Tensorflow_ODE/ODESolver.py at master · Darikiva/Tensorflow_ODE · 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
from cmath import inf
import tensorflow as tf
from tensorflow import keras
from tensorflow.keras import layers
import numpy as np
import SampleModel as sm
from tqdm import tqdm
import time
import matplotlib.pyplot as plt
from scipy.integrate import solve_ivp

tf.keras.backend.set_floatx("float64")

N = 500  # number of points to eval


class ODESolver:
    def __init__(
        self, sample_model: sm.SampleModel, nnmodel_hyperparamaters_dict
    ) -> None:
        self.sample_model = sample_model
        self.__create_model(nnmodel_hyperparamaters_dict)
        lr_schedule = keras.optimizers.schedules.ExponentialDecay(
            0.001, decay_steps=100, decay_rate=0.99
        )
        self.optimizer = tf.keras.optimizers.Adam(learning_rate=lr_schedule)

    def train(
        self, loss_threshold, logging=False, path_to_plots=None):
        bar = tqdm()
        best_model = tf.keras.models.clone_model(self.model)
        min_loss = np.Inf
        iteration_number = 0
        while min_loss > loss_threshold:
            try:
                rand_pts = np.random.uniform(
                    low=self.sample_model.T[0], high=self.sample_model.T[1], size=(N,)
                )
                rand_pts = np.sort(rand_pts)
                for i in range(self.sample_model.dim):
                    rand_pts[self.__get_pos_ind(equation_idx=i, condition_idx=0, points_amount=N)] = self.sample_model.pos[i]
                rand_pts = tf.expand_dims(rand_pts, axis=1)
                loss, start_loss = self.__train_step(rand_pts)
                if (loss) + (start_loss) < min_loss:
                    min_loss = (loss) + start_loss
                    best_model.set_weights(self.model.get_weights())
                if logging:
                    bar.set_description(
                        "Loss: {:.5f} {:.5f} min loss: {:.6f}".format(loss, start_loss, min_loss)
                    )
                    bar.update(1)
                iteration_number += 1
            except KeyboardInterrupt:
                print("ctrl-c")
                break
        self.model = best_model
        return iteration_number

    def draw_plot(self, save_path):
        gx = np.linspace(self.sample_model.T[0], self.sample_model.T[1], N)
        y = self.model(gx)
        plt.plot(gx, np.squeeze(y), "m-", label="Tensorflow")
        if hasattr(self.sample_model, "theoretical") and callable(
            getattr(self.sample_model, "theoretical")
        ):
            print("theoretical")
            plt.plot(
                gx,
                self.sample_model.theoretical(gx).T,
                "b-",
                label="Theoretical",
            )
        if np.all(self.sample_model.pos == self.sample_model.T[0]):
            sol = solve_ivp(
                self.sample_model.equation,
                self.sample_model.T,
                self.sample_model.val[0],
                dense_output=True
            )
            rk = sol.sol(gx)
            print("helo")
            plt.plot(gx, rk.T, "g-", label="Runge-Kutta")
        plt.title(self.sample_model.__doc__)
        handles, labels = plt.gca().get_legend_handles_labels()
        labels, ids = np.unique(labels, return_index=True)
        handles = [handles[i] for i in ids]
        plt.legend(handles, labels, loc="best")
        plt.savefig(save_path)
        plt.figure().clear()
        plt.close()
        plt.cla()
        plt.clf()


    def __train_step(self, x):
        with tf.GradientTape(persistent=True) as g:
            g.watch(x)
            dy = [self.model(x)]
            y = self.model(x)
            # if self.sample_model.dim >= 1:
            for i in range(self.sample_model.max_order):
                new_dy = tf.stack(
                    [g.gradient(y[:, i], x) for i in range(self.sample_model.dim)],
                    axis=1,
                )
                new_dy = tf.squeeze(new_dy, axis=2)
                dy.append(new_dy)
            wow1 = self.sample_model.tf_equation(x, *dy)
            rhs = tf.stack(wow1, axis=1)
            # else:
            # dy = g.gradient(y, x)
            # rhs = self.sample_model.tf_equation(x, y)
            loss = None
            for i in range(self.sample_model.equations_amount):
                current_equation_order = self.sample_model.orders[i]
                residual = dy[current_equation_order][i] - rhs[i]
                if not loss:
                    loss = tf.math.square(residual)
                else:
                    loss += tf.math.square(residual)
            residual = dy[0] - rhs
            loss = tf.math.square(residual)
            ind_list = []
            val = []
            for i in range(self.sample_model.equations_amount):
                for j in range(self.sample_model.orders[i]):
                    ind = self.__get_pos_ind(equation_idx=i, condition_idx=j, points_amount=N)
                    ysl = y[ind, i]
                    if [ind, i] in ind_list:
                        val[ind_list.index([ind, i])] += (tf.math.square(ysl - self.sample_model.get_val(equation_idx=i)[0])) * N
                    else:
                        ind_list.append([ind, i])
                        val.append((tf.math.square(ysl - self.sample_model.get_val(equation_idx=i)[0])) * N)
            start_loss = tf.sparse.SparseTensor(
                indices=ind_list, values=val, dense_shape=loss.shape
            )
            start_loss = tf.sparse.to_dense(start_loss)
            total_loss = loss + start_loss
            gradients = g.gradient(total_loss, self.model.trainable_variables)
            self.optimizer.apply_gradients(
                zip(gradients, self.model.trainable_variables)
            )
            return float(tf.math.reduce_mean(loss)), float(
                tf.math.reduce_mean(start_loss)
            )

    def __get_pos_ind(self, equation_idx, condition_idx, points_amount):
        return round(
            (self.sample_model.get_pos(equation_idx)[condition_idx] - self.sample_model.T[0])
            / (self.sample_model.T[1] - self.sample_model.T[0])
            * (points_amount - 1)
        )

    def __create_model(self, hyperparameters):
        layers_num_str = "layers_num"
        default_neuron_num_str = "default_neuron_num"
        default_activation_func_str = "default_activation_func"
        layers_str = "layers"

        layers_number = hyperparameters[layers_num_str]
        default_neuron_num = hyperparameters[default_neuron_num_str]
        default_activation_func = hyperparameters[default_activation_func_str]

        model = keras.models.Sequential()
        model.add(layers.InputLayer(input_shape=(1,)))

        if layers_num_str in hyperparameters:
            layers_number = hyperparameters[layers_num_str]
        for i in range(layers_number):
            neuronons_num = default_neuron_num
            activation_func = default_activation_func
            str_i = str(i)
            if str_i in hyperparameters[layers_str]:
                neuronons_num = hyperparameters[layers_str][str_i]["neuron_num"]
                activation_func = hyperparameters[layers_str][str_i]["activation_func"]
            model.add(layers.Dense(neuronons_num, activation=activation_func))
        model.add(layers.Dense((self.sample_model.dim)))
        self.model = model