Merge branch 'main' of https://github.com/lorserker/ben into main

scripts/training/opening lead/lead_nn_tf2.py

@@ -0,0 +1,70 @@
+import sys
+sys.path.append('../../../src')
+import datetime
+import numpy as np
+import os
+import logging
+from tensorflow import keras
+from batcher import Batcher
+
+model_path = './models/lead_model.h5'  # Specify the .h5 file
+seed = 1337
+batch_size = 64
+n_iterations = 10000
+display_step = 1000
+
+X_train = np.load('./lead_bin/X.npy')
+B_train = np.load('./lead_bin/B.npy')
+y_train = np.load('./lead_bin/y.npy')
+
+n_examples = X_train.shape[0]
+n_ftrs = X_train.shape[1]
+n_cards = 32
+n_bi = B_train.shape[1]
+
+n_hidden_units = 512
+
+# Define the input layers with specific shapes
+x_ftrs = keras.layers.Input(shape=(n_ftrs,), name='X')
+b_ftrs = keras.layers.Input(shape=(n_bi,), name='B')
+
+# Concatenate x_ftrs and b_ftrs
+XB = keras.layers.concatenate([x_ftrs, b_ftrs], name='XB')
+
+# Define the rest of the model architecture based on your requirements
+model = keras.Sequential()
+
+model.add(keras.layers.Dense(n_hidden_units, activation='relu', kernel_initializer='glorot_uniform', name='w1'))
+model.add(keras.layers.Dropout(0.6, name='a1'))
+model.add(keras.layers.Dense(n_hidden_units, activation='relu', kernel_initializer='glorot_uniform', name='w2'))
+model.add(keras.layers.Dropout(0.6, name='a2'))
+model.add(keras.layers.Dense(n_hidden_units, activation='relu', kernel_initializer='glorot_uniform', name='w3'))
+model.add(keras.layers.Dropout(0.6, name='a3'))
+model.add(keras.layers.Dense(32, kernel_initializer='glorot_uniform', name='w_out'))
+
+# Define the output of the model
+lead_softmax = model(XB)
+
+# Compile the model
+compiled_model = keras.Model(inputs=[x_ftrs, b_ftrs], outputs=lead_softmax)
+compiled_model.compile(optimizer=keras.optimizers.Adam(learning_rate=0.001), loss='categorical_crossentropy')
+
+# Train the Keras model
+batch = Batcher(n_examples, batch_size)
+cost_batch = Batcher(n_examples, 10000)
+
+for i in range(n_iterations):
+    x_batch, b_batch, y_batch = batch.next_batch([X_train, B_train, y_train])
+    if i % display_step == 0:
+        x_cost, b_cost, y_cost = cost_batch.next_batch([X_train, B_train, y_train])
+        c_train = compiled_model.evaluate([x_cost, b_cost], y_cost, batch_size=64, verbose=0)
+        l_train = compiled_model.predict([x_cost, b_cost], batch_size=64)
+        print('{}. c_train={}'.format(i, c_train))
+        print(np.mean(np.argmax(l_train, axis=1) == np.argmax(y_cost, axis=1)))
+
+        sys.stdout.flush()
+        compiled_model.save(model_path + '_keras.h5')  # Save the Keras model in HDF5 format
+
+    compiled_model.fit([x_batch, b_batch], y_batch, batch_size=64, epochs=1, verbose=0)
+
+compiled_model.save(model_path + '_keras.h5') 

src/binary.py

@@ -233,9 +233,11 @@ def get_lead_binary(auction, hand, binfo, vuln, ns, ew):
     n_steps = 1 + len(auction) // 4
     A = get_auction_binary(n_steps, auction, lead_index, hand, vuln, ns, ew)
 
-    p_hcp, p_shp = binfo.model(A)
-
-    b[:3] = p_hcp.reshape((-1, n_steps, 3))[:,-1,:].reshape(3)
-    b[3:] = p_shp.reshape((-1, n_steps, 12))[:,-1,:].reshape(12)
+    if binfo:
+        p_hcp, p_shp = binfo.model(A)
 
+        b[:3] = p_hcp.reshape((-1, n_steps, 3))[:,-1,:].reshape(3)
+        b[3:] = p_shp.reshape((-1, n_steps, 12))[:,-1,:].reshape(12)
+    print(x)
+    print(b)
     return x.reshape((1, -1)), b.reshape((1, -1))

src/game.py

@@ -561,7 +561,7 @@ async def main():
             rdeal = random_deal()
 
             # example of to use a fixed deal
-            # rdeal = ('AQ9.543.6.AKJ876 762.A96.KQJ42.Q2 KJ83.KJ2.T753.T5 T54.QT87.A98.943', 'S Both')
+            rdeal = ('AQ9.543.6.AKJ876 762.A96.KQJ42.Q2 KJ83.KJ2.T753.T5 T54.QT87.A98.943', 'S Both')
 
             driver.set_deal(None, *rdeal, ns, ew, False)
         else:

src/nn/leaderH5.py

@@ -0,0 +1,24 @@
+import numpy as np
+import tensorflow as tf
+from tensorflow.keras.models import load_model
+
+class LeaderH5:
+
+    def __init__(self, model_path):
+        self.model_path = model_path
+        self.model = self.init_model()
+
+    def load_model(self):
+        model = load_model(self.model_path)
+        return model
+
+    def init_model(self):
+        model = self.load_model()
+
+        def pred_fun(x, b):
+            x = np.array(x)  # Ensure that input data is in the right format
+            b = np.array(b)
+            result = model.predict([x, b])
+            result_with_softmax = tf.nn.softmax(result, axis=-1).numpy()  # Apply softmax activation
+            return result_with_softmax
+        return pred_fun