frontends.torch.avg_pool2d: initial commit (ivy-llc#8854)

ivy/functional/frontends/torch/nn/functional/pooling_functions.py

@@ -0,0 +1,68 @@
+import ivy
+from ivy.functional.frontends.tensorflow.func_wrapper import (
+    to_ivy_arrays_and_back,
+)
+
+
+def _broadcast_pooling_helper(x, pool_dims: str = "2d", name: str = "padding"):
+    dims = {"1d": 1, "2d": 2, "3d": 3}
+
+    if isinstance(x, int):
+        return tuple([x for _ in range(dims[pool_dims])])
+
+    if len(x) == 1:
+        return tuple([x[0] for _ in range(dims[pool_dims])])
+    elif len(x) == dims[pool_dims]:
+        return tuple(x)
+    elif len(x) != dims[pool_dims]:
+        raise ValueError(
+            f"`{name}` must either be a single int, "
+            f"or a tuple of {dims[pool_dims]} ints. "
+        )
+
+
+@to_ivy_arrays_and_back
+def avg_pool2d(
+    input,
+    kernel_size,
+    stride=None,
+    padding=0,
+    ceil_mode=False,
+    count_include_pad=True,
+    divisor_override=None,
+):
+    # Figure out input dims N
+    input_rank = input.ndim
+
+    if input_rank == 3:
+        # CHW
+        data_format = "CHW"
+    elif input_rank == 4:
+        # NCHW
+        data_format = "NCHW"
+
+    kernel_size = _broadcast_pooling_helper(kernel_size, "2d", name="kernel_size")
+    stride = _broadcast_pooling_helper(stride, "2d", name="stride")
+    padding = _broadcast_pooling_helper(padding, "2d", name="padding")
+    kernel_pads = list(zip(kernel_size, padding))
+
+    # Padding should be less than or equal to half of kernel size
+    if not all([pad <= kernel / 2 for kernel, pad in kernel_pads]):
+        raise ValueError(
+            "pad should be smaller than or equal to half of kernel size, "
+            f"but got padding={padding}, kernel_size={kernel_size}. "
+        )
+
+    # Figure out padding string
+    if all([pad == ivy.ceil((kernel - 1) / 2) for kernel, pad in kernel_pads]):
+        padding_str = "SAME"
+    else:
+        padding_str = "VALID"
+
+    return ivy.avg_pool2d(
+        input,
+        kernel_size,
+        stride,
+        padding_str,
+        data_format=data_format,
+    )

ivy_tests/test_ivy/test_frontends/test_torch/test_pooling_functions.py

@@ -0,0 +1,62 @@
+# global
+import ivy
+
+# local
+import ivy_tests.test_ivy.helpers as helpers
+
+from ivy_tests.test_ivy.helpers import handle_frontend_test
+
+
+# avg_pool2d
+@handle_frontend_test(
+    fn_tree="torch.nn.functional.avg_pool2d",
+    dtype_x_k_s=helpers.arrays_for_pooling(
+        min_dims=4,
+        max_dims=4,
+        min_side=1,
+        max_side=4,
+    ),
+)
+def test_torch_avg_pool2d(
+    dtype_x_k_s,
+    *,
+    as_variable,
+    num_positional_args,
+    native_array,
+    frontend,
+    fn_tree,
+    on_device,
+):
+    input_dtype, x, kernel_size, stride, padding = dtype_x_k_s
+
+    # Torch ground truth func expects input to be consistent
+    # with a channels first format i.e. NCHW
+    x[0] = x[0].reshape((x[0].shape[0], x[0].shape[-1], *x[0].shape[1:-1]))
+    x_shape = list(x[0].shape[2:])
+
+    # Torch ground truth func also takes padding input as an integer
+    # or a tuple of integers, not a string
+    padding = tuple(
+        [
+            ivy.handle_padding(x_shape[i], stride[0], kernel_size[i], padding)
+            for i in range(len(x_shape))
+        ]
+    )
+
+    helpers.test_frontend_function(
+        input_dtypes=input_dtype,
+        as_variable_flags=as_variable,
+        with_out=False,
+        num_positional_args=num_positional_args,
+        native_array_flags=native_array,
+        frontend=frontend,
+        fn_tree=fn_tree,
+        on_device=on_device,
+        input=x[0],
+        kernel_size=kernel_size,
+        stride=stride,
+        padding=padding,
+        ceil_mode=False,
+        count_include_pad=True,
+        divisor_override=None,
+    )