def eval_model(mix_speech_class,dict_spk2idx,loss_func):
mix_speech_class.training=False
print '#' * 40
eval_data_gen=prepare_data('once','valid',2,2)
acc_all,acc_line=[],[]
recall_rate_list=np.array([])
while True:
eval_data=eval_data_gen.next()
if eval_data==False:
break #如果这个epoch的生成器没有数据了，直接进入下一个epoch
mix_speech=mix_speech_class(Variable(torch.from_numpy(eval_data['mix_feas'])).cuda())
y_spk,y_map=multi_label_vector(eval_data['multi_spk_fea_list'],dict_spk2idx)
y_map=Variable(torch.from_numpy(y_map)).cuda()
y_out_batch=mix_speech.data.cpu().numpy()
acc1,acc2,all_num_batch,all_line_batch,recall_rate=count_multi_acc(y_out_batch,y_spk,alpha=-0.1,top_k_num=2)
acc_all.append(acc1)
acc_line.append(acc2)
recall_rate_list=np.append(recall_rate_list,recall_rate)
for i in range(config.BATCH_SIZE):
print 'aim:{}-->{},predict:{}'.format(eval_data['multi_spk_fea_list'][i].keys(),y_spk[i],mix_speech.data.cpu().numpy()[i][y_spk[i]])#除了输出目标的几个概率，也输出倒数四个的
print 'last 4 probility:{}'.format(mix_speech.data.cpu().numpy()[i][-5:])#除了输出目标的几个概率，也输出倒数四个的
print '\nAcc for this batch: all elements({}) acc--{},all sample({}) acc--{} recall--{}'.format(all_num_batch,acc1,all_line_batch,acc2,recall_rate)
loss=loss_func(mix_speech,y_map)
loss_sum=loss_func(mix_speech.sum(1),y_map.sum(1))
lrs.send('eval_loss',loss.data[0])
print 'loss this batch:',loss.data.cpu().numpy(),loss_sum.data.cpu().numpy()
print 'time:',time.strftime("%Y-%m-%d %H:%M:%S", time.localtime())
lrs.send('eval_sum_loss',loss_sum.data[0])
print 'Acc for eval dataset: all elements acc--{},all sample acc--{}, recall_rate--{}'.format(np.mean(acc_all),np.mean(acc_line),np.mean(recall_rate_list))
lrs.send('eval_recall_rate_aver',np.mean(recall_rate_list))
print '#'*40
def train(epoch):
e = epoch
model.train()
if config.schedule:
scheduler.step()
print("Decaying learning rate to %g" % scheduler.get_lr()[0])
if config.is_dis:
scheduler_dis.step()
lera.log({
'lr': scheduler.get_lr()[0],
})
if opt.model == 'gated':
model.current_epoch = epoch
global e, updates, total_loss, start_time, report_total, total_loss_sgm, total_loss_ss
if config.MLMSE:
global Var
train_data_gen = prepare_data('once', 'train')
# for raw_src, src, src_len, raw_tgt, tgt, tgt_len in trainloader:
while True:
train_data = train_data_gen.next()
if train_data == False:
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']
]
feas_tgt = models.rank_feas(
raw_tgt, train_data['multi_spk_fea_list']) #这里是目标的图谱
# 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。
src_len = Variable(
torch.LongTensor(config.batch_size).zero_() +
mix_speech_len).unsqueeze(0)
tgt_len = Variable(
torch.LongTensor(config.batch_size).zero_() +
len(train_data['multi_spk_fea_list'][0])).unsqueeze(0)
if config.WFM:
tmp_size = feas_tgt.size()
assert len(tmp_size) == 4
feas_tgt_square = feas_tgt * feas_tgt
feas_tgt_square_sum = torch.sum(feas_tgt_square,
dim=1,
keepdim=True).expand(tmp_size)
WFM_mask = feas_tgt_square / (feas_tgt_square_sum + 1e-10)
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:
model.zero_grad()
# optim.optimizer.zero_grad()
outputs, targets, multi_mask = model(
src, src_len, tgt,
tgt_len) #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()
if 1 and len(opt.gpus) > 1:
sgm_loss, num_total, num_correct = model.module.compute_loss(
outputs, targets, opt.memory)
else:
sgm_loss, num_total, num_correct = model.compute_loss(
outputs, targets, opt.memory)
print 'loss for SGM,this batch:', sgm_loss.data[0] / num_total
if config.unit_norm: #outputs---[len+1,bs,2*d]
assert not config.global_emb
unit_dis = (outputs[0] * outputs[1]).sum(1)
print 'unit_dis this batch:', unit_dis.data.cpu().numpy()
unit_dis = torch.masked_select(unit_dis,
unit_dis > config.unit_norm)
if len(unit_dis) > 0:
unit_dis = unit_dis.mean()
src = src.transpose(0, 1)
# expand the raw mixed-features to topk channel.
siz = src.size()
assert len(siz) == 3
topk = feas_tgt.size()[1]
x_input_map_multi = torch.unsqueeze(src, 1).expand(
siz[0], topk, siz[1], siz[2])
multi_mask = multi_mask.transpose(0, 1)
if config.WFM:
feas_tgt = x_input_map_multi.data * WFM_mask
if 1 and len(opt.gpus) > 1:
if config.MLMSE:
Var = model.module.update_var(x_input_map_multi,
multi_mask, feas_tgt)
lera.log_image(u'Var weight',
Var.data.cpu().numpy().reshape(
config.speech_fre, config.speech_fre,
1).repeat(3, 2),
clip=(-1, 1))
ss_loss = model.module.separation_loss(
x_input_map_multi, multi_mask, feas_tgt, Var)
else:
ss_loss = model.module.separation_loss(
x_input_map_multi, multi_mask, feas_tgt)
else:
ss_loss = model.separation_loss(x_input_map_multi, multi_mask,
feas_tgt)
loss = sgm_loss + 5 * ss_loss
if config.unit_norm and len(unit_dis):
print 'unit_dis masked mean:', unit_dis.data[0]
lera.log({
'unit_dis': unit_dis.data[0],
})
loss = loss + unit_dis
if config.reID:
print '#' * 30 + 'ReID part ' + '#' * 30
predict_multi_map = multi_mask * x_input_map_multi
predict_multi_map = predict_multi_map.view(
-1, mix_speech_len, speech_fre).transpose(0, 1)
tgt_reID = []
for spks in raw_tgt:
for spk in spks:
one_spk = [dict_spk2idx['<BOS>']] + [
dict_spk2idx[spk]
] + [dict_spk2idx['<EOS>']]
tgt_reID.append(one_spk)
tgt_reID = Variable(
torch.from_numpy(np.array(
tgt_reID, dtype=np.int))).transpose(0, 1).cuda()
src_len_reID = Variable(
torch.LongTensor(topk * config.batch_size).zero_() +
mix_speech_len).unsqueeze(0).cuda()
tgt_len_reID = Variable(
torch.LongTensor(topk * config.batch_size).zero_() +
1).unsqueeze(0).cuda()
outputs_reID, targets_reID, multi_mask_reID = model(
predict_multi_map, src_len_reID, tgt_reID, tgt_len_reID
) #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
if 1 and len(opt.gpus) > 1:
sgm_loss_reID, num_total_reID, _xx = model.module.compute_loss(
outputs_reID, targets_reID, opt.memory)
else:
sgm_loss_reID, num_total_reID, _xx = model.compute_loss(
outputs_reID, targets_reID, opt.memory)
print 'loss for SGM-reID mthis batch:', sgm_loss_reID.data[
0] / num_total_reID
loss = loss + sgm_loss_reID
if config.WFM:
feas_tgt = x_input_map_multi.data * WFM_mask
if 1 and len(opt.gpus) > 1:
ss_loss_reID = model.module.separation_loss(
predict_multi_map.transpose(0, 1).unsqueeze(1),
multi_mask_reID.transpose(0, 1),
feas_tgt.view(-1, 1, mix_speech_len, speech_fre))
else:
ss_loss_reID = model.separation_loss(
predict_multi_map.transpose(0, 1).unsqueeze(1),
multi_mask_reID.transpose(0, 1),
feas_tgt.view(-1, 1, mix_speech_len, speech_fre))
loss = loss + ss_loss_reID
print '#' * 30 + 'ReID part ' + '#' * 30
# dis_loss model
if config.is_dis:
dis_loss = models.loss.dis_loss(config, topk, model_dis,
x_input_map_multi, multi_mask,
feas_tgt, func_dis)
loss = loss + dis_loss
# print 'dis_para',model_dis.parameters().next()[0]
# print 'ss_para',model.parameters().next()[0]
loss.backward()
# print 'totallllllllllll loss:',loss
total_loss_sgm += sgm_loss.data[0]
total_loss_ss += ss_loss.data[0]
lera.log({
'sgm_loss': sgm_loss.data[0],
'ss_loss': ss_loss.data[0],
})
if config.reID:
lera.log({
'reID_sgm_loss': sgm_loss_reID.data[0],
'reID_ss_loss': ss_loss_reID.data[0],
})
total_loss += loss.data[0]
report_total += num_total
optim.step()
if config.is_dis:
optim_dis.step()
updates += 1
if updates % 30 == 0:
logging(
"time: %6.3f, epoch: %3d, updates: %8d, train loss this batch: %6.3f,sgm loss: %6.6f,ss loss: %6.6f\n"
% (time.time() - start_time, epoch, updates, loss /
num_total, total_loss_sgm / 30.0, total_loss_ss / 30.0))
total_loss_sgm, total_loss_ss = 0, 0
# continue
if 0 or updates % config.eval_interval == 0 and epoch > 1:
logging(
"time: %6.3f, epoch: %3d, updates: %8d, train loss: %6.5f\n"
% (time.time() - start_time, epoch, updates,
total_loss / report_total))
print('evaluating after %d updates...\r' % updates)
score = eval(epoch)
for metric in config.metric:
scores[metric].append(score[metric])
if metric == 'micro_f1' and score[metric] >= max(
scores[metric]):
save_model(log_path + 'best_' + metric +
'_checkpoint.pt')
if metric == 'hamming_loss' and score[metric] <= min(
scores[metric]):
save_model(log_path + 'best_' + metric +
'_checkpoint.pt')
model.train()
total_loss = 0
start_time = 0
report_total = 0
except RuntimeError, info:
print '**************Error occurs here************:', info
continue
if updates % config.save_interval == 1:
save_model(log_path + 'TDAA2019_{}.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'
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 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 config.schedule and scheduler.get_lr()[0] > 5e-5:
scheduler.step()
print("Decaying learning rate to %g" % scheduler.get_lr()[0])
if opt.model == 'gated':
model.current_epoch = epoch
train_data_gen = prepare_data('once', 'train')
while True:
# print '\n'
# train_data = train_data_gen.next()
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 = 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,aim_size,len,fre
padded_mixture, mixture_lengths, padded_source = train_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()
# 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。
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 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()
# 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()
outputs, targets, multi_mask, gamma = model(
src, src_len, tgt, tgt_len, dict_spk2idx,
padded_mixture) # 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())
writer.add_histogram('global gamma', gamma, updates)
if 1 and len(opt.gpus) > 1:
sgm_loss, num_total, num_correct = model.module.compute_loss(
outputs, targets, opt.memory)
else:
sgm_loss, num_total, num_correct = model.compute_loss(
outputs, targets, opt.memory)
print('loss for SGM,this batch:', sgm_loss.cpu().item())
writer.add_scalars('scalar/loss', {'sgm_loss': sgm_loss.cpu().item()},
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]
multi_mask = multi_mask.transpose(0, 1)
if config.use_tas:
if 1 and len(opt.gpus) > 1:
ss_loss = model.module.separation_tas_loss(
padded_mixture, multi_mask, padded_source, mixture_lengths)
else:
ss_loss = model.separation_tas_loss(padded_mixture, multi_mask,
padded_source,
mixture_lengths)
else:
if 1 and len(opt.gpus) > 1:
ss_loss = model.module.separation_loss(x_input_map_multi,
multi_mask, feas_tgt)
else:
ss_loss = model.separation_loss(x_input_map_multi, multi_mask,
feas_tgt)
print('loss for SS,this batch:', ss_loss.cpu().item())
writer.add_scalars('scalar/loss', {'ss_loss': ss_loss.cpu().item()},
updates)
if not config.use_tas:
loss = sgm_loss + 5 * ss_loss
else:
loss = 50 * sgm_loss + ss_loss
loss.backward()
# print 'totallllllllllll loss:',loss
total_loss_sgm += sgm_loss.cpu().item()
total_loss_ss += ss_loss.cpu().item()
# lera.log({
# 'sgm_loss': sgm_loss.cpu().item(),
# 'ss_loss': ss_loss.cpu().item(),
# 'loss:': loss.cpu().item(),
# })
# if not config.use_tas and updates>10 and updates % config.eval_interval in [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10]:
if updates > 10 and updates % config.eval_interval in [
0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10
]:
if not config.use_tas:
predicted_maps = multi_mask * x_input_map_multi
utils.bss_eval(config,
predicted_maps,
train_data['multi_spk_fea_list'],
raw_tgt,
train_data,
dst='batch_output1')
del predicted_maps, multi_mask, x_input_map_multi
else:
utils.bss_eval_tas(config,
multi_mask,
train_data['multi_spk_fea_list'],
raw_tgt,
train_data,
dst='batch_output1')
del x_input_map_multi
sdr_aver_batch, sdri_aver_batch = bss_test.cal('batch_output1/')
# 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())
#raw_input('Press to continue...')
total_loss += loss.cpu().item()
report_correct += num_correct.cpu().item()
report_total += num_total.cpu().item()
optim.step()
updates += 1
if updates % 30 == 0:
logging(
"time: %6.3f, epoch: %3d, updates: %8d, train loss this batch: %6.3f,sgm loss: %6.6f,ss loss: %6.6f,label acc: %6.6f\n"
% (time.time() - start_time, epoch, updates, loss / num_total,
total_loss_sgm / 30.0, total_loss_ss / 30.0,
report_correct / report_total))
# lera.log({'label_acc':report_correct/report_total})
writer.add_scalars('scalar/loss',
{'label_acc': report_correct / report_total},
updates)
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 / report_total))
print('evaluating after %d updates...\r' % updates)
original_bs = config.batch_size
score = eval(epoch) # eval的时候batch_size会变成1
# print 'Orignal bs:',original_bs
config.batch_size = original_bs
# print 'Now bs:',config.batch_size
for metric in config.metric:
scores[metric].append(score[metric])
# lera.log({
# 'sgm_micro_f1': score[metric],
# })
if metric == 'micro_f1' and score[metric] >= max(
scores[metric]):
save_model(log_path + 'best_' + metric + '_checkpoint.pt')
if metric == 'hamming_loss' and score[metric] <= min(
scores[metric]):
save_model(log_path + 'best_' + metric + '_checkpoint.pt')
model.train()
total_loss = 0
start_time = 0
report_total = 0
report_correct = 0
if updates % config.save_interval == 1:
save_model(log_path + 'TDAAv3_{}.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
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] > 5e-7:
scheduler.step()
print(("Decaying learning rate to %g" % scheduler.get_lr()[0]))
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,aim_size,len,fre
padded_mixture, mixture_lengths, padded_source = train_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()
topk_this_batch = int(len(raw_tgt[0]))
# 要保证底下这几个都是longTensor(长整数）
tgt_max_len = topk_this_batch + 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 = topk_this_batch # 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 = topk_this_batch # 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])
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 * IRM * 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 = topk_this_batch # 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()
# 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()
outputs, pred, targets, multi_mask, dec_enc_attn_list = model(
src,
src_len,
tgt,
tgt_len,
dict_spk2idx,
None,
mix_wav=padded_mixture
) # 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())
# 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 = topk_this_batch # 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]
multi_mask = multi_mask.transpose(0, 1)
# if config.WFM:
# feas_tgt = x_input_map_multi.data * WFM_mask
if 1 and len(opt.gpus) > 1:
sgm_loss, num_total, num_correct = model.module.compute_loss(
outputs, targets, opt.memory)
else:
sgm_loss, num_total, num_correct = model.compute_loss(
outputs, targets, opt.memory)
print(('loss for SGM,this batch:', sgm_loss.cpu().item()))
writer.add_scalars('scalar/loss', {'sgm_loss': sgm_loss.cpu().item()},
updates)
loss = sgm_loss
ss_loss = 0
loss.backward()
# print 'totallllllllllll loss:',loss
total_loss_sgm += sgm_loss.cpu().item()
lera.log({
'sgm_loss': sgm_loss.cpu().item(),
'loss:': loss.cpu().item(),
})
total_loss += loss.cpu().item()
report_correct += num_correct.cpu().item()
report_total += num_total.cpu().item()
optim.step()
updates += 1
if updates % 30 == 0:
logging(
"time: %6.3f, epoch: %3d, updates: %8d, train loss this batch: %6.3f,sgm loss: %6.6f,ss loss: %6.6f,label acc: %6.6f\n"
% (time.time() - start_time, epoch, updates, loss / num_total,
total_loss_sgm / 30.0, total_loss_ss / 30.0,
report_correct / report_total))
lera.log({'label_acc': report_correct / report_total})
writer.add_scalars('scalar/loss',
{'label_acc': report_correct / report_total},
updates)
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 / report_total))
print(('evaluating after %d updates...\r' % updates))
original_bs = config.batch_size
score = eval(epoch) # eval的时候batch_size会变成1
# print 'Orignal bs:',original_bs
config.batch_size = original_bs
# print 'Now bs:',config.batch_size
for metric in config.metric:
scores[metric].append(score[metric])
lera.log({
'sgm_micro_f1': score[metric],
})
if metric == 'micro_f1' and score[metric] >= max(
scores[metric]):
save_model(log_path + 'best_' + metric + '_checkpoint.pt')
if metric == 'hamming_loss' and score[metric] <= min(
scores[metric]):
save_model(log_path + 'best_' + metric + '_checkpoint.pt')
model.train()
total_loss = 0
start_time = 0
report_total = 0
report_correct = 0
if updates > 10 and updates % config.save_interval == 1:
save_model(log_path + 'TDAAv3_PIT_{}.pt'.format(updates))
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 train(epoch):
e = epoch
model.train()
if config.schedule:
scheduler.step()
print("Decaying learning rate to %g" % scheduler.get_lr()[0])
if opt.model == 'gated':
model.current_epoch = epoch
global e, updates, total_loss, start_time, report_total, total_loss_sgm, total_loss_ss
train_data_gen = prepare_data('once', 'train')
# for raw_src, src, src_len, raw_tgt, tgt, tgt_len in trainloader:
while True:
train_data = train_data_gen.next()
if train_data == False:
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']
]
feas_tgt = models.rank_feas(
raw_tgt, train_data['multi_spk_fea_list']) #这里是目标的图谱
# 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。
src_len = Variable(
torch.LongTensor(config.batch_size).zero_() +
mix_speech_len).unsqueeze(0)
tgt_len = Variable(
torch.LongTensor(config.batch_size).zero_() +
len(train_data['multi_spk_fea_list'][0])).unsqueeze(0)
if use_cuda:
src = src.cuda()
tgt = tgt.cuda()
src_len = src_len.cuda()
tgt_len = tgt_len.cuda()
feas_tgt = feas_tgt.cuda()
model.zero_grad()
# optim.optimizer.zero_grad()
outputs, targets, multi_mask = model(
src, src_len, tgt,
tgt_len) #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()
sgm_loss, num_total, num_correct = model.compute_loss(
outputs, targets, opt.memory)
print 'loss for SGM,this batch:', sgm_loss.data[0] / num_total
# expand the raw mixed-features to topk channel.
siz = src.size()
assert len(siz) == 3
topk = feas_tgt.size()[1]
x_input_map_multi = torch.unsqueeze(src,
1).expand(siz[0], topk, siz[1],
siz[2])
ss_loss = model.separation_loss(x_input_map_multi, multi_mask,
feas_tgt)
loss = sgm_loss + ss_loss
loss.backward()
# print 'totallllllllllll loss:',loss
total_loss_sgm += sgm_loss.data[0]
total_loss_ss += ss_loss.data[0]
total_loss += loss.data[0]
report_total += num_total
optim.step()
updates += 1
if updates % 30 == 0:
logging(
"time: %6.3f, epoch: %3d, updates: %8d, train loss this batch: %6.3f,sgm loss: %6.6f,ss loss: %6.6f\n"
% (time.time() - start_time, epoch, updates, loss / num_total,
total_loss_sgm / 30.0, total_loss_ss / 30.0))
total_loss_sgm, total_loss_ss = 0, 0
# continue
if 1 or (updates % config.eval_interval == 0) and epoch > 8:
logging(
"time: %6.3f, epoch: %3d, updates: %8d, train loss: %6.5f\n" %
(time.time() - start_time, epoch, updates,
total_loss / report_total))
print('evaluating after %d updates...\r' % updates)
score = eval(epoch)
for metric in config.metric:
scores[metric].append(score[metric])
if metric == 'micro_f1' and score[metric] >= max(
scores[metric]):
save_model(log_path + 'best_' + metric + '_checkpoint.pt')
if metric == 'hamming_loss' and score[metric] <= min(
scores[metric]):
save_model(log_path + 'best_' + metric + '_checkpoint.pt')
model.train()
total_loss = 0
start_time = 0
report_total = 0
if updates % config.save_interval == 0:
save_model(log_path + 'checkpoint_v2.pt')
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 main():
print('go to model')
print '*' * 80
spk_global_gen=prepare_data(mode='global',train_or_test='train') #写一个假的数据生成，可以用来写模型先
global_para=spk_global_gen.next()
print global_para
spk_all_list,dict_spk2idx,dict_idx2spk,mix_speech_len,speech_fre,total_frames,spk_num_total,batch_total=global_para
del spk_global_gen
num_labels=len(spk_all_list)
config.EPOCH_SIZE=batch_total
#此处顺序是 mix_speechs.shape,mix_feas.shape,aim_fea.shape,aim_spkid.shape,query.shape
#一个例子：(5, 17040) (5, 134, 129) (5, 134, 129) (5,) (5, 32, 400, 300, 3)
print 'Begin to build the maim model for Multi_Modal Cocktail Problem.'
mix_speech_class=MIX_SPEECH_classifier(speech_fre,mix_speech_len,num_labels).cuda()
print mix_speech_class
if 0 and config.Load_param:
# para_name='param_speech_WSJ0_multilabel_epoch42'
# para_name='param_speech_WSJ0_multilabel_epoch249'
# para_name='param_speech_123_WSJ0_multilabel_epoch75'
# para_name='param_speech_123_WSJ0_multilabel_epoch24'
para_name='param_speech_123onezero_WSJ0_multilabel_epoch75' #top3 召回率80%
para_name='param_speech_123onezeroag_WSJ0_multilabel_epoch80'#83.6
para_name='param_speech_123onezeroag1_WSJ0_multilabel_epoch45'
para_name='param_speech_123onezeroag2_WSJ0_multilabel_epoch40'
para_name='param_speech_123onezeroag4_WSJ0_multilabel_epoch75'
para_name='param_speech_123onezeroag3_WSJ0_multilabel_epoch40'
para_name='param_speech_123onezeroag4_WSJ0_multilabel_epoch20'
para_name='param_speech_4lstm_multilabelloss30map_epoch440'
para_name='param_speech_123onezeroag5dropout_WSJ0_multilabel_epoch20'
# mix_speech_class.load_state_dict(torch.load('params/param_speech_multilabel_epoch249'))
mix_speech_class.load_state_dict(torch.load('params/{}'.format(para_name)))
print 'Load Success:',para_name
lr_data=0.0001
optimizer = torch.optim.Adam([{'params':mix_speech_class.parameters()},
# {'params':query_video_layer.lstm_layer.parameters()},
# {'params':query_video_layer.dense.parameters()},
# {'params':query_video_layer.Linear.parameters()},
# {'params':att_layer.parameters()},
# ], lr=0.02,momentum=0.9)
], lr=lr_data)
loss_func = torch.nn.KLDivLoss() # the target label is NOT an one-hotted
# loss_func = torch.nn.MultiLabelSoftMarginLoss() # the target label is NOT an one-hotted
# loss_func = torch.nn.MSELoss() # the target label is NOT an one-hotted
# loss_func = torch.nn.CrossEntropyLoss() # the target label is NOT an one-hotted
# loss_func = torch.nn.MultiLabelMarginLoss() # the target label is NOT an one-hotted
# loss_func = torch.nn.L1Loss() # the target label is NOT an one-hotted
lrs.send({
'title': 'TDAA classifier',
'batch_size':config.BATCH_SIZE,
'batch_total':batch_total,
'epoch_size':config.EPOCH_SIZE,
'loss func':loss_func.__str__(),
'initial lr':lr_data
})
print '''Begin to calculate.'''
for epoch_idx in range(config.MAX_EPOCH):
if epoch_idx%10==0:
for ee in optimizer.param_groups:
ee['lr']/=2
lr_data=ee['lr']
acc_all,acc_line=0,0
if epoch_idx>0:
print 'recal_rate this epoch {}: {}'.format(epoch_idx,recall_rate_list.mean())
recall_rate_list=np.array([])
train_data_gen=prepare_data('once','train')
for batch_idx in range(config.EPOCH_SIZE):
continue
print '*' * 40,epoch_idx,batch_idx,'*'*40
train_data=train_data_gen.next()
if train_data==False:
break #如果这个epoch的生成器没有数据了，直接进入下一个epoch
mix_speech=mix_speech_class(Variable(torch.from_numpy(train_data['mix_feas'])).cuda())
y_spk,y_map=multi_label_vector(train_data['multi_spk_fea_list'],dict_spk2idx)
y_map=Variable(torch.from_numpy(y_map)).cuda()
y_out_batch=mix_speech.data.cpu().numpy()
acc1,acc2,all_num_batch,all_line_batch,recall_rate=count_multi_acc(y_out_batch,y_spk,alpha=-0.1,top_k_num=2)
acc_all+=acc1
acc_line+=acc2
recall_rate_list=np.append(recall_rate_list,recall_rate)
# print 'training abs norm this batch:',torch.abs(y_map-predict_map).norm().data.cpu().numpy()
for i in range(config.BATCH_SIZE):
print 'aim:{}-->{},predict:{}'.format(train_data['multi_spk_fea_list'][i].keys(),y_spk[i],mix_speech.data.cpu().numpy()[i][y_spk[i]])#除了输出目标的几个概率，也输出倒数四个的
print 'last 4 probility:{}'.format(mix_speech.data.cpu().numpy()[i][-5:])#除了输出目标的几个概率，也输出倒数四个的
print '\nAcc for this batch: all elements({}) acc--{},all sample({}) acc--{} recall--{}'.format(all_num_batch,acc1,all_line_batch,acc2,recall_rate)
lrs.send('recall_rate',recall_rate)
# continue
# if epoch_idx==0 and batch_idx<50:
# loss=loss_func(mix_speech,100*y_map)
# else:
# loss=loss_func(mix_speech,y_map)
# loss=loss_func(mix_speech,30*y_map)
loss=loss_func(mix_speech,y_map)
loss_sum=loss_func(mix_speech.sum(1),y_map.sum(1))
lrs.send('train_loss',loss.data[0])
print 'loss this batch:',loss.data.cpu().numpy(),loss_sum.data.cpu().numpy()
print 'time:',time.strftime("%Y-%m-%d %H:%M:%S", time.localtime())
# continue
# loss=loss+0.2*loss_sum
optimizer.zero_grad() # clear gradients for next train
loss.backward() # backpropagation, compute gradients
optimizer.step() # apply gradients
lrs.send('sum_loss',loss_sum.data[0])
lrs.send('lr',lr_data)
if config.Save_param and epoch_idx >= 10 and epoch_idx % 2 == 0:
try:
torch.save(mix_speech_class.state_dict(), 'params/param_speech_123onezeroag5dropout_{}_multilabel_epoch{}'.format(config.DATASET,epoch_idx))
except:
print '\n\nSave paras failed ~! \n\n\n'
# Print the Params history , that it proves well.
# print 'Parameter history:'
# for pa_gen in [{'params':mix_hidden_layer_3d.parameters()},
# {'params':query_video_layer.lstm_layer.parameters()},
# {'params':query_video_layer.dense.parameters()},
# {'params':query_video_layer.Linear.parameters()},
# {'params':att_layer.parameters()},
# ]:
# print pa_gen['params'].next()
print 'Acc for this epoch: all elements acc--{},all sample acc--{}'.format(acc_all/config.EPOCH_SIZE,acc_line/config.EPOCH_SIZE)
print '\nBegin to evaluate.'
eval_model(mix_speech_class,dict_spk2idx,loss_func)
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_complex_two_channel'])
) # bs,T,F,2 both real and imag values
raw_tgt = eval_data['batch_order']
feas_tgt = models.rank_feas(
raw_tgt,
eval_data['multi_spk_wav_list']) # 这里是目标的图谱,bs*Topk,time_len
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()
# 要保证底下这几个都是longTensor(长整数）
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)
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()
multi_mask_real, multi_mask_imag, 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
multi_mask_real = multi_mask_real.transpose(0, 1)
multi_mask_imag = multi_mask_imag.transpose(0, 1)
src_real = src[:, :, :, 0].transpose(0, 1) # bs,T,F
src_imag = src[:, :, :, 1].transpose(0, 1) # bs,T,F
print('mask size for real/imag:',
multi_mask_real.size()) # bs,topk,T,F, 已经压缩过了
print('mixture size for real/imag:', src_real.size()) # bs,T,F
predicted_maps0_real = multi_mask_real[:,
0] * src_real - multi_mask_imag[:,
0] * src_imag #bs,T,F
predicted_maps0_imag = multi_mask_real[:,
0] * src_imag + multi_mask_imag[:,
0] * src_real #bs,T,F
predicted_maps1_real = multi_mask_real[:,
1] * src_real - multi_mask_imag[:,
1] * src_imag #bs,T,F
predicted_maps1_imag = multi_mask_real[:,
1] * src_imag + multi_mask_imag[:,
1] * src_real #bs,T,F
stft_matrix_spk0 = torch.cat((predicted_maps0_real.unsqueeze(-1),
predicted_maps0_imag.unsqueeze(-1)),
3).transpose(1, 2) # bs,F,T,2
stft_matrix_spk1 = torch.cat((predicted_maps1_real.unsqueeze(-1),
predicted_maps1_imag.unsqueeze(-1)),
3).transpose(1, 2) # bs,F,T,2
wav_spk0 = models.istft_irfft(stft_matrix_spk0,
length=config.MAX_LEN,
hop_length=config.FRAME_SHIFT,
win_length=config.FRAME_LENGTH,
window='hann')
wav_spk1 = models.istft_irfft(stft_matrix_spk1,
length=config.MAX_LEN,
hop_length=config.FRAME_SHIFT,
win_length=config.FRAME_LENGTH,
window='hann')
predict_wav = torch.cat((wav_spk0.unsqueeze(1), wav_spk1.unsqueeze(1)),
1) # bs,topk,time_len
if 1 and len(opt.gpus) > 1:
ss_loss, pmt_list, max_snr_idx, *__ = model.module.separation_tas_loss(
padded_mixture, predict_wav, padded_source, mixture_lengths)
else:
ss_loss, pmt_list, max_snr_idx, *__ = model.separation_tas_loss(
padded_mixture, predict_wav, padded_source, mixture_lengths)
best_pmt = [
list(pmt_list[int(mm)].data.cpu().numpy()) for mm in max_snr_idx
]
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)
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 1 and batch_idx <= (500 / config.batch_size):
utils.bss_eval_tas(config,
predict_wav,
eval_data['multi_spk_fea_list'],
raw_tgt,
eval_data,
dst=log_path + 'batch_output')
sdr_aver_batch, snri_aver_batch = bss_test.cal(log_path +
'batch_output/')
lera.log({'SDR sample': sdr_aver_batch})
lera.log({'SI-SNRi sample': snri_aver_batch})
writer.add_scalars('scalar/loss', {
'SDR_sample': sdr_aver_batch,
'SDRi_sample': snri_aver_batch
}, updates)
SDR_SUM = np.append(SDR_SUM, sdr_aver_batch)
SDRi_SUM = np.append(SDRi_SUM, snri_aver_batch)
print(('SDR_aver_now:', SDR_SUM.mean()))
print(('SNRi_aver_now:', SDRi_SUM.mean()))
batch_idx += 1
if batch_idx > 100:
break
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_complex_two_channel'])
) # bs,T,F,2 both real and imag values
raw_tgt = train_data['batch_order']
feas_tgt = models.rank_feas(
raw_tgt,
train_data['multi_spk_wav_list']) # 这里是目标的图谱,bs*Topk,time_len
padded_mixture, mixture_lengths, padded_source = train_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()
# 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 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()
multi_mask_real, multi_mask_imag, 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
multi_mask_real = multi_mask_real.transpose(0, 1)
multi_mask_imag = multi_mask_imag.transpose(0, 1)
src_real = src[:, :, :, 0].transpose(0, 1) # bs,T,F
src_imag = src[:, :, :, 1].transpose(0, 1) # bs,T,F
print('mask size for real/imag:',
multi_mask_real.size()) # bs,topk,T,F, 已经压缩过了
print('mixture size for real/imag:', src_real.size()) # bs,T,F
predicted_maps0_real = multi_mask_real[:,
0] * src_real - multi_mask_imag[:,
0] * src_imag #bs,T,F
predicted_maps0_imag = multi_mask_real[:,
0] * src_imag + multi_mask_imag[:,
0] * src_real #bs,T,F
predicted_maps1_real = multi_mask_real[:,
1] * src_real - multi_mask_imag[:,
1] * src_imag #bs,T,F
predicted_maps1_imag = multi_mask_real[:,
1] * src_imag + multi_mask_imag[:,
1] * src_real #bs,T,F
stft_matrix_spk0 = torch.cat((predicted_maps0_real.unsqueeze(-1),
predicted_maps0_imag.unsqueeze(-1)),
3).transpose(1, 2) # bs,F,T,2
stft_matrix_spk1 = torch.cat((predicted_maps1_real.unsqueeze(-1),
predicted_maps1_imag.unsqueeze(-1)),
3).transpose(1, 2) # bs,F,T,2
wav_spk0 = models.istft_irfft(stft_matrix_spk0,
length=config.MAX_LEN,
hop_length=config.FRAME_SHIFT,
win_length=config.FRAME_LENGTH,
window='hann')
wav_spk1 = models.istft_irfft(stft_matrix_spk1,
length=config.MAX_LEN,
hop_length=config.FRAME_SHIFT,
win_length=config.FRAME_LENGTH,
window='hann')
predict_wav = torch.cat((wav_spk0.unsqueeze(1), wav_spk1.unsqueeze(1)),
1) # bs,topk,time_len
if 1 and len(opt.gpus) > 1:
ss_loss, pmt_list, max_snr_idx, *__ = model.module.separation_tas_loss(
padded_mixture, predict_wav, padded_source, mixture_lengths)
else:
ss_loss, pmt_list, max_snr_idx, *__ = model.separation_tas_loss(
padded_mixture, predict_wav, padded_source, mixture_lengths)
best_pmt = [
list(pmt_list[int(mm)].data.cpu().numpy()) for mm in max_snr_idx
]
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 epoch > 20 and updates > 5 and updates % config.eval_interval in [
0, 1, 2, 3, 4
]:
utils.bss_eval_tas(config,
predict_wav,
train_data['multi_spk_fea_list'],
raw_tgt,
train_data,
dst=log_path + 'batch_output')
sdr_aver_batch, snri_aver_batch = bss_test.cal(log_path +
'batch_output/')
lera.log({'SDR sample': sdr_aver_batch})
lera.log({'SI-SNRi sample': snri_aver_batch})
writer.add_scalars('scalar/loss', {
'SDR_sample': sdr_aver_batch,
'SDRi_sample': snri_aver_batch
}, updates)
SDR_SUM = np.append(SDR_SUM, sdr_aver_batch)
SDRi_SUM = np.append(SDRi_SUM, snri_aver_batch)
print(('SDR_aver_now:', SDR_SUM.mean()))
print(('SNRi_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 1 and updates % config.save_interval == 1:
save_model(log_path + 'Transformer_PIT_2ch_{}.pt'.format(updates))
if 0 and updates > 0 and updates % config.eval_interval == 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
def eval(epoch, test_or_valid='train'):
# 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 = [spk.keys() for spk in eval_data['multi_spk_fea_list']]
# 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_wav_list']) # 这里是目标的图谱,bs*Topk,time_len
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()
# 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 eval_data['multi_spk_fea_list']
])).unsqueeze(0)
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()
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
multi_mask = multi_mask.transpose(0, 1)
print('mask size:', multi_mask.size()) # bs,topk,T,F
predicted_maps0_spectrogram = multi_mask[:, 0] * src.transpose(
0, 1) #bs,T,F
predicted_maps1_spectrogram = multi_mask[:, 1] * src.transpose(
0, 1) #bs,T,F
if True: # Analyze the optimal assignments
predicted_spectrogram = torch.cat([
predicted_maps0_spectrogram.unsqueeze(1),
predicted_maps1_spectrogram.unsqueeze(1)
], 1)
feas_tgt_tmp = models.rank_feas(
raw_tgt,
eval_data['multi_spk_fea_list']) # 这里是目标的图谱,bs*Topk,len,fre
src = src.transpose(0, 1)
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_tmp.view(siz[0], -1, siz[1], siz[2])
angle_tgt = models.rank_feas(
raw_tgt, eval_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(
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_tmp = feas_tgt_tmp.view(siz[0], -1, siz[1],
siz[2]) * ang # bs,topk,T,F
feas_tgt_tmp = feas_tgt_tmp.cuda()
del x_input_map_multi
src = src.transpose(0, 1)
MSE_func = nn.MSELoss().cuda()
best_perms_this_batch = []
for bs_idx in range(siz[0]):
best_perms_this_sample = []
for tt in range(siz[1]): # 对每一帧
tar = feas_tgt_tmp[bs_idx, :, tt] #topk,F
est = predicted_spectrogram[bs_idx, :, tt] #topk,F
best_loss_mse_this_batch = -1
for idx, per in enumerate([[0, 1], [1, 0]]):
if idx == 0:
best_loss_mse_this_batch = MSE_func(est[per], tar)
perm_this_frame = per
predicted_spectrogram[bs_idx, :, tt] = est[per]
else:
loss = MSE_func(est[per], tar)
if loss <= best_loss_mse_this_batch:
best_loss_mse_this_batch = loss
perm_this_frame = per
predicted_spectrogram[bs_idx, :, tt] = est[per]
best_perms_this_sample.append(perm_this_frame)
best_perms_this_batch.append(best_perms_this_sample)
print(
'different assignment ratio:',
np.mean(np.min(
np.array(best_perms_this_batch).sum(1) / 751, 1)))
# predicted_maps0_spectrogram = predicted_spectrogram[:,0]
# predicted_maps1_spectrogram = predicted_spectrogram[:,1]
_mix_spec = eval_data['mix_phase'] # bs,T,F,2
angle_mix = np.angle(_mix_spec)
predicted_maps0_real = predicted_maps0_spectrogram * torch.from_numpy(
np.cos(angle_mix)).cuda() # e(ix) = cosx + isin x
predicted_maps0_imag = predicted_maps0_spectrogram * torch.from_numpy(
np.sin(angle_mix)).cuda() # e(ix) = cosx + isin x
predicted_maps1_real = predicted_maps1_spectrogram * torch.from_numpy(
np.cos(angle_mix)).cuda() # e(ix) = cosx + isin x
predicted_maps1_imag = predicted_maps1_spectrogram * torch.from_numpy(
np.sin(angle_mix)).cuda() # e(ix) = cosx + isin x
stft_matrix_spk0 = torch.cat((predicted_maps0_real.unsqueeze(-1),
predicted_maps0_imag.unsqueeze(-1)),
3).transpose(1, 2) # bs,F,T,2
stft_matrix_spk1 = torch.cat((predicted_maps1_real.unsqueeze(-1),
predicted_maps1_imag.unsqueeze(-1)),
3).transpose(1, 2) # bs,F,T,2
wav_spk0 = models.istft_irfft(stft_matrix_spk0,
length=config.MAX_LEN,
hop_length=config.FRAME_SHIFT,
win_length=config.FRAME_LENGTH,
window='hann')
wav_spk1 = models.istft_irfft(stft_matrix_spk1,
length=config.MAX_LEN,
hop_length=config.FRAME_SHIFT,
win_length=config.FRAME_LENGTH,
window='hann')
predict_wav = torch.cat((wav_spk0.unsqueeze(1), wav_spk1.unsqueeze(1)),
1) # bs,topk,time_len
if 1 and len(opt.gpus) > 1:
ss_loss, pmt_list, max_snr_idx, *__ = model.module.separation_tas_loss(
padded_mixture, predict_wav, padded_source, mixture_lengths)
else:
ss_loss, pmt_list, max_snr_idx, *__ = model.separation_tas_loss(
padded_mixture, predict_wav, padded_source, mixture_lengths)
best_pmt = [
list(pmt_list[int(mm)].data.cpu().numpy()) for mm in max_snr_idx
]
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)
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 1 and batch_idx <= (500 / config.batch_size):
utils.bss_eval_tas(config,
predict_wav,
eval_data['multi_spk_fea_list'],
raw_tgt,
eval_data,
dst=log_path + 'batch_output')
sdr_aver_batch, snri_aver_batch = bss_test.cal(log_path +
'batch_output/')
lera.log({'SDR sample': sdr_aver_batch})
lera.log({'SI-SNRi sample': snri_aver_batch})
writer.add_scalars('scalar/loss', {
'SDR_sample': sdr_aver_batch,
'SDRi_sample': snri_aver_batch
}, updates)
SDR_SUM = np.append(SDR_SUM, sdr_aver_batch)
SDRi_SUM = np.append(SDRi_SUM, snri_aver_batch)
print(('SDR_aver_now:', SDR_SUM.mean()))
print(('SNRi_aver_now:', SDRi_SUM.mean()))
batch_idx += 1
if batch_idx > 100:
break
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 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('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([])
SDRi_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:
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']
]
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(config.batch_size).zero_() +
len(eval_data['multi_spk_fea_list'][0])).unsqueeze(0)
feas_tgt = models.rank_feas(raw_tgt,
eval_data['multi_spk_fea_list']) #这里是目标的图谱
if config.WFM:
tmp_size = feas_tgt.size()
assert len(tmp_size) == 4
feas_tgt_square = feas_tgt * feas_tgt
feas_tgt_square_sum = torch.sum(feas_tgt_square,
dim=1,
keepdim=True).expand(tmp_size)
WFM_mask = feas_tgt_square / (feas_tgt_square_sum + 1e-10)
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 Exception, info:
print '**************Error eval occurs here************:', info
continue
if len(samples[0]) != 3:
print 'Wrong num of mixtures, passed.'
continue
if config.top1:
predicted_masks = torch.cat([predicted_masks, 1 - predicted_masks],
1)
# '''
# expand the raw mixed-features to topk channel.
src = src.transpose(0, 1)
siz = src.size()
assert len(siz) == 3
topk = feas_tgt.size()[1]
x_input_map_multi = torch.unsqueeze(src,
1).expand(siz[0], topk, siz[1],
siz[2])
if config.WFM:
feas_tgt = x_input_map_multi.data * WFM_mask
if 1 and len(opt.gpus) > 1:
ss_loss = model.module.separation_loss(x_input_map_multi,
predicted_masks, feas_tgt,
None)
else:
ss_loss = model.separation_loss(x_input_map_multi, predicted_masks,
feas_tgt, None)
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 0 and config.reID:
print '#' * 30 + 'ReID part ' + '#' * 30
predict_multi_map = predicted_masks * x_input_map_multi
predict_multi_map = predict_multi_map.view(-1, mix_speech_len,
speech_fre).transpose(
0, 1)
tgt_reID = Variable(torch.ones(
3, top_k * 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_reID = Variable(
torch.LongTensor(topk * config.batch_size).zero_() +
mix_speech_len).unsqueeze(0).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(
predict_multi_map,
src_len_reID,
dict_spk2idx,
tgt_reID,
beam_size=config.beam_size)
else:
samples, alignment, hiddens, predicted_masks = model.beam_sample(
predict_multi_map,
src_len_reID,
dict_spk2idx,
tgt_reID,
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 Exception, info:
print '**************Error eval occurs here************:', info
# outputs_reID, targets_reID, multi_mask_reID = model(predict_multi_map, src_len_reID, tgt_reID, tgt_len_reID) #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
if batch_idx <= (500 / config.batch_size
): #only the former batches counts the SDR
# predicted_maps=predicted_masks*x_input_map_multi
predicted_maps = predicted_masks * predict_multi_map.transpose(
0, 1).unsqueeze(1)
predicted_maps = predicted_maps.transpose(0, 1)
# predicted_maps=Variable(feas_tgt)
utils.bss_eval(config,
predicted_maps,
eval_data['multi_spk_fea_list'],
raw_tgt,
eval_data,
dst='batch_output23jo')
del predicted_maps, predicted_masks, x_input_map_multi, predict_multi_map
SDR, SDRi = bss_test.cal('batch_output23jo/')
# SDR_SUM = np.append(SDR_SUM, bss_test.cal('batch_output23jo/'))
SDR_SUM = np.append(SDR_SUM, SDR)
SDRi_SUM = np.append(SDRi_SUM, SDRi)
print 'SDR_aver_now:', SDR_SUM.mean()
print 'SDRi_aver_now:', SDRi_SUM.mean()
lera.log({'SDR sample': SDR_SUM.mean()})
lera.log({'SDRi sample': SDRi_SUM.mean()})
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))
print '#' * 30 + 'ReID part ' + '#' * 30
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'], )))
print('loading checkpoint...\n', opt.restore)
checkpoints = torch.load(opt.restore)
# cuda
use_cuda = torch.cuda.is_available() and len(opt.gpus) > 0
use_cuda = True
if use_cuda:
torch.cuda.set_device(opt.gpus[0])
torch.cuda.manual_seed(opt.seed)
print(use_cuda)
# data
print('loading data...\n')
start_time = time.time()
spk_global_gen = prepare_data(mode='global',
train_or_test='train') #写一个假的数据生成，可以用来写模型先
global_para = spk_global_gen.next()
print global_para
spk_all_list, dict_spk2idx, dict_idx2spk, mix_speech_len, speech_fre, total_frames, spk_num_total, batch_total = global_para
config.speech_fre = speech_fre
config.mix_speech_len = mix_speech_len
del spk_global_gen
num_labels = len(spk_all_list)
print('loading the global setting cost: %.3f' % (time.time() - start_time))
if opt.pretrain:
pretrain_embed = torch.load(config.emb_file)
else:
pretrain_embed = None
# model
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] > 5e-7:
scheduler.step()
print(("Decaying learning rate to %g" % scheduler.get_lr()[0]))
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,aim_size,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。
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
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()
outputs, targets, multi_mask, gamma = 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())
# writer.add_histogram('global gamma',gamma, updates)
if 1 and len(opt.gpus) > 1:
sgm_loss, num_total, num_correct = model.module.compute_loss(
outputs, targets, opt.memory)
else:
sgm_loss, num_total, num_correct = model.compute_loss(
outputs, targets, opt.memory)
print(('loss for SGM,this batch:', sgm_loss.cpu().item()))
writer.add_scalars('scalar/loss', {'sgm_loss': sgm_loss.cpu().item()},
updates)
if config.use_center_loss:
cen_alpha = 0.01
cen_loss = center_loss(outputs.view(-1, config.SPK_EMB_SIZE),
targets.view(-1))
print(('loss for SGM center loss,this batch:',
cen_loss.cpu().item()))
writer.add_scalars('scalar/loss',
{'center_loss': cen_loss.cpu().item()}, 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]
multi_mask = multi_mask.transpose(0, 1)
if config.WFM:
feas_tgt = x_input_map_multi.data * WFM_mask
if 1 and len(opt.gpus) > 1:
ss_loss = model.module.separation_loss(x_input_map_multi,
multi_mask, feas_tgt)
else:
ss_loss = model.separation_loss(x_input_map_multi, multi_mask,
feas_tgt)
print(('loss for SS,this batch:', ss_loss.cpu().item()))
writer.add_scalars('scalar/loss', {'ss_loss': ss_loss.cpu().item()},
updates)
loss = sgm_loss + 5 * ss_loss
# loss = sgm_loss
if config.use_center_loss:
loss = cen_loss * cen_alpha + loss
loss.backward()
if config.use_center_loss:
for c_param in center_loss.parameters():
c_param.grad.data *= (0.01 /
(cen_alpha * scheduler.get_lr()[0]))
# print 'totallllllllllll loss:',loss
total_loss_sgm += sgm_loss.cpu().item()
total_loss_ss += ss_loss.cpu().item()
lera.log({
'sgm_loss': sgm_loss.cpu().item(),
'ss_loss': ss_loss.cpu().item(),
'loss:': loss.cpu().item(),
})
if updates > 10 and updates % config.eval_interval in [
0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10
]:
predicted_maps = multi_mask * x_input_map_multi
# predicted_maps=Variable(feas_tgt)
utils.bss_eval(config,
predicted_maps,
train_data['multi_spk_fea_list'],
raw_tgt,
train_data,
dst='batch_output1')
# utils.bss_eval2(config, predicted_maps, train_data['multi_spk_fea_list'], raw_tgt, train_data, dst='batch_output1')
del predicted_maps, multi_mask, x_input_map_multi
sdr_aver_batch, sdri_aver_batch = bss_test.cal('batch_output1/')
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()))
total_loss += loss.cpu().item()
report_correct += num_correct.cpu().item()
report_total += num_total.cpu().item()
optim.step()
updates += 1
if updates % 30 == 0:
logging(
"time: %6.3f, epoch: %3d, updates: %8d, train loss this batch: %6.3f,sgm loss: %6.6f,ss loss: %6.6f,label acc: %6.6f\n"
% (time.time() - start_time, epoch, updates, loss / num_total,
total_loss_sgm / 30.0, total_loss_ss / 30.0,
report_correct / report_total))
lera.log({'label_acc': report_correct / report_total})
writer.add_scalars('scalar/loss',
{'label_acc': report_correct / report_total},
updates)
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 / report_total))
print(('evaluating after %d updates...\r' % updates))
original_bs = config.batch_size
score = eval(epoch) # eval的时候batch_size会变成1
# print 'Orignal bs:',original_bs
config.batch_size = original_bs
# print 'Now bs:',config.batch_size
for metric in config.metric:
scores[metric].append(score[metric])
lera.log({
'sgm_micro_f1': score[metric],
})
if metric == 'micro_f1' and score[metric] >= max(
scores[metric]):
save_model(log_path + 'best_' + metric + '_checkpoint.pt')
if metric == 'hamming_loss' and score[metric] <= min(
scores[metric]):
save_model(log_path + 'best_' + metric + '_checkpoint.pt')
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 + 'TDAAv3_{}.pt'.format(updates))
def eval(epoch):
model.eval()
reference, candidate, source, alignments = [], [], [], []
valid_mode = 'test'
print 'Valid or Test:', valid_mode
eval_data_gen = prepare_data('once', valid_mode, 2, 2)
# for raw_src, src, src_len, raw_tgt, tgt, tgt_len in validloader:
SDR_SUM = np.array([])
batch_idx = 0
global best_SDR
while True:
# for ___ in range(100):
print '-' * 30
eval_data = eval_data_gen.next()
if eval_data == False:
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']
]
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(config.batch_size).zero_() +
len(eval_data['multi_spk_fea_list'][0])).unsqueeze(0)
feas_tgt = models.rank_feas(raw_tgt,
eval_data['multi_spk_fea_list']) #这里是目标的图谱
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()
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 len(opt.gpus) > 1:
samples, alignment = model.module.sample(src, src_len)
else:
try:
samples, alignment, hiddens, predicted_masks = model.beam_sample(
src,
src_len,
dict_spk2idx,
tgt,
beam_size=config.beam_size)
except Exception, 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 channel.
siz = src.size()
assert len(siz) == 3
topk = feas_tgt.size()[1]
x_input_map_multi = torch.unsqueeze(src,
1).expand(siz[0], topk, siz[1],
siz[2])
if config.WFM:
feas_tgt = x_input_map_multi.data * WFM_mask
ss_loss = model.separation_loss(x_input_map_multi, predicted_masks,
feas_tgt)
print 'loss for ss,this batch:', 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
utils.bss_eval(config,
predicted_maps,
eval_data['multi_spk_fea_list'],
raw_tgt,
eval_data,
dst='batch_output1t2')
del predicted_maps, predicted_masks, x_input_map_multi
SDR_SUM = np.append(SDR_SUM, bss_test.cal('batch_output1t2/'))
print 'SDR_aver_now:', 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 train_recu(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] > 5e-7:
scheduler.step()
print(("Decaying learning rate to %g" % scheduler.get_lr()[0]))
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,aim_size,len,fre
if 0 and 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)
WFM_mask = WFM_mask.cuda()
feas_tgt = x_input_map_multi.data * WFM_mask
# 要保证底下这几个都是longTensor(长整数）
src_original = src.transpose(0, 1) #To T,bs,F
multi_mask_all = None
for len_idx in range(config.MIN_MIX + 2, 2, -1): #逐个分离
# len_idx=3
tgt_max_len = len_idx # 4,3,2 with bos and eos.
tgt = Variable(
torch.from_numpy(
np.array([[0] + [
dict_spk2idx[spk]
for spk in spks[-1 * (tgt_max_len - 2):]
] + 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。4,bs
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 train_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()
model.zero_grad()
outputs, targets, multi_mask, gamma = model(
src, src_len, tgt, tgt_len, dict_spk2idx,
src_original) # 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())
# writer.add_histogram('global gamma',gamma, updates)
if 1 and len(opt.gpus) > 1:
sgm_loss, num_total, num_correct = model.module.compute_loss(
outputs, targets, opt.memory)
else:
sgm_loss, num_total, num_correct = model.compute_loss(
outputs, targets, opt.memory)
print(('loss for SGM,this batch:', sgm_loss.cpu().item()))
writer.add_scalars(
'scalar/loss',
{'sgm_loss' + str(len_idx): sgm_loss.cpu().item()}, updates)
src = src_original.transpose(0, 1) #确保分离的时候用的是原始的语音
# expand the raw mixed-features to topk_max channel.
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 = len_idx - 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
# x_input_map_multi = x_input_map_multi[aim_list]
multi_mask = multi_mask.transpose(0, 1)
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
multi_mask_all = multi_mask #bs*topk,T,F
src = src * (1 - multi_mask[aim_feas]) #调整到bs为第一维,# bs,T,F
# src=src.transpose(0,1)*(1-multi_mask[aim_feas]) #调整到bs为第一维
src = src.detach() #第二轮用第一轮预测出来的剩下的谱
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
multi_mask_all[aim_feas] = multi_mask
feas_tgt = feas_tgt[aim_feas]
if 1 and len(opt.gpus) > 1:
ss_loss = model.module.separation_loss(x_input_map_multi,
multi_mask, feas_tgt)
else:
ss_loss = model.separation_loss(x_input_map_multi, multi_mask,
feas_tgt)
print(('loss for SS,this batch:', ss_loss.cpu().item()))
writer.add_scalars(
'scalar/loss',
{'ss_loss' + str(len_idx): ss_loss.cpu().item()}, updates)
loss = sgm_loss + 5 * ss_loss
loss.backward()
optim.step()
lera.log({
'sgm_loss' + str(len_idx): sgm_loss.cpu().item(),
'ss_loss' + str(len_idx): ss_loss.cpu().item(),
'loss:' + str(len_idx): loss.cpu().item(),
})
total_loss_sgm += sgm_loss.cpu().item()
total_loss_ss += ss_loss.cpu().item()
multi_mask = multi_mask_all
x_input_map_multi = torch.unsqueeze(src, 1).expand(
siz[0], 2, siz[1], siz[2]).contiguous().view(-1, siz[1], siz[2])
if updates > 10 and updates % config.eval_interval in [
0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10
]:
predicted_maps = multi_mask * x_input_map_multi
# predicted_maps=Variable(feas_tgt)
utils.bss_eval(config,
predicted_maps,
train_data['multi_spk_fea_list'],
raw_tgt,
train_data,
dst='batch_output')
del predicted_maps, multi_mask, x_input_map_multi
sdr_aver_batch, sdri_aver_batch = bss_test.cal('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()))
total_loss += loss.cpu().item()
report_correct += num_correct.cpu().item()
report_total += num_total.cpu().item()
updates += 1
if updates % 30 == 0:
logging(
"time: %6.3f, epoch: %3d, updates: %8d, train loss this batch: %6.3f,sgm loss: %6.6f,ss loss: %6.6f,label acc: %6.6f\n"
% (time.time() - start_time, epoch, updates, loss / num_total,
total_loss_sgm / 30.0, total_loss_ss / 30.0,
report_correct / report_total))
lera.log({'label_acc': report_correct / report_total})
writer.add_scalars('scalar/loss',
{'label_acc': report_correct / report_total},
updates)
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 / report_total))
print(('evaluating after %d updates...\r' % updates))
original_bs = config.batch_size
score = eval(epoch) # eval的时候batch_size会变成1
# print 'Orignal bs:',original_bs
config.batch_size = original_bs
# print 'Now bs:',config.batch_size
for metric in config.metric:
scores[metric].append(score[metric])
lera.log({
'sgm_micro_f1': score[metric],
})
if metric == 'micro_f1' and score[metric] >= max(
scores[metric]):
save_model(log_path + 'best_' + metric + '_checkpoint.pt')
if metric == 'hamming_loss' and score[metric] <= min(
scores[metric]):
save_model(log_path + 'best_' + metric + '_checkpoint.pt')
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 + 'TDAAv3_{}.pt'.format(updates))
# checkpoints = torch.load(opt.restore,map_location={'cuda:2':'cuda:0'})
checkpoints = torch.load(opt.restore)
# cuda
use_cuda = torch.cuda.is_available() and len(opt.gpus) > 0
# use_cuda = True
if use_cuda:
torch.cuda.set_device(opt.gpus[0])
torch.cuda.manual_seed(opt.seed)
print(use_cuda)
# load the global statistic of the data
print('loading data...\n')
start_time = time.time()
spk_global_gen = prepare_data(mode='global',
train_or_test='train') # 数据中的一些统计参数的读取
# global_para = spk_global_gen.next()
global_para = next(spk_global_gen)
print(global_para)
spk_all_list = global_para['all_spk'] # 所有说话人的列表
dict_spk2idx = global_para['dict_spk_to_idx']
dict_idx2spk = global_para['dict_idx_to_spk']
speech_fre = global_para['num_fre'] # 语音频率总数
total_frames = global_para['num_frames'] # 语音长度
spk_num_total = global_para['total_spk_num'] # 总计说话人数目
batch_total = global_para['total_batch_num'] # 一个epoch里多少个batch
config.speech_fre = speech_fre
mix_speech_len = total_frames
config.mix_speech_len = total_frames
def train(epoch):
e = epoch
model.train()
if config.schedule:
scheduler.step()
print("Decaying learning rate to %g" % scheduler.get_lr()[0])
if config.is_dis:
scheduler_dis.step()
lera.log({
'lr': scheduler.get_lr()[0],
})
if opt.model == 'gated':
model.current_epoch = epoch
global e, updates, total_loss, start_time, report_total, total_loss_sgm, total_loss_ss
if config.MLMSE:
global Var
train_data_gen = prepare_data('once', 'train')
# for raw_src, src, src_len, raw_tgt, tgt, tgt_len in trainloader:
while True:
train_data = train_data_gen.next()
if train_data == False:
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']
]
feas_tgt = models.rank_feas(
raw_tgt, train_data['multi_spk_fea_list']) #这里是目标的图谱
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
# 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。
src_len = Variable(
torch.LongTensor(config.batch_size).zero_() +
mix_speech_len).unsqueeze(0)
tgt_len = Variable(
torch.LongTensor(config.batch_size).zero_() +
len(train_data['multi_spk_fea_list'][0])).unsqueeze(0)
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()
# optim.optimizer.zero_grad()
outputs, targets, multi_mask = model(
src, src_len, tgt,
tgt_len) #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()
if 1 and len(opt.gpus) > 1:
sgm_loss, num_total, num_correct = model.module.compute_loss(
outputs, targets, opt.memory)
else:
sgm_loss, num_total, num_correct = model.compute_loss(
outputs, targets, opt.memory)
print 'loss for SGM,this batch:', sgm_loss.data[0] / num_total
src = src.transpose(0, 1)
# expand the raw mixed-features to topk channel.
siz = src.size()
assert len(siz) == 3
topk = feas_tgt.size()[1]
x_input_map_multi = torch.unsqueeze(src,
1).expand(siz[0], topk, siz[1],
siz[2])
multi_mask = multi_mask.transpose(0, 1)
if 1 and len(opt.gpus) > 1:
if config.MLMSE:
Var = model.module.update_var(x_input_map_multi, multi_mask,
feas_tgt)
lera.log_image(u'Var weight',
Var.data.cpu().numpy().reshape(
config.speech_fre, config.speech_fre,
1).repeat(3, 2),
clip=(-1, 1))
ss_loss = model.module.separation_loss(x_input_map_multi,
multi_mask, feas_tgt,
Var)
else:
ss_loss = model.module.separation_loss(x_input_map_multi,
multi_mask, feas_tgt)
else:
ss_loss = model.separation_loss(x_input_map_multi, multi_mask,
feas_tgt)
loss = sgm_loss + ss_loss
# dis_loss model
if config.is_dis:
dis_loss = models.loss.dis_loss(config, topk, model_dis,
x_input_map_multi, multi_mask,
feas_tgt, func_dis)
loss = loss + dis_loss
# print 'dis_para',model_dis.parameters().next()[0]
# print 'ss_para',model.parameters().next()[0]
loss.backward()
# print 'totallllllllllll loss:',loss
total_loss_sgm += sgm_loss.data[0]
total_loss_ss += ss_loss.data[0]
lera.log({
'sgm_loss': sgm_loss.data[0],
'ss_loss': ss_loss.data[0],
})
total_loss += loss.data[0]
report_total += num_total
optim.step()
if config.is_dis:
optim_dis.step()
updates += 1
if updates % 30 == 0:
logging(
"time: %6.3f, epoch: %3d, updates: %8d, train loss this batch: %6.3f,sgm loss: %6.6f,ss loss: %6.6f\n"
% (time.time() - start_time, epoch, updates, loss / num_total,
total_loss_sgm / 30.0, total_loss_ss / 30.0))
total_loss_sgm, total_loss_ss = 0, 0
# continue
if 0 or updates % config.eval_interval == 0 and epoch > 1:
logging(
"time: %6.3f, epoch: %3d, updates: %8d, train loss: %6.5f\n" %
(time.time() - start_time, epoch, updates,
total_loss / report_total))
print('evaluating after %d updates...\r' % updates)
score = eval(epoch)
for metric in config.metric:
scores[metric].append(score[metric])
if metric == 'micro_f1' and score[metric] >= max(
scores[metric]):
save_model(log_path + 'best_' + metric + '_checkpoint.pt')
if metric == 'hamming_loss' and score[metric] <= min(
scores[metric]):
save_model(log_path + 'best_' + metric + '_checkpoint.pt')
model.train()
total_loss = 0
start_time = 0
report_total = 0
if updates % config.save_interval == 1:
save_model(log_path +
'checkpoint_v2_withdis{}.pt'.format(config.is_dis))
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
# 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。
# tgt = Variable(torch.from_numpy(np.array([[0,1,2,3,102] 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。
tgt = Variable(
torch.from_numpy(
np.array([
list(range(top_k + 1)) + [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。
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()
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()
if 1 and len(opt.gpus) > 1:
outputs, pred, targets, multi_mask, dec_enc_attn_list = model(
src,
src_len,
tgt,
tgt_len,
dict_spk2idx,
None,
mix_wav=padded_mixture
) # 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
else:
outputs, pred, targets, multi_mask, dec_enc_attn_list = model(
src,
src_len,
tgt,
tgt_len,
dict_spk2idx,
None,
mix_wav=padded_mixture
) # 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
samples = list(
pred.view(config.batch_size, top_k + 1,
-1).max(2)[1].data.cpu().numpy())
'''
if 1 and len(opt.gpus) > 1:
samples, predicted_masks = model.module.beam_sample(src, src_len, dict_spk2idx, tgt, config.beam_size,None,padded_mixture)
else:
samples, predicted_masks = model.beam_sample(src, src_len, dict_spk2idx, tgt, config.beam_size, None, padded_mixture)
multi_mask = predicted_masks
samples=[samples]
# 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 = top_k
x_input_map_multi = torch.unsqueeze(src,
1).expand(siz[0], topk_max, siz[1],
siz[2])
multi_mask = multi_mask.transpose(0, 1)
if test_or_valid != 'test':
if config.use_tas:
if 1 and len(opt.gpus) > 1:
ss_loss, pmt_list, max_snr_idx, *__ = model.module.separation_tas_loss(
padded_mixture, multi_mask, padded_source,
mixture_lengths)
else:
ss_loss, pmt_list, max_snr_idx, *__ = model.separation_tas_loss(
padded_mixture, multi_mask, padded_source,
mixture_lengths)
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
utils.bss_eval_tas(config,
multi_mask,
eval_data['multi_spk_fea_list'],
raw_tgt,
eval_data,
dst=log_path + 'batch_output/')
del multi_mask, x_input_map_multi
try:
sdr_aver_batch, sdri_aver_batch = bss_test.cal(log_path +
'batch_output/')
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))
'''
import matplotlib.pyplot as plt
ax = plt.gca()
ax.invert_yaxis()
raw_src=models.rank_feas(raw_tgt,eval_data['multi_spk_fea_list'])
att_idx=0
att =dec_enc_attn_list.data.cpu().numpy()[:,att_idx] # head,topk,T
for spk in range(3):
xx=att[:,spk]
plt.matshow(xx.reshape(8,1,-1).repeat(50,1).reshape(-1,751), cmap=plt.cm.hot, vmin=0,vmax=0.05)
plt.colorbar()
plt.savefig(log_path+'batch_output/'+'spk_{}.png'.format(spk))
plt.matshow(xx.sum(0).reshape(1, 1, -1).repeat(50, 1).reshape(-1, 751), cmap=plt.cm.hot, vmin=0, vmax=0.05)
plt.colorbar()
plt.savefig(log_path + 'batch_output/' + 'spk_sum_{}.png'.format(spk))
for head in range(8):
xx=att[head]
plt.matshow(xx.reshape(3,1,-1).repeat(100,1).reshape(-1,751), 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')
# '''
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
def eval(epoch):
model.eval()
reference, candidate, source, alignments = [], [], [], []
e = epoch
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)
# for raw_src, src, src_len, raw_tgt, tgt, tgt_len in validloader:
SDR_SUM = np.array([])
batch_idx = 0
global best_SDR
while True:
# for ___ in range(2):
print '-' * 30
eval_data = eval_data_gen.next()
if eval_data == False:
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']
]
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(config.batch_size).zero_() +
len(eval_data['multi_spk_fea_list'][0])).unsqueeze(0)
feas_tgt = models.rank_feas(raw_tgt,
eval_data['multi_spk_fea_list']) #这里是目标的图谱
relitu(mix_speech_len, speech_fre, feas_tgt.numpy()[0, 0].transpose())
relitu(mix_speech_len, speech_fre, feas_tgt.numpy()[0, 1].transpose())
# 1/0
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()
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 len(opt.gpus) > 1:
samples, alignment = model.module.sample(src, src_len)
else:
samples, alignment, hiddens, predicted_masks = model.beam_sample(
src, src_len, dict_spk2idx, tgt, beam_size=config.beam_size)
# try:
# samples, alignment, hiddens, predicted_masks = model.beam_sample(src, src_len, dict_spk2idx, tgt, beam_size=config.beam_size)
# except Exception,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 channel.
siz = src.size()
assert len(siz) == 3
topk = feas_tgt.size()[1]
x_input_map_multi = torch.unsqueeze(src,
1).expand(siz[0], topk, siz[1],
siz[2])
if config.WFM:
feas_tgt = x_input_map_multi.data * WFM_mask
ss_loss = model.separation_loss(x_input_map_multi, predicted_masks,
feas_tgt)
print 'loss for ss,this batch:', ss_loss.data[0]
del ss_loss, hiddens
# '''''
if batch_idx <= (3000 / 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_eval(config,
predicted_maps,
eval_data['multi_spk_fea_list'],
raw_tgt,
eval_data,
dst='batch_output23jo')
del predicted_maps, predicted_masks, x_input_map_multi
SDR_SUM = np.append(SDR_SUM, bss_test.cal('batch_output23jo/'))
print 'SDR_aver_now:', SDR_SUM.mean()
# 1/0
raw_input('Press any key to continue......')
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
if opt.unk:
cands = []
for s, c, align in zip(source, candidate, alignments):
cand = []
for word, idx in zip(c, align):
if word == dict.UNK_WORD and idx < len(s):
try:
cand.append(s[idx])
except:
cand.append(word)
print("%d %d\n" % (len(s), idx))
else:
cand.append(word)
cands.append(cand)
candidate = cands
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
