In [1]:

    

# Data: time-serie data from smartwatch or smartwatch data
# %matplotlib inline # for plt.show()
import numpy as np
import pandas as pd
import matplotlib.pyplot as plt

# Data reading
# The smartwatch historical/time-seris data to visualize
# data_path = 'data/smartwatch_data/experimental_data_analysis/Basis_Watch_Data.csv'
# data_path = 'data/financial_data/USD_INR.csv'
data_path = 'data/bike_data/hour.csv'
data = pd.read_csv(data_path)

# Data: cleaning
# Getting rid of NaN
data = data.fillna(value=0.0)

# Showing the data file csv or comma separated value
data[:10]


    

    
Out[1]:







  

    

      

      
instant
      
dteday
      
season
      
yr
      
mnth
      
hr
      
holiday
      
weekday
      
workingday
      
weathersit
      
temp
      
atemp
      
hum
      
windspeed
      
casual
      
registered
      
cnt
    
  
  

    

      
0
      
1
      
2011-01-01
      
1
      
0
      
1
      
0
      
0
      
6
      
0
      
1
      
0.24
      
0.2879
      
0.81
      
0.0000
      
3
      
13
      
16
    
    

      
1
      
2
      
2011-01-01
      
1
      
0
      
1
      
1
      
0
      
6
      
0
      
1
      
0.22
      
0.2727
      
0.80
      
0.0000
      
8
      
32
      
40
    
    

      
2
      
3
      
2011-01-01
      
1
      
0
      
1
      
2
      
0
      
6
      
0
      
1
      
0.22
      
0.2727
      
0.80
      
0.0000
      
5
      
27
      
32
    
    

      
3
      
4
      
2011-01-01
      
1
      
0
      
1
      
3
      
0
      
6
      
0
      
1
      
0.24
      
0.2879
      
0.75
      
0.0000
      
3
      
10
      
13
    
    

      
4
      
5
      
2011-01-01
      
1
      
0
      
1
      
4
      
0
      
6
      
0
      
1
      
0.24
      
0.2879
      
0.75
      
0.0000
      
0
      
1
      
1
    
    

      
5
      
6
      
2011-01-01
      
1
      
0
      
1
      
5
      
0
      
6
      
0
      
2
      
0.24
      
0.2576
      
0.75
      
0.0896
      
0
      
1
      
1
    
    

      
6
      
7
      
2011-01-01
      
1
      
0
      
1
      
6
      
0
      
6
      
0
      
1
      
0.22
      
0.2727
      
0.80
      
0.0000
      
2
      
0
      
2
    
    

      
7
      
8
      
2011-01-01
      
1
      
0
      
1
      
7
      
0
      
6
      
0
      
1
      
0.20
      
0.2576
      
0.86
      
0.0000
      
1
      
2
      
3
    
    

      
8
      
9
      
2011-01-01
      
1
      
0
      
1
      
8
      
0
      
6
      
0
      
1
      
0.24
      
0.2879
      
0.75
      
0.0000
      
1
      
7
      
8
    
    

      
9
      
10
      
2011-01-01
      
1
      
0
      
1
      
9
      
0
      
6
      
0
      
1
      
0.32
      
0.3485
      
0.76
      
0.0000
      
8
      
6
      
14
    
  

In [2]:

    

# # Plotting the smartwatch data before scaling/batch normalization
# data[:10000]['Price'].plot()
data[: 10].plot()
plt.legend()
plt.show()


    

In [3]:

    

data_array = np.array(data)
data_array.shape, data_array.dtype
data_main = np.array(data_array[:, -1:], dtype=float)
data_main.shape, data_main.dtype

plt.plot(data_main[:100])
plt.show()


    

In [4]:

    

mean = np.mean(data_main, axis=0)
std = np.std(data_main, axis=0)
std.shape, mean.shape, std.dtype, mean.dtype

data_norm = (data_main - mean) / std
plt.plot(data_norm[:100])
plt.show()
data_norm.mean(), data_norm.std(), data_norm.var(), data_norm.shape, data_norm.dtype


    

    













    
Out[4]:





(-1.0548364452851478e-16, 1.0, 1.0, (17379, 1), dtype('float64'))

In [5]:

    

train_data = data_norm[:16000] # the last dim/variable/feature
test_data = data_norm[16000:] # the last dim/variable/feature
train_data.shape, test_data.shape
X_train = train_data[0:15999]
Y_train = train_data[1:16000]
X_train.shape, Y_train.shape

plt.plot(X_train[:100])
plt.plot(Y_train[:100])
plt.show()


    

In [6]:

    

X_valid = test_data[0:1378] 
Y_valid = test_data[1:1379]
X_valid.shape, Y_valid.shape
plt.plot(X_valid[:100])
plt.plot(Y_valid[:100])
plt.show()


    

In [7]:

    

# Model or Network
import impl.layer as l
from impl.loss import *

class GRU:
    def __init__(self, D, H, L):
        self.D = D
        self.H = H
        self.L = L
        self.losses = {'train':[], 'smooth train':[], 'valid': []}
        
        # Model params
        Z = H + D
        low, high = (-1. / np.sqrt(Z / 2.)), (1. / np.sqrt(Z / 2.)) # make sure initialized -1 < w < 1
        m = dict(
#             Wz=np.random.randn(Z, H) / np.sqrt(Z / 2.),
#             Wr=np.random.randn(Z, H) / np.sqrt(Z / 2.),
#             Wh=np.random.randn(Z, H) / np.sqrt(Z / 2.),
#             Wy=np.random.randn(H, D) / np.sqrt(H / 2.),
            Wz=np.random.uniform(size=(Z, H), low=low, high=high),
            Wr=np.random.uniform(size=(Z, H), low=low, high=high),
            Wh=np.random.uniform(size=(Z, H), low=low, high=high),
            Wy=np.random.uniform(size=(H, D), low=low, high=high),
            bz=np.zeros((1, H)),
            br=np.zeros((1, H)),
            bh=np.zeros((1, H)),
            by=np.zeros((1, D))
        )
        self.model = []
        for _ in range(self.L):
            self.model.append(m)
        
    def initial_state(self):
        return np.zeros((1, self.H))

    def forward(self, X, h, m):
        Wz, Wr, Wh, Wy = m['Wz'], m['Wr'], m['Wh'], m['Wy']
        bz, br, bh, by = m['bz'], m['br'], m['bh'], m['by']

        X_in = X.copy()
        h_in = h.copy()

        X = np.column_stack((h_in, X_in))

        hz, hz_cache = l.fc_forward(X, Wz, bz)
        hz, hz_sigm_cache = l.sigmoid_forward(hz)

        hr, hr_cache = l.fc_forward(X, Wr, br)
        hr, hr_sigm_cache = l.sigmoid_forward(hr)

        X = np.column_stack((hr * h_in, X_in))
        
        hh, hh_cache = l.fc_forward(X, Wh, bh)
        hh, hh_tanh_cache = l.tanh_forward(hh)

        # h = (1. - hz) * h_old + hz * hh
        # or
        h = ((1. - hz) * h_in) + (hz * hh)
        # or
        # h = h_in + hz (hh - h_in)

        y, y_cache = l.fc_forward(h, Wy, by)
        
        cache = (h_in, hz, hz_cache, hz_sigm_cache, hr, hr_cache, hr_sigm_cache, hh, hh_cache, hh_tanh_cache, 
                 y_cache)

        return y, h, cache

    def backward(self, dy, dh, cache):
        h_in, hz, hz_cache, hz_sigm_cache, hr, hr_cache, hr_sigm_cache, hh, hh_cache, hh_tanh_cache, y_cache = cache
        
        dh_out = dh.copy()

        dh, dWy, dby = l.fc_backward(dy, y_cache)
        dh += dh_out

        dh_in1 = (1. - hz) * dh
        dhh = hz * dh
        dhz = (hh * dh) - (h_in * dh)
        # or
        # dhz = (hh - h_in) * dh

        dhh = l.tanh_backward(dhh, hh_tanh_cache)
        dXh, dWh, dbh = l.fc_backward(dhh, hh_cache) # w_fixed/fb

        dh = dXh[:, :self.H]
        dX_in2 = dXh[:, self.H:]
        dh_in2 = hr * dh

        dhr = h_in * dh
        dhr = l.sigmoid_backward(dhr, hr_sigm_cache)
        dXr, dWr, dbr = l.fc_backward(dhr, hr_cache) # w_fixed_fb

        dhz = l.sigmoid_backward(dhz, hz_sigm_cache)
        dXz, dWz, dbz = l.fc_backward(dhz, hz_cache) # w_fixed_fb

        dX = dXr + dXz
        dh_in3 = dX[:, :self.H]
        dX_in1 = dX[:, self.H:]

        dh = dh_in1 + dh_in2 + dh_in3
        dX = dX_in1 + dX_in2

        grad = dict(Wz=dWz, Wr=dWr, Wh=dWh, Wy=dWy, bz=dbz, br=dbr, bh=dbh, by=dby)
        
        return dX, dh, grad

    def train_forward(self, X_train, h):
        ys, fc_caches, caches = [], [], []
        
        for X in X_train:
            X = X.reshape(1, -1) # X_1xn
            for layer in range(self.L):
                y, h, fc_cache = self.forward(X, h, self.model[layer])
                X = y.copy()
                fc_caches.append(fc_cache) # layers
            caches.append(fc_caches) # time
            ys.append(y)
        
        ys = np.array(ys, dtype=float).reshape(len(ys), -1) # ys_txn instead of ys_tx1xn
        
        return ys, caches
                                
    def loss_function(self, y_pred, y_train): # , alpha alpha: learning rate
        loss, dys = 0.0, []

        for y, Y in zip(y_pred, y_train):
            loss += l2_regression(y_pred=y, y_train=Y)
            dy = dl2_regression(y_pred=y, y_train=Y)
            dys.append(dy)
            
        return loss, dys
    
    def train_backward(self, dys, caches):        
        dh = np.zeros((1, self.H)) 
        grad = {key: np.zeros_like(val) for key, val in self.model[0].items()}
        grads = [] #{key: np.zeros_like(val) for key, val in self.model.items()}
        for _ in range(self.L):
            grads.append(grad)

        for t in reversed(range(len(dys))):
            dy = dys[t].reshape(1, -1) # dy_1xn
            fc_caches = caches[t]
            for layer in reversed(range(self.L)):
                dX, dh, grad = self.backward(dy, dh, fc_caches[layer])
                dy = dX.copy() # for the previous layer
                for key in grad.keys():
                    grads[layer][key] += grad[key]
                
        return dX, grads
    
    def test(self, X_seed, h, size):
        ys = []
        X = X_seed.reshape(1, -1)
        for _ in range(size):
            for layer in range(self.L):
                y, h, _ = self.forward(X, h, self.model[layer])
                X = y.copy() # previous out for the next input for prediction
            ys.append(y) # list array
        
        ys = np.array(ys, dtype=float).reshape(len(ys), -1) # ys_txn instead of ys_tx1xn
        return ys


    

In [8]:

    

def get_minibatch(X, y, minibatch_size, shuffle):
    minibatches = []

    for i in range(0, X.shape[0], minibatch_size):
        X_mini = X[i:i + minibatch_size]
        y_mini = y[i:i + minibatch_size]
        minibatches.append((X_mini, y_mini))

    return minibatches

def adam_rnn(nn, XY_train, XY_valid, alpha, mb_size, n_iter, print_after):
    X_train, y_train = XY_train
    X_valid, y_valid = XY_valid

    # Momentum
    M = [] # {key: np.zeros_like(val) for key, val in nn.model.items()}
    R = [] # {key: np.zeros_like(val) for key, val in nn.model.items()}
    for _ in range(nn.L):
        M.append({key: np.zeros_like(val) for key, val in nn.model[0].items()})
        R.append({key: np.zeros_like(val) for key, val in nn.model[0].items()})
    
    # Learning decay: suggested by Justin Jhonson in Standford
    beta1 = .9
    beta2 = .99
    state = nn.initial_state()
    smooth_loss = 1.
    minibatches = get_minibatch(X_train, y_train, mb_size, shuffle=False)
    
    # Epochs: iterating through the whole data
    for iter in range(1, n_iter + 1):
        
        # Minibatches
        for idx in range(len(minibatches)):
            
            # Train the model
            X_mini, y_mini = minibatches[idx]
            ys, caches = nn.train_forward(X_mini, state)
            loss, dys = nn.loss_function(y_pred=ys, y_train=y_mini) #, alpha=alpha
            _, grads = nn.train_backward(dys, caches)
            nn.losses['train'].append(loss)
            smooth_loss = (0.999 * smooth_loss) + (0.001 * loss)
            nn.losses['smooth train'].append(smooth_loss)
            
            # Update the model
            for layer in range(nn.L):
                for key in grads[0].keys(): # key, value, items
                    M[layer][key] = l.exp_running_avg(M[layer][key], grads[layer][key], beta1)
                    R[layer][key] = l.exp_running_avg(R[layer][key], grads[layer][key]**2, beta2)
                    m_k_hat = M[layer][key] / (1. - (beta1** iter))
                    r_k_hat = R[layer][key] / (1. - (beta2** iter))
                    nn.model[layer][key] -= alpha * m_k_hat / (np.sqrt(r_k_hat) + l.eps)

            # Validate the model (by testing)
            ys = nn.test(X_seed=X_valid[0], h=state, size=X_valid.shape[0]) # ys_tx1xn
            valid_loss, _ = nn.loss_function(y_pred=ys, y_train=Y_valid) #, alpha=alpha
            nn.losses['valid'].append(valid_loss)

        # Print the model loss/ error
        if iter % print_after == 0:
            print('Iter-{}, train loss: {:.8f}, valid loss: {:.8f}'.format(iter, loss, valid_loss))

    return nn


    

In [9]:

    

# Hyper-parameters
time_step = 128 # minibatch size: 32, 64, 128, or 256 Cache
n_iter = 100 # epochs
alpha = 1e-4 # learning_rate: 1e-3, 5e-4, 1e-4 - default choices
print_after = 1 # print training loss, valid, and test
num_hidden_units = 64 # num_hidden_units in hidden layer
num_input_units = X_train.shape[1] # X_txn: noise given by using all possible channels/ features
num_hidden_layers = 2 # number of hidden layers

# Build the network and learning it or optimizing it using SGD
# def adam_rnn(nn, X_train, y_train, alpha=0.001, mb_size=256, n_iter=2000, print_after=100):
net = GRU(D=num_input_units, H=num_hidden_units, L=num_hidden_layers)

# Start learning using BP-SGD-ADAM
adam_rnn(nn=net, XY_train=(X_train, Y_train), XY_valid=(X_valid, Y_valid), alpha=alpha, mb_size=time_step,
         n_iter=n_iter, print_after=print_after)


    

    




Iter-1, train loss: 56.02501500, valid loss: 632.87023400
Iter-2, train loss: 50.94009461, valid loss: 633.07601144
Iter-3, train loss: 46.82883779, valid loss: 634.04188393
Iter-4, train loss: 44.73668568, valid loss: 637.33993208
Iter-5, train loss: 47.99577782, valid loss: 658.62722594
Iter-6, train loss: 53.43951771, valid loss: 771.08421687
Iter-7, train loss: 52.61724792, valid loss: 2068.16043353
Iter-8, train loss: 52.25128550, valid loss: 2068.57398876
Iter-9, train loss: 52.21375281, valid loss: 2075.71791989
Iter-10, train loss: 52.16386066, valid loss: 2091.15174372
Iter-11, train loss: 52.08743263, valid loss: 2101.47297278
Iter-12, train loss: 52.00971676, valid loss: 2111.22548858
Iter-13, train loss: 51.92734709, valid loss: 2120.24381674
Iter-14, train loss: 51.84017433, valid loss: 2128.39235549
Iter-15, train loss: 51.74922681, valid loss: 2135.92133881
Iter-16, train loss: 51.65482118, valid loss: 2142.97734461
Iter-17, train loss: 51.55712648, valid loss: 2149.67609598
Iter-18, train loss: 51.45627358, valid loss: 2156.12417630
Iter-19, train loss: 51.35238192, valid loss: 2162.41831020
Iter-20, train loss: 51.24563595, valid loss: 2168.65412606
Iter-21, train loss: 51.13632686, valid loss: 2174.93259797
Iter-22, train loss: 51.02485444, valid loss: 2181.36332208
Iter-23, train loss: 50.91174830, valid loss: 2188.07235447
Iter-24, train loss: 50.79763834, valid loss: 2195.20793110
Iter-25, train loss: 50.68314822, valid loss: 2202.94664407
Iter-26, train loss: 50.56858236, valid loss: 2211.48670083
Iter-27, train loss: 50.45186784, valid loss: 2220.62822667
Iter-28, train loss: 50.32083466, valid loss: 2226.95364345
Iter-29, train loss: 50.17838883, valid loss: 2229.02520071
Iter-30, train loss: 50.04769751, valid loss: 2233.14553753
Iter-31, train loss: 49.91931188, valid loss: 2239.71198798
Iter-32, train loss: 49.78553268, valid loss: 2247.21573708
Iter-33, train loss: 49.64089739, valid loss: 2256.19986015
Iter-34, train loss: 49.48977002, valid loss: 2268.51182123
Iter-35, train loss: 49.36906602, valid loss: 2285.54376556
Iter-36, train loss: 49.15242603, valid loss: 2294.03447536
Iter-37, train loss: 49.17311230, valid loss: 2326.44628120
Iter-38, train loss: 48.84871795, valid loss: 2324.44491722
Iter-39, train loss: 48.76447508, valid loss: 2333.36524533
Iter-40, train loss: 48.58853143, valid loss: 2348.02665483
Iter-41, train loss: 48.47518833, valid loss: 2360.07563198
Iter-42, train loss: 48.18951359, valid loss: 2356.98161778
Iter-43, train loss: 48.08594051, valid loss: 2357.63510060
Iter-44, train loss: 47.93254792, valid loss: 2369.96981018
Iter-45, train loss: 47.79628312, valid loss: 2372.88424125
Iter-46, train loss: 47.52584197, valid loss: 2374.57877038
Iter-47, train loss: 47.50071071, valid loss: 2378.87751390
Iter-48, train loss: 47.25807543, valid loss: 2388.28552795
Iter-49, train loss: 47.20493285, valid loss: 2389.14385566
Iter-50, train loss: 46.92542016, valid loss: 2389.25587056
Iter-51, train loss: 46.90799380, valid loss: 2397.22086976
Iter-52, train loss: 46.65059569, valid loss: 2399.81575035
Iter-53, train loss: 46.60916276, valid loss: 2408.24530291
Iter-54, train loss: 46.34545500, valid loss: 2406.92972140
Iter-55, train loss: 46.30375705, valid loss: 2416.99379813
Iter-56, train loss: 46.07551080, valid loss: 2418.44955748
Iter-57, train loss: 46.00945201, valid loss: 2430.53169902
Iter-58, train loss: 45.77860023, valid loss: 2428.88864519
Iter-59, train loss: 45.70302141, valid loss: 2442.28004874
Iter-60, train loss: 45.52223766, valid loss: 2442.26732205
Iter-61, train loss: 45.39291298, valid loss: 2457.23236988
Iter-62, train loss: 45.20894152, valid loss: 2451.94819962
Iter-63, train loss: 45.10104422, valid loss: 2470.40962175
Iter-64, train loss: 44.98716914, valid loss: 2471.96198755
Iter-65, train loss: 44.78466601, valid loss: 2489.63584752
Iter-66, train loss: 44.57679848, valid loss: 2477.64033533
Iter-67, train loss: 44.51953193, valid loss: 2503.62335829
Iter-68, train loss: 44.38834406, valid loss: 2504.33258890
Iter-69, train loss: 44.18653449, valid loss: 2520.68340311
Iter-70, train loss: 44.00156821, valid loss: 2508.54827240
Iter-71, train loss: 43.96605877, valid loss: 2542.67591450
Iter-72, train loss: 43.74094837, valid loss: 2530.69055549
Iter-73, train loss: 43.61308720, valid loss: 2552.56367208
Iter-74, train loss: 43.37350750, valid loss: 2536.54083503
Iter-75, train loss: 43.43358889, valid loss: 2574.05920871
Iter-76, train loss: 43.10979201, valid loss: 2547.04923078
Iter-77, train loss: 43.28844235, valid loss: 2612.78211716
Iter-78, train loss: 42.93427579, valid loss: 2581.42969595
Iter-79, train loss: 42.92562851, valid loss: 2612.84553692
Iter-80, train loss: 42.65655367, valid loss: 2602.96584148
Iter-81, train loss: 42.62063605, valid loss: 2605.29771447
Iter-82, train loss: 42.47472041, valid loss: 2612.19328962
Iter-83, train loss: 42.49824418, valid loss: 2628.90039946
Iter-84, train loss: 42.46869912, valid loss: 2653.62970472
Iter-85, train loss: 41.93043931, valid loss: 2594.58665125
Iter-86, train loss: 42.37103289, valid loss: 2666.56953937
Iter-87, train loss: 41.91958849, valid loss: 2637.57429754
Iter-88, train loss: 42.00839815, valid loss: 2655.66686512
Iter-89, train loss: 41.51477473, valid loss: 2610.04898483
Iter-90, train loss: 42.08754062, valid loss: 2686.33210359
Iter-91, train loss: 41.88128136, valid loss: 2685.07856660
Iter-92, train loss: 41.47643539, valid loss: 2647.98292490
Iter-93, train loss: 41.51098648, valid loss: 2655.05469976
Iter-94, train loss: 41.34107824, valid loss: 2642.94924269
Iter-95, train loss: 41.72999178, valid loss: 2691.60231503
Iter-96, train loss: 41.59071585, valid loss: 2687.97323895
Iter-97, train loss: 41.26152452, valid loss: 2650.49474598
Iter-98, train loss: 41.16572077, valid loss: 2633.12164762
Iter-99, train loss: 41.55066647, valid loss: 2684.08422369
Iter-100, train loss: 41.45599994, valid loss: 2677.52178303






    
Out[9]:





<__main__.GRU at 0x112ca0e10>

In [10]:

    

import matplotlib.pyplot as plt

plt.plot(net.losses['train'], label='Train loss')
plt.plot(net.losses['smooth train'], label='Train smooth loss')
plt.legend()
plt.show()


    

In [11]:

    

import matplotlib.pyplot as plt

plt.plot(net.losses['valid'], label='Validation loss')
plt.legend()
plt.show()


    

In [13]:

    

import matplotlib.pyplot as plt

y_pred = net.test(X_seed=X_valid[0], h=net.initial_state(), size=X_valid.shape[0]) # ys_tx1xn
y_pred.shape, Y_valid.shape

plt.plot(y_pred[:100], label='y_pred')
plt.plot(Y_valid[:100], label='Y_valid')
# plt.plot(X_valid[:100], label='X_valid')
plt.legend()
plt.show()


    

In [ ]:

    




    

In [ ]: