Merge pull request #1 from Shakshi3104/add

Implement MarNASNet-D and MarNASNet-Ca

Author: Shakshi3104

setup.py

@@ -2,7 +2,7 @@
 
 setuptools.setup(
     name="tfmars",
-    version="1.0.0",
+    version="1.0.1",
     author="Shakshi3104",
     description="MarNASNet and famous CNN models for Sensor-based Human Activity Recognition",
     url="https://github.com/Shakshi3104/tfmars",

src/tfmars/applications/__init__.py

@@ -30,5 +30,5 @@
     "EfficientNetB0", "EfficientNetB0WithAttention",
     "EfficientNetLite0", "EfficientNetLite0WithAttention",
     "DenseNet121", "DenseNet121WithAttention",
-    "MarNASNetC", "MarNASNetA", "MarNASNetB", "MarNASNetE"
+    "MarNASNetC", "MarNASNetA", "MarNASNetB", "MarNASNetE", "MarNASNetD", "MarNASNetCa"
 ]

src/tfmars/applications/marnasnet/__init__.py

@@ -1,11 +1,14 @@
-from .marnasnet_c import MarNASNetC
+from .marnasnet_c import MarNASNetC, MarNASNetCa
 from .marnasnet_a import MarNASNetA
 from .marnasnet_b import MarNASNetB
 from .marnasnet_e import MarNASNetE
+from .marnasnet_d import MarNASNetD
 
 __all__ = [
     "MarNASNetC",
     "MarNASNetA",
     "MarNASNetB",
-    "MarNASNetE"
+    "MarNASNetE",
+    "MarNASNetD",
+    "MarNASNetCa"
 ]

src/tfmars/applications/marnasnet/marnasnet_c.py

@@ -1,7 +1,6 @@
 import tensorflow as tf
 import math
 
-
 from .blocks import ConvBlock, SkipOperation, RegularConvBlock, MBConvBlock
 from ..mobile_inverted_bottleneck import Top, Stem
 
@@ -26,6 +25,7 @@ def MarNASNetC(width_coefficient=1.0, depth_coefficient=1.0, depth_divisor=8,
     Returns
     -------
     """
+
     def round_filters(filters_, divisor=depth_divisor):
         """Round number of filters based on depth multiplier."""
         filters_ *= width_coefficient
@@ -131,3 +131,39 @@ def round_repeats(repeats_):
     # Create model
     model = tf.keras.models.Model(inputs, outputs)
     return model
+
+
+# MarNASNet-Ca
+def MarNASNetCa(include_top=True, input_shape=(256, 3), pooling=None, classes=6, classifier_activation='softmax'):
+    params = [
+        {
+            'conv_op': "MBConv",
+            'kernel_size': 5,
+            'skip_op': "identity",
+            'layers': 2,
+            'filters': 32
+        },
+        {
+            'conv_op': "Conv",
+            'kernel_size': 2,
+            'skip_op': "identity",
+            'layers': 4,
+            'filters': 82
+        },
+        {
+            'conv_op': "MBConv",
+            'kernel_size': 2,
+            'skip_op': "none",
+            'layers': 2,
+            'filters': 132
+        },
+        {
+            'conv_op': "MBConv",
+            'kernel_size': 5,
+            'skip_op': "identity",
+            'layers': 5,
+            'filters': 192
+        }
+    ]
+
+    return MarNASNetC(1.0, 1.0, 8, include_top, input_shape, pooling, classes, classifier_activation, params)

src/tfmars/applications/marnasnet/marnasnet_d.py

@@ -0,0 +1,118 @@
+import tensorflow as tf
+import math
+
+
+from .blocks import ConvBlock, SkipOperation, RegularConvBlock, MBConvBlock
+from ..mobile_inverted_bottleneck import Top, Stem
+
+
+# MarNASNet-D
+def MarNASNetD(include_top=True, input_shape=(256, 3), pooling=None, classes=6, classifier_activation='softmax',
+               params=None):
+    """
+
+    Parameters
+    ----------
+    include_top
+    input_shape
+    pooling
+    classes
+    classifier_activation
+    params
+
+    Returns
+    -------
+
+    """
+
+    # se ratio
+    se_ratio = 0.25
+
+    # inputs
+    inputs = tf.keras.layers.Input(shape=input_shape)
+
+    # Build stem
+    filters_in = 32
+    outputs = Stem(filters_in)(inputs)
+
+    if params is None:
+        params = [
+            {
+                'conv_op': "MBConv",
+                'kernel_size': 3,
+                'skip_op': "none",
+                'layers': 5,
+                'filters': int(0.75 * 76)
+            },
+            {
+                'conv_op': "MBConv",
+                'kernel_size': 5,
+                'skip_op': "identity",
+                'layers': 5,
+                'filters': int(1.25 * 88)
+            },
+            {
+                'conv_op': "MBConv",
+                'kernel_size': 2,
+                'skip_op': "identity",
+                'layers': 2,
+                'filters': int(0.75 * 100)
+            },
+            {
+                'conv_op': "MBConv",
+                'kernel_size': 2,
+                'skip_op': "none",
+                'layers': 2,
+                'filters': int(0.75 * 112)
+            }
+        ]
+
+    for i, param in enumerate(params):
+        block_id = i + 1
+
+        conv_op = ConvBlock[param["conv_op"]]
+        skip_op = SkipOperation[param["skip_op"]]
+
+        repeats = param["layers"]
+        kernel_size = param["kernel_size"]
+        filters = param["filters"]
+
+        if conv_op == ConvBlock.Conv:
+            outputs = RegularConvBlock(
+                repeats=repeats,
+                kernel_size=kernel_size,
+                filters=filters,
+                skip_op=skip_op,
+                strides=1,
+                se_ratio=se_ratio,
+                block_id=block_id
+            )(outputs)
+
+        elif conv_op == ConvBlock.MBConv:
+            outputs = MBConvBlock(
+                repeats=repeats,
+                kernel_size=kernel_size,
+                filters_in=filters_in,
+                filters_out=filters,
+                expand_ratio=1,
+                skip_op=skip_op,
+                strides=1 if i == 0 else 2,
+                se_ratio=se_ratio,
+                block_id=block_id
+            )(outputs)
+
+        filters_in = filters
+
+    # Build top
+    outputs = Top(1280, classes, classifier_activation=classifier_activation,
+                  include_top=include_top)(outputs)
+
+    if not include_top:
+        if pooling == "avg":
+            outputs = tf.keras.layers.GlobalAveragePooling1D(name="avg_pool")(outputs)
+        elif pooling == "max":
+            outputs = tf.keras.layers.GlobalMaxPooling1D(name="max_pool")(outputs)
+
+    # Create model
+    model = tf.keras.models.Model(inputs, outputs)
+    return model