constants.py

import re
import os

CARDS_IN_DECK = 52
NUM_SUITS = 4
NUM_RANKS = 13
NUM_PLAYERS = 4
PLAYS_PER_TRICK = 4

# 10 features columns. The probabilities for each of the 4 players,
# plus 6 more: legal plays, high card in trick, and point value.
# Note that legal plays is a vector we must extract in the model.
# We place it as the first column, for convenience.
INPUT_FEATURES = 10

TOTAL_SCALAR_FEATURES = CARDS_IN_DECK*INPUT_FEATURES

DECK_SHAPE = (CARDS_IN_DECK,)

MAIN_INPUT_SHAPE = (CARDS_IN_DECK,INPUT_FEATURES)

SUITS_RANKS_SHAPE = (NUM_SUITS, NUM_RANKS)

SCORES_SHAPE = DECK_SHAPE
WIN_TRICK_PROBS_SHAPE = DECK_SHAPE
MOON_CLASSES = 3
MOONPROBS_SHAPE = (CARDS_IN_DECK, MOON_CLASSES)

number_re = re.compile(r'^(\d+)([KM]?)$')

# TODO: Add a real command line interface instead of this hack using environment variables
# https://www.pivotaltracker.com/story/show/154507407
default_vals = {
    'DECK_EPOCHS': 100,   # This is really "EVALS", not "EPOCHS", since we typically evalutate multiple times per epoch
    'DECK_BATCH': 1*1024, # Batch size should start low, as we will automically double it each true epoch
    'DECK_STEPS': 64,     # batch steps per evaluation checkpoint
    'DECK_MAX_BATCH': 1024*1024,

    # These three are really booleans. Use 0 for False, 1 for True
    'DECK_SCORE': 1,
    'DECK_MOON': 1,
    'DECK_TRICK': 1,

     # The number of suits we retain redundantly in the combined layer. 0, 2 and 4 are the only reasonable inupts.
     # 0 means that the CNN is the only part of the NN that sees the "distribution" data (4 suits by 7 features)
     # 2 means that the combined NN redundantly sees the two point suits (spades and hearts)
     # 4 means that the combined NN redundantly sees all four suits as direct input
    'DECK_REDUN': 4
}

def env_val(name) :
    if name in os.environ:
        s_val = os.environ[name]
        m = number_re.match(s_val)
        if not m:
            raise ValueError('Env var {} must be integer with optional K or M suffix, given: {}'.format(name, s_val))
        val = int(m.group(1))
        if m.group(2) == 'K':
            val *= 1024
        elif m.group(2) == 'M':
            val *= 1024*1024
        print('Set {} to {}'.format(name, val))
    else:
        val = default_vals[name]
    return val

BATCH = env_val('DECK_BATCH')
MAX_BATCH = env_val('DECK_MAX_BATCH')
STEPS = env_val('DECK_STEPS')
EPOCHS = env_val('DECK_EPOCHS')
SCORE = env_val('DECK_SCORE') == 1
MOON = env_val('DECK_MOON') == 1
TRICK = env_val('DECK_TRICK') == 1

MAIN_DATA = 'main_data'
EXPECTED_SCORE = 'expected_score'
WIN_TRICK_PROB = 'win_trick_prob'
MOON_PROB = 'moon_prob'

TRAINING = 'training'
VALIDATION = 'validation'

# When we use score predictions, the range should be 0..26.0
# But for least squares regression, we use a range of 0..2.0
# This currently requires that we scale the data at specific points in the pipeline
PREDICTION_SCORE_MAX = 26.0
MODEL_SCORE_MAX = 2.0