Create evaluation_script.py

evaluation_script.py

@@ -0,0 +1,196 @@
+import chess
+import chess.engine
+import numpy as np
+import tensorflow as tf
+import time
+import os
+import datetime
+import shutil  # For zip creation
+from google.colab import files # For download trigger
+
+# --- 1. Neural Network (Policy and Value Network) ---
+class PolicyValueNetwork(tf.keras.Model):
+    def __init__(self, num_moves):
+        super(PolicyValueNetwork, self).__init__()
+        self.conv1 = tf.keras.layers.Conv2D(32, 3, activation='relu', padding='same')
+        self.flatten = tf.keras.layers.Flatten()
+        self.dense_policy = tf.keras.layers.Dense(num_moves, activation='softmax', name='policy_head')
+        self.dense_value = tf.keras.layers.Dense(1, activation='tanh', name='value_head')
+
+    def call(self, inputs):
+        x = self.conv1(inputs)
+        x = self.flatten(x)
+        policy = self.dense_policy(x)
+        value = self.dense_value(x)
+        return policy, value
+
+# --- 2. Move Encoding/Decoding (Correct and Deterministic Implementation) ---
+NUM_POSSIBLE_MOVES = 4672 # Correct value based on deterministic encoding
+NUM_INPUT_PLANES = 12
+
+# Load model weights
+policy_value_net = PolicyValueNetwork(NUM_POSSIBLE_MOVES)
+
+# dummy input for building network
+dummy_input = tf.random.normal((1, 8, 8, NUM_INPUT_PLANES))
+policy, value = policy_value_net(dummy_input)
+
+
+# Load the weights (replace 'your_model.weights.h5' with your actual file)
+try:
+    model_path = "/stockzero/models_weights/StockZero-2025-03-24-1727.weights.h5"
+    policy_value_net.load_weights(model_path)
+    print(f"Model weights loaded successfully from '{model_path}'")
+except Exception as e:
+    print(f"Error loading weights: {e}")
+
+# --- Create output directory and set output paths ---
+OUTPUT_DIR = "/content/converted_models"
+os.makedirs(OUTPUT_DIR, exist_ok=True) # Create the folder if it does not exist
+
+SAVED_MODEL_DIR = os.path.join(OUTPUT_DIR, "saved_model")
+KERAS_MODEL_PATH = os.path.join(OUTPUT_DIR, "model.keras")
+H5_MODEL_PATH = os.path.join(OUTPUT_DIR, "model_weights.h5")
+PYTORCH_MODEL_PATH = os.path.join(OUTPUT_DIR, "pytorch_model.pth")
+PYTORCH_FULL_MODEL_PATH = os.path.join(OUTPUT_DIR, "pytorch_full_model.pth")
+ONNX_MODEL_PATH = os.path.join(OUTPUT_DIR, "model.onnx")
+TFLITE_MODEL_PATH = os.path.join(OUTPUT_DIR, "model.tflite")
+BIN_FILE_PATH = os.path.join(OUTPUT_DIR, "model_weights.bin")
+NUMPY_FILE_PATH = os.path.join(OUTPUT_DIR, "model_weights.npz")
+
+
+# --- 1.  Keras/TensorFlow (SavedModel format) ---
+try:
+  tf.saved_model.save(policy_value_net, SAVED_MODEL_DIR)
+  print(f"Model saved as SavedModel to '{SAVED_MODEL_DIR}'")
+except Exception as e:
+  print(f"Error saving model as SavedModel: {e}")
+
+# --- 2. Keras .keras Format (Weights + Architecture) ---
+try:
+    policy_value_net.save(KERAS_MODEL_PATH)
+    print(f"Model saved as Keras .keras format to '{KERAS_MODEL_PATH}'")
+except Exception as e:
+    print(f"Error saving as .keras format: {e}")
+# --- 3.  Keras/TensorFlow (.h5 - Weights) ---
+try:
+    policy_value_net.save_weights(H5_MODEL_PATH)
+    print(f"Model weights saved as .h5 to '{H5_MODEL_PATH}'")
+except Exception as e:
+  print(f"Error saving model weights as .h5: {e}")
+
+
+# --- 4. PyTorch ---
+import torch
+import torch.nn as nn
+
+class PyTorchPolicyValueNetwork(nn.Module):
+    def __init__(self, num_moves):
+        super(PyTorchPolicyValueNetwork, self).__init__()
+        self.conv1 = nn.Conv2d(12, 32, kernel_size=3, padding=1) # Input 12 channels for chess
+        self.relu = nn.ReLU()
+        self.flatten = nn.Flatten()
+        self.dense_policy = nn.Linear(8*8*32, num_moves)  # Calculate size using the parameters from keras layer, after flatten output is 8*8*32
+        self.softmax = nn.Softmax(dim=1)
+        self.dense_value = nn.Linear(8*8*32, 1)
+        self.tanh = nn.Tanh()
+
+    def forward(self, x):
+        x = self.relu(self.conv1(x))
+        x = self.flatten(x)
+        policy = self.softmax(self.dense_policy(x))
+        value = self.tanh(self.dense_value(x))
+        return policy, value
+
+try:
+    pytorch_model = PyTorchPolicyValueNetwork(NUM_POSSIBLE_MOVES)
+
+    # Get Keras layers
+    keras_conv1 = policy_value_net.conv1
+    keras_dense_policy = policy_value_net.dense_policy
+    keras_dense_value = policy_value_net.dense_value
+
+    # Transfer weights from Keras to PyTorch
+    pytorch_model.conv1.weight = torch.nn.Parameter(torch.tensor(keras_conv1.kernel.numpy().transpose(3,2,0,1), dtype=torch.float32))
+    pytorch_model.conv1.bias = torch.nn.Parameter(torch.tensor(keras_conv1.bias.numpy(), dtype=torch.float32))
+
+    pytorch_model.dense_policy.weight = torch.nn.Parameter(torch.tensor(keras_dense_policy.kernel.numpy().transpose(), dtype=torch.float32))
+    pytorch_model.dense_policy.bias = torch.nn.Parameter(torch.tensor(keras_dense_policy.bias.numpy(), dtype=torch.float32))
+
+    pytorch_model.dense_value.weight = torch.nn.Parameter(torch.tensor(keras_dense_value.kernel.numpy().transpose(), dtype=torch.float32))
+    pytorch_model.dense_value.bias = torch.nn.Parameter(torch.tensor(keras_dense_value.bias.numpy(), dtype=torch.float32))
+
+    torch.save(pytorch_model.state_dict(), PYTORCH_MODEL_PATH)
+    print(f"PyTorch model weights saved to '{PYTORCH_MODEL_PATH}'")
+
+    torch.save(pytorch_model, PYTORCH_FULL_MODEL_PATH) # Save full model
+    print(f"PyTorch model saved as '{PYTORCH_FULL_MODEL_PATH}'")
+except Exception as e:
+  print(f"Error during PyTorch conversion: {e}")
+
+
+# --- 5. ONNX ---
+import tf2onnx
+try:
+    spec = (tf.TensorSpec((None, 8, 8, 12), tf.float32, name="input"),)
+    onnx_model, _ = tf2onnx.convert.from_keras(policy_value_net, input_signature=spec)
+
+    with open(ONNX_MODEL_PATH, "wb") as f:
+        f.write(onnx_model.SerializeToString())
+    print(f"Model saved as ONNX to '{ONNX_MODEL_PATH}'")
+except Exception as e:
+    print(f"Error saving model as ONNX: {e}")
+
+# --- 6. TensorFlow Lite ---
+try:
+    converter = tf.lite.TFLiteConverter.from_keras_model(policy_value_net)
+    tflite_model = converter.convert()
+
+    with open(TFLITE_MODEL_PATH, 'wb') as f:
+        f.write(tflite_model)
+    print(f"Model saved as TFLite to '{TFLITE_MODEL_PATH}'")
+except Exception as e:
+  print(f"Error converting to TFLite: {e}")
+
+
+# --- 7. Binary (.bin) format (Custom Implementation) ---
+try:
+    with open(BIN_FILE_PATH, 'wb') as f:
+       for layer in policy_value_net.layers:
+            for weight in layer.weights:
+              weight_arr = weight.numpy()
+              f.write(weight_arr.tobytes())
+    print(f"Model weights saved as .bin to '{BIN_FILE_PATH}'")
+except Exception as e:
+    print(f"Error saving model weights as .bin: {e}")
+
+# --- 8. NumPy arrays (.npz) format ---
+try:
+  all_weights = {}
+  for layer in policy_value_net.layers:
+      for i, weight in enumerate(layer.weights):
+            all_weights[f"{layer.name}_weight_{i}"] = weight.numpy()
+  np.savez(NUMPY_FILE_PATH, **all_weights)
+  print(f"Model weights saved as NumPy arrays to '{NUMPY_FILE_PATH}'")
+except Exception as e:
+  print(f"Error saving model weights as NumPy: {e}")
+
+
+# --- 9. TensorFlow.js (requires command line tool)---
+# ---   This would require the TensorFlow.js converter tool ---
+# ---  Command-Line example shown below (run in shell, not in the script) ---
+# ---    tensorflowjs_converter --input_format=tf_saved_model ./saved_model ./tfjs_model ---
+print("To convert to TensorFlow.js format, run the 'tensorflowjs_converter' command-line tool (see comments in script).")
+
+
+# --- Zip all files and create download ---
+try:
+    current_datetime = datetime.datetime.now()
+    zip_file_name = f"converted_models-{current_datetime.strftime('%Y%m%d%H%M')}"
+    zip_file_path = f"/directory/{zip_file_name}"
+    shutil.make_archive(zip_file_path, 'zip', OUTPUT_DIR) # Create zip archive
+    print(f"All converted model files zipped to '{zip_file_path}.zip'")
+    files.download(f"{zip_file_path}.zip") # Trigger download in Colab
+    print("Download should start in a moment.")
+except Exception as e:
+  print(f"Error zipping and creating download: {e}")