start implementing core image segmentation model

aimodel/src/lib/ai/components/convnext_inverse.py

@@ -0,0 +1,40 @@
+
+import tensorflow as tf
+
+from .convnext import add_convnext_block
+
+depths_dims = dict(
+	# architectures from: https://github.com/facebookresearch/ConvNeXt
+	# A ConvNet for the 2020s: https://arxiv.org/abs/2201.03545
+	convnext_i_xtiny  = (dict(depths=[3, 6, 3, 3],    dims=[528, 264, 132, 66])),
+	convnext_i_tiny   = (dict(depths=[3, 9, 3, 3],    dims=[768, 384, 192, 96])),
+	convnext_i_small  = (dict(depths=[3, 27, 3, 3],   dims=[768, 384, 192, 96])),
+	convnext_i_base   = (dict(depths=[3, 27, 3, 3],   dims=[1024, 512, 256, 128])),  
+	convnext_i_large  = (dict(depths=[3, 27, 3, 3],   dims=[1536, 768, 384, 192])),
+	convnext_i_xlarge = (dict(depths=[3, 27, 3, 3],   dims=[2048, 1024, 512, 256])),
+)
+
+
+def do_convnext_inverse(layer_in, arch_name="convnext_tiny"):
+	return convnext_inverse(layer_in,
+		depths=depths_dims[arch_name]["depths"],
+		dims=depths_dims[arch_name]["dims"]
+	)
+
+
+def convnext_inverse(layer_in, depths, dims):
+	layer_next = layer_in
+	
+	i = 0
+	for depth, dim in zip(depths, dims):
+		layer_next = block_upscale(layer_next, i, depth=depth, dim=dim)
+		i += 1
+
+
+def block_upscale(layer_in, block_number, depth, dim):
+	layer_next = layer_in
+	for i in range(depth):
+		layer_next = add_convnext_block(layer_next, dim=dim, prefix=f"cns.stage{block_number}.block.{i}")
+	
+	layer_next = tf.keras.layers.LayerNormalization(name=f"cns.stage{block_number}.end.norm", epsilon=1e-6)(layer_next)
+	layer_next = tf.keras.layers.Conv2DTranspose(name=f"cns.stage{block_number}.end.convtp", filters=dim, kernel_size=4, padding="same")(layer_next)

aimodel/src/lib/ai/model_rainfallwater_segmentation.py

@@ -0,0 +1,33 @@
+import math
+
+from loguru import logger
+import tensorflow as tf
+
+from .components.convnext_inverse import do_convnext_inverse
+
+def model_rainfallwater_segmentation(metadata, feature_dim_in, shape_water_out, batch_size=64, summary_file=None):
+	
+	layer_input = tf.keras.layers.Input(
+		shape=(feature_dim_in)
+	)
+	
+	# BEGIN
+	layer_next = tf.keras.layers.Dense(name="cns.stage.begin.dense")(layer_input)
+	layer_next = tf.keras.layers.LayerNormalisation(name="stage_begin.norm", epsilon=1e-6)(layer_next)
+	layer_next = tf.keras.layers.ReLU(name="stage_begin.relu")(layer_next)
+	
+	layer_next = do_convnext_inverse(layer_next, arch_name="convnext_i_tiny")
+	
+	# TODO: Implement projection head here
+	
+	model = tf.keras.Model(
+		inputs = layer_input,
+		outputs = layer_next
+	)
+	
+	model.compile(
+		optimizer="Adam",
+		loss="" # TODO: set this to binary cross-entropy loss
+	)
+	
+	return model