def eval_recu(epoch):
assert config.batch_size == 1
model.eval()
reference, candidate, source, alignments = [], [], [], []
e = epoch
test_or_valid = 'test'
test_or_valid = 'valid'
# test_or_valid = 'train'
print(('Test or valid:', test_or_valid))
eval_data_gen = prepare_data('once', test_or_valid, config.MIN_MIX,
config.MAX_MIX)
SDR_SUM = np.array([])
SDRi_SUM = np.array([])
batch_idx = 0
global best_SDR, Var
while True:
print(('-' * 30))
eval_data = next(eval_data_gen)
if eval_data == False:
print(('SDR_aver_eval_epoch:', SDR_SUM.mean()))
print(('SDRi_aver_eval_epoch:', SDRi_SUM.mean()))
break # 如果这个epoch的生成器没有数据了，直接进入下一个epoch
src = Variable(torch.from_numpy(eval_data['mix_feas']))
# raw_tgt = [sorted(spk.keys()) for spk in eval_data['multi_spk_fea_list']]
raw_tgt = eval_data['batch_order']
feas_tgt = models.rank_feas(
raw_tgt, eval_data['multi_spk_fea_list']) # 这里是目标的图谱
src_original = src.transpose(0, 1) #To T,bs,F
predict_multi_mask_all = None
samples_list = []
for len_idx in range(config.MIN_MIX + 2, 2, -1): #逐个分离
tgt_max_len = len_idx # 4,3,2 with bos and eos.
topk_k = len_idx - 2
tgt = Variable(torch.ones(
len_idx, config.batch_size)) # 这里随便给一个tgt，为了测试阶段tgt的名字无所谓其实。
src_len = Variable(
torch.LongTensor(config.batch_size).zero_() +
mix_speech_len).unsqueeze(0)
tgt_len = Variable(
torch.LongTensor([
tgt_max_len - 2
for one_spk in eval_data['multi_spk_fea_list']
])).unsqueeze(0)
if use_cuda:
src = src.cuda().transpose(0, 1) # to T,bs,fre
src_original = src_original.cuda() # TO T,bs,fre
tgt = tgt.cuda()
src_len = src_len.cuda()
tgt_len = tgt_len.cuda()
feas_tgt = feas_tgt.cuda()
# try:
if len(opt.gpus) > 1:
samples, alignment, hiddens, predicted_masks = model.module.beam_sample(
src, src_len, dict_spk2idx, tgt, config.beam_size,
src_original)
else:
samples, predicted_masks = model.beam_sample(
src, src_len, dict_spk2idx, tgt, config.beam_size,
src_original)
# except:
# continue
# '''
# expand the raw mixed-features to topk_max channel.
src = src_original.transpose(0, 1) #确保分离的时候用的是原始的语音
siz = src.size()
assert len(siz) == 3
topk_max = topk_k
x_input_map_multi = torch.unsqueeze(src, 1).expand(
siz[0], topk_max, siz[1], siz[2])
if 0 and config.WFM:
feas_tgt = x_input_map_multi.data * WFM_mask
if len_idx == 4:
aim_feas = list(range(0, 2 * config.batch_size,
2)) #每个samples的第一个说话人取出来
predict_multi_mask_all = predicted_masks #bs*topk,T,F
src = src * (1 - predicted_masks[aim_feas]
) #调整到bs为第一维,# bs,T,F
samples_list = samples
elif len_idx == 3:
aim_feas = list(range(1, 2 * config.batch_size,
2)) #每个samples的第二个说话人取出来
predict_multi_mask_all[aim_feas] = predicted_masks
feas_tgt = feas_tgt[aim_feas]
samples_list = [samples_list[:1] + samples]
if test_or_valid != 'test':
if 1 and len(opt.gpus) > 1:
ss_loss = model.module.separation_loss(
x_input_map_multi,
predicted_masks,
feas_tgt,
)
else:
ss_loss = model.separation_loss(x_input_map_multi,
predicted_masks, feas_tgt)
print(('loss for ss,this batch:', ss_loss.cpu().item()))
lera.log({
'ss_loss_' + str(len_idx) + test_or_valid:
ss_loss.cpu().item(),
})
del ss_loss
predicted_masks = predict_multi_mask_all
if batch_idx <= (500 / config.batch_size
): # only the former batches counts the SDR
predicted_maps = predicted_masks * x_input_map_multi
# predicted_maps=Variable(feas_tgt)
utils.bss_eval2(config,
predicted_maps,
eval_data['multi_spk_fea_list'],
raw_tgt,
eval_data,
dst='batch_output_test')
del predicted_maps, predicted_masks, x_input_map_multi
try:
sdr_aver_batch, sdri_aver_batch = bss_test.cal(
'batch_output_test/')
SDR_SUM = np.append(SDR_SUM, sdr_aver_batch)
SDRi_SUM = np.append(SDRi_SUM, sdri_aver_batch)
except (AssertionError):
print('Errors in calculating the SDR')
print(('SDR_aver_now:', SDR_SUM.mean()))
print(('SDRi_aver_now:', SDRi_SUM.mean()))
lera.log({'SDR sample' + test_or_valid: SDR_SUM.mean()})
lera.log({'SDRi sample' + test_or_valid: SDRi_SUM.mean()})
writer.add_scalars('scalar/loss',
{'SDR_sample_' + test_or_valid: sdr_aver_batch},
updates)
# raw_input('Press any key to continue......')
# '''
candidate += [
convertToLabels(dict_idx2spk, s, dict_spk2idx['<EOS>'])
for s in samples_list
]
# source += raw_src
reference += raw_tgt
print(('samples:', samples))
print(('can:{}, \nref:{}'.format(candidate[-1 * config.batch_size:],
reference[-1 * config.batch_size:])))
# alignments += [align for align in alignment]
batch_idx += 1
input('wait to continue......')
result = utils.eval_metrics(reference, candidate, dict_spk2idx,
log_path)
print((
'hamming_loss: %.8f | micro_f1: %.4f |recall: %.4f | precision: %.4f'
% (
result['hamming_loss'],
result['micro_f1'],
result['micro_recall'],
result['micro_precision'],
)))
score = {}
result = utils.eval_metrics(reference, candidate, dict_spk2idx, log_path)
logging_csv([e, updates, result['hamming_loss'], \
result['micro_f1'], result['micro_precision'], result['micro_recall'],SDR_SUM.mean()])
print(('hamming_loss: %.8f | micro_f1: %.4f' %
(result['hamming_loss'], result['micro_f1'])))
score['hamming_loss'] = result['hamming_loss']
score['micro_f1'] = result['micro_f1']
1 / 0
return score
def eval(epoch):
# config.batch_size=1
model.eval()
print '\n\n测试的时候请设置config里的batch_size为1！！！please set the batch_size as 1'
reference, candidate, source, alignments = [], [], [], []
e = epoch
test_or_valid = 'test'
print('Test or valid:', test_or_valid)
eval_data_gen = prepare_data('once', test_or_valid, config.MIN_MIX,
config.MAX_MIX)
SDR_SUM = np.array([])
SDRi_SUM = np.array([])
batch_idx = 0
global best_SDR, Var
while True:
print('-' * 30)
eval_data = eval_data_gen.next()
if eval_data == False:
print('SDR_aver_eval_epoch:', SDR_SUM.mean())
print('SDRi_aver_eval_epoch:', SDRi_SUM.mean())
break # 如果这个epoch的生成器没有数据了，直接进入下一个epoch
src = Variable(torch.from_numpy(eval_data['mix_feas']))
raw_tgt = [
sorted(spk.keys()) for spk in eval_data['multi_spk_fea_list']
]
feas_tgt = models.rank_feas(
raw_tgt, eval_data['multi_spk_fea_list']) # 这里是目标的图谱
top_k = len(raw_tgt[0])
# 要保证底下这几个都是longTensor(长整数）
# tgt = Variable(torch.from_numpy(np.array([[0]+[dict_spk2idx[spk] for spk in spks]+[dict_spk2idx['<EOS>']] for spks in raw_tgt],dtype=np.int))).transpose(0,1) #转换成数字，然后前后加开始和结束符号。
tgt = Variable(torch.ones(
top_k + 2, config.batch_size)) # 这里随便给一个tgt，为了测试阶段tgt的名字无所谓其实。
src_len = Variable(
torch.LongTensor(config.batch_size).zero_() +
mix_speech_len).unsqueeze(0)
tgt_len = Variable(
torch.LongTensor([
len(one_spk) for one_spk in eval_data['multi_spk_fea_list']
])).unsqueeze(0)
# tgt_len = Variable(torch.LongTensor(config.batch_size).zero_()+len(eval_data['multi_spk_fea_list'][0])).unsqueeze(0)
if config.WFM:
tmp_size = feas_tgt.size()
assert len(tmp_size) == 4
feas_tgt_sum = torch.sum(feas_tgt, dim=1, keepdim=True)
feas_tgt_sum_square = (feas_tgt_sum *
feas_tgt_sum).expand(tmp_size)
feas_tgt_square = feas_tgt * feas_tgt
WFM_mask = feas_tgt_square / feas_tgt_sum_square
if use_cuda:
src = src.cuda().transpose(0, 1)
tgt = tgt.cuda()
src_len = src_len.cuda()
tgt_len = tgt_len.cuda()
feas_tgt = feas_tgt.cuda()
if config.WFM:
WFM_mask = WFM_mask.cuda()
if 1 and len(opt.gpus) > 1:
samples, alignment, hiddens, predicted_masks = model.module.beam_sample(
src, src_len, dict_spk2idx, tgt, beam_size=config.beam_size)
else:
samples, alignment, hiddens, predicted_masks = model.beam_sample(
src, src_len, dict_spk2idx, tgt, beam_size=config.beam_size)
# '''
# expand the raw mixed-features to topk_max channel.
src = src.transpose(0, 1)
siz = src.size()
assert len(siz) == 3
# if samples[0][-1] != dict_spk2idx['<EOS>']:
# print '*'*40+'\nThe model is far from good. End the evaluation.\n'+'*'*40
# break
topk_max = len(samples[0]) - 1
x_input_map_multi = torch.unsqueeze(src,
1).expand(siz[0], topk_max, siz[1],
siz[2])
if config.WFM:
feas_tgt = x_input_map_multi.data * WFM_mask
if 0 and test_or_valid == 'valid':
if 1 and len(opt.gpus) > 1:
ss_loss = model.module.separation_loss(
x_input_map_multi,
predicted_masks,
feas_tgt,
)
else:
ss_loss = model.separation_loss(x_input_map_multi,
predicted_masks, feas_tgt)
print('loss for ss,this batch:', ss_loss.cpu().item())
lera.log({
'ss_loss_' + test_or_valid: ss_loss.cpu().item(),
})
del ss_loss, hiddens
# '''''
if batch_idx <= (500 / config.batch_size
): # only the former batches counts the SDR
predicted_maps = predicted_masks * x_input_map_multi
# predicted_maps=Variable(feas_tgt)
utils.bss_eval2(config,
predicted_maps,
eval_data['multi_spk_fea_list'],
raw_tgt,
eval_data,
dst='batch_output')
del predicted_maps, predicted_masks, x_input_map_multi
try:
SDR_SUM, SDRi_SUM = np.append(SDR_SUM,
bss_test.cal('batch_output/'))
except AssertionError, wrong_info:
print 'Errors in calculating the SDR', wrong_info
print('SDR_aver_now:', SDR_SUM.mean())
print('SDRi_aver_now:', SDRi_SUM.mean())
lera.log({'SDR sample' + test_or_valid: SDR_SUM.mean()})
lera.log({'SDRi sample' + test_or_valid: SDRi_SUM.mean()})
# raw_input('Press any key to continue......')
elif batch_idx == (500 / config.batch_size) + 1 and SDR_SUM.mean(
) > best_SDR: # only record the best SDR once.
print('Best SDR from {}---->{}'.format(best_SDR, SDR_SUM.mean()))
best_SDR = SDR_SUM.mean()
def eval(epoch):
model.eval()
reference, candidate, source, alignments = [], [], [], []
e = epoch
test_or_valid = 'test'
print 'Test or valid:', test_or_valid
eval_data_gen = prepare_data_aim('once', test_or_valid, config.MIN_MIX,
config.MAX_MIX)
# for raw_src, src, src_len, raw_tgt, tgt, tgt_len in validloader:
SDR_SUM = np.array([])
batch_idx = 0
global best_SDR, Var
while True:
# for ___ in range(2):
print '-' * 30
eval_data = eval_data_gen.next()
if eval_data == False:
print 'SDR_aver_eval_epoch:', SDR_SUM.mean()
break #如果这个epoch的生成器没有数据了，直接进入下一个epoch
src = Variable(torch.from_numpy(eval_data['mix_feas']))
raw_tgt = [
sorted(spk.keys()) for spk in eval_data['multi_spk_fea_list']
]
feas_tgt = models.rank_feas(raw_tgt,
eval_data['multi_spk_fea_list']) #这里是目标的图谱
top_k = len(raw_tgt[0])
# 要保证底下这几个都是longTensor(长整数）
# tgt = Variable(torch.from_numpy(np.array([[0]+[dict_spk2idx[spk] for spk in spks]+[dict_spk2idx['<EOS>']] for spks in raw_tgt],dtype=np.int))).transpose(0,1) #转换成数字，然后前后加开始和结束符号。
tgt = Variable(torch.ones(
top_k + 2, config.batch_size)) # 这里随便给一个tgt，为了测试阶段tgt的名字无所谓其实。
src_len = Variable(
torch.LongTensor(config.batch_size).zero_() +
mix_speech_len).unsqueeze(0)
tgt_len = Variable(
torch.LongTensor([
len(one_spk) for one_spk in eval_data['multi_spk_fea_list']
])).unsqueeze(0)
# tgt_len = Variable(torch.LongTensor(config.batch_size).zero_()+len(eval_data['multi_spk_fea_list'][0])).unsqueeze(0)
if config.WFM:
tmp_size = feas_tgt.size()
assert len(tmp_size) == 4
feas_tgt_sum = torch.sum(feas_tgt, dim=1, keepdim=True)
feas_tgt_sum_square = (feas_tgt_sum *
feas_tgt_sum).expand(tmp_size)
feas_tgt_square = feas_tgt * feas_tgt
WFM_mask = feas_tgt_square / feas_tgt_sum_square
if use_cuda:
src = src.cuda().transpose(0, 1)
tgt = tgt.cuda()
src_len = src_len.cuda()
tgt_len = tgt_len.cuda()
feas_tgt = feas_tgt.cuda()
if config.WFM:
WFM_mask = WFM_mask.cuda()
try:
if 1 and len(opt.gpus) > 1:
# samples, alignment = model.module.sample(src, src_len)
samples, alignment, hiddens, predicted_masks = model.module.beam_sample(
src,
src_len,
dict_spk2idx,
tgt,
beam_size=config.beam_size)
else:
samples, alignment, hiddens, predicted_masks = model.beam_sample(
src,
src_len,
dict_spk2idx,
tgt,
beam_size=config.beam_size)
# samples, alignment, hiddens, predicted_masks = model.beam_sample(src, src_len, dict_spk2idx, tgt, beam_size=config.beam_size)
except TabError, info:
print '**************Error occurs here************:', info
continue
if config.top1:
predicted_masks = torch.cat([predicted_masks, 1 - predicted_masks],
1)
# '''
# expand the raw mixed-features to topk_max channel.
src = src.transpose(0, 1)
siz = src.size()
assert len(siz) == 3
topk_max = feas_tgt.size()[1]
assert samples[0][-1] == dict_spk2idx['<EOS>']
topk_max = len(samples[0]) - 1
x_input_map_multi = torch.unsqueeze(src,
1).expand(siz[0], topk_max, siz[1],
siz[2])
if config.WFM:
feas_tgt = x_input_map_multi.data * WFM_mask
if test_or_valid == 'valid':
if 1 and len(opt.gpus) > 1:
ss_loss = model.module.separation_loss(x_input_map_multi,
predicted_masks,
feas_tgt, Var)
else:
ss_loss = model.separation_loss(x_input_map_multi,
predicted_masks, feas_tgt)
print 'loss for ss,this batch:', ss_loss.data[0]
lera.log({
'ss_loss_' + test_or_valid: ss_loss.data[0],
})
del ss_loss, hiddens
# '''''
if batch_idx <= (500 / config.batch_size
): #only the former batches counts the SDR
predicted_maps = predicted_masks * x_input_map_multi
# predicted_maps=Variable(feas_tgt)
utils.bss_eval2(config,
predicted_maps,
eval_data['multi_spk_fea_list'],
raw_tgt,
eval_data,
dst='batch_outputjaa')
del predicted_maps, predicted_masks, x_input_map_multi
SDR_SUM = np.append(SDR_SUM, bss_test.cal('batch_outputjaa/'))
print 'SDR_aver_now:', SDR_SUM.mean()
lera.log({'SDR sample': SDR_SUM.mean()})
# raw_input('Press any key to continue......')
elif batch_idx == (500 / config.batch_size) + 1 and SDR_SUM.mean(
) > best_SDR: #only record the best SDR once.
print 'Best SDR from {}---->{}'.format(best_SDR, SDR_SUM.mean())
best_SDR = SDR_SUM.mean()
# save_model(log_path+'checkpoint_bestSDR{}.pt'.format(best_SDR))
# '''
candidate += [
convertToLabels(dict_idx2spk, s, dict_spk2idx['<EOS>'])
for s in samples
]
# source += raw_src
reference += raw_tgt
print 'samples:', samples
print 'can:{}, \nref:{}'.format(candidate[-1 * config.batch_size:],
reference[-1 * config.batch_size:])
alignments += [align for align in alignment]
batch_idx += 1
def eval(epoch):
# config.batch_size=1
model.eval()
# print '\n\n测试的时候请设置config里的batch_size为1！！！please set the batch_size as 1'
reference, candidate, source, alignments = [], [], [], []
e = epoch
# test_or_valid = 'test'
test_or_valid = 'valid'
print('Test or valid:', test_or_valid)
eval_data_gen = prepare_data('once', test_or_valid, config.MIN_MIX,
config.MAX_MIX)
SDR_SUM = np.array([])
SDRi_SUM = np.array([])
batch_idx = 0
global best_SDR, Var
while True:
print('-' * 30)
eval_data = next(eval_data_gen)
if eval_data == False:
print('SDR_aver_eval_epoch:', SDR_SUM.mean())
print('SDRi_aver_eval_epoch:', SDRi_SUM.mean())
break # 如果这个epoch的生成器没有数据了，直接进入下一个epoch
src = Variable(torch.from_numpy(eval_data['mix_feas']))
raw_tgt = [
sorted(spk.keys()) for spk in eval_data['multi_spk_fea_list']
]
feas_tgt = models.rank_feas(
raw_tgt, eval_data['multi_spk_fea_list']) # 这里是目标的图谱
padded_mixture, mixture_lengths, padded_source = eval_data['tas_zip']
padded_mixture = torch.from_numpy(padded_mixture).float()
mixture_lengths = torch.from_numpy(mixture_lengths)
padded_source = torch.from_numpy(padded_source).float()
padded_mixture = padded_mixture.cuda().transpose(0, 1)
mixture_lengths = mixture_lengths.cuda()
padded_source = padded_source.cuda()
top_k = len(raw_tgt[0])
# 要保证底下这几个都是longTensor(长整数）
# tgt = Variable(torch.from_numpy(np.array([[0]+[dict_spk2idx[spk] for spk in spks]+[dict_spk2idx['<EOS>']] for spks in raw_tgt],dtype=np.int))).transpose(0,1) #转换成数字，然后前后加开始和结束符号。
tgt = Variable(torch.ones(
top_k + 2, config.batch_size)) # 这里随便给一个tgt，为了测试阶段tgt的名字无所谓其实。
src_len = Variable(
torch.LongTensor(config.batch_size).zero_() +
mix_speech_len).unsqueeze(0)
tgt_len = Variable(
torch.LongTensor([
len(one_spk) for one_spk in eval_data['multi_spk_fea_list']
])).unsqueeze(0)
# tgt_len = Variable(torch.LongTensor(config.batch_size).zero_()+len(eval_data['multi_spk_fea_list'][0])).unsqueeze(0)
if config.WFM:
tmp_size = feas_tgt.size()
#assert len(tmp_size) == 4
feas_tgt_sum = torch.sum(feas_tgt, dim=1, keepdim=True)
feas_tgt_sum_square = (feas_tgt_sum *
feas_tgt_sum).expand(tmp_size)
feas_tgt_square = feas_tgt * feas_tgt
WFM_mask = feas_tgt_square / feas_tgt_sum_square
if use_cuda:
src = src.cuda().transpose(0, 1)
tgt = tgt.cuda()
src_len = src_len.cuda()
tgt_len = tgt_len.cuda()
feas_tgt = feas_tgt.cuda()
if config.WFM:
WFM_mask = WFM_mask.cuda()
if 1 and len(opt.gpus) > 1:
samples, alignment, hiddens, predicted_masks = model.module.beam_sample(
src, src_len, dict_spk2idx, tgt, config.beam_size,
padded_mixture)
else:
samples, alignment, hiddens, predicted_masks = model.beam_sample(
src, src_len, dict_spk2idx, tgt, config.beam_size,
padded_mixture)
# '''
# expand the raw mixed-features to topk_max channel.
src = src.transpose(0, 1)
siz = src.size()
assert len(siz) == 3
# if samples[0][-1] != dict_spk2idx['<EOS>']:
# print '*'*40+'\nThe model is far from good. End the evaluation.\n'+'*'*40
# break
topk_max = len(samples[0]) - 1
x_input_map_multi = torch.unsqueeze(src,
1).expand(siz[0], topk_max, siz[1],
siz[2])
if config.WFM:
feas_tgt = x_input_map_multi.data * WFM_mask
if not config.use_tas and test_or_valid == 'valid':
if 1 and len(opt.gpus) > 1:
ss_loss = model.module.separation_loss(
x_input_map_multi,
predicted_masks,
feas_tgt,
)
else:
ss_loss = model.separation_loss(x_input_map_multi,
predicted_masks, feas_tgt)
print('loss for ss,this batch:', ss_loss.cpu().item())
# lera.log({
# 'ss_loss_' + test_or_valid: ss_loss.cpu().item(),
# })
del ss_loss, hiddens
# '''''
if batch_idx <= (100 / config.batch_size
): # only the former batches counts the SDR
if config.use_tas:
utils.bss_eval_tas(config,
predicted_masks,
eval_data['multi_spk_fea_list'],
raw_tgt,
eval_data,
dst='batch_output1')
else:
predicted_maps = predicted_masks * x_input_map_multi
utils.bss_eval2(config,
predicted_maps,
eval_data['multi_spk_fea_list'],
raw_tgt,
eval_data,
dst='batch_output1')
del predicted_maps
del predicted_masks, x_input_map_multi
try:
#SDR_SUM,SDRi_SUM = np.append(SDR_SUM, bss_test.cal('batch_output1/'))
sdr_aver_batch, sdri_aver_batch = bss_test.cal(
'batch_output1/')
SDR_SUM = np.append(SDR_SUM, sdr_aver_batch)
SDRi_SUM = np.append(SDRi_SUM, sdri_aver_batch)
except: # AssertionError,wrong_info:
print('Errors in calculating the SDR', wrong_info)
print('SDR_aver_now:', SDR_SUM.mean())
print('SDRi_aver_now:', SDRi_SUM.mean())
# lera.log({'SDR sample'+test_or_valid: SDR_SUM.mean()})
# lera.log({'SDRi sample'+test_or_valid: SDRi_SUM.mean()})
# raw_input('Press any key to continue......')
elif batch_idx == (500 / config.batch_size) + 1 and SDR_SUM.mean(
) > best_SDR: # only record the best SDR once.
print('Best SDR from {}---->{}'.format(best_SDR, SDR_SUM.mean()))
best_SDR = SDR_SUM.mean()
# save_model(log_path+'checkpoint_bestSDR{}.pt'.format(best_SDR))
# '''
candidate += [
convertToLabels(dict_idx2spk, s, dict_spk2idx['<EOS>'])
for s in samples
]
# source += raw_src
reference += raw_tgt
print('samples:', samples)
print('can:{}, \nref:{}'.format(candidate[-1 * config.batch_size:],
reference[-1 * config.batch_size:]))
alignments += [align for align in alignment]
batch_idx += 1
score = {}
result = utils.eval_metrics(reference, candidate, dict_spk2idx, log_path)
logging_csv([e, updates, result['hamming_loss'], \
result['micro_f1'], result['micro_precision'], result['micro_recall']])
print('hamming_loss: %.8f | micro_f1: %.4f' %
(result['hamming_loss'], result['micro_f1']))
score['hamming_loss'] = result['hamming_loss']
score['micro_f1'] = result['micro_f1']
return score
def eval(epoch,test_or_valid='valid'):
# config.batch_size=1
global updates,model
model.eval()
# print '\n\n测试的时候请设置config里的batch_size为1！！！please set the batch_size as 1'
reference, candidate, source, alignments = [], [], [], []
e = epoch
print(('Test or valid:', test_or_valid))
eval_data_gen = prepare_data('once', test_or_valid, config.MIN_MIX, config.MAX_MIX)
SDR_SUM = np.array([])
SDRi_SUM = np.array([])
batch_idx = 0
global best_SDR, Var
# for iii in range(2000):
while True:
print(('-' * 30))
eval_data = next(eval_data_gen)
if eval_data == False:
print(('SDR_aver_eval_epoch:', SDR_SUM.mean()))
print(('SDRi_aver_eval_epoch:', SDRi_SUM.mean()))
break # 如果这个epoch的生成器没有数据了，直接进入下一个epoch
src = Variable(torch.from_numpy(eval_data['mix_feas']))
# raw_tgt = [sorted(spk.keys()) for spk in eval_data['multi_spk_fea_list']]
raw_tgt= eval_data['batch_order']
feas_tgt = models.rank_feas(raw_tgt, eval_data['multi_spk_fea_list']) # 这里是目标的图谱
top_k = len(raw_tgt[0])
# 要保证底下这几个都是longTensor(长整数）
# tgt = Variable(torch.from_numpy(np.array([[0]+[dict_spk2idx[spk] for spk in spks]+[dict_spk2idx['<EOS>']] for spks in raw_tgt],dtype=np.int))).transpose(0,1) #转换成数字，然后前后加开始和结束符号。
tgt = Variable(torch.from_numpy(np.array([[0,1,2,102] for __ in range(config.batch_size)], dtype=np.int))).transpose(0, 1) # 转换成数字，然后前后加开始和结束符号。
src_len = Variable(torch.LongTensor(config.batch_size).zero_() + mix_speech_len).unsqueeze(0)
tgt_len = Variable(torch.LongTensor([len(one_spk) for one_spk in eval_data['multi_spk_fea_list']])).unsqueeze(0)
# tgt_len = Variable(torch.LongTensor(config.batch_size).zero_()+len(eval_data['multi_spk_fea_list'][0])).unsqueeze(0)
if config.WFM:
siz = src.size() # bs,T,F
assert len(siz) == 3
# topk_max = config.MAX_MIX # 最多可能的topk个数
topk_max = 2 # 最多可能的topk个数
x_input_map_multi = torch.unsqueeze(src, 1).expand(siz[0], topk_max, siz[1], siz[2]).contiguous().view(-1, siz[1], siz[ 2]) # bs,topk,T,F
feas_tgt_tmp = feas_tgt.view(siz[0], -1, siz[1], siz[2])
feas_tgt_square = feas_tgt_tmp * feas_tgt_tmp
feas_tgt_sum_square = torch.sum(feas_tgt_square, dim=1, keepdim=True).expand(siz[0], topk_max, siz[1], siz[2])
WFM_mask = feas_tgt_square / (feas_tgt_sum_square + 1e-15)
feas_tgt = x_input_map_multi.view(siz[0], -1, siz[1], siz[2]).data * WFM_mask # bs,topk,T,F
feas_tgt = feas_tgt.view(-1, siz[1], siz[2]) # bs*topk,T,F
WFM_mask = WFM_mask.cuda()
del x_input_map_multi
elif config.PSM:
siz = src.size() # bs,T,F
assert len(siz) == 3
# topk_max = config.MAX_MIX # 最多可能的topk个数
topk_max = 2 # 最多可能的topk个数
x_input_map_multi = torch.unsqueeze(src, 1).expand(siz[0], topk_max, siz[1], siz[2]).contiguous() # bs,topk,T,F
feas_tgt_tmp = feas_tgt.view(siz[0], -1, siz[1], siz[2])
IRM=feas_tgt_tmp/(x_input_map_multi+1e-15)
angle_tgt=models.rank_feas(raw_tgt, eval_data['multi_spk_angle_list']).view(siz[0],-1,siz[1],siz[2])
angle_mix=Variable(torch.from_numpy(np.array(eval_data['mix_angle']))).unsqueeze(1).expand(siz[0], topk_max, siz[1], siz[2]).contiguous()
ang=np.cos(angle_mix-angle_tgt)
ang=np.clip(ang,0,None)
# feas_tgt = x_input_map_multi *np.clip(IRM.numpy()*ang,0,1) # bs,topk,T,F
# feas_tgt = x_input_map_multi *IRM*ang # bs,topk,T,F
feas_tgt = feas_tgt.view(siz[0],-1,siz[1],siz[2])*ang # bs,topk,T,F
feas_tgt = feas_tgt.view(-1, siz[1], siz[2]) # bs*topk,T,F
del x_input_map_multi
if use_cuda:
src = src.cuda().transpose(0, 1)
tgt = tgt.cuda()
src_len = src_len.cuda()
tgt_len = tgt_len.cuda()
feas_tgt = feas_tgt.cuda()
if config.WFM:
WFM_mask = WFM_mask.cuda()
predicted_masks, enc_attn_list = model(src, src_len, tgt, tgt_len,
dict_spk2idx) # 这里的outputs就是hidden_outputs，还没有进行最后分类的隐层，可以直接用
print('predicted mask size:', predicted_masks.size(),'should be topk,bs,T,F') # topk,bs,T,F
# try:
# '''
# expand the raw mixed-features to topk_max channel.
src = src.transpose(0, 1)
siz = src.size()
assert len(siz) == 3
# if samples[0][-1] != dict_spk2idx['<EOS>']:
# print '*'*40+'\nThe model is far from good. End the evaluation.\n'+'*'*40
# break
topk_max = config.MAX_MIX
x_input_map_multi = torch.unsqueeze(src, 1).expand(siz[0], topk_max, siz[1], siz[2])
predicted_masks=predicted_masks.transpose(0, 1)
# if config.WFM:
# feas_tgt = x_input_map_multi.data * WFM_mask
# 注意,bs是第二维
assert predicted_masks.shape == x_input_map_multi.shape
assert predicted_masks.size(0) == config.batch_size
if 1 and len(opt.gpus) > 1:
ss_loss,best_pmt = model.module.separation_pit_loss(x_input_map_multi, predicted_masks, feas_tgt, )
else:
ss_loss,best_pmt = model.separation_pit_loss(x_input_map_multi, predicted_masks, feas_tgt)
print(('loss for ss,this batch:', ss_loss.cpu().item()))
print('best perms for this batch:', best_pmt)
lera.log({
'ss_loss_' + test_or_valid: ss_loss.cpu().item(),
})
writer.add_scalars('scalar/loss',{'ss_loss_'+test_or_valid:ss_loss.cpu().item()},updates+batch_idx)
del ss_loss
if batch_idx>10:
break
if False: #this part is to test the checkpoints sequencially.
batch_idx += 1
if batch_idx%100==0:
updates=updates+1000
opt.restore='/data1/shijing_data/2020-02-14-04:58:17/Transformer_PIT_{}.pt'.format(updates)
print('loading checkpoint...\n', opt.restore)
checkpoints = torch.load(opt.restore)
model.module.load_state_dict(checkpoints['model'])
break
continue
# '''''
if 0 and batch_idx <= (500 / config.batch_size): # only the former batches counts the SDR
predicted_maps = predicted_masks * x_input_map_multi
predicted_maps = predicted_maps.view(-1,mix_speech_len,speech_fre)
# predicted_maps=Variable(feas_tgt)
utils.bss_eval2(config, predicted_maps, eval_data['multi_spk_fea_list'], raw_tgt, eval_data,
dst='batch_output_test')
# utils.bss_eval(config, predicted_maps, eval_data['multi_spk_fea_list'], raw_tgt, eval_data,
# dst='batch_output_test')
del predicted_maps, predicted_masks, x_input_map_multi
try:
sdr_aver_batch, sdri_aver_batch= bss_test.cal('batch_output_test/')
SDR_SUM = np.append(SDR_SUM, sdr_aver_batch)
SDRi_SUM = np.append(SDRi_SUM, sdri_aver_batch)
except(AssertionError):
print('Errors in calculating the SDR')
print(('SDR_aver_now:', SDR_SUM.mean()))
print(('SRi_aver_now:', SDRi_SUM.mean()))
lera.log({'SDR sample'+test_or_valid: SDR_SUM.mean()})
lera.log({'SDRi sample'+test_or_valid: SDRi_SUM.mean()})
writer.add_scalars('scalar/loss',{'SDR_sample_'+test_or_valid:sdr_aver_batch},updates)
# raw_input('Press any key to continue......')
elif batch_idx == (200 / config.batch_size) + 1 and SDR_SUM.mean() > best_SDR: # only record the best SDR once.
print(('Best SDR from {}---->{}'.format(best_SDR, SDR_SUM.mean())))
best_SDR = SDR_SUM.mean()
# save_model(log_path+'checkpoint_bestSDR{}.pt'.format(best_SDR))
# '''
# candidate += [convertToLabels(dict_idx2spk, s, dict_spk2idx['<EOS>']) for s in samples]
# source += raw_src
# reference += raw_tgt
# print(('samples:', samples))
# print(('can:{}, \nref:{}'.format(candidate[-1 * config.batch_size:], reference[-1 * config.batch_size:])))
# alignments += [align for align in alignment]
batch_idx += 1
result = utils.eval_metrics(reference, candidate, dict_spk2idx, log_path)
print(('hamming_loss: %.8f | micro_f1: %.4f |recall: %.4f | precision: %.4f'
% (result['hamming_loss'], result['micro_f1'], result['micro_recall'], result['micro_precision'], )))
def train(epoch):
global e, updates, total_loss, start_time, report_total,report_correct, total_loss_sgm, total_loss_ss
e = epoch
model.train()
SDR_SUM = np.array([])
SDRi_SUM = np.array([])
if updates<=config.warmup: #如果不在warm阶段就正常规划
pass
elif config.schedule and scheduler.get_lr()[0]>4e-5:
scheduler.step()
print(("Decaying learning rate to %g" % scheduler.get_lr()[0],updates))
lera.log({
'lr': [group['lr'] for group in optim.optimizer.param_groups][0],
})
if opt.model == 'gated':
model.current_epoch = epoch
train_data_gen = prepare_data('once', 'train')
while True:
if updates <= config.warmup: # 如果在warm就开始warmup
tmp_lr = config.learning_rate * min(max(updates,1)** (-0.5),
max(updates,1) * (config.warmup ** (-1.5)))
for param_group in optim.optimizer.param_groups:
param_group['lr'] = tmp_lr
scheduler.base_lrs=list([group['lr'] for group in optim.optimizer.param_groups])
if updates%100==0: #记录一下
print(updates)
print("Warmup learning rate to %g" % tmp_lr)
lera.log({
'lr': [group['lr'] for group in optim.optimizer.param_groups][0],
})
train_data = next(train_data_gen)
if train_data == False:
print(('SDR_aver_epoch:', SDR_SUM.mean()))
print(('SDRi_aver_epoch:', SDRi_SUM.mean()))
break # 如果这个epoch的生成器没有数据了，直接进入下一个epoch
src = Variable(torch.from_numpy(train_data['mix_feas']))
# raw_tgt = [spk.keys() for spk in train_data['multi_spk_fea_list']]
# raw_tgt = [sorted(spk.keys()) for spk in train_data['multi_spk_fea_list']]
raw_tgt=train_data['batch_order']
feas_tgt = models.rank_feas(raw_tgt, train_data['multi_spk_fea_list']) # 这里是目标的图谱,bs*Topk,len,fre
# 要保证底下这几个都是longTensor(长整数）
tgt_max_len = config.MAX_MIX + 2 # with bos and eos.
tgt = Variable(torch.from_numpy(np.array(
[[0] + [dict_spk2idx[spk] for spk in spks] + (tgt_max_len - len(spks) - 1) * [dict_spk2idx['<EOS>']] for
spks in raw_tgt], dtype=np.int))).transpose(0, 1) # 转换成数字，然后前后加开始和结束符号。
# tgt = Variable(torch.from_numpy(np.array([[0,1,2,102] for __ in range(config.batch_size)], dtype=np.int))).transpose(0, 1) # 转换成数字，然后前后加开始和结束符号。
src_len = Variable(torch.LongTensor(config.batch_size).zero_() + mix_speech_len).unsqueeze(0)
tgt_len = Variable(
torch.LongTensor([len(one_spk) for one_spk in train_data['multi_spk_fea_list']])).unsqueeze(0)
if config.WFM:
siz = src.size() # bs,T,F
assert len(siz) == 3
# topk_max = config.MAX_MIX # 最多可能的topk个数
topk_max = 2 # 最多可能的topk个数
x_input_map_multi = torch.unsqueeze(src, 1).expand(siz[0], topk_max, siz[1], siz[2]).contiguous().view(-1, siz[1], siz[ 2]) # bs,topk,T,F
feas_tgt_tmp = feas_tgt.view(siz[0], -1, siz[1], siz[2])
feas_tgt_square = feas_tgt_tmp * feas_tgt_tmp
feas_tgt_sum_square = torch.sum(feas_tgt_square, dim=1, keepdim=True).expand(siz[0], topk_max, siz[1], siz[2])
WFM_mask = feas_tgt_square / (feas_tgt_sum_square + 1e-15)
feas_tgt = x_input_map_multi.view(siz[0], -1, siz[1], siz[2]).data * WFM_mask # bs,topk,T,F
feas_tgt = feas_tgt.view(-1, siz[1], siz[2]) # bs*topk,T,F
WFM_mask = WFM_mask.cuda()
del x_input_map_multi
elif config.PSM:
siz = src.size() # bs,T,F
assert len(siz) == 3
# topk_max = config.MAX_MIX # 最多可能的topk个数
topk_max = 2 # 最多可能的topk个数
x_input_map_multi = torch.unsqueeze(src, 1).expand(siz[0], topk_max, siz[1], siz[2]).contiguous() # bs,topk,T,F
feas_tgt_tmp = feas_tgt.view(siz[0], -1, siz[1], siz[2])
IRM=feas_tgt_tmp/(x_input_map_multi+1e-15)
angle_tgt=models.rank_feas(raw_tgt, train_data['multi_spk_angle_list']).view(siz[0],-1,siz[1],siz[2]) # bs,topk,T,F
angle_mix=Variable(torch.from_numpy(np.array(train_data['mix_angle']))).unsqueeze(1).expand(siz[0], topk_max, siz[1], siz[2]).contiguous()
ang=np.cos(angle_mix-angle_tgt)
ang=np.clip(ang,0,None)
# feas_tgt = x_input_map_multi *np.clip(IRM.numpy()*ang,0,1) # bs,topk,T,F
# feas_tgt = x_input_map_multi *IRM*ang # bs,topk,T,F
feas_tgt = feas_tgt.view(siz[0],-1,siz[1],siz[2])*ang # bs,topk,T,F
feas_tgt = feas_tgt.view(-1, siz[1], siz[2]) # bs*topk,T,F
del x_input_map_multi
elif config.frame_mask:
siz = src.size() # bs,T,F
assert len(siz) == 3
# topk_max = config.MAX_MIX # 最多可能的topk个数
topk_max = 2 # 最多可能的topk个数
x_input_map_multi = torch.unsqueeze(src, 1).expand(siz[0], topk_max, siz[1], siz[2]).contiguous() # bs,topk,T,F
feas_tgt_tmp = feas_tgt.view(siz[0], -1, siz[1], siz[2])
feas_tgt_time=torch.sum(feas_tgt_tmp,3).transpose(1,2) #bs,T,topk
for v1 in feas_tgt_time:
for v2 in v1:
if v2[0]>v2[1]:
v2[0]=1
v2[1]=0
else:
v2[0]=0
v2[1]=1
frame_mask=feas_tgt_time.transpose(1,2).unsqueeze(-1) #bs,topk,t,1
feas_tgt=x_input_map_multi*frame_mask
feas_tgt = feas_tgt.view(-1, siz[1], siz[2]) # bs*topk,T,F
if use_cuda:
src = src.cuda().transpose(0, 1)
tgt = tgt.cuda()
src_len = src_len.cuda()
tgt_len = tgt_len.cuda()
feas_tgt = feas_tgt.cuda()
model.zero_grad()
if config.use_center_loss:
center_loss.zero_grad()
# aim_list 就是找到有正经说话人的地方的标号
aim_list = (tgt[1:-1].transpose(0, 1).contiguous().view(-1) != dict_spk2idx['<EOS>']).nonzero().squeeze()
aim_list = aim_list.data.cpu().numpy()
multi_mask, enc_attn_list = model(src, src_len, tgt, tgt_len,
dict_spk2idx) # 这里的outputs就是hidden_outputs，还没有进行最后分类的隐层，可以直接用
print('mask size:', multi_mask.size()) # topk,bs,T,F
# print('mask:', multi_mask[0,0,:3:3]) # topk,bs,T,F
# writer.add_histogram('global gamma',gamma, updates)
src = src.transpose(0, 1)
# expand the raw mixed-features to topk_max channel.
siz = src.size()
assert len(siz) == 3
topk_max = config.MAX_MIX # 最多可能的topk个数
x_input_map_multi = torch.unsqueeze(src, 1).expand(siz[0], topk_max, siz[1], siz[2]).contiguous()#.view(-1, siz[1], siz[2])
# x_input_map_multi = x_input_map_multi[aim_list]
# x_input_map_multi = x_input_map_multi.transpose(0, 1) #topk,bs,T,F
multi_mask = multi_mask.transpose(0, 1)
# if config.WFM:
# feas_tgt = x_input_map_multi.data * WFM_mask
# 注意,bs是第二维
assert multi_mask.shape == x_input_map_multi.shape
assert multi_mask.size(0) == config.batch_size
if 1 and len(opt.gpus) > 1: #先ss获取Perm
ss_loss, best_pmt = model.module.separation_pit_loss(x_input_map_multi, multi_mask, feas_tgt)
else:
ss_loss, best_pmt = model.separation_pit_loss(x_input_map_multi, multi_mask, feas_tgt)
print('loss for SS,this batch:', ss_loss.cpu().item())
print('best perms for this batch:', best_pmt)
writer.add_scalars('scalar/loss',{'ss_loss':ss_loss.cpu().item()},updates)
loss = ss_loss
loss.backward()
total_loss_ss += ss_loss.cpu().item()
lera.log({
'ss_loss': ss_loss.cpu().item(),
})
if updates>3 and updates % config.eval_interval in [0, 1, 2, 3, 4, 5, 6, 7, 8, 9,]:
assert multi_mask.shape==x_input_map_multi.shape
assert multi_mask.size(0)==config.batch_size
predicted_maps = (multi_mask * x_input_map_multi).view(siz[0]*topk_max,siz[1],siz[2])
# predicted_maps=Variable(feas_tgt)
# utils.bss_eval(config, predicted_maps, train_data['multi_spk_fea_list'], raw_tgt, train_data, dst=log_path+'batch_output/')
utils.bss_eval2(config, predicted_maps, train_data['multi_spk_fea_list'], raw_tgt, train_data, dst=log_path+'batch_output')
del predicted_maps, multi_mask, x_input_map_multi
sdr_aver_batch, sdri_aver_batch= bss_test.cal(log_path+'batch_output/')
lera.log({'SDR sample': sdr_aver_batch})
lera.log({'SDRi sample': sdri_aver_batch})
writer.add_scalars('scalar/loss',{'SDR_sample':sdr_aver_batch,'SDRi_sample':sdri_aver_batch},updates)
SDR_SUM = np.append(SDR_SUM, sdr_aver_batch)
SDRi_SUM = np.append(SDRi_SUM, sdri_aver_batch)
print(('SDR_aver_now:', SDR_SUM.mean()))
print(('SDRi_aver_now:', SDRi_SUM.mean()))
# Heatmap here
# n_layer个 (head*bs) x lq x dk
'''
import matplotlib.pyplot as plt
ax = plt.gca()
ax.invert_yaxis()
raw_src=models.rank_feas(raw_tgt, train_data['multi_spk_fea_list'])
att_idx=1
att = enc_attn_list[-1].view(config.trans_n_head,config.batch_size,mix_speech_len,mix_speech_len).data.cpu().numpy()[:,att_idx]
for head in range(config.trans_n_head):
xx=att[head]
plt.matshow(xx, cmap=plt.cm.hot, vmin=0,vmax=0.05)
plt.colorbar()
plt.savefig(log_path+'batch_output/'+'head_{}.png'.format(head))
plt.matshow(raw_src[att_idx*2+0].transpose(0,1), cmap=plt.cm.hot, vmin=0,vmax=2)
plt.colorbar()
plt.savefig(log_path+'batch_output/'+'source0.png')
plt.matshow(raw_src[att_idx*2+1].transpose(0,1), cmap=plt.cm.hot, vmin=0,vmax=2)
plt.colorbar()
plt.savefig(log_path+'batch_output/'+'source1.png')
1/0
'''
total_loss += loss.cpu().item()
optim.step()
updates += 1
if updates % 30 == 0:
logging(
"time: %6.3f, epoch: %3d, updates: %8d, train loss this batch: %6.3f,ss loss: %6.6f\n"
% (time.time() - start_time, epoch, updates, loss , total_loss_ss / 30.0))
total_loss_sgm, total_loss_ss = 0, 0
# continue
if 0 and updates % config.eval_interval == 0 and epoch > 3: #建议至少跑几个epoch再进行测试，否则模型还没学到东西，会有很多问题。
logging("time: %6.3f, epoch: %3d, updates: %8d, train loss: %6.5f\n"
% (time.time() - start_time, epoch, updates, total_loss/config.eval_interval))
print(('evaluating after %d updates...\r' % updates))
eval(epoch,'valid') # eval的时候batch_size会变成1
eval(epoch,'test') # eval的时候batch_size会变成1
model.train()
total_loss = 0
start_time = 0
report_total = 0
report_correct = 0
if 1 and updates % config.save_interval == 1:
save_model(log_path + 'Transformer_PIT_{}.pt'.format(updates))
def eval(epoch):
# config.batch_size=1
model.eval()
# print '\n\n测试的时候请设置config里的batch_size为1！！！please set the batch_size as 1'
reference, candidate, source, alignments = [], [], [], []
e = epoch
test_or_valid = 'test'
test_or_valid = 'valid'
# test_or_valid = 'train'
print(('Test or valid:', test_or_valid))
eval_data_gen = prepare_data('once', test_or_valid, config.MIN_MIX, config.MAX_MIX)
SDR_SUM = np.array([])
SDRi_SUM = np.array([])
batch_idx = 0
global best_SDR, Var
# for iii in range(2000):
while True:
print(('-' * 30))
eval_data = next(eval_data_gen)
if eval_data == False:
print(('SDR_aver_eval_epoch:', SDR_SUM.mean()))
print(('SDRi_aver_eval_epoch:', SDRi_SUM.mean()))
break # 如果这个epoch的生成器没有数据了，直接进入下一个epoch
src = Variable(torch.from_numpy(eval_data['mix_feas']))
# raw_tgt = [sorted(spk.keys()) for spk in eval_data['multi_spk_fea_list']]
raw_tgt= eval_data['batch_order']
feas_tgt = models.rank_feas(raw_tgt, eval_data['multi_spk_fea_list']) # 这里是目标的图谱
top_k = len(raw_tgt[0])
# 要保证底下这几个都是longTensor(长整数）
# tgt = Variable(torch.from_numpy(np.array([[0]+[dict_spk2idx[spk] for spk in spks]+[dict_spk2idx['<EOS>']] for spks in raw_tgt],dtype=np.int))).transpose(0,1) #转换成数字，然后前后加开始和结束符号。
tgt = Variable(torch.from_numpy(np.array([[0,1,2,102] for __ in range(config.batch_size)], dtype=np.int))).transpose(0, 1) # 转换成数字，然后前后加开始和结束符号。
src_len = Variable(torch.LongTensor(config.batch_size).zero_() + mix_speech_len).unsqueeze(0)
tgt_len = Variable(torch.LongTensor([len(one_spk) for one_spk in eval_data['multi_spk_fea_list']])).unsqueeze(0)
# tgt_len = Variable(torch.LongTensor(config.batch_size).zero_()+len(eval_data['multi_spk_fea_list'][0])).unsqueeze(0)
if config.WFM:
tmp_size = feas_tgt.size()
assert len(tmp_size) == 3
feas_tgt_square = feas_tgt * feas_tgt
feas_tgt_sum_square = torch.sum(feas_tgt_square, dim=0, keepdim=True).expand(tmp_size)
WFM_mask = feas_tgt_square / (feas_tgt_sum_square + 1e-15)
if use_cuda:
src = src.cuda().transpose(0, 1)
tgt = tgt.cuda()
src_len = src_len.cuda()
tgt_len = tgt_len.cuda()
feas_tgt = feas_tgt.cuda()
if config.WFM:
WFM_mask = WFM_mask.cuda()
# try:
if 1 and len(opt.gpus) > 1:
samples, predicted_masks = model.module.pit_sample(src, src_len, dict_spk2idx, tgt,
beam_size=config.beam_size)
else:
samples, predicted_masks = model.pit_sample(src, src_len, dict_spk2idx, tgt,
beam_size=config.beam_size)
samples=samples.max(2)[1].data.cpu().numpy()
# except:
# continue
# '''
# expand the raw mixed-features to topk_max channel.
src = src.transpose(0, 1)
siz = src.size()
assert len(siz) == 3
# if samples[0][-1] != dict_spk2idx['<EOS>']:
# print '*'*40+'\nThe model is far from good. End the evaluation.\n'+'*'*40
# break
topk_max = len(samples[0]) - 1
x_input_map_multi = torch.unsqueeze(src, 1).expand(siz[0], topk_max, siz[1], siz[2])
if 1 and config.WFM:
feas_tgt = x_input_map_multi.data * WFM_mask
if test_or_valid != 'test':
if 1 and len(opt.gpus) > 1:
ss_loss = model.module.separation_loss(x_input_map_multi, predicted_masks, feas_tgt, )
else:
ss_loss = model.separation_loss(x_input_map_multi, predicted_masks, feas_tgt)
print(('loss for ss,this batch:', ss_loss.cpu().item()))
lera.log({
'ss_loss_' + test_or_valid: ss_loss.cpu().item(),
})
del ss_loss
# '''''
if 1 and batch_idx <= (500 / config.batch_size): # only the former batches counts the SDR
predicted_maps = predicted_masks * x_input_map_multi
# predicted_maps=Variable(feas_tgt)
utils.bss_eval2(config, predicted_maps, eval_data['multi_spk_fea_list'], raw_tgt, eval_data,
dst='batch_output_test')
del predicted_maps, predicted_masks, x_input_map_multi
try:
sdr_aver_batch, sdri_aver_batch= bss_test.cal('batch_output_test/')
SDR_SUM = np.append(SDR_SUM, sdr_aver_batch)
SDRi_SUM = np.append(SDRi_SUM, sdri_aver_batch)
except(AssertionError):
print('Errors in calculating the SDR')
print(('SDR_aver_now:', SDR_SUM.mean()))
print(('SDRi_aver_now:', SDRi_SUM.mean()))
lera.log({'SDR sample'+test_or_valid: SDR_SUM.mean()})
lera.log({'SDRi sample'+test_or_valid: SDRi_SUM.mean()})
writer.add_scalars('scalar/loss',{'SDR_sample_'+test_or_valid:sdr_aver_batch},updates)
# raw_input('Press any key to continue......')
elif batch_idx == (200 / config.batch_size) + 1 and SDR_SUM.mean() > best_SDR: # only record the best SDR once.
print(('Best SDR from {}---->{}'.format(best_SDR, SDR_SUM.mean())))
best_SDR = SDR_SUM.mean()
# save_model(log_path+'checkpoint_bestSDR{}.pt'.format(best_SDR))
# '''
candidate += [convertToLabels(dict_idx2spk, s, dict_spk2idx['<EOS>']) for s in samples]
# source += raw_src
reference += raw_tgt
print(('samples:', samples))
print(('can:{}, \nref:{}'.format(candidate[-1 * config.batch_size:], reference[-1 * config.batch_size:])))
# alignments += [align for align in alignment]
batch_idx += 1
result = utils.eval_metrics(reference, candidate, dict_spk2idx, log_path)
print(('hamming_loss: %.8f | micro_f1: %.4f |recall: %.4f | precision: %.4f'
% (result['hamming_loss'], result['micro_f1'], result['micro_recall'], result['micro_precision'], )))
score = {}
result = utils.eval_metrics(reference, candidate, dict_spk2idx, log_path)
logging_csv([e, updates, result['hamming_loss'], \
result['micro_f1'], result['micro_precision'], result['micro_recall'],SDR_SUM.mean()])
print(('hamming_loss: %.8f | micro_f1: %.4f'
% (result['hamming_loss'], result['micro_f1'])))
score['hamming_loss'] = result['hamming_loss']
score['micro_f1'] = result['micro_f1']
1/0
return score
