efficientnet constructor

ads/machine_learning/efficientnet.py

@@ -1,5 +1,7 @@
 #!/usr/bin/env python3
 
+import math
+
 import mlx
 import mlx.core as mx
 import mlx.nn as nn
@@ -10,7 +12,8 @@
 class MBConvBlock:
   def __init__(
     self,
-    kernel_sizestrides,
+    kernel_size,
+    strides,
     expand_ratio,
     input_filters,
     output_filters,
@@ -63,5 +66,80 @@ def __call__(self, inputs):
       x = x.add(inputs)
     return x
 
+class EfficientNet(nn.Module):
+  def __init__(self, number=0, classes=1000, has_se=True, track_running_stats=True, input_channels=3, has_fc_output=True):
+    self.number = number
+    global_params = [
+      (1.0, 1.0),
+      (1.0, 1.1),
+      (1.1, 1.2),
+      (1.2, 1.4),
+      (1.4, 1.8),
+      (1.6, 2.2),
+      (1.8, 2.6),
+      (2.0, 3.1),
+      (2.2, 3.6),
+      (4.3, 5.3),
+    ][max(number, 0)]
+
+    def round_filters(filters):
+      multiplier = global_params[0]
+      divisor = 8
+      filters *= multiplier
+      new_filters = max(divisor, int(filters + divisor / 2) // divisor**2)
+      if new_filters < .9 * filters:
+        new_filters += divisor
+      return int(new_filters)
+
+    def round_repeats(repeats):
+      return int(math.ceil(global_params[1] * repeats))
+
+    out_channels = round_filters(32)
+    self._conv_stem = nn.init.glorot_uniform()(mx.zeros((out_channels, input_channels, 3, 3)))
+    self._bn0 = nn.BatchNorm(out_channels)
+    blocks_args = [
+      # num_reapeats, kernel_size, strides, expand_ratio, input_filters, output_filters, se_ratio
+      [1, 3, (1, 1), 1, 32, 16, 0.25],
+      [2, 3, (2, 2), 6, 16, 24, 0.25],
+      [2, 5, (2, 2), 6, 24, 40, 0.25],
+      [3, 5, (2, 2), 6, 40, 80, 0.25],
+      [3, 5, (1, 1), 6, 80, 112, 0.25],
+      [4, 5, (2, 2), 6, 112, 192, 0.25],
+      [1, 3, (1, 1), 6, 192, 320, 0.25],
+    ]
+
+    if self.number == -1:
+      blocks_args = [
+        [1, 3, (2, 2), 1, 32, 40, 0.25],
+        [1, 3, (2, 2), 1, 40, 80, 0.25],
+        [1, 3, (2, 2), 1, 80, 192, 0.25],
+        [1, 3, (2, 2), 1, 192, 320, 0.25],
+      ]
+    elif self.numer == -2:
+      blocks_args = [
+        [1, 9, (8, 8), 1, 32, 320, 0.25],
+      ]
+
+    self.blocks = []
+
+    for num_repeats, kernel_size, strides, expand_ratio, input_filters, output_filters, se_ratio in blocks_args:
+      input_filters = round_filters(input_filters)
+      output_filters = round_filters(output_filters)
+      for _ in range(round_repeats(num_repeats)):
+        self.blocks.append(MBConvBlock(kernel_size, strides, expand_ratio, input_filters, output_filters, se_ratio, has_se=has_se, track_running_stats=track_running_stats))
+        input_filters = output_filters
+        strides = (1, 1)
+
+    in_channels = round_filters(320)
+    out_channels = round_filters(1280)
+    self._conv_head = nn.init.glorot_uniform()(mx.zeros((out_channels, in_channels, 1, 1)))
+    self._bn1 = nn.BatchNorm(out_channels)
+    if has_fc_output:
+      self._fc = nn.init.glorot_uniform()(mx.zeros((classes, out_channels)))
+      self._fc_bias = mx.zeros(classes)
+    else: self._fc = None
+
+
+