pyerualjetwork 4.0.5__py3-none-any.whl
Sign up to get free protection for your applications and to get access to all the features.
- pyerualjetwork/__init__.py +71 -0
- pyerualjetwork/activation_functions.py +367 -0
- pyerualjetwork/activation_functions_cuda.py +363 -0
- pyerualjetwork/data_operations.py +439 -0
- pyerualjetwork/data_operations_cuda.py +463 -0
- pyerualjetwork/help.py +16 -0
- pyerualjetwork/loss_functions.py +21 -0
- pyerualjetwork/loss_functions_cuda.py +21 -0
- pyerualjetwork/metrics.py +190 -0
- pyerualjetwork/metrics_cuda.py +165 -0
- pyerualjetwork/model_operations.py +408 -0
- pyerualjetwork/model_operations_cuda.py +414 -0
- pyerualjetwork/plan.py +681 -0
- pyerualjetwork/plan_cuda.py +677 -0
- pyerualjetwork/planeat.py +734 -0
- pyerualjetwork/planeat_cuda.py +736 -0
- pyerualjetwork/ui.py +22 -0
- pyerualjetwork/visualizations.py +799 -0
- pyerualjetwork/visualizations_cuda.py +799 -0
- pyerualjetwork-4.0.5.dist-info/METADATA +95 -0
- pyerualjetwork-4.0.5.dist-info/RECORD +23 -0
- pyerualjetwork-4.0.5.dist-info/WHEEL +5 -0
- pyerualjetwork-4.0.5.dist-info/top_level.txt +1 -0
@@ -0,0 +1,439 @@
|
|
1
|
+
from tqdm import tqdm
|
2
|
+
import numpy as np
|
3
|
+
from colorama import Fore, Style
|
4
|
+
import sys
|
5
|
+
import math
|
6
|
+
|
7
|
+
def encode_one_hot(y_train, y_test=None, summary=False):
|
8
|
+
"""
|
9
|
+
Performs one-hot encoding on y_train and y_test data.
|
10
|
+
|
11
|
+
Args:
|
12
|
+
y_train (numpy.ndarray): Train label data.
|
13
|
+
y_test (numpy.ndarray): Test label data one-hot encoded. (optional).
|
14
|
+
summary (bool): If True, prints the class-to-index mapping. Default: False
|
15
|
+
|
16
|
+
Returns:
|
17
|
+
tuple: One-hot encoded y_train and (if given) y_test.
|
18
|
+
"""
|
19
|
+
if len(y_train) < 256:
|
20
|
+
if y_train.dtype != np.uint8:
|
21
|
+
y_train = np.array(y_train, copy=False).astype(np.uint8, copy=False)
|
22
|
+
elif len(y_train) <= 32767:
|
23
|
+
if y_train.dtype != np.uint16:
|
24
|
+
y_train = np.array(y_train, copy=False).astype(np.uint16, copy=False)
|
25
|
+
else:
|
26
|
+
if y_train.dtype != np.uint32:
|
27
|
+
y_train = np.array(y_train, copy=False).astype(np.uint32, copy=False)
|
28
|
+
|
29
|
+
if y_test is not None:
|
30
|
+
if len(y_test) < 256:
|
31
|
+
if y_test.dtype != np.uint8:
|
32
|
+
y_test = np.array(y_test, copy=False).astype(np.uint8, copy=False)
|
33
|
+
elif len(y_test) <= 32767:
|
34
|
+
if y_test.dtype != np.uint16:
|
35
|
+
y_test = np.array(y_test, copy=False).astype(np.uint16, copy=False)
|
36
|
+
else:
|
37
|
+
if y_test.dtype != np.uint32:
|
38
|
+
y_test = np.array(y_test, copy=False).astype(np.uint32, copy=False)
|
39
|
+
|
40
|
+
classes = np.unique(y_train)
|
41
|
+
class_count = len(classes)
|
42
|
+
|
43
|
+
class_to_index = {cls: idx for idx, cls in enumerate(classes)}
|
44
|
+
|
45
|
+
if summary:
|
46
|
+
print("Class-to-index mapping:")
|
47
|
+
for cls, idx in class_to_index.items():
|
48
|
+
print(f" {idx}: {cls}")
|
49
|
+
|
50
|
+
y_train_encoded = np.zeros((y_train.shape[0], class_count))
|
51
|
+
for i, label in enumerate(y_train):
|
52
|
+
y_train_encoded[i, class_to_index[label]] = 1
|
53
|
+
|
54
|
+
if y_test is not None:
|
55
|
+
y_test_encoded = np.zeros((y_test.shape[0], class_count))
|
56
|
+
for i, label in enumerate(y_test):
|
57
|
+
y_test_encoded[i, class_to_index[label]] = 1
|
58
|
+
return y_train_encoded, y_test_encoded
|
59
|
+
|
60
|
+
return y_train_encoded
|
61
|
+
|
62
|
+
|
63
|
+
def decode_one_hot(encoded_data):
|
64
|
+
"""
|
65
|
+
Decodes one-hot encoded data to original categorical labels.
|
66
|
+
|
67
|
+
Args:
|
68
|
+
encoded_data (numpy.ndarray): One-hot encoded data with shape (n_samples, n_classes).
|
69
|
+
|
70
|
+
Returns:
|
71
|
+
numpy.ndarray: Decoded categorical labels with shape (n_samples,).
|
72
|
+
"""
|
73
|
+
|
74
|
+
decoded_labels = np.argmax(encoded_data, axis=1)
|
75
|
+
|
76
|
+
return decoded_labels
|
77
|
+
|
78
|
+
|
79
|
+
def split(X, y, test_size, random_state, dtype=np.float32):
|
80
|
+
"""
|
81
|
+
Splits the given X (features) and y (labels) data into training and testing subsets.
|
82
|
+
|
83
|
+
Args:
|
84
|
+
X (numpy.ndarray): Features data.
|
85
|
+
y (numpy.ndarray): Labels data.
|
86
|
+
test_size (float or int): Proportion or number of samples for the test subset.
|
87
|
+
random_state (int or None): Seed for random state.
|
88
|
+
dtype (numpy.dtype): Data type for the arrays. np.float32 by default. Example: np.float64 or np.float16. [fp32 for balanced devices, fp64 for strong devices, fp16 for weak devices: not reccomended!]
|
89
|
+
|
90
|
+
Returns:
|
91
|
+
tuple: x_train, x_test, y_train, y_test as ordered training and testing data subsets.
|
92
|
+
"""
|
93
|
+
X = X.astype(dtype)
|
94
|
+
if len(y) < 256:
|
95
|
+
if y.dtype != np.uint8:
|
96
|
+
y = np.array(y, copy=False).astype(np.uint8, copy=False)
|
97
|
+
elif len(y) <= 32767:
|
98
|
+
if y.dtype != np.uint16:
|
99
|
+
y = np.array(y, copy=False).astype(np.uint16, copy=False)
|
100
|
+
else:
|
101
|
+
if y.dtype != np.uint32:
|
102
|
+
y = np.array(y, copy=False).astype(np.uint32, copy=False)
|
103
|
+
|
104
|
+
num_samples = X.shape[0]
|
105
|
+
|
106
|
+
if isinstance(test_size, float):
|
107
|
+
test_size = int(test_size * num_samples)
|
108
|
+
elif isinstance(test_size, int):
|
109
|
+
if test_size > num_samples:
|
110
|
+
raise ValueError(
|
111
|
+
"test_size cannot be larger than the number of samples.")
|
112
|
+
else:
|
113
|
+
raise ValueError("test_size should be float or int.")
|
114
|
+
|
115
|
+
if random_state is not None:
|
116
|
+
np.random.seed(random_state)
|
117
|
+
|
118
|
+
indices = np.arange(num_samples)
|
119
|
+
np.random.shuffle(indices)
|
120
|
+
|
121
|
+
test_indices = indices[:test_size]
|
122
|
+
train_indices = indices[test_size:]
|
123
|
+
|
124
|
+
x_train, x_test = X[train_indices], X[test_indices]
|
125
|
+
y_train, y_test = y[train_indices], y[test_indices]
|
126
|
+
|
127
|
+
return x_train, x_test, y_train, y_test
|
128
|
+
|
129
|
+
|
130
|
+
def manuel_balancer(x_train, y_train, target_samples_per_class, dtype=np.float32):
|
131
|
+
"""
|
132
|
+
Generates synthetic examples to balance classes to the specified number of examples per class.
|
133
|
+
|
134
|
+
Arguments:
|
135
|
+
x_train -- Input dataset (examples) - NumPy array format
|
136
|
+
y_train -- Class labels (one-hot encoded) - NumPy array format
|
137
|
+
target_samples_per_class -- Desired number of samples per class
|
138
|
+
dtype (numpy.dtype): Data type for the arrays. np.float32 by default. Example: np.float64 or np.float16. [fp32 for balanced devices, fp64 for strong devices, fp16 for weak devices: not reccomended!]
|
139
|
+
|
140
|
+
Returns:
|
141
|
+
x_balanced -- Balanced input dataset (NumPy array format)
|
142
|
+
y_balanced -- Balanced class labels (one-hot encoded, NumPy array format)
|
143
|
+
"""
|
144
|
+
from .ui import loading_bars
|
145
|
+
|
146
|
+
bar_format = loading_bars()[0]
|
147
|
+
|
148
|
+
x_train = np.array(x_train, copy=False).astype(dtype, copy=False)
|
149
|
+
if len(y_train[0]) < 256:
|
150
|
+
if y_train.dtype != np.uint8:
|
151
|
+
y_train = np.array(y_train, copy=False).astype(np.uint8, copy=False)
|
152
|
+
elif len(y_train[0]) <= 32767:
|
153
|
+
if y_train.dtype != np.uint16:
|
154
|
+
y_train = np.array(y_train, copy=False).astype(np.uint16, copy=False)
|
155
|
+
else:
|
156
|
+
if y_train.dtype != np.uint32:
|
157
|
+
y_train = np.array(y_train, copy=False).astype(np.uint32, copy=False)
|
158
|
+
|
159
|
+
classes = np.arange(y_train.shape[1])
|
160
|
+
class_count = len(classes)
|
161
|
+
|
162
|
+
x_balanced = []
|
163
|
+
y_balanced = []
|
164
|
+
|
165
|
+
for class_label in tqdm(range(class_count),leave=False, ascii="▱▰",
|
166
|
+
bar_format=bar_format,desc='Augmenting Data',ncols= 52):
|
167
|
+
class_indices = np.where(np.argmax(y_train, axis=1) == class_label)[0]
|
168
|
+
num_samples = len(class_indices)
|
169
|
+
|
170
|
+
if num_samples > target_samples_per_class:
|
171
|
+
|
172
|
+
selected_indices = np.random.choice(class_indices, target_samples_per_class, replace=False)
|
173
|
+
x_balanced.append(x_train[selected_indices])
|
174
|
+
y_balanced.append(y_train[selected_indices])
|
175
|
+
|
176
|
+
else:
|
177
|
+
|
178
|
+
x_balanced.append(x_train[class_indices])
|
179
|
+
y_balanced.append(y_train[class_indices])
|
180
|
+
|
181
|
+
if num_samples < target_samples_per_class:
|
182
|
+
|
183
|
+
samples_to_add = target_samples_per_class - num_samples
|
184
|
+
additional_samples = np.zeros((samples_to_add, x_train.shape[1]))
|
185
|
+
additional_labels = np.zeros((samples_to_add, y_train.shape[1]))
|
186
|
+
|
187
|
+
for i in range(samples_to_add):
|
188
|
+
|
189
|
+
random_indices = np.random.choice(class_indices, 2, replace=False)
|
190
|
+
sample1 = x_train[random_indices[0]]
|
191
|
+
sample2 = x_train[random_indices[1]]
|
192
|
+
|
193
|
+
|
194
|
+
synthetic_sample = sample1 + (sample2 - sample1) * np.random.rand()
|
195
|
+
|
196
|
+
additional_samples[i] = synthetic_sample
|
197
|
+
additional_labels[i] = y_train[class_indices[0]]
|
198
|
+
|
199
|
+
|
200
|
+
x_balanced.append(additional_samples)
|
201
|
+
y_balanced.append(additional_labels)
|
202
|
+
|
203
|
+
x_balanced = np.vstack(x_balanced)
|
204
|
+
y_balanced = np.vstack(y_balanced)
|
205
|
+
|
206
|
+
return x_balanced.astype(dtype), y_balanced.astype(dtype)
|
207
|
+
|
208
|
+
|
209
|
+
def auto_balancer(x_train, y_train, dtype=np.float32):
|
210
|
+
|
211
|
+
"""
|
212
|
+
Function to balance the training data across different classes.
|
213
|
+
|
214
|
+
Arguments:
|
215
|
+
x_train (list): Input data for training.
|
216
|
+
y_train (list): Labels corresponding to the input data. one-hot encoded.
|
217
|
+
dtype (numpy.dtype): Data type for the arrays. np.float32 by default. Example: np.float64 or np.float16. [fp32 for balanced devices, fp64 for strong devices, fp16 for weak devices: not reccomended!]
|
218
|
+
|
219
|
+
Returns:
|
220
|
+
tuple: A tuple containing balanced input data and labels.
|
221
|
+
"""
|
222
|
+
from .ui import loading_bars
|
223
|
+
|
224
|
+
bar_format = loading_bars()[0]
|
225
|
+
|
226
|
+
x_train = np.array(x_train, copy=False).astype(dtype, copy=False)
|
227
|
+
if len(y_train[0]) < 256:
|
228
|
+
if y_train.dtype != np.uint8:
|
229
|
+
y_train = np.array(y_train, copy=False).astype(np.uint8, copy=False)
|
230
|
+
elif len(y_train[0]) <= 32767:
|
231
|
+
if y_train.dtype != np.uint16:
|
232
|
+
y_train = np.array(y_train, copy=False).astype(np.uint16, copy=False)
|
233
|
+
else:
|
234
|
+
if y_train.dtype != np.uint32:
|
235
|
+
y_train = np.array(y_train, copy=False).astype(np.uint32, copy=False)
|
236
|
+
|
237
|
+
classes = np.arange(y_train.shape[1])
|
238
|
+
class_count = len(classes)
|
239
|
+
|
240
|
+
try:
|
241
|
+
ClassIndices = {i: np.where(y_train[:, i] == 1)[
|
242
|
+
0] for i in range(class_count)}
|
243
|
+
classes = [len(ClassIndices[i]) for i in range(class_count)]
|
244
|
+
|
245
|
+
if len(set(classes)) == 1:
|
246
|
+
print(Fore.WHITE + "INFO: Data have already balanced. from: auto_balancer" + Style.RESET_ALL)
|
247
|
+
return x_train, y_train
|
248
|
+
|
249
|
+
MinCount = min(classes)
|
250
|
+
|
251
|
+
BalancedIndices = []
|
252
|
+
for i in tqdm(range(class_count),leave=False, ascii="▱▰",
|
253
|
+
bar_format= bar_format, desc='Balancing Data',ncols=70):
|
254
|
+
if len(ClassIndices[i]) > MinCount:
|
255
|
+
SelectedIndices = np.random.choice(
|
256
|
+
ClassIndices[i], MinCount, replace=False)
|
257
|
+
else:
|
258
|
+
SelectedIndices = ClassIndices[i]
|
259
|
+
BalancedIndices.extend(SelectedIndices)
|
260
|
+
|
261
|
+
BalancedInputs = [x_train[idx] for idx in BalancedIndices]
|
262
|
+
BalancedLabels = [y_train[idx] for idx in BalancedIndices]
|
263
|
+
|
264
|
+
permutation = np.random.permutation(len(BalancedInputs))
|
265
|
+
BalancedInputs = np.array(BalancedInputs)[permutation]
|
266
|
+
BalancedLabels = np.array(BalancedLabels)[permutation]
|
267
|
+
|
268
|
+
print(Fore.GREEN + "Data Succesfully Balanced from: " + str(len(x_train)
|
269
|
+
) + " to: " + str(len(BalancedInputs)) + ". from: auto_balancer " + Style.RESET_ALL)
|
270
|
+
except:
|
271
|
+
print(Fore.RED + "ERROR: Inputs and labels must be same length check parameters")
|
272
|
+
sys.exit()
|
273
|
+
|
274
|
+
return BalancedInputs.astype(dtype, copy=False), BalancedLabels.astype(dtype, copy=False)
|
275
|
+
|
276
|
+
|
277
|
+
def synthetic_augmentation(x_train, y_train, dtype=np.float32):
|
278
|
+
"""
|
279
|
+
Generates synthetic examples to balance classes with fewer examples.
|
280
|
+
|
281
|
+
Arguments:
|
282
|
+
x -- Input dataset (examples) - array format
|
283
|
+
y -- Class labels (one-hot encoded) - array format
|
284
|
+
dtype (numpy.dtype): Data type for the arrays. np.float32 by default. Example: np.float64 or np.float16. [fp32 for balanced devices, fp64 for strong devices, fp16 for weak devices: not reccomended!]
|
285
|
+
|
286
|
+
Returns:
|
287
|
+
x_balanced -- Balanced input dataset (array format)
|
288
|
+
y_balanced -- Balanced class labels (one-hot encoded, array format)
|
289
|
+
"""
|
290
|
+
from .ui import loading_bars
|
291
|
+
|
292
|
+
bar_format = loading_bars()[0]
|
293
|
+
|
294
|
+
x_train = x_train.astype(dtype)
|
295
|
+
if len(y_train[0]) < 256:
|
296
|
+
if y_train.dtype != np.uint8:
|
297
|
+
y_train = np.array(y_train, copy=False).astype(np.uint8, copy=False)
|
298
|
+
elif len(y_train[0]) <= 32767:
|
299
|
+
if y_train.dtype != np.uint16:
|
300
|
+
y_train = np.array(y_train, copy=False).astype(np.uint16, copy=False)
|
301
|
+
else:
|
302
|
+
if y_train.dtype != np.uint32:
|
303
|
+
y_train = np.array(y_train, copy=False).astype(np.uint32, copy=False)
|
304
|
+
|
305
|
+
x = x_train
|
306
|
+
y = y_train
|
307
|
+
|
308
|
+
classes = np.arange(y_train.shape[1])
|
309
|
+
class_count = len(classes)
|
310
|
+
|
311
|
+
class_distribution = {i: 0 for i in range(class_count)}
|
312
|
+
for label in y:
|
313
|
+
class_distribution[np.argmax(label)] += 1
|
314
|
+
|
315
|
+
max_class_count = max(class_distribution.values())
|
316
|
+
|
317
|
+
x_balanced = list(x)
|
318
|
+
y_balanced = list(y)
|
319
|
+
|
320
|
+
|
321
|
+
for class_label in tqdm(range(class_count), leave=False, ascii="▱▰",
|
322
|
+
bar_format=bar_format,desc='Augmenting Data',ncols= 52):
|
323
|
+
class_indices = [i for i, label in enumerate(
|
324
|
+
y) if np.argmax(label) == class_label]
|
325
|
+
num_samples = len(class_indices)
|
326
|
+
|
327
|
+
if num_samples < max_class_count:
|
328
|
+
while num_samples < max_class_count:
|
329
|
+
|
330
|
+
random_indices = np.random.choice(
|
331
|
+
class_indices, 2, replace=False)
|
332
|
+
sample1 = x[random_indices[0]]
|
333
|
+
sample2 = x[random_indices[1]]
|
334
|
+
|
335
|
+
synthetic_sample = sample1 + \
|
336
|
+
(np.array(sample2) - np.array(sample1)) * np.random.rand()
|
337
|
+
|
338
|
+
x_balanced.append(synthetic_sample.tolist())
|
339
|
+
y_balanced.append(y[class_indices[0]])
|
340
|
+
|
341
|
+
num_samples += 1
|
342
|
+
|
343
|
+
|
344
|
+
return np.array(x_balanced).astype(dtype, copy=False), np.array(y_balanced).astype(dtype, copy=False)
|
345
|
+
|
346
|
+
|
347
|
+
def standard_scaler(x_train=None, x_test=None, scaler_params=None, dtype=np.float32):
|
348
|
+
"""
|
349
|
+
Standardizes training and test datasets. x_test may be None.
|
350
|
+
|
351
|
+
Args:
|
352
|
+
x_train: numpy.ndarray
|
353
|
+
|
354
|
+
x_test: numpy.ndarray (optional)
|
355
|
+
|
356
|
+
scaler_params (optional for using model)
|
357
|
+
|
358
|
+
dtype (numpy.dtype): Data type for the arrays. np.float32 by default. Example: np.float64 or np.float16. [fp32 for balanced devices, fp64 for strong devices, fp16 for weak devices: not reccomended!]
|
359
|
+
|
360
|
+
Returns:
|
361
|
+
list:
|
362
|
+
Scaler parameters: mean and std
|
363
|
+
tuple
|
364
|
+
Standardized training and test datasets
|
365
|
+
"""
|
366
|
+
if x_train is not None and scaler_params is None and x_test is not None:
|
367
|
+
x_train = x_train.astype(dtype, copy=False)
|
368
|
+
x_test = x_test.astype(dtype, copy=False)
|
369
|
+
|
370
|
+
mean = np.mean(x_train, axis=0)
|
371
|
+
std = np.std(x_train, axis=0)
|
372
|
+
|
373
|
+
train_data_scaled = (x_train - mean) / std
|
374
|
+
test_data_scaled = (x_test - mean) / std
|
375
|
+
|
376
|
+
train_data_scaled = np.nan_to_num(train_data_scaled, nan=0)
|
377
|
+
test_data_scaled = np.nan_to_num(test_data_scaled, nan=0)
|
378
|
+
|
379
|
+
scaler_params = [mean, std]
|
380
|
+
|
381
|
+
return scaler_params, train_data_scaled, test_data_scaled
|
382
|
+
|
383
|
+
if scaler_params is None and x_train is None and x_test is not None:
|
384
|
+
return x_test.astype(dtype, copy=False) # sample data not scaled
|
385
|
+
|
386
|
+
if scaler_params is not None:
|
387
|
+
x_test = x_test.astype(dtype, copy=False)
|
388
|
+
scaled_data = (x_test - scaler_params[0]) / scaler_params[1]
|
389
|
+
scaled_data = np.nan_to_num(scaled_data, nan=0)
|
390
|
+
|
391
|
+
return scaled_data # sample data scaled
|
392
|
+
|
393
|
+
|
394
|
+
def normalization(
|
395
|
+
Input, # num: Input data to be normalized.
|
396
|
+
dtype=np.float32):
|
397
|
+
"""
|
398
|
+
Normalizes the input data using maximum absolute scaling.
|
399
|
+
|
400
|
+
Args:
|
401
|
+
Input (num): Input data to be normalized.
|
402
|
+
dtype (numpy.dtype): Data type for the arrays. np.float32 by default. Example: np.float64 or np.float16. [fp32 for balanced devices, fp64 for strong devices, fp16 for weak devices: not reccomended!]
|
403
|
+
|
404
|
+
Returns:
|
405
|
+
(num) Scaled input data after normalization.
|
406
|
+
"""
|
407
|
+
|
408
|
+
MaxAbs = np.max(np.abs(Input.astype(dtype, copy=False)))
|
409
|
+
return (Input / MaxAbs)
|
410
|
+
|
411
|
+
|
412
|
+
def find_closest_factors(a):
|
413
|
+
|
414
|
+
root = int(math.sqrt(a))
|
415
|
+
|
416
|
+
for i in range(root, 0, -1):
|
417
|
+
if a % i == 0:
|
418
|
+
j = a // i
|
419
|
+
return i, j
|
420
|
+
|
421
|
+
|
422
|
+
def batcher(x_test, y_test, batch_size=1):
|
423
|
+
|
424
|
+
y_labels = np.argmax(y_test, axis=1)
|
425
|
+
|
426
|
+
sampled_x, sampled_y = [], []
|
427
|
+
|
428
|
+
for class_label in np.unique(y_labels):
|
429
|
+
|
430
|
+
class_indices = np.where(y_labels == class_label)[0]
|
431
|
+
|
432
|
+
num_samples = int(len(class_indices) * batch_size)
|
433
|
+
|
434
|
+
sampled_indices = np.random.choice(class_indices, num_samples, replace=False)
|
435
|
+
|
436
|
+
sampled_x.append(x_test[sampled_indices])
|
437
|
+
sampled_y.append(y_test[sampled_indices])
|
438
|
+
|
439
|
+
return np.concatenate(sampled_x), np.concatenate(sampled_y)
|