Refactor autoencoder model (#8)

reflects API changes made in chess-position parent repo

.gitignore

@@ -1,4 +1,7 @@
+cogram_test.ipynb
+
 # Data and models
+data/*
 data/pgn/*
 !data/pgn/README.md
 !data/pgn/lichess_db_standard_rated_2013-01-truncated.pgn

config.yaml

@@ -1,7 +1,28 @@
 preprocess:
-  pgn_path: "./data/pgn/lichess_db_standard_rated_2013-01.pgn"
-  save_path: "./data/tensors/lichess_db_standard_rated_2013-01_v2.h5"
-  game_filter: "none"
-  game_processor: "positions_to_tensor"
+  pgn_path: ./data/pgn/lichess_db_standard_rated_2013-01.pgn
+  save_path: ./data/tensors/lichess_db_standard_rated_2013-01_v2.h5
+  game_filter: none
+  game_processor: positions_to_tensor
   chunk_size: 100000
-  number_games: 100
+  number_games: 100
+train:
+  data:
+    format: tensor
+    train_dir: ./data/tensors/train
+    train_batch_size: 256
+    test_dir: ./data/tensors/test
+    test_batch_size: 256
+  model:
+    loss: binary_crossentropy
+    train_steps_per_epoch: 20
+    test_steps_per_epoch: 5
+    tf_callbacks: [checkpoints]
+    save_dir: ./models
+evaluate:
+  data:
+    test_dir: ./data/tensors/test
+    test_batch_size: 64
+    model_dir: ./models
+  eval:
+    number_examples: 10
+    result_dir: ./results

evaluate_model.py

@@ -1,3 +1,4 @@
+import chess
 import yaml
 import json
 from pathlib import Path
@@ -6,75 +7,86 @@
 import tensorflow as tf
 from tensorflow import keras
 
-from chesspos.preprocessing import SampleGenerator, bitboard_to_board
-from chesspos.models import DenseAutoencoder, CnnAutoencoder
+from chesspos.preprocessing import SampleGenerator, tensor_to_board, board_to_tensor
+from chesspos.models import ResnetAutoencoder
 
-evaluate_data_params = yaml.safe_load(open(Path(__file__).with_name('params.yaml')))['evaluate']['data']
-evaluate_eval_params = yaml.safe_load(open(Path(__file__).with_name('params.yaml')))['evaluate']['eval']
+evaluate_data_params = yaml.safe_load(open(Path(__file__).with_name('params_tensor.yaml')))['evaluate']['data']
+evaluate_eval_params = yaml.safe_load(open(Path(__file__).with_name('params_tensor.yaml')))['evaluate']['eval']
+model_params = yaml.safe_load(open(Path(__file__).with_name('params_tensor.yaml')))['train']['model']
+
+sample_preprocessor = lambda samples: tuple([samples.reshape((*samples.shape, 1)), samples.reshape((*samples.shape, 1))])
 
 test_generator = SampleGenerator(
-    sample_dir = evaluate_data_params['test_dir'],
-    batch_size = evaluate_data_params['test_batch_size']
+	sample_dir = evaluate_data_params['test_dir'],
+	sample_preprocessor = sample_preprocessor,
+	batch_size = evaluate_data_params['test_batch_size'],
+	sample_type = np.float32,
 )
-test_generator.set_subsampling_functions(['singlets'])
-test_generator.construct_generator()
 
-autoencoder = CnnAutoencoder(
-    input_size = None,
-    embedding_size = None,
-    hidden_layers = None,
-    loss = None,
-    train_generator = None,
-    test_generator = test_generator,
-	save_dir = evaluate_data_params['model_dir'],
-    train_steps_per_epoch = None,
-    test_steps_per_epoch = None,
-    tf_callbacks = None
+autoencoder = ResnetAutoencoder(
+    train_generator=test_generator,
+	test_generator=test_generator,
+	train_steps_per_epoch=model_params['train_steps_per_epoch'],
+	test_steps_per_epoch=model_params['test_steps_per_epoch'],
+	save_dir=model_params['save_dir'],
+	output_to_board = tensor_to_board,
+	board_to_input = board_to_tensor,
+	loss = model_params['loss'],
+	tf_callbacks = model_params['tf_callbacks'] + [
+		keras.callbacks.EarlyStopping(monitor='val_loss', min_delta=0.03, patience=5, verbose=1, mode='min', restore_best_weights=True),
+		#keras.callbacks.TensorBoard(log_dir=model_params['save_dir'], histogram_freq=1, write_images=True, embeddings_freq=1)
+	]
 )
 
 autoencoder.load()
 examples = evaluate_eval_params['number_examples']
 examples_out = ""
 
-best_samples = autoencoder.get_best_samples(examples, 10*examples)
-for i in range(examples):
-    best_sample_bitboard = best_samples[i]['position'].reshape((773, -1))
-    best_sample_position = bitboard_to_board(best_sample_bitboard)
-    examples_out += str(best_samples[i]['loss']) + '\n\n'
-    examples_out += autoencoder.compare_sample_to_prediction(best_sample_bitboard)
+start_board = chess.Board()
+prediction = autoencoder.predict_from_board([start_board])
+prediction = prediction[0,:,:,:,0]
+prediction = autoencoder.binarize_array(prediction)
+print(prediction[:,:,0])
+print(prediction[:,:,1])
+print(tensor_to_board(prediction).__str__())
+
+# best_samples = autoencoder.get_best_samples(examples)
+# for i in range(examples):
+#     best_sample_input = best_samples[i]['input']
+#     examples_out += str(best_samples[i]['loss']) + '\n\n'
+#     examples_out += autoencoder.compare_input_to_prediction(best_sample_input)
 
-worst_samples = autoencoder.get_worst_samples(examples, 10)
-for i in range(examples):
-    worst_sample_bitboard = worst_samples[i]['position'].reshape((773, -1))
-    worst_sample_position = bitboard_to_board(worst_sample_bitboard)
-    examples_out += str(worst_samples[i]['loss']) + '\n'
-    examples_out += autoencoder.compare_sample_to_prediction(worst_sample_bitboard)
+# worst_samples = autoencoder.get_worst_samples(examples)
+# for i in range(examples):
+#     worst_sample_input = worst_samples[i]['input']
+#     examples_out += str(worst_samples[i]['loss']) + '\n'
+#     examples_out += autoencoder.compare_input_to_prediction(worst_sample_input)
 
-print(examples_out)
+#print(examples_out)
 
-with open(evaluate_eval_params['result_dir']+'/examples.out', 'w') as file:
-    file.write(examples_out)
+# with open(evaluate_eval_params['result_dir']+'/examples.out', 'w') as file:
+#     file.write(examples_out)
 
 
-with open(evaluate_eval_params['result_dir']+'/scores.json', 'w') as file:
-    json.dump(
-		{
-			'train_loss': autoencoder.train_history['loss'][-1],
-			'test_loss': autoencoder.train_history['val_loss'][-1]
-		},
-		file,
-		indent=2
-	)
+# with open(evaluate_eval_params['result_dir']+'/scores.json', 'w') as file:
+#     json.dump(
+# 		{
+# 			'train_loss': autoencoder.train_history['loss'][-1],
+# 			'test_loss': autoencoder.train_history['val_loss'][-1]
+# 		},
+# 		file,
+# 		indent=2
+# 	)
 
-with open(evaluate_eval_params['result_dir']+'/test_scores.json', 'w') as file:
-    json.dump(
-		{
-			'val_loss': [{
-				'epoch': i,
-				'train_loss': autoencoder.train_history['loss'][i],
-				'val_loss':  autoencoder.train_history['val_loss'][i]
-			} for i in range(len(autoencoder.train_history['val_loss']))]
-		},
-		file,
-		indent=2
-	)
+# with open(evaluate_eval_params['result_dir']+'/test_scores.json', 'w') as file:
+#     json.dump(
+# 		{
+# 			'val_loss': [{
+# 				'epoch': i,
+# 				'train_loss': autoencoder.train_history['loss'][i],
+# 				'val_loss':  autoencoder.train_history['val_loss'][i]
+# 			} for i in range(len(autoencoder.train_history['val_loss']))]
+# 		},
+# 		file,
+# 		indent=2
+# 	)

extract_positions.py

@@ -1,11 +1,12 @@
+from typing import Callable
 import h5py
 import yaml
 import logging
 import argparse
 from pathlib import Path
-from chesspos.preprocessing import (
-	PgnExtractor, get_game_processor, get_game_filter
-)
+import chess
+
+from chesspos.preprocessing import *
 
 if __name__ == "__main__":
 
@@ -16,11 +17,21 @@
 	args = parser.parse_args()
 	params = yaml.safe_load(open(Path(__file__).with_name(args.config)))['preprocess']
 
+	game_filter = None
+	if params['game_filter'] == 'custom':
+		def custom_filter(header: chess.pgn.Headers):
+			return filter_by_elo(header, white_elo_range=[2600,4000], black_elo_range=[2600,4000]) and \
+				filter_by_time_control(header, time_range=[5,60])
+		game_filter = custom_filter
+	else:
+		game_filter = get_game_filter(params['game_filter'])
+
+
 	position_extractor = PgnExtractor(
 		pgn_path=params['pgn_path'],
 		save_path=params['save_path'],
 		game_processor=get_game_processor(params['game_processor']),
-		game_filter=get_game_filter(params['game_filter']),
+		game_filter=game_filter,
 		chunk_size=params['chunk_size'],
 		log_level=logging.INFO
 	)