def main():
import sys
import config as cfg
with cfg.ConfigBoundary():
t_train = test_backward(cfg.gene_ver, cfg.gene_value)
t_eval = test_forward(cfg.gene_ver, cfg.gene_value)
print("Train: %fbat/s" % (1 / t_train, ))
print("Eval: %fbat/s" % (1 / t_eval, ))
def count(ver, gene_int, n_ch):
import tensorflow as tf
from mir_util import infer
import config as cfg
import netop
with cfg.ConfigBoundary(gene_ver=ver, gene_value=gene_int):
batch_size = 1
graph = tf.Graph()
run_meta = tf.compat.v1.RunMetadata()
with graph.as_default():
x_mixed = tf.compat.v1.placeholder(tf.float32, shape=(batch_size, 64, cfg.frame_size // 2, 1), name="x_mixed")
y_mixed = tf.compat.v1.placeholder(tf.float32, shape=(batch_size, 64, cfg.frame_size // 2, n_ch), name="y_mixed")
y_pred = infer(x_mixed, n_ch, True)
n_forward_flop = tf.compat.v1.profiler.profile(graph, run_meta=run_meta, cmd="op", options=tf.compat.v1.profiler.ProfileOptionBuilder.float_operation()).total_float_ops
y_output = tf.multiply(x_mixed, y_pred)
loss_fn = tf.reduce_mean(input_tensor=tf.abs(y_mixed - y_output) , name="loss0")
global_step = tf.Variable(0, dtype=tf.int32, trainable=False, name="global_step")
optimizer = tf.compat.v1.train.AdamOptimizer(learning_rate=1e-4).minimize(loss_fn, global_step=global_step)
n_total_flop = tf.compat.v1.profiler.profile(graph, run_meta=run_meta, cmd="op", options=tf.compat.v1.profiler.ProfileOptionBuilder.float_operation()).total_float_ops
total_parameters = netop.count_parameter()
return total_parameters, n_forward_flop
import tensorflow as tf
import os, time
import numpy as np
import random
from mir_util import infer
import pickle
import config as cfg
with cfg.ConfigBoundary():
# Load data
batch_size = cfg.MIR2Config.batch_size
cache_path = cfg.mir2_cache_path
n_feature = cfg.frame_size // 2
assert os.path.exists(cache_path), "Dataset cache not found"
print("* Read cached spectrograms")
mixed_list, vocal_list, inst_list, n_sample = pickle.load(
open(cache_path, "rb"))
print("* Number of training examples: %d" % (n_sample, ))
# Model
print("* Initialize network")
tf.compat.v1.random.set_random_seed(0x41526941)
p_input = tf.compat.v1.placeholder(tf.float32,
shape=(batch_size, 64, n_feature, 1),
name="p_input")
p_target = tf.compat.v1.placeholder(tf.float32,
shape=(batch_size, 64, n_feature, 2),
name="p_target")
v_pred = infer(p_input, 2, True)
if isinstance(v_pred, list):
def main():
import multiprocessing as mp
import numpy as np
import tensorflow as tf
import os, sys
from mir_util import infer, to_spec, to_wav_file
import scipy.signal as sp
import config as cfg
sys.path.append("../lib")
from eval_util import bss_eval
from common import loadWav
import redirect, simpleopt
step_idx = int(simpleopt.get("step"))
n_eval = simpleopt.get("first", None)
with cfg.ConfigBoundary():
batch_size = 1
n_feature = cfg.frame_size // 2
# Model
print("* Initialize network")
p_input = tf.compat.v1.placeholder(tf.float32,
shape=(batch_size, 64, n_feature,
1),
name="p_input")
v_pred = infer(p_input, 2, False)
if isinstance(v_pred, list):
v_pred = v_pred[-1]
v_pred = tf.clip_by_value(v_pred, 0.0, 1.0) * p_input
x_input = np.zeros((batch_size, 64, n_feature, 1), dtype=np.float32)
with tf.compat.v1.Session(config=cfg.sess_cfg) as sess:
# Initialized, Load state
sess.run(tf.compat.v1.global_variables_initializer())
print("* Load checkpoint")
ckpt_path = os.path.join(cfg.MIR2Config.ckpt_path,
"checkpoint-%d" % (step_idx, ))
tf.compat.v1.train.Saver().restore(sess, ckpt_path)
print(" :Loaded: `%s`" % (ckpt_path, ))
os.makedirs("./eval_output", exist_ok=True)
name_list = []
ret_list = []
with mp.Pool(processes=1, initializer=worker_main) as pool:
for (root, _, file_list) in os.walk(cfg.mir_wav_path):
file_list = sorted(f for f in file_list if not (
f.startswith("abjones") or f.startswith("amy")))
if n_eval is not None:
file_list = file_list[:int(n_eval)]
for i_file, filename in enumerate(file_list):
print("[%03d/%03d] SEND: `%s`" % (
i_file + 1,
len(file_list),
filename,
))
name_list.append(filename)
path = os.path.join(root, filename)
mixed_wav, sr_orig = loadWav(path)
gt_wav_vocal = mixed_wav[:, 1]
gt_wav_inst = mixed_wav[:, 0]
mixed_wav = np.sum(mixed_wav, axis=1)
mixed_wav_orig = mixed_wav
gt_wav_vocal_orig = gt_wav_vocal
gt_wav_inst_orig = gt_wav_inst
gt_wav_vocal = sp.resample_poly(
gt_wav_vocal, cfg.sr, sr_orig).astype(np.float32)
gt_wav_inst = sp.resample_poly(
gt_wav_inst, cfg.sr, sr_orig).astype(np.float32)
mixed_wav = sp.resample_poly(
mixed_wav, cfg.sr, sr_orig).astype(np.float32)
mixed_spec = to_spec(mixed_wav)
mixed_spec_mag = np.abs(mixed_spec)
mixed_spec_phase = np.angle(mixed_spec)
max_tmp = np.max(mixed_spec_mag)
mixed_spec_mag = mixed_spec_mag / max_tmp
src_len = mixed_spec_mag.shape[0]
start_idx = 0
y_est_inst = np.zeros((src_len, n_feature),
dtype=np.float32)
y_est_vocal = np.zeros((src_len, n_feature),
dtype=np.float32)
while start_idx + 64 < src_len:
x_input[0, :, :,
0] = mixed_spec_mag[start_idx:start_idx +
64, :n_feature]
y_output = sess.run(v_pred,
feed_dict={p_input: x_input})
if start_idx == 0:
y_est_inst[start_idx:start_idx +
64, :] = y_output[0, :, :, 0]
y_est_vocal[start_idx:start_idx +
64, :] = y_output[0, :, :, 1]
else:
y_est_inst[start_idx + 16:start_idx +
48, :] = y_output[0, 16:48, :, 0]
y_est_vocal[start_idx + 16:start_idx +
48, :] = y_output[0, 16:48, :, 1]
start_idx += 32
x_input[0, :, :,
0] = mixed_spec_mag[src_len -
64:src_len, :n_feature]
y_output = sess.run(v_pred,
feed_dict={p_input: x_input})
src_start = src_len - start_idx - 16
y_est_inst[start_idx +
16:src_len, :] = y_output[0, 64 -
src_start:64, :,
0]
y_est_vocal[start_idx +
16:src_len, :] = y_output[0, 64 -
src_start:64, :,
1]
y_est_inst *= max_tmp
y_est_vocal *= max_tmp
y_wav_inst = to_wav_file(
y_est_inst, mixed_spec_phase[:, :n_feature])
y_wav_vocal = to_wav_file(
y_est_vocal, mixed_spec_phase[:, :n_feature])
#saveWav("inst.wav", y_wav_inst, cfg.sr)
#saveWav("vocal.wav", y_wav_vocal, cfg.sr)
# upsample to original samprate
y_wav_inst_orig = sp.resample_poly(
y_wav_inst, sr_orig, cfg.sr).astype(np.float32)
y_wav_vocal_orig = sp.resample_poly(
y_wav_vocal, sr_orig, cfg.sr).astype(np.float32)
ret_list.append(
pool.apply_async(bss_eval, (
mixed_wav_orig,
gt_wav_inst_orig,
gt_wav_vocal_orig,
y_wav_inst_orig,
y_wav_vocal_orig,
)))
with redirect.ConsoleAndFile(
"./eval_output/mir2_%s_%d_step%d.txt" %
(cfg.gene_ver, cfg.gene_value, step_idx)) as r:
gnsdr = 0.0
gsir = 0.0
gsar = 0.0
total_len = 0
for name, ret in zip(name_list, ret_list):
nsdr, sir, sar, lens = ret.get()
printstr = name + " " + str(nsdr) + " " + str(
sir) + " " + str(sar)
r.print(printstr)
total_len += lens
gnsdr += nsdr * lens
gsir += sir * lens
gsar += sar * lens
r.print("Final results")
r.print("GNSDR [Accompaniments, voice]")
r.print(gnsdr / total_len)
r.print("GSIR [Accompaniments, voice]")
r.print(gsir / total_len)
r.print("GSAR [Accompaniments, voice]")
r.print(gsar / total_len)
def main():
import multiprocessing as mp
import numpy as np
import tensorflow as tf
import os, sys
import config as cfg
#import librosa
from mir_util import infer, to_spec, to_wav_file
import scipy.signal as sp
sys.path.append("../lib")
from eval_util import bss_eval_sdr_framewise
from common import loadWav
import redirect, simpleopt
import pandas as pd
step_idx = int(simpleopt.get("step"))
n_eval = simpleopt.get("first", None)
with cfg.ConfigBoundary():
batch_size = 1
n_feature = cfg.frame_size // 2
# Model
print("* Initialize network")
p_input = tf.compat.v1.placeholder(tf.float32,
shape=(batch_size, 64, n_feature,
1),
name="p_input")
v_pred = infer(p_input, 2, False)
if isinstance(v_pred, list):
v_pred = v_pred[-1]
v_pred = tf.clip_by_value(v_pred, 0.0, 1.0) * p_input
x_input = np.zeros((batch_size, 64, n_feature, 1), dtype=np.float32)
with tf.compat.v1.Session(config=cfg.sess_cfg) as sess:
# Initialized, Load state
sess.run(tf.compat.v1.global_variables_initializer())
print("* Load checkpoint")
ckpt_path = os.path.join(cfg.DSD2Config.ckpt_path,
"checkpoint-%d" % (step_idx, ))
tf.compat.v1.train.Saver().restore(sess, ckpt_path)
print(" :Loaded: `%s`" % (ckpt_path, ))
os.makedirs("./eval_output", exist_ok=True)
name_list = []
ret_list = []
with mp.Pool(processes=8, initializer=worker_main) as pool:
for (root, dir_list, _) in os.walk(
os.path.join(cfg.dsd_path, "Mixtures", "Test")):
dir_list = sorted(dir_list)
if n_eval is not None:
dir_list = dir_list[:int(n_eval)]
for i_dir, d in enumerate(dir_list):
print("[%02d/%02d] STG1: `%s`" % (
i_dir + 1,
len(dir_list),
d,
))
name_list.append(d)
filename_vocal = os.path.join(cfg.dsd_path, "Sources",
"Test", d, "vocals.wav")
filename_mix = os.path.join(cfg.dsd_path, "Mixtures",
"Test", d, "mixture.wav")
import time
t = time.time()
mixed_wav_orig, sr_orig = loadWav(
filename_mix
) #librosa.load(filename_mix, sr=None, mono=True)
mixed_wav_orig = np.sum(mixed_wav_orig, axis=1)
gt_wav_vocal_orig, _ = loadWav(
filename_vocal
) #librosa.load(filename_vocal, sr=None, mono=True)[0]
gt_wav_vocal_orig = np.sum(gt_wav_vocal_orig, axis=1)
gt_wav_inst_orig = mixed_wav_orig - gt_wav_vocal_orig
mixed_wav = sp.resample_poly(
mixed_wav_orig, cfg.sr, sr_orig
).astype(
np.float32
) #librosa.load(filename_mix, sr=cfg.sr, mono=True)[0]
gt_wav_vocal = sp.resample_poly(
gt_wav_vocal_orig, cfg.sr, sr_orig
).astype(
np.float32
) #librosa.load(filename_vocal, sr=cfg.sr, mono=True)[0]
gt_wav_inst = mixed_wav - gt_wav_vocal
mixed_spec = to_spec(mixed_wav)
mixed_spec_mag = np.abs(mixed_spec)
mixed_spec_phase = np.angle(mixed_spec)
max_tmp = np.max(mixed_spec_mag)
mixed_spec_mag = mixed_spec_mag / max_tmp
src_len = mixed_spec_mag.shape[0]
start_idx = 0
y_est_inst = np.zeros((src_len, n_feature),
dtype=np.float32)
y_est_vocal = np.zeros((src_len, n_feature),
dtype=np.float32)
while start_idx + 64 < src_len:
x_input[0, :, :,
0] = mixed_spec_mag[start_idx:start_idx +
64, :n_feature]
y_output = sess.run(v_pred,
feed_dict={p_input: x_input})
if start_idx == 0:
y_est_inst[start_idx:start_idx +
64, :] = y_output[0, :, :, 0]
y_est_vocal[start_idx:start_idx +
64, :] = y_output[0, :, :, 1]
else:
y_est_inst[start_idx + 16:start_idx +
48, :] = y_output[0, 16:48, :, 0]
y_est_vocal[start_idx + 16:start_idx +
48, :] = y_output[0, 16:48, :, 1]
start_idx += 32
x_input[0, :, :,
0] = mixed_spec_mag[src_len -
64:src_len, :n_feature]
y_output = sess.run(v_pred,
feed_dict={p_input: x_input})
src_start = src_len - start_idx - 16
y_est_inst[start_idx +
16:src_len, :] = y_output[0, 64 -
src_start:64, :,
0]
y_est_vocal[start_idx +
16:src_len, :] = y_output[0, 64 -
src_start:64, :,
1]
y_est_inst *= max_tmp
y_est_vocal *= max_tmp
y_wav_inst = to_wav_file(
y_est_inst, mixed_spec_phase[:, :n_feature])
y_wav_vocal = to_wav_file(
y_est_vocal, mixed_spec_phase[:, :n_feature])
#saveWav("inst.wav", y_wav_inst, cfg.sr)
#saveWav("vocal.wav", y_wav_vocal, cfg.sr)
#upsample to original SR
y_wav_inst_orig = sp.resample_poly(
y_wav_inst, sr_orig, cfg.sr).astype(
np.float32
) #librosa.resample(y_wav_inst, cfg.sr, sr_orig)
y_wav_vocal_orig = sp.resample_poly(
y_wav_vocal, sr_orig, cfg.sr).astype(
np.float32
) #librosa.resample(y_wav_vocal, cfg.sr, sr_orig)
ret_list.append(
pool.apply_async(bss_eval_sdr_framewise, (
np.array([gt_wav_inst_orig, gt_wav_vocal_orig],
dtype=np.float32),
np.array([y_wav_inst_orig, y_wav_vocal_orig],
dtype=np.float32),
)))
head_list = [
"method", "track", "target", "metric", "score", "time"
]
row_list = []
out_path = "./old_fw/dsd2_%s_%d_step%d.json" % (
cfg.gene_ver, cfg.gene_value, step_idx)
method_name = "dsd2_%s_%d_step%d" % (cfg.gene_ver,
cfg.gene_value, step_idx)
for name, ret in zip(name_list, ret_list):
print(name)
sdr, sir, sar = ret.get()
for i, v in enumerate(sdr[0]):
row_list.append((
method_name,
name,
"accompaniment",
"SDR",
v,
i,
))
for i, v in enumerate(sir[0]):
row_list.append((
method_name,
name,
"accompaniment",
"SIR",
v,
i,
))
for i, v in enumerate(sar[0]):
row_list.append((
method_name,
name,
"accompaniment",
"SAR",
v,
i,
))
for i, v in enumerate(sdr[1]):
row_list.append((
method_name,
name,
"vocals",
"SDR",
v,
i,
))
for i, v in enumerate(sir[1]):
row_list.append((
method_name,
name,
"vocals",
"SIR",
v,
i,
))
for i, v in enumerate(sar[1]):
row_list.append((
method_name,
name,
"vocals",
"SAR",
v,
i,
))
out = pd.DataFrame(row_list, columns=head_list).reset_index()
print(out)
out.to_json(out_path)
def main_estimate(pool):
import numpy as np
import tensorflow as tf
import os, sys, pathlib
sys.path.append("../lib")
import config as cfg
#import librosa
from mir_util import infer, to_spec, to_wav_file
import scipy.signal as sp
import musdb, museval
import simpleopt
sys.is_train = False
step_idx = int(simpleopt.get("step"))
n_eval = simpleopt.get("first", None)
if n_eval is not None:
n_eval = int(n_eval)
assert n_eval > 0
sound_sample_root = simpleopt.get("sound-out", None)
source = simpleopt.get("source")
if source == "vocals":
source = None
with cfg.ConfigBoundary():
if source is None:
model_name = "mus2f_%s_%d_step%d" % (
cfg.gene_ver,
cfg.gene_value,
step_idx,
)
else:
model_name = "mus2f_%s_%d_step%d_%s" % (
cfg.gene_ver,
cfg.gene_value,
step_idx,
source,
)
model_name_nosrc = "mus2f_%s_%d_step%d" % (
cfg.gene_ver,
cfg.gene_value,
step_idx,
)
if sound_sample_root is None:
sound_sample_root = "./sound_output_mus2f/{}".format(
model_name_nosrc)
pathlib.Path(sound_sample_root).mkdir(parents=True, exist_ok=True)
ckpt_path = cfg.MUS2FConfig.ckpt_path
if source != "vocals":
ckpt_path = "{}_{}".format(ckpt_path, source)
batch_size = 1
n_feature = 5644 // 2
# Model
print("* Initialize network")
p_input = tf.compat.v1.placeholder(tf.float32,
shape=(batch_size, 64, n_feature,
1),
name="p_input")
v_pred = infer(p_input, 2, False)
if isinstance(v_pred, list):
v_pred = v_pred[-1]
with tf.compat.v1.Session(config=cfg.sess_cfg) as sess:
# Initialized, Load state
sess.run(tf.compat.v1.global_variables_initializer())
print("* Load checkpoint")
ckpt_path = os.path.join(ckpt_path, "checkpoint-%d" % (step_idx, ))
tf.compat.v1.train.Saver().restore(sess, ckpt_path)
print(" :Loaded: `%s`" % (ckpt_path, ))
os.makedirs("./eval_output", exist_ok=True)
name_list = []
ret_list = []
mus = musdb.DB(root=cfg.mus_root_path,
download=False,
subsets="test",
is_wav=True)
mus_trk_list = list(mus.tracks)
mus_trk_list.sort(key=lambda x: x.name)
assert len(mus_trk_list) > 0
if n_eval is not None:
mus_trk_list = mus_trk_list[:n_eval]
results = museval.EvalStore()
for i_song, track in enumerate(mus_trk_list):
print("[%02d/%02d] Estimate: `%s`" % (
i_song + 1,
len(mus_trk_list),
track.name,
))
voc_ch_list = []
inst_ch_list = []
for i_channel in range(2):
print(" :Channel #%d" % (i_channel, ))
name_list.append(track.name + " Channel %d" %
(i_channel, ))
mixed_wav = track.audio[:, i_channel]
mixed_spec = to_spec(mixed_wav,
len_frame=5644,
len_hop=5644 // 4)
mixed_spec_mag = np.abs(mixed_spec)
mixed_spec_phase = np.angle(mixed_spec)
max_tmp = np.max(mixed_spec_mag)
mixed_spec_mag = mixed_spec_mag / max_tmp
src_len = mixed_spec_mag.shape[0]
start_idx = 0
y_est_inst = np.zeros((src_len, n_feature),
dtype=np.float32)
y_est_vocal = np.zeros((src_len, n_feature),
dtype=np.float32)
x_input = np.zeros((batch_size, 64, n_feature, 1),
dtype=np.float32)
while start_idx + 64 < src_len:
x_input[0, :, :,
0] = mixed_spec_mag[start_idx:start_idx +
64, :n_feature]
y_output = sess.run(v_pred,
feed_dict={p_input: x_input})
if start_idx == 0:
y_est_inst[start_idx:start_idx +
64, :] = y_output[0, :, :, 0]
y_est_vocal[start_idx:start_idx +
64, :] = y_output[0, :, :, 1]
else:
y_est_inst[start_idx + 16:start_idx +
48, :] = y_output[0, 16:48, :, 0]
y_est_vocal[start_idx + 16:start_idx +
48, :] = y_output[0, 16:48, :, 1]
start_idx += 32
x_input[0, :, :,
0] = mixed_spec_mag[src_len -
64:src_len, :n_feature]
y_output = sess.run(v_pred, feed_dict={p_input: x_input})
src_start = src_len - start_idx - 16
y_est_inst[start_idx +
16:src_len, :] = y_output[0,
64 - src_start:64, :,
0]
y_est_vocal[start_idx +
16:src_len, :] = y_output[0,
64 - src_start:64, :,
1]
y_est_inst *= max_tmp
y_est_vocal *= max_tmp
y_wav_inst = to_wav_file(y_est_inst,
mixed_spec_phase[:, :n_feature],
len_hop=5644 // 4)
y_wav_vocal = to_wav_file(y_est_vocal,
mixed_spec_phase[:, :n_feature],
len_hop=5644 // 4)
voc_ch_list.append(y_wav_vocal.reshape(
y_wav_vocal.size, 1))
inst_ch_list.append(y_wav_inst.reshape(y_wav_inst.size, 1))
del y_wav_inst, y_wav_vocal, y_est_inst, y_est_vocal, src_start, x_input, y_output, mixed_spec_mag, max_tmp, mixed_spec_phase, mixed_spec, mixed_wav
estimates = {
source: np.concatenate(voc_ch_list, axis=1),
}
del voc_ch_list, inst_ch_list
if sound_sample_root:
mus.save_estimates(estimates, track, sound_sample_root)
del estimates, i_song, track