From 28cc011c31b0dde05eddb5ea2d170e8c0c7fa78a Mon Sep 17 00:00:00 2001
From: Yigit Sever
Date: Fri, 3 Jul 2020 00:19:30 +0300
Subject: Clean supervised matching

---
 learn_and_predict.py | 78 ++++++++++++++++++++++++++++++++++------------------
 1 file changed, 52 insertions(+), 26 deletions(-)

diff --git a/learn_and_predict.py b/learn_and_predict.py
index 4c094d7..907d4be 100644
--- a/learn_and_predict.py
+++ b/learn_and_predict.py
@@ -4,10 +4,11 @@ import csv
 import keras
 import keras.backend as K
 import numpy as np
+from Helpers import Data, Get_Embedding
 from keras.layers import LSTM, Embedding, Input, Lambda, concatenate
 from keras.models import Model
-
-from Helpers import Data, Get_Embedding
+from lapjv import lapjv
+from sklearn.model_selection import train_test_split
 
 
 def get_learning_rate(epoch=None, model=None):
@@ -39,8 +40,8 @@ def main(args):
 
     x_train = data.x_train
     y_train = data.y_train
-    x_predict = data.x_val
-    y_predict = data.y_val
+    x_test = data.x_val
+    y_test = data.y_val
     vocab_size = data.vocab_size
     max_len = data.max_len
 
@@ -49,13 +50,23 @@ def main(args):
     y = np.bincount(x.astype(np.int32))
     ii = np.nonzero(y)[0]
     assert ii == 1
-    assert y[ii] == 1000  # hardcoded for now
+    assert y[ii] == 1000  # hardcoded, sorry
+
+    # separating the train-test set and the set used for our task
+    (
+        source_train,
+        source_validate,
+        target_train,
+        target_validate,
+        Y_train,
+        Y_validate,
+    ) = train_test_split(x_train[0], x_train[1], y_train)
 
     if not batch:
         print(f"Source Lang: {source_lang}")
         print(f"Target Lang: {target_lang}")
         print(f"Using {len(x_train[0])} pairs to learn")
-        print(f"Predicting {len(y_predict)} pairs")
+        print(f"Predicting {len(Y_validate)} pairs")
         print(f"Vocabulary size: {vocab_size}")
         print(f"Maximum sequence length: {max_len}")
 
@@ -82,7 +93,7 @@ def main(args):
     input_1 = embed_layer(seq_1)
     input_2 = embed_layer(seq_2)
 
-    l1 = LSTM(units=hidden_size)
+    l1 = LSTM(units=300, use_bias=True, dropout=0.05, recurrent_dropout=0.1)
 
     l1_out = l1(input_1)
     l2_out = l1(input_2)
@@ -95,7 +106,7 @@ def main(args):
 
     model = Model(inputs=[seq_1, seq_2], outputs=[main_output])
 
-    opt = keras.optimizers.Adadelta(lr=learning_rate, clipnorm=1.25)
+    opt = keras.optimizers.Adadelta(lr=learning_rate, clipnorm=1.2)
 
     model.compile(optimizer=opt, loss="mean_squared_error", metrics=["accuracy"])
     model.summary()
@@ -104,45 +115,60 @@ def main(args):
         monitor="val_accuracy", patience=5, verbose=1, factor=0.5, min_lr=0.0001
     )
 
-    history = model.fit(
-        x_train,
-        y_train,
-        validation_data=(x_predict, y_predict),
+    model.fit(
+        x=[source_train, target_train],
+        y=Y_train,
+        validation_data=([source_validate, target_validate], Y_validate),
         epochs=num_iters,
         batch_size=32,
-        verbose=1,
+        verbose=0,
         callbacks=[adjuster],
     )
 
-    target_sents = x_predict[1]
+    cost_matrix = None
+
+    target_sents = x_test[1]
     precision_at_one = 0
-    precision_at_ten = 0
-    for index, sent in enumerate(x_predict[0]):
+    for index, sent in enumerate(x_test[0]):
         source_sents = np.array([sent] * 1000)
         to_predict = [source_sents, target_sents]
         preds = model.predict(to_predict)
-        ind = np.argpartition(preds.ravel(), -10)[-10:]
-        if index in ind:
-            precision_at_ten += 1
+
+        if index == 0:
+            cost_matrix = np.c_[preds]
+        else:
+            cost_matrix = np.c_[cost_matrix, preds]
+
         if np.argmax(preds.ravel()) == index:
             precision_at_one += 1
 
-    training_samples = len(x_train[0])
-    validation_samples = len(y_predict)
+    cost_matrix = cost_matrix * -1000
+
+    row_ind, col_ind, a = lapjv(cost_matrix, verbose=False)
+
+    result = zip(row_ind, col_ind)
+    hit_one = len([x for x, y in result if x == y])
+
+    matching_pa1 = hit_one / 1000 * 100
+
+    training_samples = len(Y_train)
+    validation_samples = len(Y_validate)
+    test_samples = len(y_test)
     fields = [
         source_lang,
         target_lang,
         training_samples,
         validation_samples,
+        test_samples,
         precision_at_one,
-        precision_at_ten,
+        matching_pa1,
     ]
 
     if not batch:
         print(f"Supervised Retrieval {source_lang} - {target_lang}")
         print(f"P@1: {precision_at_one/1000}")
     else:
-        with open("supervised.csv", "a") as f:
+        with open("supervised_matching.csv", "a") as f:
             writer = csv.writer(f)
             writer.writerow(fields)
 
@@ -186,7 +212,7 @@ if __name__ == "__main__":
         "--hidden_size",
         type=int,
         help="Number of units in LSTM layer.",
-        default=50,
+        default=300,
     )
     parser.add_argument(
         "-b",
@@ -196,14 +222,14 @@ if __name__ == "__main__":
         + "running in a single instance (output the results)",
     )
     parser.add_argument(
-        "-n", "--num_iters", type=int, help="Number of iterations/epochs.", default=7
+        "-n", "--num_iters", type=int, help="Number of iterations/epochs.", default=30
     )
     parser.add_argument(
         "-lr",
         "--learning_rate",
         type=float,
         help="Learning rate for optimizer.",
-        default=1.0,
+        default=0.5,
     )
 
     args = parser.parse_args()
-- 
cgit v1.2.3-61-g4310