Added max_pool and avg_pool functionalities

keras/src/backend/mlx/nn.py

@@ -1,4 +1,7 @@
 import builtins
+import math
+import operator
+from itertools import accumulate
 
 import mlx.core as mx
 import mlx.nn as nn
@@ -122,16 +125,172 @@ def log_softmax(x, axis=-1):
     return x - mx.logsumexp(x, axis=axis, keepdims=True)
 
 
+def _calculate_padding(input_shape, pool_size, strides):
+    ndim = len(input_shape)
+
+    padding = ()
+    for d in range(ndim):
+        pad = max(0, (pool_size[d] - 1) - ((input_shape[d] - 1) % strides[d]))
+        padding = padding + (pad,)
+
+    return [(p // 2, (p + 1) // 2) for p in padding]
+
+
+def _non_overlapping_sliding_windows(x, shape, window_shape):
+    # Compute the intermediate shape
+    new_shape = [shape[0]]
+    for s, w in zip(shape[1:], window_shape):
+        new_shape.append(s // w)
+        new_shape.append(w)
+    new_shape.append(shape[-1])
+
+    last_axis = len(new_shape) - 1
+    axis_order = [
+        0,
+        *range(1, last_axis, 2),
+        *range(2, last_axis, 2),
+        last_axis,
+    ]
+
+    x = x.reshape(new_shape)
+    x = x.transpose(axis_order)
+    return x
+
+
+def _sliding_windows(x, window_shape, window_strides):
+    if x.ndim < 3:
+        raise ValueError(
+            f"To extract sliding windows at least 1 spatial dimension "
+            f"(3 total) is needed but the input only has {x.ndim} dimensions."
+        )
+
+    spatial_dims = x.shape[1:-1]
+    if not (len(spatial_dims) == len(window_shape) == len(window_strides)):
+        raise ValueError(
+            f"To extract sliding windows the window shapes and strides must "
+            f"have the same number of spatial dimensions as the signal but "
+            f"the signal has {len(spatial_dims)} dims and the window shape "
+            f"has {len(window_shape)} and strides have {len(window_strides)}."
+        )
+
+    shape = x.shape
+    if all(
+        window == stride and size % window == 0
+        for size, window, stride in zip(
+            spatial_dims, window_shape, window_strides
+        )
+    ):
+        return _non_overlapping_sliding_windows(x, shape, window_shape)
+
+    strides = list(
+        reversed(list(accumulate(reversed(shape + (1,)), operator.mul)))
+    )[1:]
+
+    # Compute the output shape
+    final_shape = [shape[0]]
+    final_shape += [
+        (size - window) // stride + 1
+        for size, window, stride in zip(
+            spatial_dims, window_shape, window_strides
+        )
+    ]
+    final_shape += window_shape
+    final_shape += [shape[-1]]
+
+    # Compute the output strides
+    final_strides = strides[:1]
+    final_strides += [
+        og_stride * stride
+        for og_stride, stride in zip(strides[1:-1], window_strides)
+    ]
+    final_strides += strides[1:-1]
+    final_strides += strides[-1:]  # should always be [1]
+
+    return mx.as_strided(x, final_shape, final_strides)
+
+
+def _pool(
+    inputs, pool_size, strides, padding, padding_value, data_format, pooling_fn
+):
+    if padding not in ("same", "valid"):
+        raise ValueError(
+            f"Invalid padding '{padding}', must be 'same' or 'valid'."
+        )
+
+    if data_format == "channels_first":
+        # mlx expects channels_last
+        inputs = inputs.transpose(0, *range(2, inputs.ndim), 1)
+
+    if padding == "same":
+        pads = _calculate_padding(inputs.shape[1:-1], pool_size, strides)
+
+        if any(p[1] > 0 for p in pads):
+            inputs = mx.pad(
+                inputs,
+                [(0, 0)] + pads + [(0, 0)],
+                constant_values=padding_value,
+            )
+
+    inputs = _sliding_windows(inputs, pool_size, strides)
+
+    axes = tuple(range(-len(pool_size) - 1, -1, 1))
+    result = pooling_fn(inputs, axes)
+
+    if data_format == "channels_first":
+        result = result.transpose(0, -1, *range(1, result.ndim - 1))
+    return result
+
+
 def max_pool(
     inputs, pool_size, strides=None, padding="valid", data_format=None
 ):
-    raise NotImplementedError("MLX backend doesn't support max pooling yet")
+    inputs = convert_to_tensor(inputs)
+    data_format = standardize_data_format(data_format)
+    num_spatial_dims = inputs.ndim - 2
+    pool_size = standardize_tuple(pool_size, num_spatial_dims, "pool_size")
+    strides = pool_size if strides is None else strides
+    strides = standardize_tuple(strides, num_spatial_dims, "strides")
+
+    return _pool(
+        inputs, pool_size, strides, padding, -mx.inf, data_format, mx.max
+    )
 
 
 def average_pool(
     inputs, pool_size, strides=None, padding="valid", data_format=None
 ):
-    raise NotImplementedError("MLX backend doesn't support average pooling yet")
+    inputs = convert_to_tensor(inputs)
+    data_format = standardize_data_format(data_format)
+    num_spatial_dims = inputs.ndim - 2
+    pool_size = standardize_tuple(pool_size, num_spatial_dims, "pool_size")
+    strides = pool_size if strides is None else strides
+    strides = standardize_tuple(strides, num_spatial_dims, "strides")
+
+    # Create a pool by applying the sum function in each window
+    pooled = _pool(
+        inputs, pool_size, strides, padding, 0.0, data_format, mx.sum
+    )
+    if padding == "valid":
+        # No padding needed. Divide by the size of the pool which gives
+        # the average
+        return pooled / math.prod(pool_size)
+    else:
+        # Create a tensor of ones of the same shape of inputs.
+        # Then create a pool, padding by zero and using sum as function.
+        # This will create a tensor of the smae dimensions as pooled tensor
+        # with values being the sum.
+        # By dividing pooled by windows_counts, we get the average while
+        # skipping the padded values.
+        window_counts = _pool(
+            mx.ones(inputs.shape, inputs.dtype),
+            pool_size,
+            strides,
+            padding,
+            0.0,
+            data_format,
+            mx.sum,
+        )
+        return pooled / window_counts
 
 
 def conv(