/
test.py
186 lines (156 loc) · 7.18 KB
/
test.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
import os
import sys
import yaml
import torch
import argparse
import warnings
import importlib
import torchaudio
import pandas as pd
from tqdm import tqdm
from sigfig import round
import pytorch_lightning as pl
from torch.utils.data import DataLoader
from src.metrics import ALLMetricsTracker
from src.utils.parser_utils import parse_args_as_dict
from src.datas.avspeech_dataset import AVSpeechDataset
from src.losses import PITLossWrapper, pairwise_neg_sisdr
warnings.filterwarnings("ignore")
torch.set_float32_matmul_precision("high")
class TestModule(pl.LightningModule):
def __init__(self, conf):
super(TestModule, self).__init__()
self.conf = conf
self.conf["videonet"] = conf.get("videonet", {})
self.conf["videonet"]["model_name"] = conf["videonet"].get("model_name", None)
self.exp_dir = os.path.abspath(os.path.join("../experiments/audio-visual", conf["log"]["exp_name"]))
sys.path.append(os.path.dirname(self.exp_dir))
models_module = importlib.import_module(os.path.basename(self.exp_dir) + ".models")
videomodels = importlib.import_module(os.path.basename(self.exp_dir) + ".models.videomodels")
AVNet = getattr(models_module, "AVNet")
model_path = os.path.join(self.exp_dir, "best_model.pth")
self.audiomodel = AVNet.from_pretrain(model_path, **self.conf["audionet"])
self.videomodel = None
if self.conf["videonet"]["model_name"]:
self.videomodel = videomodels.get(self.conf["videonet"]["model_name"])(**self.conf["videonet"], print_macs=False)
self.loss_func = PITLossWrapper(pairwise_neg_sisdr, pit_from="pw_mtx")
self.ex_save_dir = os.path.join(self.exp_dir, "results_new/")
os.makedirs(self.ex_save_dir, exist_ok=True)
self.metrics = ALLMetricsTracker(save_file=os.path.join(self.ex_save_dir, "metrics.csv"))
self.test_step_outputs = {"mix": [], "sources": [], "est_sources": [], "key": []}
def test_step(self, batch, batch_idx):
mix, sources, target_mouths, key = batch
sources = sources.unsqueeze(1)
mouth_emb = self.videomodel(target_mouths.float()) if self.videomodel is not None else None
est_sources = self.audiomodel(mix, mouth_emb)
loss, reordered_sources = self.loss_func(est_sources, sources, return_ests=True)
self.log("test_loss", loss, prog_bar=True)
self.test_step_outputs["mix"].append(mix)
self.test_step_outputs["sources"].append(sources)
self.test_step_outputs["est_sources"].append(reordered_sources)
self.test_step_outputs["key"].append(key)
return loss
def on_test_epoch_end(self):
with torch.no_grad():
mix = [x for x in torch.cat(self.test_step_outputs["mix"], dim=0)]
sources = [x for x in torch.cat(self.test_step_outputs["sources"], dim=0)]
est_sources = [x for x in torch.cat(self.test_step_outputs["est_sources"], dim=0)]
key = [item for sublist in self.test_step_outputs["key"] for item in sublist]
pbar = tqdm(range(len(mix)))
for idx in pbar:
self.metrics(mix=mix[idx], clean=sources[idx], estimate=est_sources[idx], key=key[idx])
if idx < self.conf["n_save_ex"]:
self._save_audio_example(idx, mix[idx], sources[idx], est_sources[idx])
if not (idx % 10):
pbar.set_postfix(self.metrics.get_mean())
self.metrics.final()
mean, std = self.metrics.get_mean(), self.metrics.get_std()
keys = list(mean.keys() & std.keys())
order = ["si-snr_i", "sdr_i", "pesq", "stoi", "si-snr", "sdr"]
def get_order(k):
try:
ind = order.index(k)
return ind
except ValueError:
return 100
self.audiomodel.get_MACs()
self.videomodel.get_MACs()
results_dict = []
results_dict.append(("Model", self.conf["log"]["exp_name"]))
results_dict.append(("MACs and Params", self.audiomodel.macs_parms))
results_dict.append(("Videomodel MACs", self.videomodel.macs))
results_dict.append(("Videomodel Params", self.videomodel.number_of_parameters))
keys.sort(key=get_order)
for k in keys:
m, s = round(mean[k], 4), round(std[k], 3)
results_dict.append((k, str(m) + " ± " + str(s)))
print(f"{k}\tmean: {m} std: {s}")
for k, v in self.conf["audionet"].items():
if isinstance(v, dict):
results_dict.extend([(k + "_" + kk, vv) for kk, vv in v.items()])
else:
results_dict.append((k, v))
df = pd.DataFrame.from_records(results_dict, columns=["Key", "Value"])
df.to_csv(os.path.join(self.ex_save_dir, "results.csv"), encoding="utf-8", index=False)
def _save_audio_example(self, idx, mix_np, sources_np, est_sources_np):
examples_dir = os.path.join(self.ex_save_dir, "examples")
if not os.path.exists(examples_dir):
os.makedirs(examples_dir)
est_sources_np = est_sources_np[0].cpu().unsqueeze(0)
torchaudio.save(os.path.join(self.ex_save_dir, "examples", str(idx) + "_est.wav"), est_sources_np, 16000)
sources_np = sources_np[0].cpu().unsqueeze(0)
torchaudio.save(os.path.join(self.ex_save_dir, "examples", str(idx) + "_gt.wav"), sources_np, 16000)
mix_np = mix_np.cpu().unsqueeze(0)
torchaudio.save(os.path.join(self.ex_save_dir, "examples", str(idx) + "_mix.wav"), mix_np, 16000)
def test_dataloader(self):
test_set = AVSpeechDataset(
self.conf["test_dir"],
n_src=self.conf["data"]["nondefault_nsrc"],
sample_rate=self.conf["data"]["sample_rate"],
segment=None,
normalize_audio=self.conf["data"]["normalize_audio"],
)
data_loader = DataLoader(
test_set,
shuffle=False,
batch_size=self.conf["training"]["batch_size"] * 2,
num_workers=self.conf["training"]["num_workers"],
)
return data_loader
def main(conf):
model = TestModule(conf)
trainer = pl.Trainer(
default_root_dir=model.exp_dir,
devices=[0],
accelerator="auto",
sync_batchnorm=True,
)
trainer.test(model)
if __name__ == "__main__":
parser = argparse.ArgumentParser()
parser.add_argument(
"-t",
"--test-dir",
type=str,
default="data-preprocess/LRS2/tt",
help="Test directory including the json files",
)
parser.add_argument(
"-c",
"--conf-dir",
type=str,
default="../experiments/audio-visual/RTFS-Net/LRS2/4_layers/conf.yaml",
help="Full path to save best validation model",
)
parser.add_argument(
"--n-save-ex",
type=int,
default=-1,
help="Number of audio examples to save, -1 means none",
)
args = parser.parse_args()
with open(args.conf_dir) as f:
def_conf = yaml.safe_load(f)
arg_dic = parse_args_as_dict(parser)
def_conf.update(arg_dic["main_args"])
main(def_conf)