diff --git a/mlair/model_modules/model_class.py b/mlair/model_modules/model_class.py
index 04293c8f5c0ecf73e7d51fe6ae97e8ff7db96252..96cfdccf9e29f0b4cf780432f2a14a4406a60580 100644
--- a/mlair/model_modules/model_class.py
+++ b/mlair/model_modules/model_class.py
@@ -121,8 +121,6 @@ __author__ = "Lukas Leufen, Felix Kleinert"
__date__ = '2020-05-12'
import tensorflow.keras as keras
-import tensorflow as tf
-import numpy as np
from mlair.model_modules import AbstractModelClass
from mlair.model_modules.inception_model import InceptionModelBase
@@ -598,9 +596,9 @@ class MyUnet(AbstractModelClass):
def __init__(self, input_shape: list, output_shape: list):
super().__init__(input_shape[0], output_shape[0])
- self.filters = 32
- self.lstm_units = 64
- self.kernel_size = (3,1)# (3,1)
+ self.first_filter_size = self._input_shape[-1] # 16
+ self.lstm_units = 64 * 2
+ self.kernel_size = (3, 1) # (3,1)
self.activation = "elu"
self.pool_size = (2, 1)
@@ -612,8 +610,6 @@ class MyUnet(AbstractModelClass):
self.dense_units = 32*2
self.initial_lr = 0.001
- print(f"tf.__version__ = {tf.__version__}")
- print(f"np.__version__ = {np.__version__}")
# apply to model
self.set_model()
@@ -625,12 +621,12 @@ class MyUnet(AbstractModelClass):
pad_size = PadUtils.get_padding_for_same(self.kernel_size)
c1 = Padding2D("SymPad2D")(padding=pad_size)(input_train)
- c1 = keras.layers.Conv2D(16, self.kernel_size, activation=self.activation,
+ c1 = keras.layers.Conv2D(self.first_filter_size, self.kernel_size, activation=self.activation,
kernel_initializer=self.kernel_initializer, kernel_regularizer=self.kernel_regularizer,
bias_regularizer=self.bias_regularizer)(c1)
c1 = keras.layers.Dropout(0.1)(c1)
c1 = Padding2D("SymPad2D")(padding=pad_size)(c1)
- c1 = keras.layers.Conv2D(16, self.kernel_size, activation=self.activation,
+ c1 = keras.layers.Conv2D(self.first_filter_size, self.kernel_size, activation=self.activation,
kernel_initializer=self.kernel_initializer, name='c1',
kernel_regularizer=self.kernel_regularizer,
bias_regularizer=self.bias_regularizer)(c1)
@@ -638,12 +634,12 @@ class MyUnet(AbstractModelClass):
# p1 = keras.layers.MaxPooling2D(self.pool_size)(c1)
c2 = Padding2D("SymPad2D")(padding=pad_size)(p1)
- c2 = keras.layers.Conv2D(32, self.kernel_size, activation=self.activation,
+ c2 = keras.layers.Conv2D(self.first_filter_size * 2, self.kernel_size, activation=self.activation,
kernel_initializer=self.kernel_initializer, kernel_regularizer=self.kernel_regularizer,
bias_regularizer=self.bias_regularizer)(c2)
c2 = keras.layers.Dropout(0.1)(c2)
c2 = Padding2D("SymPad2D")(padding=pad_size)(c2)
- c2 = keras.layers.Conv2D(32, self.kernel_size, activation=self.activation,
+ c2 = keras.layers.Conv2D(self.first_filter_size * 2, self.kernel_size, activation=self.activation,
kernel_initializer=self.kernel_initializer, name='c2',
kernel_regularizer=self.kernel_regularizer,
bias_regularizer=self.bias_regularizer)(c2)
@@ -651,12 +647,12 @@ class MyUnet(AbstractModelClass):
# p2 = keras.layers.MaxPooling2D(self.pool_size)(c2)
c3 = Padding2D("SymPad2D")(padding=pad_size)(p2)
- c3 = keras.layers.Conv2D(64, self.kernel_size, activation=self.activation,
+ c3 = keras.layers.Conv2D(self.first_filter_size * 4, self.kernel_size, activation=self.activation,
kernel_initializer=self.kernel_initializer, kernel_regularizer=self.kernel_regularizer,
bias_regularizer=self.bias_regularizer)(c3)
c3 = keras.layers.Dropout(0.2)(c3)
c3 = Padding2D("SymPad2D")(padding=pad_size)(c3)
- c3 = keras.layers.Conv2D(64, self.kernel_size, activation=self.activation,
+ c3 = keras.layers.Conv2D(self.first_filter_size * 4, self.kernel_size, activation=self.activation,
kernel_initializer=self.kernel_initializer, name='c3',
kernel_regularizer=self.kernel_regularizer,
bias_regularizer=self.bias_regularizer)(c3)
@@ -665,14 +661,14 @@ class MyUnet(AbstractModelClass):
### own LSTM Block ###
ls1 = keras.layers.Reshape((p3.shape[1], p3.shape[-1]))(p3)
- ls1 = keras.layers.LSTM(64*2, return_sequences=True)(ls1)
- ls1 = keras.layers.LSTM(64*2, return_sequences=True)(ls1)
+ ls1 = keras.layers.LSTM(self.lstm_units, return_sequences=True)(ls1)
+ ls1 = keras.layers.LSTM(self.lstm_units, return_sequences=True)(ls1)
c4 = keras.layers.Reshape((p3.shape[1], 1, -1))(ls1)
### own 2nd LSTM Block ###
ls2 = keras.layers.Reshape((c3.shape[1], c3.shape[-1]))(c3)
- ls2 = keras.layers.LSTM(64 * 2, return_sequences=True)(ls2)
- ls2 = keras.layers.LSTM(64 * 2, return_sequences=True)(ls2)
+ ls2 = keras.layers.LSTM(self.lstm_units, return_sequences=True)(ls2)
+ ls2 = keras.layers.LSTM(self.lstm_units, return_sequences=True)(ls2)
c4_2 = keras.layers.Reshape((c3.shape[1], 1, -1))(ls2)
# c4 = Padding2D("SymPad2D")(padding=pad_size)(p3)
@@ -709,13 +705,13 @@ class MyUnet(AbstractModelClass):
u7 = keras.layers.concatenate([u7, cn3], name="u7_c3")
c7 = u7
# c7 = Padding2D("SymPad2D")(padding=pad_size)(u7)
- c7 = keras.layers.Conv2D(64, self.kernel_size, activation=self.activation,
+ c7 = keras.layers.Conv2D(self.first_filter_size * 4, self.kernel_size, activation=self.activation,
kernel_initializer=self.kernel_initializer, kernel_regularizer=self.kernel_regularizer,
bias_regularizer=self.bias_regularizer)(c7)
c7 = keras.layers.concatenate([c7, c4_2], name="Concat_2nd_LSTM")
c7 = keras.layers.Dropout(0.2)(c7)
c7 = Padding2D("SymPad2D")(padding=pad_size)(c7)
- c7 = keras.layers.Conv2D(64, self.kernel_size, activation=self.activation,
+ c7 = keras.layers.Conv2D(self.first_filter_size * 4, self.kernel_size, activation=self.activation,
kernel_initializer=self.kernel_initializer, name='c7_to_u8',
kernel_regularizer=self.kernel_regularizer,
bias_regularizer=self.bias_regularizer)(c7)
@@ -726,12 +722,12 @@ class MyUnet(AbstractModelClass):
# u8 = c3
u8 = keras.layers.concatenate([u8, c2], name="u8_c2")
c8 = Padding2D("SymPad2D")(padding=pad_size)(u8)
- c8 = keras.layers.Conv2D(32, self.kernel_size, activation=self.activation,
+ c8 = keras.layers.Conv2D(self.first_filter_size * 2, self.kernel_size, activation=self.activation,
kernel_initializer=self.kernel_initializer, kernel_regularizer=self.kernel_regularizer,
bias_regularizer=self.bias_regularizer)(c8)
c8 = keras.layers.Dropout(0.1)(c8)
c8 = Padding2D("SymPad2D")(padding=pad_size)(c8)
- c8 = keras.layers.Conv2D(32, self.kernel_size, activation=self.activation,
+ c8 = keras.layers.Conv2D(self.first_filter_size * 2, self.kernel_size, activation=self.activation,
kernel_initializer=self.kernel_initializer, name='c8_to_u9',
kernel_regularizer=self.kernel_regularizer,
bias_regularizer=self.bias_regularizer)(c8)
@@ -741,12 +737,12 @@ class MyUnet(AbstractModelClass):
u9 = c8
u9 = keras.layers.concatenate([u9, c1], name="u9_c1")
c9 = Padding2D("SymPad2D")(padding=pad_size)(u9)
- c9 = keras.layers.Conv2D(16, self.kernel_size, activation=self.activation,
+ c9 = keras.layers.Conv2D(self.first_filter_size, self.kernel_size, activation=self.activation,
kernel_initializer=self.kernel_initializer, kernel_regularizer=self.kernel_regularizer,
bias_regularizer=self.bias_regularizer)(c9)
c9 = keras.layers.Dropout(0.1)(c9)
c9 = Padding2D("SymPad2D")(padding=pad_size)(c9)
- c9 = keras.layers.Conv2D(16, self.kernel_size, activation=self.activation,
+ c9 = keras.layers.Conv2D(self.first_filter_size, self.kernel_size, activation=self.activation,
kernel_initializer=self.kernel_initializer, name='c9',
kernel_regularizer=self.kernel_regularizer,
bias_regularizer=self.bias_regularizer)(c9)