afp_ai/load_models.py at main · anserko/afp_ai · 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
import numpy as np
import tensorflow as tf
from tensorflow.keras import layers, models
import json

from build_model_lstm import build_model_lstm_3
from build_model_classifier import build_model_classifier_3
from get_encoder_decoder import get_encoder_decoder

def load_encoder_decoder_model(model_ed_name, model_ed_path):
	with open(f'{model_ed_path}/{model_ed_name}_settings.json', 'r') as f:
		ed_settings_dict = json.load(f)

	input_shape = ed_settings_dict['input_shape']
	dense_units = ed_settings_dict['dense_units']
	ifBatchNorm = ed_settings_dict['ifBatchNorm']
	kernel_size = ed_settings_dict['kernel_size']
	nn_blocks = ed_settings_dict['nn_blocks']

	ifEncoder = True
	ifDecoder = True
	ifSummary = False
	loss = 'mean_squared_error'
	optimizer = tf.keras.optimizers.legacy.Adadelta(learning_rate=0.1, name="Adadelta")

	model_settings = (input_shape, dense_units,ifBatchNorm,kernel_size,nn_blocks)
	model_weights_path = f'{model_ed_path}/{model_ed_name}/{model_ed_name}'

	model_encoder, model_decoder = get_encoder_decoder(
		model_ed_name,
		model_weights_path,
		model_settings,
		loss, optimizer,
		ifEncoder=ifEncoder,
		ifDecoder=ifDecoder,
		ifSummary=ifSummary
		)

	return model_encoder, model_decoder

def load_lstm_model(model_lstm_name, model_lstm_path):

	with open(f'{model_lstm_path}/{model_lstm_name}_settings.json', 'r') as f:
		lstm_settings_dict_test = json.load(f)

	unit_numb = lstm_settings_dict_test['unit_numb']
	cells_list = lstm_settings_dict_test['cells_list']
	ifDense = lstm_settings_dict_test['ifDense']
	many_to_many = lstm_settings_dict_test['many_to_many']
	stateful = lstm_settings_dict_test['stateful']

	input_shape = (None, unit_numb)

	model = build_model_lstm_3(
		input_shape,
		cells_list,
		ifDense=ifDense,
		ifDropout=False,
		many_to_many=many_to_many
		)
	model.stateful = stateful

	loss = 'mean_squared_error'
	optimizer = tf.keras.optimizers.legacy.Adadelta(learning_rate=0.1, name="Adadelta")
	model.compile(loss=loss, optimizer=optimizer)

	model.load_weights(f'{model_lstm_path}/{model_lstm_name}/{model_lstm_name}').expect_partial()

	return model

def load_classification_model(model_name, model_path, ifProbability):

	with open(f'{model_path}/{model_name}_settings.json', 'r') as f:
		classification_settings_dict_test = json.load(f)

	optimizer = 'adam'
	loss = tf.keras.losses.SparseCategoricalCrossentropy(from_logits=True)

	#model = model = build_model_classifier_3(classification_settings_dict_test)
	model = build_model_classifier_3(classification_settings_dict_test)
	model.compile(
		loss=loss,
		optimizer=optimizer,
		metrics=['accuracy']
	)

	model.load_weights(f'{model_path}/{model_name}/{model_name}').expect_partial()

	if ifProbability:
		output_model = tf.keras.Sequential([model, tf.keras.layers.Softmax()])
	else:
		output_model = model
	return output_model

def load_models(unit_numb, stride_step, lstm_pars_dict, classifier=False, ifProbability=True):

	models_dict={}

	#load encoder decoder models
	model_ed_name = f'model_ed_2_ps_2_bn_True_du_{unit_numb}'
	model_ed_path = f'saved_models/stride_{stride_step}/ed'
	model_encoder, model_decoder = load_encoder_decoder_model(model_ed_name, model_ed_path)

	#load lstm models
	stateful = lstm_pars_dict['stateful']
	window_size = lstm_pars_dict['window_size']
	window_size_predicted = lstm_pars_dict['window_size_predicted']
	overlap = lstm_pars_dict['overlap']
	lstm_type = lstm_pars_dict['lstm_type']

	if window_size_predicted==1:
		if lstm_type == 'large_quasi_state_win':
			cells_list = [unit_numb*4]
		else:
			cells_list = [unit_numb*2]
	else:
		if lstm_type == 'large_quasi_state_win':
			cells_list = [unit_numb*4,unit_numb*4,unit_numb*4]
		else:
			cells_list = [unit_numb*2,unit_numb*2]
	cells_list_str = '_'.join(str(x) for x in cells_list)

	model_lstm_name = f'model_lstm_{lstm_type}_{window_size}_{window_size_predicted}_ovrp_{overlap}_un_{unit_numb}_c_{cells_list_str}'
	#model_lstm_name = f'model_lstm_large_quasi_state_win_{window_size}_{window_size_predicted}_ovrp_{overlap}_un_{unit_numb}_c_{cells_list_str}'
	#model_lstm_name = f'model_lstm_quasi_state_win_{window_size}_{window_size_predicted}_ovrp_{overlap}_un_{unit_numb}_c_{cells_list_str}'
	model_lstm_path = f'saved_models/stride_{stride_step}/lstm/lstm_state_{stateful}'
	model_lstm = load_lstm_model(model_lstm_name, model_lstm_path)


	models_dict['model_encoder'] = model_encoder
	models_dict['model_decoder'] = model_decoder
	models_dict['model_lstm'] = model_lstm

	if classifier:
		model_classifier_name = f'model_class_str_{stride_step}_un_{unit_numb}'
		model_classifier_path = f'saved_models/stride_{stride_step}/classify'
		model_classifier = load_classification_model(model_classifier_name, model_classifier_path, ifProbability=ifProbability)
		models_dict['model_classifier'] = model_classifier

	return models_dict