Uncertainty_deep_learning/Bayesian_NN.py at master · deepakdalakoti/Uncertainty_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
import numpy as np
from torch.autograd import Variable
import tensorflow as tf
from tensorflow import keras
from tensorflow.keras.layers import Input, Dense, Dropout
from tensorflow.keras.callbacks import EarlyStopping, ModelCheckpoint, ReduceLROnPlateau
from util import do_inverse_norm, ReduceLROnPlateau
import tensorflow_probability as tfp

tf.keras.backend.set_floatx('float32')

class Bayesian_net():

    def __init__(self, inF, outF, H, lr=1e-4,  problem='regression'):
        self.inF= inF
        self.outF = outF
        self.H = H
        self.problem = problem
        self.lr = lr
        self.model_fn = self.base_model_regression
        self.train_step = self.train_step_regression
        self.loss_fn = self.loss_reg
        self.pred = self.pred_regression
        self.optimizer = tf.keras.optimizers.Adam(self.lr, beta_1=0.9, beta_2=0.999)
        if(problem=='classification'):
            self.model_fn = self.base_model_classification
            self.train_step = self.train_step_classification
            self.loss_fn = self.loss_class
            self.pred = self.pred_classification

        self.model = self.model_fn()
        return


    def base_model_regression(self):
        # I am modelling a case where there is no aleatoric uncertainty in the data
        # so no need for modelling sigma
        inputs = Input(shape=(self.inF,))
        x = tfp.layers.DenseFlipout(self.H,kernel_prior_fn=tfp.layers.default_multivariate_normal_fn,
                           kernel_divergence_fn=self.kernel_divergence_fn,activation='relu')(inputs)
        x = tfp.layers.DenseFlipout(self.H,kernel_prior_fn=tfp.layers.default_multivariate_normal_fn,
                           kernel_divergence_fn=self.kernel_divergence_fn,activation='relu')(x)
        x = tfp.layers.DenseFlipout(self.H,kernel_prior_fn=tfp.layers.default_multivariate_normal_fn,
                           kernel_divergence_fn=self.kernel_divergence_fn,activation='relu')(x)
        mu = tfp.layers.DenseFlipout(self.outF,kernel_prior_fn=tfp.layers.default_multivariate_normal_fn,
                           kernel_divergence_fn=self.kernel_divergence_fn,activation='linear')(x)
        #sigma = tfp.layers.DenseFlipout(self.outF,kernel_prior_fn=tfp.layers.default_multivariate_normal_fn,
        #                   kernel_divergence_fn=self.kernel_divergence_fn,activation='softplus')(x)

        #model = keras.Model(inputs=inputs, outputs=[mu, sigma])
        model = keras.Model(inputs=inputs, outputs=mu)
        return model

    def base_model_classification(self):

        inputs = Input(shape=(self.inF,))
        x = tfp.layers.DenseFlipout(self.H,kernel_prior_fn=tfp.layers.default_multivariate_normal_fn,
                           kernel_divergence_fn=self.kernel_divergence_fn,activation='relu')(inputs)
        x = tfp.layers.DenseFlipout(self.H,kernel_prior_fn=tfp.layers.default_multivariate_normal_fn,
                           kernel_divergence_fn=self.kernel_divergence_fn,activation='relu')(x)
        x = tfp.layers.DenseFlipout(self.H,kernel_prior_fn=tfp.layers.default_multivariate_normal_fn,
                           kernel_divergence_fn=self.kernel_divergence_fn,activation='relu')(x)
        x = tfp.layers.DenseFlipout(self.outF,kernel_prior_fn=tfp.layers.default_multivariate_normal_fn,
                           kernel_divergence_fn=self.kernel_divergence_fn,activation='softmax')(x)
        model = keras.Model(inputs=inputs, outputs=x)

        return model

    def kernel_divergence_fn(self, q, p, _):
        return  tfp.distributions.kl_divergence(q, p)

    def loss_reg(self, model, xtrain, ytrain, kl_weight, training):
        # NLL loss
        # For case where there is no aleatoric uncertainty NLL will come out to be just MSE
        #mu, sigma = model(xtrain, training=training)
        #var = sigma + 1e-6
        #NLL = tf.math.log(var)*0.5 + 0.5*tf.math.divide(tf.math.square(ytrain-mu),var)
        #total_loss = NLL + kl_weight*sum(model.losses)
        mu = model(xtrain, training=training)
        #var = sigma + 1e-6
        NLL = tf.reduce_mean(tf.math.square(ytrain-mu))
        KL_loss = kl_weight*sum(model.losses)
        total_loss = NLL + KL_loss

        return total_loss, NLL, KL_loss

    def loss_class(self, model, xtrain, ytrain, kl_weight, training):
        #
        ypred = model(xtrain, training=training)
        loss_fn = tf.keras.losses.CategoricalCrossentropy()
        total_loss = loss_fn(ytrain, ypred) + kl_weight*sum(model.losses)
        return total_loss

    @tf.function
    def train_step_regression(self, model, xtrain, ytrain, kl_weight):
        with tf.GradientTape() as tape:
            loss, NLL, KL_loss = self.loss_fn(model,xtrain, ytrain, kl_weight, True)
        grad = tape.gradient(loss, model.trainable_variables)
        self.optimizer.apply_gradients(zip(grad, model.trainable_variables))
        return  loss, NLL, KL_loss

    @tf.function
    def train_step_classification(self, model, xtrain, ytrain, kl_weight):
        with tf.GradientTape() as tape:
            loss = self.loss_fn(model, xtrain, ytrain, kl_weight, True)
        grad = tape.gradient(loss, model.trainable_variables)
        self.optimizer.apply_gradients(zip(grad, model.trainable_variables))
        return loss

    def train(self, batch_size, epochs, xtrain, ytrain, kl_weight=1e-3, validation_data=None):
        train_dataset = tf.data.Dataset.from_tensor_slices((xtrain.astype(np.float32), ytrain.astype(np.float32)))
        train_dataset = train_dataset.shuffle(buffer_size=xtrain.shape[0], reshuffle_each_iteration=True).batch(batch_size)
        train_loss = []
        NLL_loss = []
        KL_loss = []
        valid_loss = []
        red_lr = ReduceLROnPlateau(self.optimizer, 0.8, 10, 1e-5)
        epoch_loss_avg = tf.keras.metrics.Mean()
        NLL_loss_avg = tf.keras.metrics.Mean()
        KL_loss_avg = tf.keras.metrics.Mean()
        for i in range(epochs):
            epoch_loss_avg.reset_states()
            NLL_loss_avg.reset_states()
            KL_loss_avg.reset_states()
            for x, y in train_dataset:
                loss, NLL, KL = self.train_step(self.model, x, y, kl_weight)
                epoch_loss_avg.update_state(loss)
                NLL_loss_avg.update_state(NLL)
                KL_loss_avg.update_state(KL)
            train_loss.append(epoch_loss_avg.result().numpy())
            NLL_loss.append(NLL_loss_avg.result().numpy())
            KL_loss.append(KL_loss_avg.result().numpy())
            red_lr.on_epoch_end(train_loss[-1], i)
            if(validation_data):
                valid_loss.append(np.mean(self.loss_fn(self.model, validation_data[0], validation_data[1], kl_weight, False).numpy()))
                print("Step {} loss {} valid_loss {}".format(i, epoch_loss_avg.result(), valid_loss[i]))
            else:
                print("Step {} loss {}".format(i, epoch_loss_avg.result()))

        return [train_loss, NLL_loss, KL_loss], valid_loss


    def pred_regression(self, xdata, y, norm, Nsamp):
        pred_all = np.zeros([Nsamp,xdata.shape[0], self.outF])
        for i in range(Nsamp):
            #mu, sigma = self.model(xdata)
            #pred_all[i,:,:]  = np.array([mu, sigma]).T
            pred_all[i,:,0:self.outF] = self.model(xdata)

        # Mean, std as in paper
        pred_all  = do_inverse_norm(y, pred_all, norm)
        pred_mean = np.mean(pred_all, axis=0)
        #pred_std = np.sqrt(np.mean(pred_all[:,:,0]**2 + pred_all[:,:,1]+1e-6,axis=0) - pred_mean**2 )
        pred_std = np.sqrt(np.mean(pred_all**2, axis=0) - pred_mean**2 )
        #pred_mean = do_inverse_norm(y, pred_mean, norm)
        #pred_std = pred_std*np.std(y)

        return pred_all, pred_mean, pred_std

    def pred_classification(self, xdata, y, norm, Nsamp):
        pred_all = np.zeros([Nsamp,xdata.shape[0]])
        for i in range(Nsamp):
            pred_all[i,:] = self.model(xdata)

        pred_all = do_inverse_norm(y, pred_all, norm)
        pred_mean = np.mean(pred_all,axis=0)
        pred_std = np.std(pred_all,axis=0)
        return pred_all, pred_mean, pred_std

    def predict(self, xdata, y, norm, Nsamp):
        pred_all, pred_mean, pred_std = self.pred(xdata, y, norm, Nsamp)
        return pred_all, pred_mean, pred_std