# -*- coding: utf-8 -*-
__version__ = '2.5.1'
__author__ = "Avinash Kak (kak@purdue.edu)"
__date__ = '2024-September-30'
__url__ = 'https://engineering.purdue.edu/kak/distDLS/DLStudio-2.5.1.html'
__copyright__ = "(C) 2024 Avinash Kak. Python Software Foundation."
__doc__ = '''
You are looking at the DataPrediction module file in the DLStudio platform.
For the overall documentation on DLStudio, visit:
https://engineering.purdue.edu/kak/distDLS/
DATA PREDICTION:
Let's say you have a sequence of observations recorded at regular intervals.
These could, for example, be the price of a stock share recorded every hour;
the hourly recordings of electrical load at your local power utility company;
the mean average temperature recorded on an annual basis; and so on. We want
to use the past observations to predict the value of the next one. Solving
these types of problems is the focus of the DataPrediction co-class of
DLStudio.
As a problem, data prediction has much in common with text analytics and
seq2seq processing, in the sense that the prediction at the next time instant
must be based on all previous observations in a manner similar to what we do
in text analytics where the next word is understood taking into account all
the previous words in a sentence. However, there are three significant
differences between purely numerical data prediction problems and text-based
problems:
1) Data Normalization: As you know by this time, neural networks require that
your input data be normalized to the [0,1] interval, assuming it consists
of non-negative numbers, or the [-1,1] interval otherwise. This is
dictated by the design of the data processing components made available by
the commonly used platforms. For example, a commonly used activation
function, Sigmoid, saturates out outside the [0,1] interval; and the
activation function Tanh saturates out outside the [-1,1] interval. For
multi-channel inputs, with each channel being represented by a different
axis of the input tensor, such normalization is generally carried out
separately for each channel.
When solving a problem like text analytics, after you have normalized the
input data (which is likely to consist of the numeric embeddings for the
input words), you can forget about it. You don't have that luxury when
solving a data prediction problem. As you would expect, the next value
predicted by an algorithm must be at the same scale as the original input
input data. This requires that the output of a neural-network-based
prediction algorithm must be "inverse-normalized". And that, in turn,
requires remembering the normalization parameters used in each channel of
the input data.
2) Input Data Chunking: The notion of a sentence that is important in text
analytics does not carry over to the data prediction problem. In general,
you would want a prediction to be made using ALL of the past observations.
You are also likely to believe that the longer the time span over which the
past observations are available, the greater the likelihood that the
prediction made for the next time instant will be accurate. When the
sequential data available for training a predictor is arbitrarily long, as
is the case with numerical data in general, you would need to decide how to
"chunk" the data --- that is, how to extract sub-sequences from the data
for the purpose of training a neural network.
3) Data Conditioning: Data prediction problems give you data as a time series
--- meaning data values recorded at regular intervals. A datafile could be
as rudimentary as consisting of two comma separated columns, one containing
the "datetime" of the observation and other the actual value for what's
being observed. To illustrate what I am saying, let's consider the entries
in one of the data files for the utility power load prediction problem at:
https://www.kaggle.com/robikscube/hourly-energy-consumption
The dataset consists of over 10-years worth of hourly electric load
recordings made available by several utilities in the east and the midwest
of the United States. Shown below is the start of a datafile typical
datafile:
2013-12-31 01:00:00, 1861.0
2013-12-31 02:00:00, 1835.0
2013-12-31 03:00:00, 1841.0
2013-12-31 04:00:00, 1872.0
2013-12-31 05:00:00, 1934.0
2013-12-31 06:00:00, 1995.0
2013-12-31 07:00:00, 2101.0
... ...
... ...
As you can see, the datafile has two comma separated columns: the left
column is the datetime and the right the electrical load recorded in
megawatts at that time. As far as the data is concerned, obviously there's
not much to be done with the actual load observations in the right column
since those values are purely one-dimensional.
However, it is an entirely different matter with regard to the datetime in
the left column. You see, for the case of predicting the electric power
load at utility company, the time of the day matters, as does the day of
the week, what season it is (day of the year), etc. What that means that
it would make no sense to encode the datetime of the observations as a
one-dimensional entity. You would need to encode it so that it has multiple
degrees of freedom in a learning framework. In a GRU-based solution for
the electric load prediction problem by Gabriel Loye, the datetime is
encoded into four numbers as follows:
-- Hour of the day with an integer value between 0 and 23
-- Day of the week as an integer value between 1 and 7
-- Month of the year as an integer value between 1 and 12.
-- Day of the year as an integer value between 1 and 365
Data conditioning here would refer to extracting this four dimensional
information from the datetime column shown above. As it turns out, the
Pandas module comes in handy for that, as demonstrated by Gabriel Loye in
his code that you can download from:
https://blog.floydhub.com/gru-with-pytorch/
---------------------
Now that you understand how the data prediction problem differs from, say, the
problem of text analytics, it is time for me to state my main goal in defining
the DataPrediction class in this file. I actually have two goals:
(a) To deepen your understanding of a GRU. At this point, your understanding
of a GRU is likely to be based on calling PyTorch's GRU in your own code.
Using a pre-programmed implementation for a GRU makes life easy and you
also get a piece of highly optimized code that you can just call in your
own code. However, with a pre-programmed GRU, you are unlikely to get
insights into how such an RNN is actually implemented.
(b) To demonstrate how you can use a Recurrent Neural Network (RNN) for data
prediction taking into account the data normalization, chunking, and
conditioning issues mentioned earlier.
To address the first goal above, the DataPrediction class presented in this
file is based on my pmGRU (Poor Man's GRU). This GRU is my implementation of
the "Minimal Gated Unit" GRU variant that was first presented by Joel Heck and
Fathi Salem in their paper "Simplified Minimal Gated Unit Variations for
Recurrent Neural Networks". Its hallmark is that it combines the Update and
the Reset gates of a regular GRU into a single gate called the Forget Gate.
You could say that pmGRU is a lightweight version of a regular GRU and its use
may therefore lead to a slight loss of accuracy in the predictions. As to
whether that loss would be acceptable would depend obviously on the
application context and the needs of the customer. You will find it
educational to compare the performance you get with my pmGRU-based
implementation with Gabriel Loye's implementation that uses PyTorch's GRU for
the same dataset. And after you have seen my implementation of pmGRU, you
should yourself be able to translate the regular GRU gating equations into
your own implementation --- just as a learning exercise.
For measuring the accuracy of his data prediction networks, Gabriel Loye used
the Symmetric Mean Absolute Percentage Error (sMAPE) that is the sum of the
absolute difference between the predicted and the actual values divided by the
average of the predicted and the actual values:
100 | P_t - GT_t |
sMAPE = ----- \SUM_{t=1}^N ---------------------------
N | P_t + GT_t | / 2.0
where N is the number of predictions made in a file. P_t is the value
predicted by the network at inference time and GT_t is the ground-truth value
that actually occurred at that time. I have used the same metric for
estimating the accuracy of the prediction results produced by pmGRU.
@endofdocs
'''
from DLStudio import DLStudio
import os
import time
import numpy as np
import pandas as pd
import matplotlib.pyplot as plt
import torch
import torch.nn as nn
from torch.utils.data import TensorDataset, DataLoader
from sklearn.preprocessing import MinMaxScaler
import random
import sys
class DataPrediction(object):
def __init__(self, *args, **kwargs ):
if args:
raise ValueError(
'''DataPrediction constructor can only be called with keyword arguments for
the following keywords: dataroot, hidden_size, input_size, output_size, sequence_length,
path_saved_model, use_gpu, ngpu, dlstudio, device, and debug''')
allowed_keys = 'dataroot','path_saved_model','hidden_size','input_size','output_size','sequence_length',\
'use_gpu','ngpu','dlstudio','debug'
keywords_used = kwargs.keys()
for keyword in keywords_used:
if keyword not in allowed_keys:
raise SyntaxError(keyword + ": Wrong keyword used --- check spelling")
learning_rate=input_size=hidden_size=output_size=sequence_length=dataroot=path_saved_model=use_gpu=ngpu=debug=None
if 'ngpu' in kwargs : ngpu = kwargs.pop('ngpu')
if 'dlstudio' in kwargs : dlstudio = kwargs.pop('dlstudio')
if 'debug' in kwargs : debug = kwargs.pop('debug')
if 'input_size' in kwargs : input_size = kwargs.pop('input_size')
if 'hidden_size' in kwargs : hidden_size = kwargs.pop('hidden_size')
if 'output_size' in kwargs : output_size = kwargs.pop('output_size')
if 'sequence_length' in kwargs : sequence_length = kwargs.pop('sequence_length')
if input_size:
self.input_size = input_size
if hidden_size:
self.hidden_size = hidden_size
if output_size:
self.output_size = output_size
if sequence_length:
self.sequence_length = sequence_length
if ngpu:
self.ngpu = ngpu
if dlstudio:
self.dls = dlstudio
self.training_sequences = [] ## Array of sequences extracted from the sequential observations
self.training_predictions = [] ## The corresponding predictions for each of the training sequences
self.test_sequences = {} ## We store the test sequences in dicts whose key values are the data files
self.test_gt_predictions = {} ## 'gt' stands for 'ground truth'. The predictions for the test sequences are
## are similary stored in dicts
self.predictions_scaling_params = {} ## This dict stores the scaling params for the test sequences extracted from
## a file. These params are used to inverse-scale the final predictions
if debug:
self.debug = debug
self.device = torch.device("cuda:0" if (torch.cuda.is_available() and ngpu > 0) else "cpu")
class pmGRU(nn.Module):
"""
This GRU implementation is based primarily on a "Minimal Gated" version of a GRU as described in
"Simplified Minimal Gated Unit Variations for Recurrent Neural Networks" by Joel Heck and Fathi
Salem. The Wikipedia page on "Gated_recurrent_unit" has a summary presentation of the equations
proposed by Heck and Salem.
"""
def __init__(self, predictor):
super(DataPrediction.pmGRU, self).__init__()
self.input_size = predictor.input_size
self.hidden_size = predictor.hidden_size
self.output_size = predictor.output_size
self.batch_size = predictor.dls.batch_size
## for forget gate:
self.project1 = nn.Sequential( nn.Linear(self.input_size + self.hidden_size, self.hidden_size), nn.Sigmoid() )
## for interim out:
self.project2 = nn.Sequential( nn.Linear( self.input_size + self.hidden_size, self.hidden_size), nn.Tanh() )
## for final out
self.project3 = nn.Sequential( nn.Linear( self.hidden_size, self.output_size ), nn.Tanh() )
def forward(self, x, h, sequence_end=False):
combined1 = torch.cat((x, h), 1)
forget_gate = self.project1(combined1)
interim = forget_gate * h
combined2 = torch.cat((x, interim), 1)
output_interim = self.project2( combined2 )
output = (1 - forget_gate) * h + forget_gate * output_interim
if sequence_end == False:
return output, output
else:
final_out = self.project3(output)
return final_out, final_out
def init_hidden(self):
weight = next(self.parameters()).data
hidden = weight.new(self.batch_size, self.hidden_size).zero_()
return hidden
def construct_dataframes_from_datafiles(self):
"""
The goal is to store contents of each CSV file in a Pandas DataFrame. A DataFrame is a two-dimensional
data structure with rows and columns, like an Excel spreadsheet. Each column is represented by a separate
Pandas Series object. A DataFrame object has two axes: “Axis 0” and “Axis 1”. “Axis 0” points in the
direction of increasing row-index values and "Axis 1" points in the direction of increasing column index.
"""
dataframes = {}
datafiles = os.listdir( self.dls.dataroot )
print("\n\n[construct dataframes] datafiles: ", datafiles)
print("\n[construct dataframes] file header and a few rows: ")
print(pd.read_csv(self.dls.dataroot + datafiles[0]).head())
for file in datafiles:
## Setting 'parse_dates' to '[0]' tells Pandas that datetime is in column indexed 0.
df = pd.read_csv(self.dls.dataroot + file, parse_dates=[0], encoding='utf-8')
print("\n[construct dataframes] first 10 rows of the df dataframe: ")
print(df[:10])
## The following four statements create additional columns in the tabular data stored the DataFrame
## instance "df". The first statement adds a column with its header set to "hour". The values in this
## column will be set to an integer between 0 and 23 depending on the hour at which the electrical
## load was measured. And so on. I have taken these statements from the code by Gabriel Loye.
df['hour'] = df.apply(lambda x: x['Datetime'].hour, axis=1)
df['dayofweek'] = df.apply(lambda x: x['Datetime'].dayofweek, axis=1)
df['month'] = df.apply(lambda x: x['Datetime'].month, axis=1)
df['dayofyear'] = df.apply(lambda x: x['Datetime'].dayofyear, axis=1)
## Now sort the rows based on the values in the 'Datetime' column and then drop the 'Datetime' column:
df = df.sort_values("Datetime").drop("Datetime", axis=1)
print("\n[construct dataframes] Shape of the DataFrame object for file %s: " % file, df.shape) ## like: (32896, 5)
dataframes[file] = df
return dataframes
def data_normalizer(self, dataframes):
"""
For neural-network based processing, we want to normalize (meaning to scale) each column of the data in each
DataFrame object to the [0,1] interval. The data normalization step changes the value x IN EACH CHANNEL of
the input data into
x - x_min
--------------------
x_max - x_min
where x_min is the minimum and x_max the maximum of all the values IN THAT CHANNEL. So, logically speaking,
the data normalization is applied SEPARATELY to each channel. In our case, each channel of the input corresponds
to each column of the DataFrame object. This is most easily done first constructing an instance of MinMaxScaler
from the "sklearn.preprocessing" module as shown in Line (A) below and calling on its "fit_transform()" method
as shown in Line (C). Using this functionality of "sklearn.preprocessing" saves you from having to write your
own code that would calculate the x_min and x_max values for each column of a DataFrame object.
The data normalization as explained above implies that if we wanted to reverse the scaling for the column
that is being subject to data prediction, the x_min and x_max values for that column must be remembered for
its use at the and of the learning process. These values are returned by the "fit()" method of a MinMaxScaler
instance, as shown in Line (D). The MinMaxScaler instance constructed in Line (B) is for extracting the x_min
and the x_max values for the main observation column that is focus of predictions.
To illustrate the normalization step with an example, consider the following 3-column (the same thing as
3-channel) input data:
X = [ [1.0 -1.0 2.0]
[2.0 0. 0. ]
[0. 1.0 -1.] ]
The min values in each column are:
X.min(axis=0) = [0.0 -1.0 -1.]
Subtracting the values in each column by its min will turn all values into POSITIVE NUMBERS:
X - X.min(axis=0) = [ [1. 0.0 3.0]
[2. 1.0 1.0]
[0. 2.0 0.0] ]
Next, we need to apply the scaling multiplier to these numbers. For each column, the scaling multiplier
is
1.0
X_std = ---------------------------------
X.max(axis=0) - X.min(axis=0)
= [ 0.5 , 0.5 , 0.33333333]
The multiplication
( X - X.min(axis=0) ) * X_std
gives a normalized version of the data with each of its columns separately mapped to the [0,1] interval.
The explanation given so far in this comment block applies to the role of the data_scaler object defined
Line (A) below. About the role of a similar data scaler object, gt_predictions_scaler, in Line (B),
that is for the purpose of remembering the x_min and x_max values for that column of the input data
that is subject to predictions. As already stated, we need those values for inverting the data scaling
step at the output of the prediction network.
"""
dataframes_normalized = {}
for file in dataframes:
df = dataframes[file]
data_scaler = MinMaxScaler() ## (A)
gt_predictions_scaler = MinMaxScaler() ## (B)
data = data_scaler.fit_transform(df.values) ## (C)
gt_predictions_scaler.fit(df.iloc[:,0].values.reshape(-1,1)) ## (D)
## save the MinMaxScaler instance
self.predictions_scaling_params[file] = gt_predictions_scaler
dataframes_normalized[file] = data
return dataframes_normalized
def construct_sequences_from_data(self, dataframes):
"""
The purpose of this method is to construct sequences from the electrical-load data in the DataFrame objects
created for each file by the method "construct_dataframes_from_datafiles()". The length of the sequences
extracted on a running-basis from the data is set by the constructor parameter "sequence_length".
"""
for file in dataframes:
num_records_in_file = len(dataframes[file])
sequence_length = self.sequence_length
## Note that the maximum number of sequences that can be extracted from a file is the number of observations
## in the file minus the sequence length. If there are, say, 1000 observations in a file and the sequence
## length is 50, you will be able to construct a maximum of 950 sequences from that file. The shape of each
## sequence will also depend on how the datetime for each observation was encoded. As currently programmed
## in the method "construct_dataframes_from_datafiles()", the datetime is encoded into 4 values. Therefore,
## for the case of 1000 observations in a file and with sequence length set to 50, the shape of the tensor
## that stores all the sequences will be "950 x 5".
sequences_from_file = np.zeros((num_records_in_file - sequence_length, sequence_length, dataframes[file].shape[1]))
predictions_from_file = np.zeros(num_records_in_file - sequence_length)
predictions_from_file = predictions_from_file.reshape(-1, 1)
for i in range(sequence_length, num_records_in_file):
sequences_from_file[i - sequence_length] = dataframes[file][i-sequence_length : i]
predictions_from_file[i - sequence_length] = dataframes[file][i,0]
## We save 10% of the sequences extracted from each data file for the final evaluation of the prediction model:
test_portion = int(0.1*len(sequences_from_file))
if len(self.training_sequences) == 0:
self.training_sequences = sequences_from_file[:-test_portion]
self.training_predictions = predictions_from_file[:-test_portion]
else:
self.training_sequences = np.concatenate((self.training_sequences, sequences_from_file[:-test_portion]))
self.training_predictions = np.concatenate((self.training_predictions, predictions_from_file[:-test_portion]))
## We store the test sequences and their associated labels in the two dicts shown below that are keyed
## to the file names.
self.test_sequences[file] = (sequences_from_file[-test_portion:])
self.test_gt_predictions[file] = (predictions_from_file[-test_portion:])
def set_dataloader(self):
train_data = TensorDataset(torch.from_numpy(self.training_sequences), torch.from_numpy(self.training_predictions))
train_loader = DataLoader(train_data, shuffle=True, batch_size=self.dls.batch_size, drop_last=True)
return train_loader
def run_code_for_training_data_predictor(self, train_loader, model):
"""
The 'model' parameter of the method is set to an instance of pmGRU. See the following script
power_load_prediction_with_pmGRU.py
in the ExamplesDataPrediction directory of the DLStudio distribution for how to construct an instance
of pmGRU and how to then supply that instance to this 'run_code_for_training' method.
"""
device = self.dls.device
model.to(device)
learning_rate = self.dls.learning_rate
criterion = nn.MSELoss()
optimizer = torch.optim.Adam(model.parameters(), lr=learning_rate)
print("\n\nStarting Training")
start_time = time.perf_counter()
epochs = self.dls.epochs
loss_tally = []
for epoch in range(epochs):
avg_loss = 0.
running_loss = 0.0
for counter,data in enumerate(train_loader):
x,pred = data
optimizer.zero_grad()
## Shape of x: (batch_size, sequence_length, input_size)
## For power-load dataset and for the constructor params chosen, the shape of x is (1024,90,5)
pred = pred.to(device).float()
h = model.init_hidden().data.to(device)
## The following loop steps through each element of the input sequence, one element at a
## time. Note that x.shape[1] is the sequence_length, which is 90 as set by the constructor
## in the script in the ExamplesDataPrediction directory.
for load_sample_index in range(x.shape[1]):
input_sample = x[:,load_sample_index,:]
input_sample = torch.squeeze(input_sample)
input_sample = input_sample.to(device).float()
if load_sample_index < x.shape[1] - 1:
out, h = model(input_sample, h)
elif load_sample_index == x.shape[1] - 1:
out, h = model(input_sample, h, sequence_end=True)
loss = criterion(out, pred)
loss.backward()
optimizer.step()
running_loss += loss.item()
if counter%200 == 199:
current_time = time.perf_counter()
elapsed_time = current_time - start_time
avg_loss = running_loss / float(200)
loss_tally.append(avg_loss)
print("\n[epoch:%d/%d iter=%4d elapsed_time=%3d secs] Avg Loss: %.8f" % (epoch+1, epochs, counter+1, elapsed_time, avg_loss))
running_loss = 0.0
self.save_model(model)
print("\nFinished Training\n")
plt.figure(figsize=(10,5))
plt.title("Loss")
plt.plot(loss_tally)
plt.xlabel("Training iterations over 5 epochs")
plt.ylabel("loss")
plt.legend()
plt.savefig("loss_vs_iterations_predictor_training.png")
plt.show()
return model
def run_code_for_evaluation_on_unseen_data(self, model):
"""
We now evaluate the quality of the pmGRU based prediction model leanred from the training data. The
evaluation is carried out on the UNSEEN test data. Recall that when the normalized load data in each
file was digested by the method "construct_sequences_from_data()", we saved 10% of the sequences
extracted from each data file for the final evaluation of the prediction model.
"""
hidden_size = self.hidden_size
device = self.dls.device
with torch.no_grad():
model.eval()
model.to(device)
all_predictions = [] ## For storing all unseen-data predictions made by the network
all_gt_predictions = [] ## For storing the corresponding ground-truth predictions
for file in sorted(self.test_sequences):
start_time = time.perf_counter()
print("\n\n\n[eval] file: ", file)
## We access the MinMaxScaler instance for "unscaling" the predictions:
scaler = self.predictions_scaling_params[file]
minval, maxval = torch.zeros(1,1), torch.zeros(1,1)
minval[0,0], maxval[0,0] = scaler.data_min_[0], scaler.data_max_[0]
minval, maxval = minval.to(device), maxval.to(device)
testdata = TensorDataset(torch.from_numpy(self.test_sequences[file]),
torch.from_numpy(self.test_gt_predictions[file]))
testdata_loader = DataLoader(testdata, batch_size=self.dls.batch_size, drop_last=True)
for x, gt_predictions in testdata_loader: ## x is of shape (batch_size, sequence_length, input_size)
gt_predictions = gt_predictions.to(device)
gt_predictions_unscaled = ( gt_predictions*(maxval - minval) + minval ).reshape(-1)
all_gt_predictions += gt_predictions_unscaled.tolist()
h = model.init_hidden().data.to(device)
for load_sample_index in range(x.shape[1]):
input_sample = x[:,load_sample_index,:]
input_sample = torch.squeeze(input_sample)
if load_sample_index < x.shape[1] - 1:
out, h = model(input_sample.to(device).float(), h)
elif load_sample_index == x.shape[1] - 1:
out, h = model(input_sample.to(device).float(), h, sequence_end=True)
out_unscaled = ( out*(maxval - minval) + minval ).reshape(-1)
all_predictions += out_unscaled.tolist()
current_time = time.perf_counter()
evaluation_time = current_time - start_time
print("evaluation time taken: ", evaluation_time)
print("\ntotal number of predictions made by the network for evaluation: ", len(all_predictions))
print("\ntotal number of all ground-truth predictions: ", len(all_gt_predictions))
print("\nfirst 100 predictions made by network: ", all_predictions[:100])
print("\nfirst 100 ground-truth predictions: ", all_gt_predictions[:100])
"""
Following Gabriel Loye, I'll use Symmetric Mean Absolute Percentage Error (sMAPE) to evaluate prediction
accuracy. sMAPE is "the sum of the absolute difference between the predicted and actual values divided
by the average of the predicted and actual value, therefore giving a percentage measuring the amount of error."
Here is the formula for sMAPE:
100 | P_t - GT_t |
sMAPE = ----- \SUM_{t=1}^n ---------------------------
n | P_t + GT_t | / 2.0
where n is the number of predictions made in a file. Note P_t is the value predicted by the network at
inference time and GT_t is the ground-truth load value that actually occurred at that time.
"""
sMAPE = 0
for i in range(len(all_predictions)):
sMAPE += np.mean( abs(all_predictions[i] - all_gt_predictions[i]) / (all_gt_predictions[i] + all_predictions[i])/2)/len(all_predictions)
print("\n\nsMAPE: %.8f" % (sMAPE*100))
return all_predictions, all_gt_predictions
def display_sample_predictions(self, predictions, predictions_gt):
"""
The two arguments 'predictions' and the corresponding 'predictions_gt' are for solely the unseen test data.
"""
plt.figure(figsize=(14,10))
plt.subplot(2,2,1)
plt.plot(predictions[:100], "-o", color="g", label="Predicted")
plt.plot(predictions_gt[:100], color="b", label="Actual")
plt.ylabel('Energy Consumption (MW)')
plt.legend()
plt.subplot(2,2,2)
plt.plot(predictions[30000-50:30000], "-o", color="g", label="Predicted")
plt.plot(predictions_gt[30000-50:30000], color="b", label="Actual")
plt.ylabel('Energy Consumption (MW)')
plt.legend()
plt.subplot(2,2,3)
plt.plot(predictions[60000-50:60000], "-o", color="g", label="Predicted")
plt.plot(predictions_gt[60000-50:60000], color="b", label="Actual")
plt.ylabel('Energy Consumption (MW)')
plt.legend()
plt.subplot(2,2,4)
plt.plot(predictions[90000-50:90000], "-o", color="g", label="Predicted")
plt.plot(predictions_gt[90000-50:90000], color="b", label="Actual")
plt.ylabel('Energy Consumption (MW)')
plt.legend()
plt.show()
def save_model(self, model):
'''
Save the trained prediction network to a disk file
'''
torch.save(model.state_dict(), self.dls.path_saved_model)
#_________________________ End of DataPrediciton Class Definition ___________________________
#_________________________________ Test code follows ___________________________________
if __name__ == '__main__':
pass