test_nn/test_functional/test_vision.py

@@ -10,6 +10,47 @@
 # --- Helpers --- #
 # --------------- #
 
+#pixel_unshuffle
+@handle_frontend_test(
+    fn_tree="paddle.nn.functional.pixel_unshuffle",
+    dtype_and_x=helpers.dtype_and_values(
+        available_dtypes=["float32", "float64"],
+        min_value=0,
+        min_num_dims=4,
+        max_num_dims=4,
+        min_dim_size=3,
+    ),
+    downscale_factor=helpers.ints(min_value=1),
+    data_format=st.sampled_from(["NCHW", "NHWC"]),
+)
+
+def test_paddle_pixel_unshuffle(
+    *,
+    dtype_and_x,
+    downscale_factor,
+    data_format,
+    on_device,
+    fn_tree,
+    frontend,
+    test_flags,
+    backend_fw,
+):
+    input_dtype, x = dtype_and_x
+    if data_format == "NCHW":
+        assume(ivy.shape(x[0])[1] % (downscale_factor**2) == 0)
+    else:
+        assume(ivy.shape(x[0])[3] % (downscale_factor**2) == 0)
+    helpers.test_frontend_function(
+        input_dtypes=input_dtype,
+        frontend=frontend,
+        test_flags=test_flags,
+        fn_tree=fn_tree,
+        on_device=on_device,
+        x=x[0],
+        downscale_factor=downscale_factor,
+        data_format=data_format,
+        backend_to_test=backend_fw,
+    )
 
 @st.composite
 def _affine_grid_helper(draw):