💰

Dockerfile

@@ -1,5 +1,5 @@
 FROM ubuntu:22.04
 
 RUN apt update && apt install -y vim make python3 python3-pip python-is-python3 gcc g++ golang wget unzip git
-RUN pip install pandas matplotlib scikit-learn
+RUN pip install pandas matplotlib scikit-learn tensorflow
 CMD "bash"

dataset.jenkinsfile

@@ -8,7 +8,7 @@ node {
 			checkout([$class: 'GitSCM', branches: [[name: 'ztm']], extensions: [], userRemoteConfigs: [[url: 'https://git.wmi.amu.edu.pl/s452639/ium_452639']]])
 			sh 'cd src; ./prepare-ztm-data.sh'
 
-			archiveArtifacts artifacts: 'src/stop_times.normalized.tsv,src/stop_times.train.tsv,src/stop_times.test.tsv,src/stop_times.valid.tsv',
+			archiveArtifacts artifacts: 'src/stop_times.normalized.tsv,src/stop_times.train.tsv,src/stop_times.test.tsv,src/stop_times.valid.tsv,src/stop_times.categories.tsv',
 				followSymlinks: false
 		}
 	}

run.sh

@@ -3,4 +3,4 @@
 set -xe
 
 docker build -t ium .
-docker run -it ium
+docker run -v .:/ium/ -it ium

src/.gitignore

@@ -0,0 +1,2 @@
+model.keras
+pics

src/split_train_valid_test.py

@@ -4,7 +4,7 @@ from sklearn.model_selection import train_test_split
 
 data = pd.read_csv('./stop_times.normalized.tsv', sep='\t')
 
-train, test = train_test_split(data, test_size=0.5)
+train, test = train_test_split(data, test_size=0.8)
 valid, test = train_test_split(test, test_size=0.5)
 
 train.to_csv('stop_times.train.tsv', sep='\t')

src/tf_test.py

@@ -0,0 +1,15 @@
+from tf_train import *
+import numpy as np
+
+def test():
+    global model, le
+    test_x, test_y, _ = load_data('./stop_times.test.tsv', le)
+    test_x = tf.convert_to_tensor(test_x, dtype=tf.float32)
+    test_y = tf.convert_to_tensor(test_y)
+
+    model = tf.keras.models.load_model('model.keras')
+    pd.DataFrame(model.predict(test_x), columns=le.classes_).to_csv('stop_times.predictions.tsv', sep='\t')
+
+
+if __name__ == "__main__":
+    test()

src/tf_train.py

@@ -22,20 +22,22 @@ def load_data(path: str, le: LabelEncoder):
 
 num_classes = len(le.classes_)
 
-model = tf.keras.Sequential([
-    tf.keras.layers.Input(shape=(2,)),
-    tf.keras.layers.Dense(4 * num_classes, activation='relu'),
-    tf.keras.layers.Dense(4 * num_classes, activation='relu'),
-    tf.keras.layers.Dense(4 * num_classes, activation='relu'),
-    tf.keras.layers.Dense(num_classes, activation='softmax')
-])
-
-model.compile(optimizer='adam',
-              loss=tf.keras.losses.SparseCategoricalCrossentropy(),
-              metrics=['accuracy'])
 
 def train():
-    global model, le
+    global le
+
+    model = tf.keras.Sequential([
+        tf.keras.layers.Input(shape=(2,)),
+        tf.keras.layers.Dense(4 * num_classes, activation='relu'),
+        tf.keras.layers.Dense(4 * num_classes, activation='relu'),
+        tf.keras.layers.Dense(4 * num_classes, activation='relu'),
+        tf.keras.layers.Dense(num_classes, activation='softmax')
+    ])
+
+    model.compile(optimizer='adam',
+                  loss=tf.keras.losses.SparseCategoricalCrossentropy(),
+                  metrics=['accuracy'])
+
     train_x, train_y, _ = load_data('./stop_times.train.tsv', le)
     train_x = tf.convert_to_tensor(train_x, dtype=tf.float32)
     train_y = tf.convert_to_tensor(train_y)
@@ -50,22 +52,7 @@ def train():
     with open('history', 'w') as f:
         print(repr(history), file=f)
 
-    model.save_weights('model.ckpt')
+    model.save('model.keras')
 
-def test():
-    global model, le
-    test_x, test_y, _ = load_data('./stop_times.test.tsv', le)
-    test_x = tf.convert_to_tensor(test_x, dtype=tf.float32)
-    test_y = tf.convert_to_tensor(test_y)
-    model.load_weights('model.ckpt')
-    model.evaluate(test_x, test_y)
-
-SUBCOMMANDS = {
-    "test": test,
-    "train": train,
-}
-
-import sys
-assert len(sys.argv) == 2
-assert sys.argv[1] in SUBCOMMANDS.keys()
-SUBCOMMANDS[sys.argv[1]]()
+if __name__ == "__main__":
+    train()