From 8d7309567179761994a29a396fb772016ffce409 Mon Sep 17 00:00:00 2001
From: Julian Linke <jlinke-tug@aim-gpu6.spsc.tugraz.at>
Date: Thu, 27 Jul 2023 20:23:23 +0200
Subject: [PATCH] initial commit
---
local/codebook_combine_arrays.py | 63 +++++++++++++
local/codebook_freqs.py | 136 +++++++++++++++++++++++++++
local/create_xlsr_new.py | 23 +++++
local/plot_pca_similarities.py | 156 +++++++++++++++++++++++++++++++
local/prepare_data.py | 53 +++++++++++
local/similarity_matrix.py | 96 +++++++++++++++++++
path.sh | 2 +
run.sh | 93 ++++++++++++++++++
8 files changed, 622 insertions(+)
create mode 100644 local/codebook_combine_arrays.py
create mode 100644 local/codebook_freqs.py
create mode 100644 local/create_xlsr_new.py
create mode 100644 local/plot_pca_similarities.py
create mode 100644 local/prepare_data.py
create mode 100644 local/similarity_matrix.py
create mode 100755 path.sh
create mode 100755 run.sh
diff --git a/local/codebook_combine_arrays.py b/local/codebook_combine_arrays.py
new file mode 100644
index 0000000..bde5ca8
--- /dev/null
+++ b/local/codebook_combine_arrays.py
@@ -0,0 +1,63 @@
+# Author: Julian Linke (linke@tugraz.at)
+# SPSC TU Graz (July 2023)
+
+import os, sys
+import fairseq
+import torch, torchaudio
+import numpy as np
+import json
+
+def main(exp_path, json_path):
+ # load json ...
+ # dict with {"corpusA_speakingstyle": [spk1 spk2,...],
+ # "corpusB_speakingstyle": [spk1 spk2,...], ...}
+ with open(json_path, 'r') as f:
+ corpora = json.load(f)
+
+ cnt = 0
+ for corpus in corpora:
+ freqN_path = os.path.join(exp_path, corpus)
+ for freqN_vec in os.listdir(freqN_path):
+ if 'freqN' in freqN_vec:
+ cnt = cnt+1
+ print('counted {} freqN-files'.format(cnt))
+
+ freqN_convs = np.zeros((cnt, 320**2), dtype=np.float32)
+ spks_vec = np.zeros((cnt,), dtype=object)
+ idx = 0
+ print('... start combining freqN_*.npy-files coming from path {}'.format(exp_path))
+ for corpus in corpora:
+ style = corpus.split("_")[1] # second entry is always style
+ freqN_path = os.path.join(exp_path, corpus)
+ for freqN_vec in os.listdir(freqN_path):
+ if 'freqN' in freqN_vec:
+ #print(freqN_vec)
+ split = freqN_vec.split('_')[1].replace('.npy','')
+ print('read and append split {} ...'.format(os.path.join(freqN_path,freqN_vec)))
+ # combine
+ freqN_convs[idx,:] = np.load(os.path.join(freqN_path,freqN_vec))
+ spks_vec[idx] = f'{split}{style}'
+ idx = idx + 1
+ freq_path, splits_path = os.path.join(exp_path, 'splits_freqs'), os.path.join(exp_path, 'splits_labels')
+ print('\nwrite {}.npy and {}.npy'.format(freq_path, splits_path))
+ np.save(freq_path, freqN_convs)
+ np.save(splits_path, spks_vec)
+ print('\nwrite {}.tsv and {}.tsv'.format(freq_path, splits_path))
+ np.savetxt(freq_path+'.tsv', freqN_convs, delimiter='\t')
+ np.savetxt(splits_path+'.tsv', spks_vec, fmt='%s', delimiter='\t')
+
+if __name__ == "__main__":
+ ############################# EXP PATH ###########################
+ try:
+ exp_path = sys.argv[1]
+ print("exp path is: " + exp_path)
+ except:
+ print("ERROR: data_path not specified")
+ ############################# JSON PATH ###########################
+ try:
+ json_path = sys.argv[2]
+ print("json path is: " + json_path)
+ except:
+ print("ERROR: json_path not specified")
+
+ main(exp_path, json_path)
diff --git a/local/codebook_freqs.py b/local/codebook_freqs.py
new file mode 100644
index 0000000..0b7e2ea
--- /dev/null
+++ b/local/codebook_freqs.py
@@ -0,0 +1,136 @@
+# Author: Julian Linke (linke@tugraz.at)
+# SPSC TU Graz (July 2023)
+
+import os, sys
+import fairseq
+import torch, torchaudio
+import numpy as np
+import json
+
+def print_list(l):
+ for s in l:
+ print(s)
+
+def count_freq(codebook_all_indexes, freq):
+ seq = []
+ for t in range(0, len(codebook_all_indexes)):
+ codebook_idx_at_time_t = codebook_all_indexes[t].item() # tensor(integer).item()
+ if VERBOSE:
+ print('(Verbose) frame={}: used codebook entry: {}'.format(t+1, codebook_idx_at_time_t))
+ seq.append(codebook_idx_at_time_t)
+ freq[codebook_idx_at_time_t] = freq[codebook_idx_at_time_t] + 1
+ return freq, seq
+
+def calc_codebook_indexes(audio_path, freq, N):
+ x, fs = torchaudio.load(audio_path)
+ x = x.to(device) # torch.Size([1, 57120]) [1 x Samples]
+
+ if np.shape(x)[1] > 512:
+ C = model.quantize(x)
+ quantized_features = C[0][0] # torch.Size([178, 768]) [T x d]
+ codebook_G2_indices = C[1] # torch.Size([1, 178, 2]) [1 x T x G]; G=2
+ codebook_all_indexes = model.quantizer.to_codebook_index(codebook_G2_indices)[0] # torch.Size([178]) [T]
+ Nwav = len(codebook_all_indexes)
+ freq, seq = count_freq(codebook_all_indexes, freq)
+ N = N + Nwav
+ print('feature vectors: {}/{} (file/all)'.format(Nwav, N))
+ else:
+ print('WARNING: Input size of file is {} (smaller than Kernel size), skip ...'.format(len(x)))
+ return freq, Nwav
+
+def main(exp_path, lst_path, json_path, model_path, VERBOSE):
+ # set model and device global
+ global model, device
+ device = torch.device('cuda')
+ # load existing model
+ model, cfg, task = fairseq.checkpoint_utils.load_model_ensemble_and_task([model_path])
+ model = model[0]
+ model = model.to(device)
+ # load json ...
+ # dict with {"corpusA_speakingstyle": [spk1 spk2,...],
+ # "corpusB_speakingstyle": [spk1 spk2,...], ...}
+ with open(json_path, 'r') as f:
+ corpora = json.load(f)
+ # corpus loop
+ for corpus in corpora.keys():
+ # LOGFILE:
+ if VERBOSE:
+ sys.stdout = open(os.path.join(exp_path, 'logs', "codebook_freqs_{}.log".format(corpus)),"w")
+ # speaker list and style
+ spks = corpora[corpus]
+ style = corpus.split("_")[1] # second entry is always style
+ # prepare numpy arrays
+ freqN_spks = np.zeros((len(spks), 320**2), dtype=np.float32) # [SPKS x 120400]
+ spks_vec = np.zeros((len(spks),), dtype=object) # [SPKS x 1]
+ # speaker loop
+ for idx, spk in enumerate(spks):
+ print(f"\n--- speaker {spk} in corpus {corpus} ---")
+ freq, N = dict.fromkeys(range(1, 320**2+1), 0), 0
+ processed_files = []
+ # extract frequencies per speaker
+ with open(lst_path, 'r') as tsv:
+ rows = tsv.readlines()
+ for row in rows:
+ uttID, audio_path = row.split()
+ corpus_match = ('_').join(audio_path.split('/')[1].split('_')[1:]) # DATA/data_corpus_speakingstyle/spk/*wav
+ spk_match = audio_path.split('/')[2] # DATA/data_corpus_speakingstyle/spk/*wav
+ if corpus == corpus_match and spk == spk_match:
+ print('\nread wav-file {}'.format(audio_path))
+ freq, Nwav = calc_codebook_indexes(audio_path, freq, N)
+ N = N + Nwav
+ processed_files.append(audio_path)
+ print(f'(DONE) Found {N} observations for speaker {spk} ...')
+ # combine frequencies
+ freqN_vec = np.zeros((1, 320**2), dtype=np.float32)
+ os.system('mkdir -p {}'.format(os.path.join(exp_path, 'txt', corpus)))
+ os.system('mkdir -p {}'.format(os.path.join(exp_path, 'numpy', corpus)))
+ with open(os.path.join(exp_path, 'txt', corpus, 'freq_{}.txt'.format(spk)), 'w') as ffreq, \
+ open(os.path.join(exp_path, 'txt', corpus, 'freqN_{}.txt'.format(spk)), 'w') as ffreqN:
+ for i, code_entry in enumerate(freq.keys()):
+ ffreq.write('{}\t{}\n'.format(code_entry, freq[code_entry]))
+ ffreqN.write('{}\t{}\n'.format(code_entry, freq[code_entry]/N))
+ freqN_vec[0,i] = freq[code_entry]/N
+ np.save(os.path.join(exp_path, 'numpy', corpus, 'freqN_{}'.format(spk,style)), freqN_vec)
+ # combine frequencies per corpus
+ print(f"... speaker {spk} is column {idx} of array freq_{corpus}.npy!")
+ freqN_spks[idx,:] = freqN_vec
+ spks_vec[idx] = f'{spk}{style}'
+ # write combined frequencies
+ np.save(os.path.join(exp_path, 'numpy', corpus, 'freq_{}'.format(corpus)), freqN_spks)
+ np.save(os.path.join(exp_path, 'numpy', corpus, 'spkIDs_{}'.format(corpus)), spks_vec)
+
+if __name__ == "__main__":
+ ############################# EXP PATH ###########################
+ try:
+ exp_path = sys.argv[1]
+ print("\nexp path is: " + exp_path)
+ except:
+ print("ERROR: data_path not specified")
+ ############################# LIST PATH ###########################
+ try:
+ lst_path = sys.argv[2]
+ print("list path is: " + lst_path)
+ except:
+ print("ERROR: lst_path not specified")
+ ############################# JSON PATH ###########################
+ try:
+ json_path = sys.argv[3]
+ print("json path is: " + json_path)
+ except:
+ print("ERROR: json_path not specified")
+ ############################# MODEL PATH ###########################
+ try:
+ model_path = sys.argv[4]
+ print("model path is: " + model_path)
+ except:
+ print("ERROR: model_path not specified")
+ ############################# VERBOSE #############################
+ global VERBOSE
+ try:
+ VERBOSE = int(sys.argv[5])
+ print("VERBOSE is " + str(VERBOSE) + "\n")
+ except:
+ print("VERBOSE is not specified, default is 0!")
+ VERBOSE = 0
+
+ main(exp_path, lst_path, json_path, model_path, VERBOSE)
diff --git a/local/create_xlsr_new.py b/local/create_xlsr_new.py
new file mode 100644
index 0000000..0b0589a
--- /dev/null
+++ b/local/create_xlsr_new.py
@@ -0,0 +1,23 @@
+# https://github.com/facebookresearch/fairseq/issues/3741
+
+from omegaconf import DictConfig, OmegaConf, open_dict
+import torch
+
+cp_path = 'model/xlsr_53_56k.pt'
+cp = torch.load(cp_path)
+cfg = DictConfig(cp['cfg'])
+dd = OmegaConf.to_container(cfg, resolve=True)
+for k,v in dd.items():
+ if not isinstance(v, dict):
+ continue
+ for key, _ in v.items():
+ if key.split("_")[:2] == ["eval", "wer"]:
+ print(k,key)
+with open_dict(cfg):
+ cfg.task.pop('eval_wer')
+ cfg.task.pop('eval_wer_config')
+ cfg.task.pop('eval_wer_tokenizer')
+ cfg.task.pop('eval_wer_post_process')
+ cfg.task.pop('autoregressive')
+cp['cfg'] = cfg
+torch.save(cp, 'model/xlsr_53_56k_new.pt')
\ No newline at end of file
diff --git a/local/plot_pca_similarities.py b/local/plot_pca_similarities.py
new file mode 100644
index 0000000..431e072
--- /dev/null
+++ b/local/plot_pca_similarities.py
@@ -0,0 +1,156 @@
+# Author: Julian Linke (linke@tugraz.at)
+# SPSC TU Graz (July 2023)
+
+import os, sys
+import json
+import numpy as np
+import matplotlib
+import matplotlib.pyplot as plt
+from sklearn.metrics.pairwise import cosine_similarity
+from sklearn.metrics.pairwise import euclidean_distances
+from sklearn.metrics import pairwise_distances
+import faiss
+from mpl_toolkits.mplot3d import Axes3D
+import matplotlib.patches as patches
+
+# List of predefined colors for each corpus (you can add more if you like)
+colors = ['tab:red', 'tab:blue',
+ 'tab:green', 'tab:orange',
+ 'tab:purple', 'tab:brown',
+ 'tab:pink', 'tab:gray',
+ 'tab:olive', 'tab:cyan']
+# markersize
+MS = 30
+
+def main(X_path, y_path, pcaA_path, pcab_path, out_path, json_path):
+ # load json ...
+ # dict with {"corpusA_speakingstyle": [spk1 spk2,...],
+ # "corpusB_speakingstyle": [spk1 spk2,...], ...}
+ with open(json_path, 'r') as f:
+ corpora = json.load(f)
+
+ X = np.load(X_path)
+ spks = np.load(y_path, allow_pickle=True)
+
+ A = np.load(pcaA_path)
+ b = np.load(pcab_path)
+ print('\nspkIDs:\n{}'.format(spks))
+ print('\ninput matrix:\n{}'.format(X))
+
+ labels = list(spks)
+ fig, ax = plt.subplots(figsize=(15,15))
+ cax = ax.matshow(X, interpolation='nearest')
+ ax.grid()
+ plt.xticks(range(len(labels)), labels, rotation=90);
+ plt.yticks(range(len(labels)), labels);
+ fig.colorbar(cax, ticks=[0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, .8, .9, 1])
+ fig_path = os.path.join(out_path,'{}.png'.format(X_path.split('/')[-1].replace('.npy','')))
+ print('\nsave fig {}'.format(fig_path))
+ plt.savefig(fig_path)
+
+ colordict = {}
+ for corpus in corpora:
+ colordict[corpus] = []
+ assert len(corpora.keys()) <= len(colors), 'Not enough colors for corpora'
+ for idx, corpus in enumerate(corpora):
+ spks_tmp = corpora[corpus]
+ for spk in spks_tmp:
+ colordict[corpus].append(colors[idx])
+ print(colordict)
+ col = sum([lst for lst in colordict.values()], [])
+
+ print("\nProject with PCA (3 dimensions)...")
+ X_proj = np.dot(X, A) + b # [styles x 320^2] * [320^2 x 2] or [styles x corpora] + [corpora x 2]
+ print(f"\nshape(X_proj) = {np.shape(X_proj)}")
+ print('mean(X_proj,0): {}'.format(np.mean(X_proj,0)))
+
+ # 3D plot:
+ fig = plt.figure()
+ ax = fig.add_subplot(111, projection='3d')
+ ax.scatter(X_proj[:,0], X_proj[:,1], X_proj[:,2], c=col)
+ # Set the labels and title
+ ax.set_xlabel('PCA1')
+ ax.set_ylabel('PCA2')
+ ax.set_zlabel('PCA3')
+ # Create the legend
+ for k,v in colordict.items():
+ ax.scatter([], [], [], c=v[0], label=k, alpha=1, s=MS)
+ # Shrink current axis by 20%
+ box = ax.get_position()
+ ax.set_position([box.x0, box.y0, box.width * 0.85, box.height])
+ # Put a legend to the right of the current axis
+ ax.legend(loc='center left', bbox_to_anchor=(1.175, .5))
+ # Save the plot
+ fig_path = os.path.join(out_path,'scatter3D_{}.png'.format(X_path.split('/')[-1].replace('.npy','')))
+ print('\nsave fig {}'.format(fig_path))
+ plt.savefig(fig_path)
+ fig_path = os.path.join(out_path,'scatter3D_{}.eps'.format(X_path.split('/')[-1].replace('.npy','')))
+ print('save fig {}'.format(fig_path))
+ plt.savefig(fig_path, format='eps')
+
+ plt.rcParams.update({'font.size': 50})
+ for proj in [(0, 1), (0, 2), (1, 2)]:
+ x, y = X_proj[:,proj[0]], X_proj[:,proj[1]]
+ fig, ax = plt.subplots(figsize=(18,16))
+
+ for i, txt in enumerate(spks):
+ print(f"speaker {txt}: x = {x[i]}, y = {y[i]}")
+ ax.plot(x[i], y[i], 'o', color=col[i], markersize=MS, alpha=.5, label=col[i])
+
+ handles, labels = plt.gca().get_legend_handles_labels()
+ by_label = dict(zip(labels, handles))
+
+ cols = list(by_label.keys())
+ for old, new in zip(cols, corpora.keys()):
+ by_label[new] = by_label.pop(old)
+ ax.legend(by_label.values(), by_label.keys(), fontsize=30)
+ plt.xlabel(f'PCA{str(proj[0]+1)}')
+ plt.ylabel(f'PCA{str(proj[1]+1)}')
+ plt.grid()
+ plt.tight_layout()
+ fig_path = os.path.join(out_path,'scatter_proj{}_{}.png'.format(''.join([str(proj[0]),str(proj[1])]), X_path.split('/')[-1].replace('.npy','')))
+ print('\nsave fig {}'.format(fig_path))
+ plt.savefig(fig_path)
+ fig_path = os.path.join(out_path,'scatter_proj{}_{}.eps'.format(''.join([str(proj[0]),str(proj[1])]), X_path.split('/')[-1].replace('.npy','')))
+ print('save fig {}'.format(fig_path))
+ plt.savefig(fig_path, format='eps')
+
+if __name__ == "__main__":
+ ############################# FEATURES PATH ###########################
+ try:
+ X_path = sys.argv[1]
+ print("input matrix path is: " + X_path)
+ except:
+ print("ERROR: X_path not specified")
+ ############################# LABELS PATH ###########################
+ try:
+ y_path = sys.argv[2]
+ print("label vector path is: " + y_path)
+ except:
+ print("ERROR: y_path not specified")
+ ############################# PCA A PATH ###########################
+ try:
+ pcaA_path = sys.argv[3]
+ print("pca matrix A path is: " + pcaA_path)
+ except:
+ print("ERROR: pcaA_path not specified")
+ ############################# PCA b PATH ###########################
+ try:
+ pcab_path = sys.argv[4]
+ print("pca vector b path is: " + pcab_path)
+ except:
+ print("ERROR: pcab_path not specified")
+ ############################# OUT PATH ###########################
+ try:
+ out_path = sys.argv[5]
+ print("output path is: " + out_path)
+ except:
+ print("ERROR: out_path not specified")
+ ############################# JSON PATH ###########################
+ try:
+ json_path = sys.argv[6]
+ print("json path is: " + json_path)
+ except:
+ print("ERROR: json_path not specified")
+
+ main(X_path, y_path, pcaA_path, pcab_path, out_path, json_path)
diff --git a/local/prepare_data.py b/local/prepare_data.py
new file mode 100644
index 0000000..1643678
--- /dev/null
+++ b/local/prepare_data.py
@@ -0,0 +1,53 @@
+# Author: Julian Linke (linke@tugraz.at)
+# SPSC TU Graz (July 2023)
+
+import os
+import json
+import argparse
+
+def get_speaker_ids_and_lst_lines(corpus_dir):
+ spk_ids = []
+ lst_lines = []
+ for spk in os.listdir(corpus_dir):
+ spk_dir = os.path.join(corpus_dir, spk)
+ if os.path.isdir(spk_dir):
+ spk_ids.append(spk)
+ for audio_file in os.listdir(spk_dir):
+ if audio_file.endswith('.wav') or audio_file.endswith('.flac'):
+ uttID = audio_file.replace('.wav','').replace('.flac','')
+ audio_path = os.path.join(spk_dir, audio_file)
+ lst_line = f"{uttID} {audio_path}\n"
+ lst_lines.append(lst_line)
+ return spk_ids, lst_lines
+
+def process_DATA_directory(DATA_dir):
+ spk_dict = {}
+ lst_lines = []
+ for corpus in os.listdir(DATA_dir):
+ corpus_dir = os.path.join(DATA_dir, corpus)
+ if os.path.isdir(corpus_dir):
+ corpus_name = '_'.join(corpus.split('_')[1:])
+ spk_ids, new_lst_lines = get_speaker_ids_and_lst_lines(corpus_dir)
+ spk_dict[corpus_name] = spk_ids
+ lst_lines.extend(new_lst_lines)
+ return spk_dict, lst_lines
+
+def write_output_files(output_lst_path, output_json_path, lst_lines, spk_dict):
+ with open(output_lst_path, 'w') as f:
+ f.writelines(lst_lines)
+
+ with open(output_json_path, 'w') as f:
+ json.dump(spk_dict, f, indent=4)
+
+def main(output_lst_path, output_json_path, DATA_dir):
+ spk_dict, lst_lines = process_DATA_directory(DATA_dir)
+ #print(lst_lines)
+ write_output_files(output_lst_path, output_json_path, lst_lines, spk_dict)
+
+if __name__ == "__main__":
+ parser = argparse.ArgumentParser(description='Generate DATA.lst and DATA.json files.')
+ parser.add_argument('--output_lst_path', required=True, help='Path to the output .lst file.')
+ parser.add_argument('--output_json_path', required=True, help='Path to the output .json file.')
+ parser.add_argument('--DATA_dir', required=True, help='Path to the DATA directory.')
+ args = parser.parse_args()
+ main(args.output_lst_path, args.output_json_path, args.DATA_dir)
diff --git a/local/similarity_matrix.py b/local/similarity_matrix.py
new file mode 100644
index 0000000..9912ee4
--- /dev/null
+++ b/local/similarity_matrix.py
@@ -0,0 +1,96 @@
+# Author: Julian Linke (linke@tugraz.at)
+# SPSC TU Graz (July 2023)
+
+import os, sys
+import json
+import numpy as np
+from sklearn.metrics.pairwise import cosine_similarity
+
+from scipy.stats import entropy
+from numpy.linalg import norm
+
+def JSD(P, Q):
+ _P = P / norm(P, ord=1)
+ _Q = Q / norm(Q, ord=1)
+ _M = 0.5 * (_P + _Q)
+ return 1 - (0.5 * (entropy(_P, _M) + entropy(_Q, _M)))
+
+def JSD_similarity(X):
+ sim = np.zeros((np.shape(X)[0], np.shape(X)[0]), dtype=np.float32)
+ print('\ncalculate JSD similarity matrix of X: {} ...'.format(X.shape))
+ for row, x in enumerate(X):
+ for col, _ in enumerate(sim):
+ #print('calculate JSD similarity of features ({}, {})'.format(row, col))
+ sim[row, col] = JSD(x, X[col,:])
+ return sim
+
+def main(X_path, y_path, out_path, json_path):
+ # load json ...
+ # dict with {"corpusA_speakingstyle": [spk1 spk2,...],
+ # "corpusB_speakingstyle": [spk1 spk2,...], ...}
+ with open(json_path, 'r') as f:
+ corpora = json.load(f)
+
+ # LOAD MATRIX AND SPLITS
+ X = np.load(X_path)
+ splits = np.load(y_path, allow_pickle=True)
+ print('\ninput splits: {}\n... len: {}'.format(splits, len(splits)))
+ corpora_lengths = {key: len(value) for key, value in corpora.items()}
+
+ # SORT:
+ d_splits, col = {}, []
+ for idx, split in enumerate(splits):
+ d_splits[split] = X[idx,:]
+ X = np.zeros((np.shape(X)), dtype=np.float32)
+ splits = np.zeros((np.shape(splits)), dtype=object)
+ i = 0
+ for corpus in corpora:
+ for split in corpora[corpus]:
+ style = corpus.split("_")[1] # second entry is always style
+ split = f'{split}{style}'
+ if split in d_splits:
+ X[i,:] = d_splits[split]
+ splits[i] = split
+ i = i + 1
+ else: print('Wrong corpora entry: {}?'.format(split))
+ print('\nsorted splits: {}\n... len: {}'.format(splits, len(splits)))
+
+ print('\nwrite sorted {} and {}'.format(X_path, y_path))
+ np.save(X_path, X)
+ np.save(y_path, splits)
+ print('write sorted {} and {}'.format(X_path, y_path))
+ np.savetxt(X_path.replace('.npy','.tsv'), X, delimiter='\t')
+ np.savetxt(y_path.replace('.npy','.tsv'), splits, fmt='%s', delimiter='\t')
+
+ # Calculate similarity matrix and save:
+ X_sim = JSD_similarity(X)
+ np.save(os.path.join(out_path,'similarity_matrix'), X_sim)
+ np.savetxt(os.path.join(out_path,'similarity_matrix.tsv'), X_sim, delimiter='\t')
+
+if __name__ == "__main__":
+ ############################# FEATURES PATH ###########################
+ try:
+ X_path = sys.argv[1]
+ print("input matrix path is: " + X_path)
+ except:
+ print("ERROR: X_path not specified")
+ ############################# LABELS PATH ###########################
+ try:
+ y_path = sys.argv[2]
+ print("label vector path is: " + y_path)
+ except:
+ print("ERROR: y_path not specified")
+ ############################# OUT PATH ###########################
+ try:
+ out_path = sys.argv[3]
+ print("output path is: " + out_path)
+ except:
+ print("ERROR: out_path not specified")
+ ############################# JSON PATH ###########################
+ try:
+ json_path = sys.argv[4]
+ print("json path is: " + json_path)
+ except:
+ print("ERROR: json_path not specified")
+
+ main(X_path, y_path, out_path, json_path)
diff --git a/path.sh b/path.sh
new file mode 100755
index 0000000..d7c0ec1
--- /dev/null
+++ b/path.sh
@@ -0,0 +1,2 @@
+export CWD=$(pwd)
+export FAIRSEQ=../fairseq
\ No newline at end of file
diff --git a/run.sh b/run.sh
new file mode 100755
index 0000000..47b0e31
--- /dev/null
+++ b/run.sh
@@ -0,0 +1,93 @@
+#!/bin/bash
+#set -x
+
+# Author: Julian Linke (linke@tugraz.at)
+# SPSC TU Graz (July 2023)
+
+set -e -o pipefail
+. path.sh
+. conda.sh
+
+if [[ $# -eq 0 ]] ; then
+ echo 'ERROR: this run-script requires an argument: stage=?'
+ exit 1
+fi
+
+## runDATA/STAGE
+runDATA=$1
+stage=$2
+VERBOSE=1 # write logs for codebook frequency extraction?
+printf "\n### STAGE ###\n"
+printf "stage: %d\n" $stage
+printf "### STAGE ###\n"
+
+## DIRS/PATHS
+model_path=model/xlsr_53_56k_new.pt
+exp_dir=exp_$runDATA
+
+## STAGE 0: DELETE AND RUN ALL STAGES
+if [ $stage == 0 ]; then
+ printf "\n... Delete old experiment and run all ...\n"
+ rm -rf ${exp_dir}
+fi
+
+## print:
+printf "\nCWD: %s" "$CWD"
+printf "\nFAIRSEQ: %s" "$FAIRSEQ"
+printf "\nrunDATA: %s" "$runDATA"
+printf "\nmodel_path: %s" "$model_path"
+printf "\nexp_dir: %s\n\n" "$exp_dir"
+
+## CREATE EXPERIMENT FOLDER
+mkdir -p $exp_dir
+mkdir -p $exp_dir/logs
+mkdir -p $exp_dir/data
+mkdir -p $exp_dir/plots
+mkdir -p $exp_dir/txt
+mkdir -p $exp_dir/numpy
+mkdir -p $exp_dir/numpy/pca
+
+## PREPARE DATA
+if [ $stage == 1 ] || [ $stage == 0 ]; then
+ printf "\n... Prepare data (*lst and *json) ...\n"
+ python3 local/prepare_data.py --output_lst_path $exp_dir/data/${runDATA}.lst \
+ --output_json_path $exp_dir/data/${runDATA}.json \
+ --DATA_dir ${runDATA}
+fi
+
+if [ $stage == 2 ] || [ $stage == 0 ]; then
+ printf "\n... Count frequencies of codebooks ...\n"
+ python3 local/codebook_freqs.py $exp_dir \
+ $exp_dir/data/${runDATA}.lst \
+ $exp_dir/data/${runDATA}.json \
+ $model_path \
+ $VERBOSE
+ printf "\n... Combine Arrays ... \n"
+ python3 local/codebook_combine_arrays.py $exp_dir/numpy \
+ $exp_dir/data/${runDATA}.json
+fi
+
+if [ $stage == 3 ] || [ $stage == 0 ]; then
+ printf "\n... Similarity Matrix ...\n"
+ python3 local/similarity_matrix.py \
+ $exp_dir/numpy/splits_freqs.npy \
+ $exp_dir/numpy/splits_labels.npy \
+ $exp_dir/numpy \
+ $exp_dir/data/${runDATA}.json
+fi
+
+if [ $stage == 4 ] || [ $stage == 0 ]; then
+ printf "\n... PCA of similarity matrix ...\n"
+ python3 $FAIRSEQ/examples/wav2vec/unsupervised/scripts/pca.py \
+ $exp_dir/numpy/similarity_matrix.npy \
+ --output $exp_dir/numpy/pca \
+ --dim 3
+ printf "\n... PLOT ...\n"
+ python3 local/plot_pca_similarities.py \
+ $exp_dir/numpy/similarity_matrix.npy \
+ $exp_dir/numpy/splits_labels.npy \
+ $exp_dir/numpy/pca/3_pca_A.npy \
+ $exp_dir/numpy/pca/3_pca_b.npy \
+ $exp_dir/plots \
+ $exp_dir/data/${runDATA}.json
+fi
\ No newline at end of file
--
GitLab