ntakouris

def lr_scheduler(epoch, lr, warmup_epochs=15, decay_epochs=100, initial_lr=1e-6, base_lr=1e-3, min_lr=5e-5):
    if epoch <= warmup_epochs:
        pct = epoch / warmup_epochs
        return ((base_lr - initial_lr) * pct) + initial_lr

    if epoch > warmup_epochs and epoch < warmup_epochs+decay_epochs:
        pct = 1 - ((epoch - warmup_epochs) / decay_epochs)
        return ((base_lr - min_lr) * pct) + min_lr

    return min_lr

ntakouris

class ModelTrunk(keras.Model):
      def __init__(self, name='ModelTrunk', time2vec_dim=1, num_heads=2, head_size=128, ff_dim=None, num_layers=1, dropout=0, **kwargs):
        super().__init__(name=name, **kwargs)
        self.time2vec = Time2Vec(kernel_size=time2vec_dim)
        if ff_dim is None:
            ff_dim = head_size
        self.dropout = dropout
        self.attention_layers = [AttentionBlock(num_heads=num_heads, head_size=head_size, ff_dim=ff_dim, dropout=dropout) for _ in range(num_layers)]

        

ntakouris

from tensorflow_addons.layers import MultiHeadAttention

class AttentionBlock(keras.Model):
    def __init__(self, name='AttentionBlock', num_heads=2, head_size=128, ff_dim=None, dropout=0, **kwargs):
        super().__init__(name=name, **kwargs)

        if ff_dim is None:
            ff_dim = head_size

        self.attention = MultiHeadAttention(num_heads=num_heads, head_size=head_size, dropout=dropout)

ntakouris

class Time2Vec(keras.layers.Layer):
    def __init__(self, kernel_size=1):
        super(Time2Vec, self).__init__(trainable=True, name='Time2VecLayer')
        self.k = kernel_size
    
    def build(self, input_shape):
        # trend
        self.wb = self.add_weight(name='wb',shape=(input_shape[1],),initializer='uniform',trainable=True)
        self.bb = self.add_weight(name='bb',shape=(input_shape[1],),initializer='uniform',trainable=True)
        # periodic

ntakouris

    @staticmethod
    def gh_jacobian(y, x):
      # x must be a vector
      # y must be matrix with dimensions [batch_size, n], n > 1. if n == 1 use the _vec version
      J = tf.map_fn(lambda m: tf.gradients(y[:,m:m+1], x)[0], tf.range(tf.shape(y)[-1]), tf.float32)
      #J = tf.Print(J, [tf.shape(J)], "J shape = ", summarize=-1)
      J = tf.transpose(tf.squeeze(J), perm = [1,0,2])
      return J

    @staticmethod

ntakouris

        with tf.gfile.GFile(modelfilename, 'rb') as f:
            graph_def = tf.GraphDef()
            graph_def.ParseFromString(f.read())
            self.graph_def = graph_def

        with tf.Graph().as_default() as g1:
            #gi_name = model_name+'/'+gen_input
            gi = tf.placeholder(tf.float32,
                              [None, self.z_dim],
                              name='g1_gi')

ntakouris

DIR="$( cd "$( dirname "${BASH_SOURCE[0]}" )" >/dev/null 2>&1 && pwd )"

docker run -d -v $DIR/grafana:/var/lib/grafana --name=grafana -p 3000:3000 grafana/grafana 

ntakouris

ntakouris

IntelliJ IDEAPyCharm   
from typing import Dict, Text

import tensorflow as tf
from absl import logging
from tensorflow.keras.layers import (LSTM, Activation, Concatenate, Dense)
import kerastuner
from rnn.constants import (INPUT_FEATURE_KEYS, PREDICT_FEATURE_KEYS,
                           HP_HIDDEN_LATENT_DIM,
                           HP_HIDDEN_LAYER_NUM, HP_LR,

ntakouris

from typing import Tuple, Text, Dict
import tensorflow as tf
from tensorflow.keras.layers import Input, Concatenate, Reshape, Dense
from input_fn_utils import transformed_name


def get_input_graph(input_feature_keys, input_window_size) -> Tuple[Input, tf.keras.layers.Layer]:
    """ 
        Creates the named input layers, strips the column names and provides
        them as a plain tensor.