Time series prediction, end-to-end

This notebook illustrates several models to find the next value of a time-series:

  1. Linear
  2. DNN
  3. CNN
  4. RNN
In [ ]:
# You must update BUCKET, PROJECT, and REGION to proceed with the lab
BUCKET = 'cloud-training-demos-ml'
PROJECT = 'cloud-training-demos'
REGION = 'us-central1'
SEQ_LEN = 50
In [ ]:
import os
os.environ['BUCKET'] = BUCKET
os.environ['PROJECT'] = PROJECT
os.environ['REGION'] = REGION
os.environ['SEQ_LEN'] = str(SEQ_LEN)
os.environ['TFVERSION'] = '1.15'

Simulate some time-series data

Essentially a set of sinusoids with random amplitudes and frequencies.

In [ ]:
import warnings

import tensorflow as tf
In [ ]:
import numpy as np
import seaborn as sns

def create_time_series():
  freq = (np.random.random()*0.5) + 0.1  # 0.1 to 0.6
  ampl = np.random.random() + 0.5  # 0.5 to 1.5
  noise = [np.random.random()*0.3 for i in range(SEQ_LEN)] # -0.3 to +0.3 uniformly distributed
  x = np.sin(np.arange(0,SEQ_LEN) * freq) * ampl + noise
  return x

flatui = ["#9b59b6", "#3498db", "#95a5a6", "#e74c3c", "#34495e", "#2ecc71"]
for i in range(0, 5):
  sns.tsplot( create_time_series(), color=flatui[i%len(flatui)] );  # 5 series
In [ ]:
def to_csv(filename, N):
  with open(filename, 'w') as ofp:
    for lineno in range(0, N):
      seq = create_time_series()
      line = ",".join(map(str, seq))
      ofp.write(line + '\n')

import os
except OSError:

np.random.seed(1) # makes data generation reproducible

to_csv('data/sines/train-1.csv', 1000)  # 1000 sequences
to_csv('data/sines/valid-1.csv', 250)
In [ ]:
!head -5 data/sines/*-1.csv

Train model locally

Make sure the code works as intended.

Please remember to update the "--model=" variable on the last line of the command

You may ignore any tensorflow deprecation warnings.

Note: This step will be complete when you see a message similar to the following: "INFO : tensorflow :Loss for final step: N.NNN...N"

In [ ]:
rm -rf $OUTDIR
gcloud ai-platform local train \
   --module-name=sinemodel.task \
   --package-path=${PWD}/sinemodel \
   -- \
   --train_data_path="${DATADIR}/train-1.csv" \
   --eval_data_path="${DATADIR}/valid-1.csv"  \
   --output_dir=${OUTDIR} \
   --model=linear --train_steps=10 --sequence_length=$SEQ_LEN

Cloud AI Platform

Now to train on Cloud AI Platform with more data.

In [ ]:
import shutil
shutil.rmtree('data/sines', ignore_errors=True)
np.random.seed(1) # makes data generation reproducible
for i in range(0,10):
  to_csv('data/sines/train-{}.csv'.format(i), 1000)  # 1000 sequences
  to_csv('data/sines/valid-{}.csv'.format(i), 250)
In [ ]:
gsutil -m rm -rf gs://${BUCKET}/sines/*
gsutil -m cp data/sines/*.csv gs://${BUCKET}/sines
In [ ]:
for MODEL in linear dnn cnn rnn rnn2 rnnN; do
  JOBNAME=sines_${MODEL}_$(date -u +%y%m%d_%H%M%S)
  gsutil -m rm -rf $OUTDIR
  gcloud ai-platform jobs submit training $JOBNAME \
     --region=$REGION \
     --module-name=sinemodel.task \
     --package-path=${PWD}/sinemodel \
     --job-dir=$OUTDIR \
     --scale-tier=BASIC \
     --runtime-version=$TFVERSION \
     -- \
     --train_data_path="gs://${BUCKET}/sines/train*.csv" \
     --eval_data_path="gs://${BUCKET}/sines/valid*.csv"  \
     --output_dir=$OUTDIR \
     --train_steps=3000 --sequence_length=$SEQ_LEN --model=$MODEL

Monitor training with TensorBoard

To activate TensorBoard within the JupyterLab UI navigate to "File" - "New Launcher". Then double-click the 'Tensorboard' icon on the bottom row.

TensorBoard 1 will appear in the new tab. Navigate through the three tabs to see the active TensorBoard. The 'Graphs' and 'Projector' tabs offer very interesting information including the ability to replay the tests.

You may close the TensorBoard tab when you are finished exploring.


When I ran it, these were the RMSEs that I got for different models. Your results will vary:

Model Sequence length # of steps Minutes RMSE
linear 50 3000 10 min 0.150
dnn 50 3000 10 min 0.101
cnn 50 3000 10 min 0.105
rnn 50 3000 11 min 0.100
rnn2 50 3000 14 min 0.105
rnnN 50 3000 15 min 0.097


You can see there is a significant improvement when switching from the linear model to non-linear models. But within the the non-linear models (DNN/CNN/RNN) performance for all is pretty similar.

Perhaps it's because this is too simple of a problem to require advanced deep learning models. In the next lab we'll deal with a problem where an RNN is more appropriate.

Copyright 2017 Google Inc. Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with the License. You may obtain a copy of the License at http://www.apache.org/licenses/LICENSE-2.0 Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License