Codemaster

codenames/big_glove_to_w2v_learning.py

@@ -0,0 +1,69 @@
+import tensorflow as tf
+from tensorflow.keras import layers, models
+import numpy as np
+import matplotlib.pyplot as plt
+from sklearn.model_selection import train_test_split
+
+from codenames.game import Game
+
+# Custom loss function using cosine distance
+def cosine_distance_loss(y_true, y_pred):
+    # Calculate the cosine similarity and convert it to a distance
+    y_true = tf.math.l2_normalize(y_true, axis=-1)
+    y_pred = tf.math.l2_normalize(y_pred, axis=-1)
+    return 1 - tf.reduce_sum(y_true * y_pred, axis=-1)
+
+def learn_vector_relationship(dict1, dict2, key_list, epochs=1024, batch_size=32, validation_split=0.2):
+    # Extract input (X) and output (Y) data based on the key list
+    X = np.array([dict1[key.lower()] for key in key_list])
+    Y = np.array([dict2[key.lower()] for key in key_list])
+
+    # Split the data into training and validation sets
+    X_train, X_val, Y_train, Y_val = train_test_split(X, Y, test_size=validation_split)
+
+    # Define a simple feedforward neural network
+    model = models.Sequential()
+    model.add(layers.Input(shape=(X.shape[1],)))  # Input layer with number of units matching X's dimensionality
+    model.add(layers.Dense(300, activation='relu'))  # Hidden layer with 128 units
+    model.add(layers.Dense(Y.shape[1]))  # Output layer with number of units matching Y's dimensionality
+
+    # Compile the model
+    model.compile(optimizer='adam', loss=cosine_distance_loss)
+
+    # Train the model with validation
+    history = model.fit(X_train, Y_train, validation_data=(X_val, Y_val),
+                        epochs=epochs, batch_size=batch_size)
+
+    # Plot the training and validation loss over epochs
+    plt.plot(history.history['loss'], label='Training Loss')
+    plt.plot(history.history['val_loss'], label='Validation Loss')
+    plt.title('Model Loss Over Epochs learning w2v from glove 300')
+    plt.xlabel('Epoch')
+    plt.ylabel('Loss (MSE)')
+    plt.legend()
+    plt.savefig("learning_loss_from_big_glove_to_w2v.png")
+    plt.show()
+
+    # Return the trained model
+    return model
+
+if __name__ == "__main__":
+    # Example usage:
+    glove_vectors = Game.load_glove_vecs("players/glove.6B.300d.txt")
+    w2v_vectors = Game.load_w2v("players/GoogleNews-vectors-negative300.bin")
+    wordlist = []
+    with open('combine_words.txt') as infile:
+        for line in infile:
+            wordlist.append(line.rstrip().lower())
+    model = learn_vector_relationship(glove_vectors, w2v_vectors, wordlist)
+
+    # Prepare the input matrix for all words in the wordlist
+    input_vectors = np.array([glove_vectors[word.lower()] for word in wordlist])
+
+    # Predict all vectors at once
+    predicted_vectors = model.predict(input_vectors)
+
+    # Create the translated_vecs dictionary
+    translated_vecs = {word.lower(): predicted_vectors[i] for i, word in enumerate(wordlist)}
+    model.save('glove300_to_w2v_model.keras')
+