sbrl-github/research-rainfallradar
1
0
Fork 0
mirror of https://github.com/sbrl/research-rainfallradar synced 2024-11-26 02:43:02 +00:00
Code Issues Projects Releases Packages Wiki Activity

drop activation function in last layers

Browse source
This commit is contained in:
Starbeamrainbowlabs 2022-12-12 17:20:04 +00:00
parent bcd2f1251e
commit dbf8f5617c
Signed by: sbrl
GPG key ID: 1BE5172E637709C2
2 changed files with 6 additions and 5 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

7
aimodel/src/lib/ai/model_rainfallwater_mono.py
View file

@ -69,12 +69,13 @@ def model_rainfallwater_mono(metadata, model_arch_enc="convnext_xtiny", model_ar
# TODO: An attention layer here instead of a dense layer, with a skip connection perhaps? # TODO: An attention layer here instead of a dense layer, with a skip connection perhaps?
logger.warning("Warning: TODO implement attention from https://ieeexplore.ieee.org/document/9076883") logger.warning("Warning: TODO implement attention from https://ieeexplore.ieee.org/document/9076883")
layer_next = tf.keras.layers.Dense(32, activation="gelu")(layer_next) # Ref https://stackoverflow.com/a/54031459/1460422, no activation functions in segmentation head!
layer_next = tf.keras.layers.Dense(32)(layer_next)
# LOSS cross entropy # LOSS cross entropy
# layer_next = tf.keras.layers.Conv2D(water_bins, activation="gelu", kernel_size=1, padding="same")(layer_next) # layer_next = tf.keras.layers.Conv2D(water_bins, activation="gelu", kernel_size=1, padding="same")(layer_next)
# layer_next = tf.keras.layers.Softmax(axis=-1)(layer_next) # layer_next = tf.keras.layers.Softmax(axis=-1)(layer_next)
# LOSS dice # LOSS dice
layer_next = tf.keras.layers.Conv2D(1, activation="gelu", kernel_size=1, padding="same")(layer_next) layer_next = tf.keras.layers.Conv2D(1, kernel_size=1, padding="same")(layer_next)
layer_next = tf.keras.layers.Reshape(layer_next.shape[1:-1])(layer_next) # Strip the channels layer_next = tf.keras.layers.Reshape(layer_next.shape[1:-1])(layer_next) # Strip the channels
model = tf.keras.Model( model = tf.keras.Model(
@ -92,7 +93,7 @@ def model_rainfallwater_mono(metadata, model_arch_enc="convnext_xtiny", model_ar
# loss=LossCrossentropy(batch_size=batch_size), # loss=LossCrossentropy(batch_size=batch_size),
loss=LossDice(), loss=LossDice(),
# loss=tf.keras.losses.CategoricalCrossentropy(), # loss=tf.keras.losses.CategoricalCrossentropy(),
metrics=[tf.keras.metrics.CategoricalAccuracy()] metrics=[tf.keras.metrics.BinaryAccuracy()]
) )
return model return model

4
aimodel/src/lib/ai/model_rainfallwater_segmentation.py
View file

@ -44,8 +44,8 @@ def model_rainfallwater_segmentation(metadata, shape_water_out, model_arch="conv
# TODO: An attention layer here instead of a dense layer, with a skip connection perhaps? # TODO: An attention layer here instead of a dense layer, with a skip connection perhaps?
logger.warning("Warning: TODO implement attention from https://ieeexplore.ieee.org/document/9076883") logger.warning("Warning: TODO implement attention from https://ieeexplore.ieee.org/document/9076883")
layer_next = tf.keras.layers.Dense(32, activation="relu")(layer_next) layer_next = tf.keras.layers.Dense(32)(layer_next)
layer_next = tf.keras.layers.Conv2D(water_bins, activation="relu", kernel_size=1, padding="same")(layer_next) layer_next = tf.keras.layers.Conv2D(water_bins, kernel_size=1, padding="same")(layer_next)
layer_next = tf.keras.layers.Softmax(axis=-1)(layer_next) layer_next = tf.keras.layers.Softmax(axis=-1)(layer_next)
model = tf.keras.Model( model = tf.keras.Model(