def eval_openset(self):
self.sequential_extract(self.valid_test,
f"{self.project_dir}/tmp/test.h5")
if self.valid_enroll:
self.sequential_extract(self.valid_enroll,
f"{self.project_dir}/tmp/enroll.h5")
enroll_embedding = f"{self.project_dir}/tmp/enroll.h5"
else:
enroll_embedding = f"{self.project_dir}/tmp/test.h5"
if self.valid_target:
self.sequential_extract(self.valid_target,
f"{self.project_dir}/tmp/target.h5")
data_target = h52dict(f"{self.project_dir}/tmp/target.h5")
transform_lst = [PCA(whiten=True)]
for transform in transform_lst:
transform.fit_transform(data_target["X"])
else:
transform_lst = None
if self.score_paras is None:
self.score_paras = {}
scorer = Scorer(
comp_minDCF=False,
enroll=enroll_embedding,
test=f"{self.project_dir}/tmp/test.h5",
ndx_file=self.valid_trial_list,
transforms=transform_lst,
**self.score_paras,
)
eer = scorer.batch_cosine_score()
with open(f"{self.logger_dir}/validation.log", "a") as f:
f.write(f"{self.epoch} EER is {eer}\n")
os.mkdir(train_para["project_dir"])
with open(train_para["project_dir"] + "/models/train_config.yaml",
"a") as f:
yaml.dump(train_para, f, default_flow_style=False)
"""
Calling sequence: Trainer.dist_train() -->
Trainer._dist_train() --> SemiTrainer.train() -->
SemiTrainer._semi_train()
"""
SemiTrainer(**train_para).dist_train()
if run_mode == "scoring":
print('Scoring...')
nclasses = Trainer.get_n_classes(train_para['train_file'],
train_para['min_utts_per_spk'],
train_para['sample_length_range'][1])
if net_type == 'DenseNet':
densenet_para = {'n_classes': nclasses, 'mfcc_dim': mfcc_dim}
model = Model(**densenet_para)
else:
model = Model(nclasses, mfcc_dim)
model.load_state_dict(
torch.load(f"{PRJ_DIR}/models/model_final")['model_state_dict'])
model.to(torch.device(torch.cuda.current_device()))
sequential_extract(model, f"{PRJ_DIR}/h5/mfcc_enroll.h5",
score_para['enroll'])
sequential_extract(model, f"{PRJ_DIR}/h5/mfcc_test.h5", score_para['test'])
scorer = Scorer(**score_para)
scorer.batch_cosine_score()
