populationDescent/NN_models_FMNIST.py at main · abhi0220/populationDescent · 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
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
import random
import math
import matplotlib.pyplot as plt
from matplotlib.pyplot import figure
import scipy
from scipy.special import softmax
import numpy as np

# Typing
# import typing
# from typing import TypeVar, Generic
# from collections.abc import Callable

from tqdm import tqdm
from collections import namedtuple
# from sklearn.cluster import KMeans
import statistics
import dataclasses
from dataclasses import dataclass
from tensorflow.keras.optimizers import Adam
from tensorflow.keras import datasets, layers, models
from tensorflow.keras.layers import Conv2D, MaxPooling2D, Flatten, Dense
#import keras.backend as K
import copy
# from copy import deepcopygit
import tensorflow as tf


NN_Individual = namedtuple("NN_Individual", ["nn", "opt_obj", "LR_constant", "reg_constant"])

# Testing population descent
def new_pd_NN_individual_FMNIST_without_regularization():


	# model #4 for FMNIST without regularization (for ESGD model comparison)
	model_num = "4_no_reg"
	FM_input_shape = (28, 28, 1)

	model = tf.keras.Sequential([
	tf.keras.layers.Conv2D(filters=64, kernel_size=(3,3), strides=(2,2), dilation_rate=(1,1), activation='relu', input_shape=FM_input_shape),
	tf.keras.layers.Conv2D(filters=128, kernel_size=(3,3), strides=(2,2), dilation_rate=(1,1), activation='relu'),
	tf.keras.layers.Conv2D(filters=256, kernel_size=(3,3), dilation_rate=(1,1), activation='relu'),


	tf.keras.layers.Flatten(),
	tf.keras.layers.Dense(1024),
	tf.keras.layers.Activation('relu'),
	# tf.keras.layers.Dropout(0.5),
	tf.keras.layers.Dense(10, activation='softmax')
	])


	print(model.summary())


	optimizer = tf.keras.optimizers.legacy.Adam(learning_rate=1e-3) # 1e-3 (for FMNIST, CIFAR)
	LR_constant = 10**(np.random.normal(-4, 2))
	reg_constant = 10**(np.random.normal(0, 2))

	# creating NN object with initialized parameters
	NN_object = NN_Individual(model, optimizer, LR_constant, reg_constant)

	return NN_object, model_num


# Testing population descent
def new_pd_NN_individual_FMNIST_with_regularization():
	# model #4 with regularization
	model_num = "4_with_reg"
	FM_input_shape = (28, 28, 1)
	model = tf.keras.Sequential([
	tf.keras.layers.Conv2D(filters=64, kernel_size=(3,3), strides=(2,2), dilation_rate=(1,1), activation='relu', input_shape=FM_input_shape),
	tf.keras.layers.Conv2D(filters=128, kernel_size=(3,3), strides=(2,2), dilation_rate=(1,1), activation='relu'),
	tf.keras.layers.Conv2D(filters=256, kernel_size=(3,3), dilation_rate=(1,1), activation='relu'),


	tf.keras.layers.Flatten(),
	tf.keras.layers.Dense(1024, kernel_regularizer=tf.keras.regularizers.l2(l=.001)),
	tf.keras.layers.Activation('relu'),
	# tf.keras.layers.Dropout(0.5),
	tf.keras.layers.Dense(10, activation='softmax')
	])


	print(model.summary())


	# optimizer = tf.keras.optimizers.legacy.Adam() # 1e-3 (for FMNIST, CIFAR)
	# optimizer = tf.keras.optimizers.legacy.Adam(learning_rate=1e-3) # 1e-3 (for FMNIST, CIFAR)

	optimizer = tf.keras.optimizers.legacy.Adam(learning_rate=1e-3) # 1e-3 (for FMNIST, CIFAR)
	LR_constant = 10**(np.random.normal(-4, 2))
	reg_constant = 10**(np.random.normal(0, 2))

	# creating NN object with initialized parameters
	NN_object = NN_Individual(model, optimizer, LR_constant, reg_constant)

	return NN_object, model_num


	# # model #4 with more regularization
	# model_num = "4_with_reg"
	# FM_input_shape = (28, 28, 1)
	# model = tf.keras.Sequential([
	# tf.keras.layers.Conv2D(filters=64, kernel_size=(3,3), strides=(2,2), dilation_rate=(1,1), activation='relu', input_shape=FM_input_shape, kernel_regularizer=tf.keras.regularizers.l2(l=.001)),
	# tf.keras.layers.Conv2D(filters=128, kernel_size=(3,3), strides=(2,2), dilation_rate=(1,1), activation='relu', kernel_regularizer=tf.keras.regularizers.l2(l=.001)),
	# tf.keras.layers.Conv2D(filters=256, kernel_size=(3,3), dilation_rate=(1,1), activation='relu', kernel_regularizer=tf.keras.regularizers.l2(l=.001)),


	# tf.keras.layers.Flatten(),
	# tf.keras.layers.Dense(1024, kernel_regularizer=tf.keras.regularizers.l2(l=.001)),
	# tf.keras.layers.Activation('relu'),
	# # tf.keras.layers.Dropout(0.5),
	# tf.keras.layers.Dense(10, activation='softmax')
	# ])


# Testing Hyperparameter search
def new_hps_NN_individual_FMNIST():

	# regularization_amount = [0.001]
	# learning_rate = [1e-3]

	# regularization_amount = [0.001, 0.01, 0.1]
	# learning_rate = [0.001, 0.01, 0.1]

	# regularization_amount = [0.01, 0.001, 0.0001, 0.00001, 0.000001]

	regularization_amount = [0.1]
	learning_rate = [0.01, 0.001, 0.0001, 0.00001, 0.000001]

	# regularization_amount = [0.01, 0.001, 0.0001, 0.00001, 0.000001, 5e-1, 5e-2, 5e-3, 5e-4, 5e-5]
	# learning_rate = [0.01, 0.001, 0.0001, 0.00001, 0.000001, 5e-1, 5e-2, 5e-3, 5e-4, 5e-5]


	population = []
	reg_list = []

	for r in range(len(regularization_amount)):
		for l in range(len(learning_rate)):


			# # model #4 without regularization (for ESGD model comparison)
			model_num = "4_no_reg; 5 models"
			FM_input_shape = (28, 28, 1)
			model = tf.keras.Sequential([
			tf.keras.layers.Conv2D(filters=64, kernel_size=(3,3), strides=(2,2), dilation_rate=(1,1), activation='relu', input_shape=FM_input_shape),
			tf.keras.layers.Conv2D(filters=128, kernel_size=(3,3), strides=(2,2), dilation_rate=(1,1), activation='relu'),
			tf.keras.layers.Conv2D(filters=256, kernel_size=(3,3), dilation_rate=(1,1), activation='relu'),


			tf.keras.layers.Flatten(),
			tf.keras.layers.Dense(1024),
			tf.keras.layers.Activation('relu'),
			tf.keras.layers.Dropout(0.5),
			tf.keras.layers.Dense(10, activation='softmax')
			])


			# # # model #4 with regularization (for ESGD model comparison)
			# model_num = "4_with_reg; 25 models"
			# FM_input_shape = (28, 28, 1)
			# model = tf.keras.Sequential([
			# tf.keras.layers.Conv2D(filters=64, kernel_size=(3,3), strides=(2,2), dilation_rate=(1,1), activation='relu', input_shape=FM_input_shape),
			# tf.keras.layers.Conv2D(filters=128, kernel_size=(3,3), strides=(2,2), dilation_rate=(1,1), activation='relu'),
			# tf.keras.layers.Conv2D(filters=256, kernel_size=(3,3), dilation_rate=(1,1), activation='relu'),


			# tf.keras.layers.Flatten(),
			# tf.keras.layers.Dense(1024, kernel_regularizer=tf.keras.regularizers.l2(l=regularization_amount[r])),
			# tf.keras.layers.Activation('relu'),
			# # tf.keras.layers.Dropout(0.5),
			# tf.keras.layers.Dense(10, activation='softmax')
			# ])


			optimizer = tf.keras.optimizers.legacy.Adam(learning_rate=learning_rate[l])

			model.compile(optimizer=optimizer,
			         loss=tf.keras.losses.SparseCategoricalCrossentropy(from_logits=True),
			         metrics=['accuracy'])

			population.append(model)
			reg_list.append(regularization_amount[r])

	population = np.array(population)


	return population, reg_list, model_num