train.py

#The main training function
#For event-based scenario generation and spatial scenario generation, implement the code with labels or 
#reshape the imput samples to spatio-temporal samples respectively.
#16 is the maximum value for wind capacity we use. Change to the customized max value for normalized data
#import ipdb
import os
import pandas as pd
import numpy as np
from model import *
from util import *
from load import load_wind, load_solar_data, load_wind_data_spatial #Change the data source for other tasks
from numpy import shape
import csv
import matplotlib.pyplot as plt

n_epochs = 70 #Number of overall training epochs on training data
learning_rate = 0.0002 
batch_size = 32
image_shape = [24,24,1] #The shape for input data
dim_z = 100 #input dimension for samples
dim_W1 = 1024 #first layer neurons
dim_W2 = 128 #second layer neurons
dim_W3 = 64 #third layer#16 is the maximum value for wind capacity we use. Change to your max value here
dim_channel = 1 #reserved for future use if multi=channels
mu, sigma = 0, 0.1 # input Gaussian
events_num=5 #kind of events

visualize_dim=32
generated_dim=32


#Comment out corresponding part to reproduce the results for 
#wind_events_generation, solar_events_generation, spatial_generation respectively
trX, trY=load_wind()
#trX, trY=load_solar()
#trX, trY=load_spatial()

print("shape of training samples ", shape(trX))
print("Training data loaded")

dcgan_model = GAN(
    dim_y=events_num # Change parameters based on number of events
    #change paprameters here for model revision
    #dim_z: the dimension for input noise
    #W1,W2,W3: the dimension for convolutional layers
        )
print("W_DCGAN model initialized")


#Z_tf,Y_tf: placeholder
#image_tf: image placeholder
#d_cost_tf, g_cost_tf: discriminator and generator cost#16 is the maximum value for wind capacity we use. Change to your max value here
#p_real, p_gen: the output of discriminator to judge real/generated
Z_tf, Y_tf, image_tf, d_cost_tf, g_cost_tf, p_real, p_gen = dcgan_model.build_model()
sess = tf.InteractiveSession()
saver = tf.train.Saver(max_to_keep=10)

discrim_vars = filter(lambda x: x.name.startswith('discrim'), tf.trainable_variables())
gen_vars = filter(lambda x: x.name.startswith('gen'), tf.trainable_variables())
discrim_vars = [i for i in discrim_vars]
gen_vars = [i for i in gen_vars]

train_op_discrim = (tf.train.RMSPropOptimizer(learning_rate=1e-4).minimize(-d_cost_tf, var_list=discrim_vars))
train_op_gen = (tf.train.RMSPropOptimizer(learning_rate=1e-4).minimize(g_cost_tf, var_list=gen_vars))

Z_tf_sample, Y_tf_sample, image_tf_sample = dcgan_model.samples_generator(batch_size=visualize_dim)
tf.initialize_all_variables().run()

Zs = np.random.normal(mu, sigma, size=[batch_size, dim_z]).astype(np.float32)
Y_np_sample = OneHot(np.random.randint(events_num, size=[visualize_dim]), n=events_num)
iterations = 0
k = 4 #control the balance of training D and G

gen_loss_all=[]#16 is the maximum value for wind capacity we use. Change to your max value here
P_real=[]
P_fake=[]
P_distri=[]
discrim_loss=[]


#begin training
for epoch in range(n_epochs):
    print("epoch" + str(epoch))
    index = np.arange(len(trY))
    np.random.shuffle(index)
    trX = trX[index]
    trY = trY[index]
    trY2 = OneHot(trY, n=events_num)

    for start, end in zip(
            range(0, len(trY), batch_size),
            range(batch_size, len(trY), batch_size)
            ):

        Xs = trX[start:end].reshape([-1, 24, 24, 1])
        Ys = trY2[start:end]

        #use uniform or Gaussian distribution data to generate adversarial samples
        Zs = np.random.normal(mu, sigma, size=[batch_size, dim_z]).astype(np.float32)

        #for each iteration, generate g and d respectively, k=2
        if np.mod( iterations, k) == 0:
            _, gen_loss_val = sess.run(
                    [train_op_gen, g_cost_tf],
                    feed_dict={
                        Z_tf:Zs,
                        Y_tf:Ys,
                        image_tf: Xs
                        })
            discrim_loss_val, p_real_val, p_gen_val = sess.run([d_cost_tf,p_real,p_gen], feed_dict={Z_tf:Zs, image_tf:Xs, Y_tf:Ys})

            '''print("=========== updating G ==========")
            print("iteration:", iterations)
            print("gen loss:", gen_loss_val)
            print("discrim loss:", discrim_loss_val)'''

        else:
            _, discrim_loss_val = sess.run(
                    [train_op_discrim, d_cost_tf],
                    feed_dict={
                        Z_tf:Zs,
                        Y_tf:Ys,
                        image_tf:Xs
                        })

            '''print("=========== updating D ==========")
            print("iteration:", iterations)
            print("gen loss:", gen_loss_val)
            print("discrim loss:", discrim_loss_val)'''

            gen_loss_val, p_real_val, p_gen_val = sess.run([g_cost_tf, p_real, p_gen],
                                                       feed_dict={Z_tf: Zs, image_tf: Xs, Y_tf: Ys})
        P_real.append(p_real_val.mean())
        P_fake.append(p_gen_val.mean())
        '''gen_loss_val, p_real_val, p_gen_val = sess.run([g_cost_tf, p_real, p_gen],
                                                       feed_dict={Z_tf: Zs, image_tf: fake_data, Y_tf: Ys})'''
        discrim_loss.append(discrim_loss_val)


        if np.mod(iterations, 1000) == 0:
            print("iterations ", iterations)
            print("Average P(real)=", p_real_val.mean())
            print("Average P(gen)=", p_gen_val.mean())
            print("Discrim loss:", discrim_loss_val)
            Y_np_sample = OneHot(np.random.randint(5, size=[visualize_dim]), n=events_num)
            Z_np_sample = np.random.normal(mu, sigma, size=[batch_size, dim_z]).astype(np.float32)
            generated_samples = sess.run(
                image_tf_sample,
                feed_dict={
                    Z_tf_sample: Z_np_sample,
                    Y_tf_sample: Y_np_sample
                })
            generated_samples=generated_samples.reshape([-1,576])
            generated_samples = generated_samples * 16 #16 is the maximum value for wind capacity we use. Change to your max value here
            csvfile=file('%s.csv' %iterations, 'wb')
            writer=csv.writer(csvfile)
            writer.writerows(generated_samples)

        iterations += 1

Y_np_sample = OneHot(np.random.randint(5, size=[visualize_dim]), n=events_num) 
Zs = np.random.normal(mu, sigma, size=[batch_size, dim_z]).astype(np.float32)
generated_samples = sess.run(
    image_tf_sample,
    feed_dict={
        Z_tf_sample: Z_np_sample,
        Y_tf_sample: Y_np_sample
    })
generated_samples=generated_samples.reshape([-1,576])
generated_samples = generated_samples * 16 #16 is the maximum value for wind capacity we use. Change to your max value here
csvfile=file('sample1.csv', 'wb')
writer=csv.writer(csvfile)
writer.writerows(generated_samples)
csvfile=file('label1.csv', 'wb')
writer=csv.writer(csvfile)
writer.writerows(Y_np_sample)


#plot the loss and P_real as well as P_fake
print("P_real",P_real)
print("P_fake",P_fake)

plt.plot(P_real,label="real")
plt.plot(P_fake,label="fake")
plt.legend()
plt.show()

plt.plot(discrim_loss,label="discrim_loss")
plt.legend()
plt.show()