Bayesian_Deep_Learning/HW3_Bayesian_Neural_Network_Variational_Inference.py at master · Ramtin92/Bayesian_Deep_Learning · GitHub

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
from __future__ import division
import pickle
import copy
import time
import matplotlib.pyplot as plt
import seaborn as sns
import numpy as np
import pandas as pd
import autograd.numpy as ag_np
import autograd
from scipy.stats import norm
from autograd.scipy import stats as ag_stats

"""
 # m_tilda = 0.0 # tilda corresponds to weights
 # s_tilda = 1.0
 # m_bar = 0.0  # bar corresponds to bias
 # s_bar = 1.0
 """


x_trains = np.asarray([-2.,    -1.8,   -1.,  1.,  1.8,     2.])
y_trains = np.asarray([-3.,  0.2224,    3.,  3.,  0.2224, -3.])

learning_rate = 1e-5
ITERATIONS = 100

L = 2
nn_hidden_layer = 10

def make_nn_params_as_list_of_dicts(
        n_hiddens_per_layer_list=[],
        n_dims_input=1,
        n_dims_output=1,
        weight_fill_func=np.zeros,
        bias_fill_func=np.zeros):
    nn_param_list = []
    n_hiddens_per_layer_list = [n_dims_input] + n_hiddens_per_layer_list + [n_dims_output]
    for n_in, n_out in zip(n_hiddens_per_layer_list[:-1], n_hiddens_per_layer_list[1:]):
        nn_param_list.append(
            dict(
                w=weight_fill_func((n_in, n_out)),
                b=bias_fill_func((n_out,)),
            ))
    return nn_param_list


def predict_y_given_x_with_NN(x=None, nn_param_list=None, activation_func=ag_np.tanh):
    for layer_id, layer_dict in enumerate(nn_param_list):
        if layer_id == 0:
            if x.ndim > 1:
                in_arr = x
            else:
                if x.size == nn_param_list[0]['w'].shape[0]:
                    in_arr = x[ag_np.newaxis ,:]
                else:
                    in_arr = x[: ,ag_np.newaxis]
        else:
            in_arr = activation_func(out_arr)
        out_arr = ag_np.dot(in_arr, layer_dict['w']) + layer_dict['b']
    return ag_np.squeeze(out_arr)


def get_all_weights_and_biases(nn_param_list_of_dict):
    all_weights = []
    all_biases = []
    for ll, layer_dict in enumerate(nn_param_list_of_dict):
        all_weights.extend(layer_dict['w'].flatten())
        all_biases.extend(layer_dict['b'].flatten())
    return all_weights + all_biases


def convert_list_to_dict(list_weights_and_biases):
    dict_list = []
    dict_list.append({'w': np.array(list_weights_and_biases[0:10]).reshape(1, 10),
                      'b':np.array(list_weights_and_biases[20:30]).reshape(10)})
    dict_list.append({'w': np.array(list_weights_and_biases[10:20]).reshape(10, 1),
                      'b': np.array(list_weights_and_biases[30])})
    return dict_list


def estimate_variational_inference_loss_function(landa, nn_params):
    return -1 *(log_pdf_prior(nn_params) + log_pdf_likelihood(nn_params) - log_pdf_approximate_posterior(landa, nn_params))


def estimate_gradient_variational_loss_functtion(samples):
    pass

def log_pdf_approximate_posterior(landa, nn_params):
    all_weights_and_biases = get_all_weights_and_biases(nn_params)
    log_pdf_weights_biases = []
    for i in range(len(all_weights_and_biases)):
        log_pdf_weights_biases.append(ag_stats.norm.logpdf(all_weights_and_biases[i], landa[2*i], ag_np.exp(landa[2*i+1])))
    log_pdf_approximate_posterior = ag_np.sum(log_pdf_weights_biases)
    return log_pdf_approximate_posterior


calc_gradient_log_pdf_approximate_posterior = autograd.grad(log_pdf_approximate_posterior)


def log_pdf_prior(nn_params):
    all_weights_and_biases = get_all_weights_and_biases(nn_params)
    log_pdf_weights_biases_prior = ag_stats.norm.logpdf(all_weights_and_biases, 0.0, 1)
    log_pdf_prior = ag_np.sum(log_pdf_weights_biases_prior)
    #print "log_pdf_prior", log_pdf_prior
    return log_pdf_prior

def log_pdf_likelihood(nn_params):
    yhat_predictions = predict_y_given_x_with_NN(x_trains, nn_params, activation_func=ag_np.tanh)
    #print "yhat_predictions", yhat_predictions
    log_pdf_likelihood = ag_stats.norm.logpdf(y_trains, yhat_predictions, 0.1)
    log_pdf_likelihood_sum = ag_np.sum(log_pdf_likelihood)
    #print "log_pdf_likelihood_sum",log_pdf_likelihood_sum
    return log_pdf_likelihood_sum


def main():

    time_start = time.time()
    nn_params = make_nn_params_as_list_of_dicts(n_hiddens_per_layer_list=[10], n_dims_input=1, n_dims_output=1)
    all_weights_and_biases = get_all_weights_and_biases(nn_params)

    estimated_loss_mean_in_all_iteration = []

    n_samples = 150
    landa = np.zeros(62, dtype=float)

    for index in range(62):
        if index % 2 ==0:
            landa[index] = np.random.normal(0.0, 1.0, 1)
        else:
            landa[index] = 1

    for iteration in range(ITERATIONS):
        estimated_loss_list_each_iteration = []

        grad_loss = []
        my_samples = []

        for _ in range(n_samples):
            if iteration == 0:
                for i in range(0, len(all_weights_and_biases)):
                    all_weights_and_biases[i] = np.random.normal(landa[2*i], 1, 1)
                nn_params = convert_list_to_dict(all_weights_and_biases)
                my_samples.append(nn_params)
            else:
                for i in range(0, len(all_weights_and_biases)):
                    all_weights_and_biases[i] = np.random.normal(landa[2*i], ag_np.exp(landa[2*i+1]), 1)
                nn_params = convert_list_to_dict(all_weights_and_biases)
                my_samples.append(nn_params)

        for sample in my_samples:
            #print "sample", sample
            estimated_loss = estimate_variational_inference_loss_function(landa, sample)
            #print "estimated_loss", estimated_loss
            estimated_loss_list_each_iteration.append(estimated_loss)
            grad_with_respect_to_landa_diff_samples = calc_gradient_log_pdf_approximate_posterior(landa, sample)
            #print 'grad_with_respect_to_landa_diff_samples', grad_with_respect_to_landa_diff_samples
            grad_loss.append(grad_with_respect_to_landa_diff_samples * estimated_loss)


        grad_los_mean = np.mean(grad_loss, axis= 0)
        landa += learning_rate * grad_los_mean
        #print "grad_loss_mean", grad_los_mean
        estimated_loss_mean_in_all_iteration.append(np.mean(estimated_loss_list_each_iteration, axis=0))

    _, subplot_grid = plt.subplots(nrows=1, ncols=1, sharex=True, sharey=False, figsize=(6, 8), squeeze=False)
    subplot_grid[0, 0].plot([x for x in range(ITERATIONS)], estimated_loss_mean_in_all_iteration)
    subplot_grid[0, 0].set_xlabel('iterations')
    subplot_grid[0, 0].set_ylabel('estimated_loss')
    plt.title('estimated loss in 2000 iterations')
    plt.savefig('estimated_loss1.png')


    grad_with_respect_to_landa_diff_samples = -1 * np.array(grad_with_respect_to_landa_diff_samples).reshape(10, 20)
    grad_with_respect_to_landa_diff_samples = np.mean(grad_with_respect_to_landa_diff_samples, axis=0)


if __name__=="__main__":
    main()