-
Notifications
You must be signed in to change notification settings - Fork 2
Expand file tree
/
Copy pathrun.py
More file actions
221 lines (187 loc) · 10.2 KB
/
run.py
File metadata and controls
221 lines (187 loc) · 10.2 KB
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
import os
import numpy as np
import pandas as pd
from distances import mut_normalize_sequences, librosa_dtw
import argparse
from projections import *
from data_frame_builders import build_cmn_df, build_kr_df, CMN, KR, SPN
TSNE_COL=['tsne-1', 'tsne-2', 'tsne-3']
def measure_dist(filename, sub, speakers, feature_columns, data_group, verbose=False):
normalize = True
normalize_by_len = True
normalization_func = mut_normalize_sequences
dist_df = pd.DataFrame(columns=["speaker1", "speaker2", "group", "part", "sent"])
feature_columns = feature_columns[0]
distance_func, func_name = librosa_dtw, 'Distance'
if data_group == KR:
reading_task = ""
elif data_group in [CMN, SPN]:
reading_task = speakers[0].split("_")[5]
for sp_id, speaker1 in enumerate(speakers[:-1]):
for speaker2 in speakers[sp_id+1:]:
if data_group == KR:
spkr1_sent_lng = speaker1.split("_")[2]
spkr2_sent_lng = speaker2.split("_")[2]
speaker1_l1 = speaker1.split("_")[1][:1]
speaker2_l1 = speaker2.split("_")[1][:1]
elif data_group in [CMN, SPN]:
speaker1_l1 = speaker1.split("_")[3][:2]
speaker2_l1 = speaker2.split("_")[3][:2]
spkr1_sent_lng = speaker1.split("_")[4][:2]
spkr2_sent_lng = speaker2.split("_")[4][:2]
speaker1_sentence = speaker1.split("_")[-1]
speaker2_sentence = speaker2.split("_")[-1]
speaker1_reading_task = speaker1.split("_")[5]
speaker2_reading_task = speaker2.split("_")[5]
if speaker1_sentence != speaker2_sentence or speaker1_reading_task != speaker2_reading_task:
continue
group = 'BT' if speaker1_l1[0] != speaker2_l1[0] else speaker1_l1[0]
if spkr1_sent_lng != spkr2_sent_lng:
continue
if verbose:
print(f"Measuring distance for spkr {speaker1} and {speaker2}")
sent_id = speaker1_sentence if data_group in [CMN, SPN] else speaker1.split("_")[2]
dist_row = {"speaker1": speaker1, "speaker2": speaker2,
"group": group, "part": reading_task, "sent": sent_id}
features_speaker1 = sub[(sub.speaker == speaker1)][feature_columns].to_numpy()
features_speaker2 = sub[(sub.speaker == speaker2)][feature_columns].to_numpy()
features_speaker1, features_speaker2 = normalization_func(features_speaker1, features_speaker2, normalize)
distance = distance_func(features_speaker1, features_speaker2)
if type(distance) == tuple and len(distance) > 0:
distance = distance[0]
if normalize_by_len:
distance = distance/(len(features_speaker1)+len(features_speaker2))
dist_row[f"{func_name}"] = distance
dist_df = pd.concat([dist_df, pd.DataFrame([dist_row])], ignore_index=True)
out_name = f"{filename}.csv"
dist_df.to_csv(out_name)
def fill_projection_in_df(df_subset,ids, tsne_results):
'''
Fill the t-SNE results into the dataframe for each available dimension.
'''
df_subset.loc[ids,('tsne-1')] = tsne_results[:,0]
df_subset.loc[ids,('tsne-2')] = tsne_results[:,1]
if tsne_results.shape[1] == 3:
df_subset.loc[ids,('tsne-3')] = tsne_results[:,2]
return df_subset
def construct_filename(args, projection_type):
# create output file name based on the argments
portion_str = f"_portion{args.portion}" if args.project else ""
if args.data_group in [CMN,SPN]:
task = args.reading_task.lower()
if not os.path.exists(os.path.join(args.output_path, f"{task}")):
os.mkdir(os.path.join(args.output_path, f"{task}"))
filename = os.path.join(
args.output_path, f"{args.data}_{projection_type}_layer{args.layer}_{task}{portion_str}")
else:
filename = os.path.join(args.output_path,
f"{args.data}_{projection_type}_layer{args.layer}{portion_str}")
return filename
def project_samples(args, data_subset):
'''
Project the samples using the specified method.
'''
if not args.pca and not args.umap:
projection_result = tsne(data_subset,
init=args.projection_init_method,
early_exaggeration=args.early_exagg,
lr=args.tsne_lr,
n_comp=args.num_projection_components,
perplexity=args.tsne_perplexity,
iters=args.tsne_iter,
seed=args.seed,)
elif args.pca:
projection_result = kpca(data_subset,
n_comp=args.num_projection_components,
kernel=args.kernel,
random_state=args.seed,
degree=args.degree,
)
elif args.umap:
projection_result = umap(
data_subset,
n_comp=args.num_projection_components,
n_neighbors=args.umap_nei,
min_dist=args.min_umap_dist,
init=args.umap_init,
random_state=args.seed,
densmap=args.densmap,
dens_lambda=args.d_lamb,
dens_frac=args.d_frac,
)
return projection_result
def main(args):
'''
Main function to run the distance evaluation.
'''
use_dim_reduction = args.project
projection_type = 'full_dim' if not use_dim_reduction else ('tsne' if not args.umap and not args.pca else ('umap' if args.umap else 'kpca'))
if args.data_group == KR:
data_subset, df, speakers_len, hubert_feature_columns = build_kr_df(args)
else:
data_subset, df, speakers_len, hubert_feature_columns = build_cmn_df(args)
df['new_id'] = list(range(df.shape[0]))
filename = construct_filename(args, projection_type)
if args.verbose:
print(
f"Running distance evaluation for layer {args.layer} with projection {use_dim_reduction} and projection type {projection_type}")
print("File will be saved to ", filename)
ids = np.array(df.index)
if use_dim_reduction:
projection_result = project_samples(args, data_subset)
df = fill_projection_in_df(df,ids,projection_result)
print(df.columns)
feature_columns = hubert_feature_columns if not args.project else [f"tsne-{i+1}" for i in range(projection_result.shape[1])],
measure_dist(filename, df, df.speaker.unique(), feature_columns, args.data_group, args.verbose)
parser = argparse.ArgumentParser(description='Parser for distance evaluation',)
parser.add_argument('--layer', type=int, required=True, help='HuBERT Layer to use for the distance evaluation')
parser.add_argument('--project', action='store_true', help='Project the data to lower dimensions')
parser.add_argument('--data', type=str, required=True, help='Type of data group: cmn, kr')
parser.add_argument('--projection_init_method', type=str, default='pca', help='Initialization method for t-SNE and Umap')
parser.add_argument("--tsne_iter", type=int, default=1000, help="Number of iterations for t-SNE")
parser.add_argument("--early_exagg", type=float, default=12.0, help="Early exaggeration for t-SNE")
parser.add_argument("--tsne_perplexity", type=int, default=50, help="Perplexity for t-SNE")
parser.add_argument('--portion', type=int, default=1, help='Portion (an integer between 1-9) of the data to use - only relevant if projecting')
parser.add_argument("--tsne_lr", type=str, default="auto", help="Learning rate for t-SNE")
parser.add_argument("--pca", action="store_true", help="Use Kernel-PCA for dimensionality reduction")
parser.add_argument("--kernel", type=str, default="rbf", help="Kernel to use for Kernel-PCA")
parser.add_argument("--degree", type=int, default=3, help="Degree of the polynomial kernel for Kernel-PCA")
parser.add_argument("--umap", action="store_true", help="Use UMAP for dimensionality reduction")
parser.add_argument("--umap_nei", type=int, default=15, help="Number of neighbors for UMAP")
parser.add_argument("--min_umap_dist", type=float, default=0.1, help="Minimum distance for UMAP")
parser.add_argument("--umap_init", type=str, default="spectral", help="Initialization method for UMAP")
parser.add_argument("--densmap", action="store_true", help="Use densmap for UMAP")
parser.add_argument("--d_lamb", type=float, default=0.0, help="Lambda for densmap")
parser.add_argument("--d_frac", type=float, default=0.3, help="Fraction for densmap")
parser.add_argument('--file_pref', type=str, default="", help="Prefix for output file name")
parser.add_argument('--tsv_name', type=str, required=True, help="Name of the tsv file")
parser.add_argument('--output_path', type=str, required=True,
help="Path where distance outputs will be stored (csv format)")
parser.add_argument('--feat_path', type=str, required=True,
help="Path to where Hubert's .npy feature files are stored")
parser.add_argument('--reading_task', type=str, default=None, help="Reading task for the data: HT1 or HT2 (not relevant for Korean)")
parser.add_argument('--verbose', action='store_true', help='Print verbose output')
parser.add_argument('--run_all_data', action='store_true')
parser.add_argument('--num_projection_components', type=int, default=3, help='Number of components to project to')
parser.add_argument("--seed", type=int, default=0, help="Seed for reproducibility")
args = parser.parse_args()
if __name__ == '__main__':
if args.project:
print(f'evaluating distance with {args.num_projection_components} dimensions')
else:
print('evaluating distance with all dimensions')
if args.tsne_lr != "auto":
args.tsne_lr = float(args.tsne_lr)
if args.data.lower() == 'cmn':
args.data_group = CMN
elif args.data.lower() == 'kr':
args.data_group = KR
elif args.data.lower() == 'spn':
args.data_group = SPN
else:
raise Exception(
f"Invalid data passed {args.data} | Possible values are 'cmn', 'spn' or 'kr'")
if args.data_group != KR and args.reading_task is None and not args.project:
raise Exception(
f"For ALLSTAR dataset you must specify reading_task and portion")
main(args)