-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathanalyse_masterarbeit_paul.py
278 lines (217 loc) · 13.1 KB
/
analyse_masterarbeit_paul.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
import pandas as pd
import torch
import numpy as np
import scipy as sp
import scipy.stats
from paule import models
#CONDITION = 'recording-embedder'
CONDITION = 'segment-embedder'
OBJECTIVES = ('acoustic', 'semvec', 'init-seg')
LABEL_VECTORS = pd.read_pickle('data/label_vectors_vectors_checked.pkl')
# load embedder for evaluation
embedder = models.MelEmbeddingModel_MelSmoothResidualUpsampling(mel_smooth_layers=3).double()
#embedder.load_state_dict(torch.load("../paule/paule/pretrained_models/embedder/model_recorded_embed_model_3_4_180_8192_rmse_lr_00001_400.pt", map_location=torch.device('cpu')))
embedder.load_state_dict(torch.load("../paule/paule/pretrained_models/embedder/model_synthesized_embed_model_3_4_180_8192_rmse_lr_00001_400.pt", map_location=torch.device('cpu')))
# load and concatenate data
dats = dict()
for objective in OBJECTIVES:
dat1 = pd.read_pickle(f'results_20210911_0838_masterarbeit_paul/dat_{objective}.pickle')
dat2 = pd.read_pickle(f'results_20210911_1012_masterarbeit_paul/dat_{objective}.pickle')
dat3 = pd.read_pickle(f'results_20210910_2201_masterarbeit_paul/dat_{objective}.pickle')
dats[objective] = pd.concat((dat1, dat2, dat3))
del dat1, dat2, dat3
def mel_wasserstein_distance(mel1, mel2):
"""
perform the 1d Wasserstein distance function over mel bands, time points and energy
:param mel1: np.array
log-mel spectrogram (seq_length, mel channels)
:param: mel1: np.array
log-mel spectrogram (seq_length, mel channels)
:return mean_time_dist, meean_mel_dist, energy_dist: np.float, np.float, np.float
average 1d distance over time, mel_channel and energy
"""
assert mel1.shape == mel2.shape
if isinstance(mel1, np.ndarray):
mel1 = torch.from_numpy(mel1)
if isinstance(mel2, np.ndarray):
mel2 = torch.from_numpy(mel2)
time_dist = []
mel_dist = []
for time_point in range(mel1.shape[0]):
time_dist.append(sp.stats.wasserstein_distance(mel1[time_point],mel2[time_point]))
for mel_channel in range(mel1.shape[1]):
mel_dist.append(sp.stats.wasserstein_distance(mel1[:,mel_channel], mel2[:,mel_channel]))
energy1 = torch.mean(mel1, axis=1)
energy2 = torch.mean(mel2, axis=1)
energy_dist = sp.stats.wasserstein_distance(energy1, energy2)
#return np.mean(time_dist), np.mean(mel_dist), energy_dist
return np.mean((np.mean(time_dist), np.mean(mel_dist), energy_dist))
def get_semvec(mel):
mel = mel.copy()
mel.shape = (1,) + mel.shape
mel = torch.from_numpy(mel)
seq_length = mel.shape[1]
with torch.no_grad():
semvec = embedder(mel, (torch.tensor(seq_length),))
semvec = semvec.cpu().numpy()
semvec.shape = (semvec.shape[1],)
return semvec
def rmse(array1, array2):
# eps = 1e-6
return np.sqrt(np.mean((array2 - array1) ** 2) + 1e-6)
def rmse_mel_seg(row):
rec_mel = row.rec_mel
seg_mel = row.seg_mel
half_shift = int((seg_mel.shape[0] - rec_mel.shape[0]) / 2)
seg_mel = seg_mel[half_shift:-half_shift]
if rec_mel.shape[0] < seg_mel.shape[0]:
seg_mel = seg_mel[:rec_mel.shape[0]]
return rmse(seg_mel, rec_mel)
def mel_wasserstein_distance_mel_seg(row):
rec_mel = row.rec_mel
seg_mel = row.seg_mel
half_shift = int((seg_mel.shape[0] - rec_mel.shape[0]) / 2)
seg_mel = seg_mel[half_shift:-half_shift]
if rec_mel.shape[0] < seg_mel.shape[0]:
seg_mel = seg_mel[:rec_mel.shape[0]]
return mel_wasserstein_distance(seg_mel, rec_mel)
def rank(label, vector):
corr = np.array([np.correlate(vector, vec2) for vec2 in LABEL_VECTORS['vector']])
return int(sp.stats.rankdata(-corr, method='min')[LABEL_VECTORS.label == 'Problem'])
for dat in dats.values():
dat['rec_vec'] = None
dat['seg_vec'] = None
dat['inv_vec'] = None
dat['prod_vec_acoustic'] = None
dat['prod_vec_acoustic_semvec'] = None
dat['prod_vec_semvec'] = None
dat['rmse_mel_rec'] = None
dat['rmse_mel_seg'] = None
dat['rmse_mel_inv'] = None
dat['rmse_mel_acoustic'] = None
dat['rmse_mel_acoustic_semvec'] = None
dat['rmse_mel_semvec'] = None
dat['rmse_vec_rec'] = None
dat['rmse_vec_seg'] = None
dat['rmse_vec_inv'] = None
dat['rmse_vec_acoustic'] = None
dat['rmse_vec_acoustic_semvec'] = None
dat['rmse_vec_semvec'] = None
dat['wasser_mel_rec'] = None
dat['wasser_mel_seg'] = None
dat['wasser_mel_inv'] = None
dat['wasser_mel_acoustic'] = None
dat['wasser_mel_acoustic_semvec'] = None
dat['wasser_mel_semvec'] = None
# vec (semantic vectors)
dat['rec_vec'] = dat.rec_mel.apply(get_semvec)
dat['seg_vec'] = dat.seg_mel.apply(get_semvec)
dat['inv_vec'] = dat.inv_mel.apply(get_semvec)
dat['prod_vec_acoustic'] = dat.prod_mel_acoustic.apply(get_semvec)
dat['prod_vec_acoustic_semvec'] = dat.prod_mel_acoustic_semvec.apply(get_semvec)
dat['prod_vec_semvec'] = dat.prod_mel_semvec.apply(get_semvec)
# rmse mel
dat['rmse_mel_rec'] = dat.apply(lambda row: rmse(row.rec_mel, row.rec_mel), axis=1)
dat['rmse_mel_inv'] = dat.apply(lambda row: rmse(row.rec_mel, row.inv_mel), axis=1)
dat['rmse_mel_acoustic'] = dat.apply(lambda row: rmse(row.rec_mel, row.prod_mel_acoustic), axis=1)
dat['rmse_mel_acoustic_semvec'] = dat.apply(lambda row: rmse(row.rec_mel, row.prod_mel_acoustic_semvec), axis=1)
dat['rmse_mel_semvec'] = dat.apply(lambda row: rmse(row.rec_mel, row.prod_mel_semvec), axis=1)
dat['rmse_mel_seg'] = dat.apply(rmse_mel_seg, axis=1)
# rmse vec
dat['rmse_vec_rec'] = dat.apply(lambda row: rmse(row.vector, row.rec_vec), axis=1)
dat['rmse_vec_inv'] = dat.apply(lambda row: rmse(row.vector, row.inv_vec), axis=1)
dat['rmse_vec_acoustic'] = dat.apply(lambda row: rmse(row.vector, row.prod_vec_acoustic), axis=1)
dat['rmse_vec_acoustic_semvec'] = dat.apply(lambda row: rmse(row.vector, row.prod_vec_acoustic_semvec), axis=1)
dat['rmse_vec_semvec'] = dat.apply(lambda row: rmse(row.vector, row.prod_vec_semvec), axis=1)
dat['rmse_vec_seg'] = dat.apply(lambda row: rmse(row.vector, row.seg_vec), axis=1)
# wasserstein mel
dat['wasser_mel_rec'] = dat.apply(lambda row: mel_wasserstein_distance(row.rec_mel, row.rec_mel), axis=1)
dat['wasser_mel_inv'] = dat.apply(lambda row: mel_wasserstein_distance(row.rec_mel, row.inv_mel), axis=1)
dat['wasser_mel_acoustic'] = dat.apply(lambda row: mel_wasserstein_distance(row.rec_mel, row.prod_mel_acoustic), axis=1)
dat['wasser_mel_acoustic_semvec'] = dat.apply(lambda row: mel_wasserstein_distance(row.rec_mel, row.prod_mel_acoustic_semvec), axis=1)
dat['wasser_mel_semvec'] = dat.apply(lambda row: mel_wasserstein_distance(row.rec_mel, row.prod_mel_semvec), axis=1)
dat['wasser_mel_seg'] = dat.apply(mel_wasserstein_distance_mel_seg, axis=1)
# correlations and rank of target word
dat['corr_vec_rec'] = dat.apply(lambda row: 1 - float(np.correlate(row.vector, row.rec_vec)), axis=1)
dat['corr_vec_inv'] = dat.apply(lambda row: 1 - float(np.correlate(row.vector, row.inv_vec)), axis=1)
dat['corr_vec_acoustic'] = dat.apply(lambda row: 1 - float(np.correlate(row.vector, row.prod_vec_acoustic)), axis=1)
dat['corr_vec_acoustic_semvec'] = dat.apply(lambda row: 1 - float(np.correlate(row.vector, row.prod_vec_acoustic_semvec)), axis=1)
dat['corr_vec_semvec'] = dat.apply(lambda row: 1 - float(np.correlate(row.vector, row.prod_vec_semvec)), axis=1)
dat['corr_vec_seg'] = dat.apply(lambda row: 1 - float(np.correlate(row.vector, row.seg_vec)), axis=1)
dat['rank_vec_rec'] = dat.apply(lambda row: rank(row.label, row.rec_vec), axis=1)
dat['rank_vec_inv'] = dat.apply(lambda row: rank(row.label, row.inv_vec), axis=1)
dat['rank_vec_acoustic'] = dat.apply(lambda row: rank(row.label, row['prod_vec_acoustic']), axis=1)
dat['rank_vec_acoustic_semvec'] = dat.apply(lambda row: rank(row.label, row['prod_vec_acoustic_semvec']), axis=1)
dat['rank_vec_semvec'] = dat.apply(lambda row: rank(row.label, row['prod_vec_semvec']), axis=1)
dat['rank_vec_seg'] = dat.apply(lambda row: rank(row.label, row.seg_vec), axis=1)
#dat.to_pickle(f'data/exp_seg_vs_record_metrics_{CONDITION}_{OBJECTIVE}.pickle')
import matplotlib.pyplot as plt
import ptitprince as pt
import seaborn as sns
import numpy as np
import pandas as pd
#dat = pd.read_pickle(f'data/exp_seg_vs_record_metrics_{CONDITION}_{OBJECTIVE}.pickle')
for objective, dat in dats.items():
df = pd.DataFrame({
'file': np.tile(dat['file'], 6),
'group': np.repeat(['rec', 'inv', 'semvec', 'acoustic semvec', 'acoustic', 'segment'], dat.shape[0]),
'rMSE loss between true and resynth semantic vectors': dat['rmse_vec_rec'].to_list() + dat['rmse_vec_inv'].to_list() + dat['rmse_vec_semvec'].to_list() + dat['rmse_vec_acoustic_semvec'].to_list() + dat['rmse_vec_acoustic'].to_list() + dat['rmse_vec_seg'].to_list(),
'1 - corr between true and resynth semantic vector': dat['corr_vec_rec'].to_list() + dat['corr_vec_inv'].to_list() + dat['corr_vec_semvec'].to_list() + dat['corr_vec_acoustic_semvec'].to_list() + dat['corr_vec_acoustic'].to_list() + dat['corr_vec_seg'].to_list(),
'rank of resynth semantic vector': dat['rank_vec_rec'].to_list() + dat['rank_vec_inv'].to_list() + dat['rank_vec_semvec'].to_list() + dat['rank_vec_acoustic_semvec'].to_list() + dat['rank_vec_acoustic'].to_list() + dat['rank_vec_seg'].to_list(),
'rMSE loss between true and resynth acoustics': dat['rmse_mel_rec'].to_list() + dat['rmse_mel_inv'].to_list() + dat['rmse_mel_semvec'].to_list() + dat['rmse_mel_acoustic_semvec'].to_list() + dat['rmse_mel_acoustic'].to_list() + dat['rmse_mel_seg'].to_list(),
'Wasserstein Distance between true and resynth acoustics': dat['wasser_mel_rec'].to_list() + dat['wasser_mel_inv'].to_list() + dat['wasser_mel_semvec'].to_list() + dat['wasser_mel_acoustic_semvec'].to_list() + dat['wasser_mel_acoustic'].to_list() + dat['wasser_mel_seg'].to_list()})
# plot a bar chart
plt.figure(figsize=(14, 8))
ax = sns.barplot(x="group", y="rMSE loss between true and resynth acoustics", data=df[df.group.str.contains('inv|acoustic|acoustc_semvec|semvec')], estimator=np.mean, ci=95, capsize=.2, color='lightblue')
ax.set_title(f'{objective}')
#ax.set_ylim((0.16, 0.185))
plt.savefig(f'plots/masterthesis_paul_{objective}_mean_acoustic.pdf')
# plot a bar chart
plt.figure(figsize=(14, 8))
ax = sns.barplot(x="group", y="rMSE loss between true and resynth semantic vector", data=df[df.group.str.contains('inv|acoustic|acoustc_semvec|semvec')], estimator=np.mean, ci=95, capsize=.2, color='lightblue')
ax.set_title(f'{objective}')
#ax.set_ylim((0.16, 0.185))
plt.savefig(f'plots/masterthesis_paul_{objective}_mean_semvec.pdf')
# plots including all data points ----
# both plots in one figure
plt.clf()
fig, (ax1, ax2) = plt.subplots(1, 2, figsize=(15, 5))
ort = "h"; pal = "Set2"; sigma = .2
pt.RainCloud(x='group', y='rMSE loss between true and resynth acoustics', data=df, palette=pal, bw=sigma, width_viol=.6, ax=ax1, orient=ort, box_showfliers=False)
pt.RainCloud(x='group', y='rMSE loss between true and resynth semantic vectors', data=df, palette=pal, bw=sigma, width_viol=.6, ax=ax2, orient=ort, box_showfliers=False)
#ax1.set_xlim((-0.0001, 0.4))
#ax2.set_xlim((-0.0001, 0.07))
fig.savefig(f'plots/masterthesis_paul_{objective}_rmse.pdf')
plt.clf()
fig, (ax1, ax2) = plt.subplots(1, 2, figsize=(15, 5))
ort = "h"; pal = "Set2"; sigma = .2
pt.RainCloud(x='group', y='1 - corr between true and resynth semantic vector', data=df, palette=pal, bw=sigma, width_viol=.6, ax=ax1, orient=ort, box_showfliers=False)
pt.RainCloud(x='group', y='rank of resynth semantic vector', data=df, palette=pal, bw=sigma, width_viol=.6, ax=ax2, orient=ort, box_showfliers=False)
#ax1.set_xlim((-0.0001, 1.0))
fig.savefig(f'plots/masterthesis_paul_{objective}_corr-rank.pdf')
plt.clf()
fig, (ax1, ax2) = plt.subplots(1, 2, figsize=(15, 5))
ort = "h"; pal = "Set2"; sigma = .2
pt.RainCloud(x='group', y='Wasserstein Distance between true and resynth acoustics', data=df, palette=pal, bw=sigma, width_viol=.6, ax=ax1, orient=ort, box_showfliers=False)
#pt.RainCloud(x='group', y='rMSE loss between true and resynth semantic vectors', data=df, palette=pal, bw=sigma, width_viol=.6, ax=ax2, orient=ort, box_showfliers=False)
#ax1.set_xlim((-0.0001, 0.3))
#ax2.set_xlim((-0.0001, 0.07))
fig.savefig(f'plots/masterthesis_paul_{objective}_wasserstein.pdf')
# small t-tests
rvs1 = df[df.group == 'acoustic semvec']['rMSE loss between true and resynth acoustics']
rvs2 = df[df.group == 'inv']['rMSE loss between true and resynth acoustics']
sp.stats.ttest_rel(rvs1, rvs2)
# play all words
import sounddevice as sd
import time
print(f"condition: {CONDITION}")
for index, row in dat.iterrows():
print(f"{row['label']} (inv, planned acoustic_semvec, rec, seg)")
sd.play(row['inv_sig'] * 4, row['inv_sr'], blocking=True)
time.sleep(0.1)
sd.play(row['prod_sig_acoustic_semvec'] * 4, 44100, blocking=True)
time.sleep(0.1)
sd.play(row['rec_sig'], row['rec_sr'], blocking=True)
time.sleep(0.1)
sd.play(row['seg_sig'], row['seg_sr'], blocking=True)
time.sleep(0.5)