Python utils.product() Examples

def make_onehot(feature, planes):
    onehot_features = np.zeros(feature.shape + (planes,), dtype=np.uint8)
    capped = np.minimum(feature, planes)
    onehot_index_offsets = np.arange(0, product(
        onehot_features.shape), planes) + capped.ravel()
    # A 0 is encoded as [0,0,0,0], not [1,0,0,0], so we'll
    # filter out any offsets that are a multiple of $planes
    # A 1 is encoded as [1,0,0,0], not [0,1,0,0], so subtract 1 from offsets
    nonzero_elements = (capped != 0).ravel()
    nonzero_index_offsets = onehot_index_offsets[nonzero_elements] - 1
    onehot_features.ravel()[nonzero_index_offsets] = 1
    return onehot_features 

def make_onehot(feature, planes):
    onehot_features = np.zeros(feature.shape + (planes,), dtype=np.uint8)
    capped = np.minimum(feature, planes)
    onehot_index_offsets = np.arange(0, product(
        onehot_features.shape), planes) + capped.ravel()
    # A 0 is encoded as [0,0,0,0], not [1,0,0,0], so we'll
    # filter out any offsets that are a multiple of $planes
    # A 1 is encoded as [1,0,0,0], not [0,1,0,0], so subtract 1 from offsets
    nonzero_elements = (capped != 0).ravel()
    nonzero_index_offsets = onehot_index_offsets[nonzero_elements] - 1
    onehot_features.ravel()[nonzero_index_offsets] = 1
    return onehot_features 

def make_onehot(feature, planes):
    onehot_features = np.zeros(feature.shape + (planes,), dtype=np.uint8)
    capped = np.minimum(feature, planes)
    onehot_index_offsets = np.arange(0, product(onehot_features.shape), planes) + capped.ravel()
    # A 0 is encoded as [0,0,0,0], not [1,0,0,0], so we'll
    # filter out any offsets that are a multiple of $planes
    # A 1 is encoded as [1,0,0,0], not [0,1,0,0], so subtract 1 from offsets
    nonzero_elements = (capped != 0).ravel()
    nonzero_index_offsets = onehot_index_offsets[nonzero_elements] - 1
    onehot_features.ravel()[nonzero_index_offsets] = 1
    return onehot_features 

def make_onehot(feature, planes):
    onehot_features = np.zeros(feature.shape + (planes,), dtype=np.uint8)
    capped = np.minimum(feature, planes)
    onehot_index_offsets = np.arange(0, product(onehot_features.shape), planes) + capped.ravel()
    # A 0 is encoded as [0,0,0,0], not [1,0,0,0], so we'll
    # filter out any offsets that are a multiple of $planes
    # A 1 is encoded as [1,0,0,0], not [0,1,0,0], so subtract 1 from offsets
    nonzero_elements = (capped != 0).ravel()
    nonzero_index_offsets = onehot_index_offsets[nonzero_elements] - 1
    onehot_features.ravel()[nonzero_index_offsets] = 1
    return onehot_features 

def make_onehot(feature, planes):
    onehot_features = np.zeros(feature.shape + (planes,), dtype=np.uint8)
    capped = np.minimum(feature, planes)
    onehot_index_offsets = np.arange(0, utils.product(
        onehot_features.shape), planes) + capped.ravel()
    # A 0 is encoded as [0,0,0,0], not [1,0,0,0], so we'll
    # filter out any offsets that are a multiple of $planes
    # A 1 is encoded as [1,0,0,0], not [0,1,0,0], so subtract 1 from offsets
    nonzero_elements = (capped != 0).ravel()
    nonzero_index_offsets = onehot_index_offsets[nonzero_elements] - 1
    onehot_features.ravel()[nonzero_index_offsets] = 1
    return onehot_features 

def few_liberties_feature(position):
    feature = position.get_liberties()
    onehot_features = np.zeros(feature.shape + (3,), dtype=np.uint8)
    onehot_index_offsets = np.arange(0, utils.product(
        onehot_features.shape), 3) + feature.ravel()
    nonzero_elements = ((feature != 0) & (feature <= 3)).ravel()
    nonzero_index_offsets = onehot_index_offsets[nonzero_elements] - 1
    onehot_features.ravel()[nonzero_index_offsets] = 1
    return onehot_features 

def make_onehot(feature, planes):
    onehot_features = np.zeros(feature.shape + (planes,), dtype=np.uint8)
    capped = np.minimum(feature, planes)
    onehot_index_offsets = np.arange(0, product(onehot_features.shape), planes) + capped.ravel()
    # A 0 is encoded as [0,0,0,0], not [1,0,0,0], so we'll
    # filter out any offsets that are a multiple of $planes
    # A 1 is encoded as [1,0,0,0], not [0,1,0,0], so subtract 1 from offsets
    nonzero_elements = (capped != 0).ravel()
    nonzero_index_offsets = onehot_index_offsets[nonzero_elements] - 1
    onehot_features.ravel()[nonzero_index_offsets] = 1
    return onehot_features 

def set_up_network(self):
        # a global_step variable allows epoch counts to persist through multiple training sessions
        global_step = tf.Variable(0, name="global_step", trainable=False)
        x = tf.placeholder(tf.float32, [None, go.N, go.N, self.num_input_planes])
        y = tf.placeholder(tf.float32, shape=[None, go.N ** 2])

        #convenience functions for initializing weights and biases
        def _weight_variable(shape, name):
            # If shape is [5, 5, 20, 32], then each of the 32 output planes
            # has 5 * 5 * 20 inputs.
            number_inputs_added = utils.product(shape[:-1])
            stddev = 1 / math.sqrt(number_inputs_added)
            # http://neuralnetworksanddeeplearning.com/chap3.html#weight_initialization
            return tf.Variable(tf.truncated_normal(shape, stddev=stddev), name=name)

        def _conv2d(x, W):
            return tf.nn.conv2d(x, W, strides=[1,1,1,1], padding="SAME")

        # initial conv layer is 5x5
        W_conv_init = _weight_variable([5, 5, self.num_input_planes, self.k], name="W_conv_init")
        h_conv_init = tf.nn.relu(_conv2d(x, W_conv_init), name="h_conv_init")

        # followed by a series of 3x3 conv layers
        W_conv_intermediate = []
        h_conv_intermediate = []
        _current_h_conv = h_conv_init
        for i in range(self.num_int_conv_layers):
            with tf.name_scope("layer"+str(i)):
                W_conv_intermediate.append(_weight_variable([3, 3, self.k, self.k], name="W_conv"))
                h_conv_intermediate.append(tf.nn.relu(_conv2d(_current_h_conv, W_conv_intermediate[-1]), name="h_conv"))
                _current_h_conv = h_conv_intermediate[-1]

        W_conv_final = _weight_variable([1, 1, self.k, 1], name="W_conv_final")
        b_conv_final = tf.Variable(tf.constant(0, shape=[go.N ** 2], dtype=tf.float32), name="b_conv_final")
        h_conv_final = _conv2d(h_conv_intermediate[-1], W_conv_final)

        logits = tf.reshape(h_conv_final, [-1, go.N ** 2]) + b_conv_final
        output = tf.nn.softmax(logits)

        log_likelihood_cost = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(logits=logits, labels=y))

        train_step = tf.train.AdamOptimizer(1e-4).minimize(log_likelihood_cost, global_step=global_step)
        was_correct = tf.equal(tf.argmax(logits, 1), tf.argmax(y, 1))
        accuracy = tf.reduce_mean(tf.cast(was_correct, tf.float32))

        weight_summaries = tf.summary.merge([
            tf.summary.histogram(weight_var.name, weight_var)
            for weight_var in [W_conv_init] +  W_conv_intermediate + [W_conv_final, b_conv_final]],
            name="weight_summaries"
        )
        activation_summaries = tf.summary.merge([
            tf.summary.histogram(act_var.name, act_var)
            for act_var in [h_conv_init] + h_conv_intermediate + [h_conv_final]],
            name="activation_summaries"
        )
        saver = tf.train.Saver()

        # save everything to self.
        for name, thing in locals().items():
            if not name.startswith('_'):
                setattr(self, name, thing)