Ready for final training: 130 epoch, Adam, lr=1e-8

Author: ArbinTimilsina

analyze_model.py

@@ -6,7 +6,7 @@
 from keras import backend as K
 from keras.models import load_model
 from tools.plotting_tools import plot_feature_label_prediction
-from tools.loss_metrics_tools import intersection_over_union, mean_iou, focal_loss
+from tools.loss_metrics_tools import intersection_over_union, mean_iou
 from tools.data_tools import DataSequence, get_data_generator, preprocess_feature, preprocess_label
 
 def argument_parser():
@@ -79,8 +79,7 @@ def main():
 
     # Get the model
     model_path = os.path.join("saved_models", "model_and_weights.hdf5")
-    model = load_model(model_path, custom_objects={"loss": focal_loss(),
-                                                   "mean_iou": mean_iou})
+    model = load_model(model_path, custom_objects={"mean_iou": mean_iou})
 
     # Make comparision plots
     generator_testing = get_data_generator(FEATURE_FILE_TESTING, LABEL_FILE_TESTING)

batch_jobs/submit_batch_jobs_fnal_tev.sh

@@ -47,7 +47,7 @@ echo "Running python analyze_model.py"
 echo "JOB $SLURM_JOB_ID is running on $SLURM_JOB_NODELIST "
 echo "*********************************************************"
 echo ""
-singularity exec --bind /data/arbint --nv  /data/arbint/DeepLearningWithProtoDUNE.img python analyze_model.py -p 5 -s Development
+singularity exec --bind /data/arbint --nv  /data/arbint/DeepLearningWithProtoDUNE.img python analyze_model.py -p 50 -s Development
 
 echo "*********************************************************"
 echo "All done. Exiting"

configurations/fnal_tev_configuration.ini

@@ -21,7 +21,7 @@ BATCH_SIZE = 2
 NUM_TRAINING = 2637
 NUM_VALIDATION = 246
 NUM_TESTING = 62
-NUM_EPOCHS = 120
+NUM_EPOCHS = 130
 
 [DEVELOPMENT]
 NUM_TRAINING = 40

tools/tiramisu_model.py

@@ -1,3 +1,15 @@
+"""
+Implementation of the One Hundred Layers Tiramisu as described in
+The One Hundred Layers Tiramisu: Fully Convolutional DenseNets for Semantic Segmentation.
+
+Tiramisu is a Fully Convolutional Networks (FCN) network based on DenseNet architecture.
+
+Papers:
+Tiramisu: https://arxiv.org/pdf/1611.09326.pdf
+FCN: https://people.eecs.berkeley.edu/~jonlong/long_shelhamer_fcn.pdf
+DenseNet: https://arxiv.org/abs/1608.06993
+"""
+
 from keras.models import Model
 from keras.layers import Reshape
 from keras.regularizers import l2
@@ -36,6 +48,7 @@ def transition_dn(x, dr, wd):
     #x = conv_relu_bn(x, x.get_shape().as_list()[-1], size=1, dr=dr, wd=wd)
     #return MaxPooling2D(strides=(2, 2))(x)
 
+    # Try stride 2 1x1 convolution instead
     return conv_relu_bn(x, x.get_shape().as_list()[-1], size=1, dr=dr, wd=wd, stride=2)
 
 def down_path(x, num_layers, growth_rate, dr, wd):
@@ -58,7 +71,7 @@ def up_path(added, skips, num_layers, growth_rate, dr, wd):
         x, added = dense_block(n, x, growth_rate, dr, wd)
     return x
 
-""""
+"""
 init_num_filter: initial number of filters
 num_layers_per_block: list of number of layers in each dense block
 growth_rate: number of filters to add per dense block

train_model.py

@@ -6,12 +6,12 @@
 import configparser
 from keras.layers import Input
 from keras import backend as K
+from keras.optimizers import Adam
 from keras.utils import plot_model
 from tools.data_tools import DataSequence
+from tools.loss_metrics_tools import mean_iou
 from tools.callbacks import PredictionsCallback
-from keras.optimizers import SGD, RMSprop, Adam
 from tools.tiramisu_model import get_tiramisu_model
-from tools.loss_metrics_tools import mean_iou, focal_loss
 from tools.plotting_tools import plot_history, plot_feature_label_prediction
 from keras.callbacks import EarlyStopping, ReduceLROnPlateau, ModelCheckpoint
 
@@ -123,25 +123,11 @@ def main():
         except:
             print("Old weights couldn't be loaded successfully, will continue!")
 
-    learning_rate = 1.0e-6;
-    decay_rate = learning_rate/NUM_EPOCHS
-    print("Decay rate is set to {}.".format(decay_rate))
-
-    test = 1
-    if test == 1:
-        model.compile(optimizer=SGD(lr=learning_rate), loss='categorical_crossentropy', metrics=['accuracy', mean_iou])
-    elif test == 2:
-        model.compile(optimizer=SGD(lr=learning_rate, decay=decay_rate), loss='categorical_crossentropy', metrics=['accuracy', mean_iou])
-    elif test == 3:
-        model.compile(optimizer=RMSprop(lr=learning_rate), loss='categorical_crossentropy', metrics=['accuracy', mean_iou])
-    elif test == 4:
-        model.compile(optimizer=RMSprop(lr=learning_rate), loss=focal_loss(), metrics=['accuracy', mean_iou])
-    elif test == 5:
-        model.compile(optimizer=Adam(lr=learning_rate), loss='categorical_crossentropy', metrics=['accuracy', mean_iou])
-    else:
-        print("\nError: Test is not in the range.")
-        print("Exiting!\n")
-        sys.exit(1)
+    learning_rate = 1.0e-8;
+    #decay_rate = learning_rate/NUM_EPOCHS
+    #print("Decay rate is set to {}.".format(decay_rate))
+
+    model.compile(optimizer=Adam(lr=learning_rate), loss='categorical_crossentropy', metrics=['accuracy', mean_iou])
 
     # Print model summary
     #model.summary()