-
Notifications
You must be signed in to change notification settings - Fork 4
/
outGraphs.py
290 lines (250 loc) · 12.7 KB
/
outGraphs.py
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
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
import random
from typing import List, Dict, Tuple, Union
import matplotlib.pyplot as plt
import numpy as np
from scipy.signal import savgol_filter
from sklearn.decomposition import PCA
from sklearn.preprocessing import StandardScaler
from sklearn.metrics import classification_report
from tensorflow.python.framework.ops import EagerTensor
from peakConvDeconv import recoverPeakAreas
def getHistPlot(history: Dict[str, List], title: str = '', annotate: bool = True,
marker: str = 'o') -> plt.Figure:
histPlot: plt.Figure = plt.figure()
histAx: plt.Axes = histPlot.add_subplot()
if marker is None:
histAx.plot(history["loss"], label='training')
histAx.plot(history["val_loss"], label='validation')
else:
histAx.plot(history["loss"], label='training', marker=marker)
histAx.plot(history["val_loss"], label='validation', marker=marker)
histAx.set_ylabel("Loss")
histAx.set_xlabel("Epochs")
xAxis = np.arange(len(history["loss"]))
if annotate:
histAx.set_xticks(xAxis)
for i, (train, val) in enumerate(zip(history["loss"], history["val_loss"])):
histAx.annotate(f'{round(train, 3)}', xy=(xAxis[i], train), textcoords='data')
histAx.annotate(f'{round(val, 3)}', xy=(xAxis[i], val), textcoords='data')
histAx.legend()
histAx.set_title(title)
return histPlot
def getSpectraComparisons(origSpecs: 'EagerTensor', noisySpecs: 'EagerTensor', recSpecs: 'EagerTensor',
wavenumbers: np.ndarray, title: str = '', randomIndSeed: Union[None, int] = 42,
includeSavGol: bool = True) -> Tuple[plt.Figure, plt.Figure]:
"""
Used for creating an overview of the spectra reconstruction
:param origSpecs: Tensor of original spectra
:param noisySpecs: Tensor of noisy (or distorted) spectra
:param recSpecs: Tensor of recunstructed spectra
:param wavenumbers: Wavenumbers to use for the plots
:param title: Title to give the spectra overview
:param randomIndSeed: if True, random spectra are used for plotting. If an integer is provided, it is used as
seed for the random index selection. If None,
:param includeSavGol:
:return:
"""
origSpecs: np.ndarray = tensor_to_npy2D(origSpecs)
noisySpecs: np.ndarray = tensor_to_npy2D(noisySpecs)
recSpecs: np.ndarray = tensor_to_npy2D(recSpecs)
wavenumbers = np.linspace(wavenumbers[0], wavenumbers[-1], origSpecs.shape[1])
plotIndices = []
savgolLength, savgolOrder = 15, 1
corrs = np.zeros((len(recSpecs), 2))
fig: plt.Figure = plt.figure(figsize=(14, 7))
for step in ["step1", "step2"]:
if step == "step2" and plotIndices == []:
if randomIndSeed is not None:
random.seed(randomIndSeed if type(randomIndSeed) == int else 42)
plotIndices = random.sample(range(corrs.shape[0]), 4)
else:
indGoodNN = np.argwhere(corrs[:, 0] > 90).flatten()
indBadNN = np.argwhere(corrs[:, 0] < 50).flatten()
plotIndices = list(indGoodNN[-2:]) + list(indBadNN[-2:])
for i in range(len(recSpecs)):
orig = origSpecs[i]
noisy = noisySpecs[i]
reconst = recSpecs[i]
if step == "step1":
corrNN = np.corrcoef(orig, reconst)[0, 1] * 100
if np.isnan(corrNN):
corrNN = 0
corrs[i, 0] = corrNN
savgol = savgol_filter(noisy, window_length=savgolLength, polyorder=savgolOrder)
corrSavGol = np.corrcoef(orig, savgol)[0, 1] * 100
corrs[i, 1] = corrSavGol
elif step == "step2":
if i in plotIndices:
orig -= orig.min()
orig /= orig.max()
noisy -= noisy.min()
noisy /= noisy.max()
reconst -= reconst.min()
reconst /= reconst.max()
corrNN = corrs[i, 0]
if np.isnan(corrNN):
corrNN = 0.0
plotNumber = plotIndices.index(i) + 1
ax: plt.Axes = fig.add_subplot(2, 2, plotNumber)
ax.plot(wavenumbers, noisy, color='blue')
ax.plot(wavenumbers, orig - 1, color='orange')
ax.plot(wavenumbers, reconst - 2, color='green')
if includeSavGol:
savgol = savgol_filter(noisy, window_length=savgolLength, polyorder=savgolOrder)
savgol -= savgol.min()
savgol /= savgol.max()
corrSavGol = corrs[i, 1]
ax.plot(wavenumbers, savgol - 3, color='red')
if np.isnan(corrSavGol):
corrSavGol = 0.0
ax.set_title(f"Neural net: {round(corrNN)} % Correlation\nSavGol Filter: {round(corrSavGol)} % Correlation", fontsize=13)
else:
ax.set_title(f"Neural net: {round(corrNN)} % Correlation", fontsize=13)
ax.set_xlabel("Wavenumbers (cm-1)", fontsize=12)
ax.set_yticks([])
plotNumber += 1
lines = tuple(ax.lines)
if includeSavGol:
fig.legend(lines, ('Input', 'Target', 'Neural Net', 'Savitzky-Golay'), fontsize=12)
else:
fig.legend(lines, ('Input', 'Target', 'Neural Net'), fontsize=12)
fig.suptitle(title, fontsize=15)
fig.tight_layout()
if includeSavGol:
summary = title + f'\nmean NN: {round(np.mean(corrs[:, 0]))}, mean savgol: {round(np.mean(corrs[:, 1]))}'
else:
summary = title + f'\nmean NN: {round(np.mean(corrs[:, 0]))}'
boxfig: plt.Figure = plt.figure(figsize=(4, 5))
box_ax: plt.Axes = boxfig.add_subplot()
if includeSavGol:
box_ax.boxplot(corrs, labels=['Neuronal\nNet', 'Savitzky-\nGolay'], widths=[0.6, 0.6], showfliers=False)
else:
box_ax.boxplot(corrs[:, 0], labels=['Neuronal\nNet'], widths=[0.6], showfliers=True)
box_ax.set_title(summary, fontsize=12)
box_ax.set_ylabel('Pearson Correlation (%)')
box_ax.set_ylim(min([0, corrs.min()]), 100)
boxfig.tight_layout()
return fig, boxfig
def getCorrelationPCAPlot(noisyTest: 'EagerTensor', reconstructed: 'EagerTensor',
true: 'EagerTensor', noisyTrain: 'EagerTensor') -> plt.Figure:
noisyTest: np.ndarray = tensor_to_npy2D(noisyTest)
reconstructed: np.ndarray = tensor_to_npy2D(reconstructed)
true: np.ndarray = tensor_to_npy2D(true)
noisyTrain: np.ndarray = tensor_to_npy2D(noisyTrain)
corrs: List[float] = []
for specTrue, specReconst in zip(true, reconstructed):
corrs.append(np.corrcoef(specTrue, specReconst)[0, 1])
corrs: np.ndarray = np.array(corrs)
numNoisyTest, numNoisyTrain = noisyTest.shape[0], noisyTrain.shape[0]
noisyTestPlusnoisyTrain: np.ndarray = np.vstack((noisyTest, noisyTrain))
standardScaler: StandardScaler = StandardScaler()
standardScaler.fit(noisyTestPlusnoisyTrain)
pca: PCA = PCA(n_components=3, random_state=42)
princComps: np.ndarray = pca.fit_transform(noisyTestPlusnoisyTrain)
princComps -= princComps.min()
princComps /= princComps.max()
fig: plt.Figure = plt.figure()
ax1: plt.Axes = fig.add_subplot(projection='3d')
plot = ax1.scatter(princComps[:numNoisyTest, 0], princComps[:numNoisyTest, 1], princComps[:numNoisyTest, 2], c=corrs, alpha=0.5)
ax1.set_title('PCA Map of Testing data', fontsize=14)
cb = fig.colorbar(plot)
cb.set_label("Correlation Reconstruction -> Target", fontsize=12)
fig.tight_layout()
return fig
def getCorrelationToTrainDistancePlot(noisy: 'EagerTensor', noisyEncoded: 'EagerTensor', reconstructed: 'EagerTensor',
true: 'EagerTensor', trainEncoded: 'EagerTensor', numClosestPoints: int = 5) -> plt.Figure:
noisy: np.ndarray = tensor_to_npy2D(noisy)
noisyEncoded: np.ndarray = tensor_to_npy2D(noisyEncoded)
reconstructed: np.ndarray = tensor_to_npy2D(reconstructed)
true: np.ndarray = tensor_to_npy2D(true)
trainEncoded: np.ndarray = tensor_to_npy2D(trainEncoded)
corrs: np.ndarray = np.zeros(noisy.shape[0])
for i, (specTrue, specReconst) in enumerate(zip(true, reconstructed)):
corrs[i] = np.corrcoef(specTrue, specReconst)[0, 1]
minDistances: np.ndarray = np.zeros_like(corrs)
plotDistances: np.ndarray = np.zeros_like(corrs)
origCorrs: np.ndarray = np.zeros_like(corrs)
for i in range(noisyEncoded.shape[0]):
distances = np.linalg.norm(trainEncoded - noisyEncoded[i, :], axis=1)
avgMinDist = np.mean(np.sort(distances)[:numClosestPoints])
minDistances[i] = avgMinDist
plotDistances[i] = (1000*avgMinDist**2 + corrs[i]**2)**0.5
origCorrs[i] = np.corrcoef(noisy[i, :], true[i, :])[0, 1]
fig = plt.figure()
ax = fig.add_subplot()
plot = ax.scatter(minDistances, corrs, c=origCorrs, alpha=1.0, label='testing data')
ax.set_xlabel(f"Average Distance to {numClosestPoints} closest Training Point", fontsize=12)
ax.set_ylabel("Correlation Reconstruction -> Target", fontsize=12)
ax.set_title("Distance of Testing to Training data", fontsize=14)
cb = fig.colorbar(plot)
cb.set_label("Correlation Input -> Target", fontsize=12)
fig.tight_layout()
return fig
def getPeakAreaBoxPlot(peakParams: List[List[Tuple[float, float, float]]],
reconstSpecs: np.ndarray, noisySpecs: np.ndarray) -> plt.Figure:
areaAccuraciesNN: List[float] = getDeconvolutionAccuracies(peakParams, reconstSpecs)
savGolSpecs = np.zeros_like(noisySpecs)
for i in range(noisySpecs.shape[0]):
savGolSpecs[i, :] = savgol_filter(noisySpecs[i, :], window_length=21, polyorder=4)
areaAccuraciesSG: List[float] = getDeconvolutionAccuracies(peakParams, savGolSpecs)
fig: plt.Figure = plt.figure()
ax: plt.Axes = fig.add_subplot()
ax.boxplot(np.vstack((areaAccuraciesNN, areaAccuraciesSG)).transpose(),
labels=['Neuronal\nNet', 'Savitzky-\nGolay'], widths=[0.6, 0.6], showfliers=False)
ax.set_ylabel("Accuracy of recovered peak area")
ax.set_title(f"Mean Area Accuracies:\n"
f"{round(np.mean(areaAccuraciesNN))} % Neural Net, "
f"{round(np.mean(areaAccuraciesSG))} % SG Filter")
return fig
def getDeconvolutionAccuracies(peakParamsList: List[List[Tuple[float, float, float]]], reconstSpecs: np.ndarray) -> List[float]:
areaAccuracies: List[float] = []
for peakParams, reconstSpec in zip(peakParamsList, reconstSpecs):
recoveredAreas = recoverPeakAreas(reconstSpec, peakParams)
origAreas = [i[2] for i in peakParams]
for recovered, orig in zip(recoveredAreas, origAreas):
error = abs(recovered - orig) / orig
areaAccuracies.append(100 - (error * 100))
return areaAccuracies
def tensor_to_npy2D(tensor: EagerTensor) -> np.ndarray:
"""
Converts a tensor into 2D numpy array
:param tensor:
:return: (NxM) nparray of N sampes with M features
"""
if type(tensor) == np.ndarray:
arr: np.ndarray = tensor
elif type(tensor) == EagerTensor:
arr: np.ndarray = tensor.numpy()
if len(arr.shape) == 3:
arr = arr.reshape((arr.shape[0], arr.shape[1]))
return arr
def getSpecCorrelation(reconstructedSpecs: 'EagerTensor', origNames: List[str], dbSpecs: np.ndarray, dbNames: List[str]):
"""
Gets Quality of Spectra Reconstruction by running a correlation to database spectra.
:param reconstructedSpecs: Eager Tensor of reconstructed Spectra
:param origNames: Expected Names for each reconstructed Spectrum
:param dbSpecs: (NxM) array of database Specs (M spectra with N wavenumbers)
:param dbNames: Names of spectra in database
:return:
"""
specs: np.ndarray = tensor_to_npy2D(reconstructedSpecs)
dbSpecs = dbSpecs.copy().transpose()
predictedNames: List[str] = []
for i in range(specs.shape[0]):
predictedName: str = getPredictionForSpec(specs[i, :], dbSpecs, dbNames)
predictedNames.append(predictedName)
report = classification_report(origNames, predictedNames)
return predictedNames, report
def getPredictionForSpec(intensities: np.ndarray, dbSpectra: np.ndarray, dbNames: List[str], thresh: float = 0.00) -> str:
assert dbSpectra.shape[0] == len(dbNames)
assert dbSpectra.shape[1] == len(intensities)
numDBSpecs: int = len(dbNames)
corrs: np.ndarray = np.zeros(numDBSpecs)
for i in range(numDBSpecs):
corrs[i] = np.corrcoef(intensities, dbSpectra[i, :])[0, 1]
maxCorr = corrs.max()
if maxCorr < thresh:
assignment = 'unknown'
else:
assignment = dbNames[np.argmax(corrs)]
return assignment