WillKoehrsen

history = model.fit(X_train,  y_train, 
                    batch_size=2048, epochs=150,
                    callbacks=callbacks,
                    validation_data=(X_valid, y_valid))

WillKoehrsen

from keras.callbacks import EarlyStopping, ModelCheckpoint

# Create callbacks
callbacks = [EarlyStopping(monitor='val_loss', patience=5),
             ModelCheckpoint('../models/model.h5'), save_best_only=True, 
                             save_weights_only=False)]

WillKoehrsen

# Load in embeddings
glove_vectors = '/home/ubuntu/.keras/datasets/glove.6B.100d.txt'
glove = np.loadtxt(glove_vectors, dtype='str', comments=None)

# Extract the vectors and words
vectors = glove[:, 1:].astype('float')
words = glove[:, 0]

# Create lookup of words to vectors
word_lookup = {word: vector for word, vector in zip(words, vectors)}

WillKoehrsen

from keras.models import Sequential
from keras.layers import LSTM, Dense, Dropout, Masking, Embedding

model = Sequential()

# Embedding layer
model.add(
    Embedding(input_dim=num_words,
              input_length = training_length,
              output_dim=100,

WillKoehrsen

features = []
labels = []

training_length = 50

# Iterate through the sequences of tokens
for seq in sequences:

    # Create multiple training examples from each sequence
    for i in range(training_length, len(seq)):

WillKoehrsen

from tpot import TPOTClassifier
from sklearn.model_selection import TimeSeriesSplit

# Cross validation object
tss = TimeSeriesSplit(n_splits = 3)

# Make the tpot search model
tpot_pipeline = TPOTClassifier(generations = 10, population_size = 10, 
                               cv = tss, scoring = 'f1')
# Find best model

WillKoehrsen

from sklearn.ensemble import RandomForestClassifier

# Create the model with 100 trees and default hyperparameters
model = RandomForestClassifier(n_estimators=100)

# Fit on training data
model.fit(train, train_labels)

# Make predictions on holdout test data
predictions = model.predict(test)

WillKoehrsen

import featuretools as ft

# Primitives for deep feature synthesis
trans_primitives = ['weekend', 'cum_sum', 'day', 'month', 'diff', 'time_since_previous']
agg_primitives = ['sum', 'time_since_last', 'avg_time_between', 'all', 'mode', 
                  'num_unique', 'min', 'last', 'mean', 'percent_true', 
                  'max', 'std', 'count']

# Perform deep feature synthesis 
feature_matrix, feature_names = ft.dfs(entityset=es, 

WillKoehrsen

import numpy as np
import random
random.seed(100)

def generate_batch(pairs, n_positive = 50, negative_ratio = 1.0):
    """Generate batches of samples for training. 
       Random select positive samples
       from pairs and randomly select negatives."""
    
    # Create empty array to hold batch

WillKoehrsen

__________________________________________________________________________________________________
Layer (type)                    Output Shape         Param #     Connected to                     
==================================================================================================
book (InputLayer)               (None, 1)            0                                            
__________________________________________________________________________________________________
link (InputLayer)               (None, 1)            0                                            
__________________________________________________________________________________________________
book_embedding (Embedding)      (None, 1, 50)        1851000     book[0][0]                       
__________________________________________________________________________________________________