extract.py

def extract_train_data(winner, replayName):
	with open('replays/' + replayName) as replay_file:
		game = json.load(replay_file)
	height = game['height']
	width = game['width']
	player_id = get_id(game, winner)
	print "Winner:", winner, "(", player_id, ")"
	num_frames = game['num_frames']
	dataX = []
	dataY = []
	weights = []
	frames = game['frames']
	moves = game['moves']
	for i, frame in enumerate(frames):
		if i == num_frames - 1 or i > limit_n_frames:
			break # No moves were made at this frame
		total_squares = 0 # How many squares we own
		for y, row in enumerate(frame):
			for x, square in enumerate(row):
				owner = square[0]
				if owner == player_id:
					total_squares += 1
		# Now construct the training data
		for y, row in enumerate(frame):
			for x, square in enumerate(row):
				owner = square[0]
				strength = square[1]
				move = moves[i][y][x]
				if owner == player_id:
					if move == 0:
						if random.randint(0,still_downsample) != 0:
							continue # We downsample
					data = np.zeros(shape=(4, look_area, look_area))
					found_enemy = False # True if we found at least one enemy 
					for oi, oX in enumerate(range(x - look_distance, x + look_distance + 1)):
						for oj, oY in enumerate(range(y - look_distance, y + look_distance + 1)):
							oX = (oX + width) % width
							oY = (oY + height) % height
							owner = frame[oY][oX][0]
							if owner != player_id:
								found_enemy = True
							strength = frame[oY][oX][1]
							production = game['productions'][oY][oX]
							data[0][oj][oi] = float(strength) / float(max_strength)
							data[1][oj][oi] = float(production) / float(max_production)
							data[2][oj][oi] = int(owner == player_id)
							data[3][oj][oi] = int(owner not in (0, player_id))
					if found_enemy:
						dataX.append(data)
						dataY.append(move)
						weights.append(float(1) / float(total_squares)) # TODO: This is incorrect. total_squares is not the right number
						#weights.append(1)
	return dataX, dataY, weights