def train(epoch):
print('\nEpoch: %d' % epoch)
print("Train")
net.train()
train_loss = 0
correct_count = 0
total = 0
flag = 0
for batch_idx, (input1, target1, input2,
target2) in enumerate(train_loader):
input1, target1, input2, target2 = input1.to(device), target1.to(
device), input2.to(device), target2.to(device)
#inputs, targets = inputs.to(device), targets.to(device)
optimizer.zero_grad()
output1 = net(input1).float()
output2 = net(input2).float()
target1 = target1.float()
target2 = target2.float()
##
if output1 > output2 and target1 > target2:
if epoch == 5:
print(output1, output2, target1, target2)
correct = True
elif output2 > output1 and target2 > target1:
if epoch == 5:
print(output1, output2, target1, target2)
correct = True
elif output1 == output2 and target1 == target2:
if epoch == 5:
print(output1, output2, target1, target2)
correct = True
else:
correct = False
#(like count1, like count2]
## Have to write the criterion function
if flag <= 5:
print(flag, ":", output1, output2, target1, target2)
flag += 1
loss = criterion(output1, output2, target1, target2)
loss.backward()
optimizer.step()
train_loss += loss.item()
# _, predicted = outputs.max(1)
# total += targets.size(0)
# correct += predicted.eq(targets).sum().item()
# loss.data[0]
total += 1
correct_count += 1 if correct else 0
# print(batch_idx, len(train_loader), 'Loss: %.3f | Acc: %.3f%% (%d/%d)'
# % (train_loss/(batch_idx+1), 100.*correct/total, correct, total))
print("Total Loss: %.3f | Acc: %.3f" %
(train_loss / (batch_idx + 1), 100. * correct_count / total))
lera.log('train_loss', train_loss / (batch_idx + 1))
lera.log('train_acc', 100. * correct_count / total)
def test(epoch):
print("Validation")
global best_acc
net.eval()
test_loss = 0
correct_count = 0
total = 0
with torch.no_grad():
for batch_idx, (input1, target1, input2,
target2) in enumerate(val_loader):
input1, target1, input2, target2 = input1.to(device), target1.to(
device), input2.to(device), target2.to(device)
output1 = net(input1).float()
output2 = net(input2).float()
target1 = target1.float()
target2 = target2.float()
loss = criterion(output1, output2, target1, target2)
test_loss += loss.item()
# _, predicted = outputs.max(1)
# total += targets.size(0)
# correct += predicted.eq(targets).sum().item()
# print(batch_idx, len(val_loader), 'Loss: %.3f | Acc: %.3f%% (%d/%d)'
# % (test_loss/(batch_idx+1), 100.*correct/total, correct, total))
if output1 > output2 and target1 > target2:
correct = True
elif output2 > output1 and target2 > target1:
correct = True
elif output1 == output2 and target1 == target2:
correct = True
else:
correct = False
total += 1
correct_count += 1 if correct else 0
print("Total Loss: %.3f | Acc: %.3f" %
(test_loss / (batch_idx + 1), 100. * correct_count / total))
lera.log('val_loss', test_loss / (batch_idx + 1))
lera.log('val_acc', 100. * correct_count / total)
# Save checkpoint.
acc = 100. * correct_count / total
if acc > best_acc:
print('Saving..')
state = {
'net': net.state_dict(),
'acc': acc,
'epoch': epoch,
}
if not os.path.isdir('checkpoint'):
os.mkdir('checkpoint')
torch.save(state, './checkpoint/ckpt.pth')
best_acc = acc
def train(config):
lera.log_hyperparams({
"title": "hw1",
"epoch": config.epochs,
"lr": config.lr
})
dataset = img_dataset("./dataset/train", "train")
dataloader = torch.utils.data.DataLoader(dataset=dataset,
batch_size=config.bs,
shuffle=True,
drop_last=True)
net = Classifier(num_classes=13).cuda()
net.load_state_dict(
torch.load(
join(f"{config.weight_path}",
f"{config.pre_epochs}_classifier.pth")))
criterion = nn.CrossEntropyLoss().cuda()
optimizer = optim.Adam(net.parameters(), lr=config.lr)
for epoch in range(config.epochs):
for _, data in enumerate(dataloader, 0):
optimizer.zero_grad()
net.train()
inputs, labels = data
inputs = inputs.cuda()
labels = labels.cuda()
outputs = net(inputs)
loss = criterion(outputs, labels)
loss.backward()
optimizer.step()
_, predicted = torch.max(outputs.data, 1)
correct_counts = predicted.eq(labels.data.view_as(predicted))
train_acc = torch.sum(correct_counts).item() / predicted.size(0)
lera.log({"loss": loss.item(), "acc": train_acc})
print("epoch:{}/{}, loss:{}, acc:{:02f}".format(
epoch + 1 + config.pre_epochs,
config.epochs + config.pre_epochs,
loss.item(),
train_acc,
))
if (epoch + 1 + config.pre_epochs) % 10 == 0:
torch.save(
net.state_dict(),
join(
f"{config.weight_path}",
f"{epoch + 1 + config.pre_epochs}_classifier.pth",
),
)
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 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):
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, 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_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 config.buffer_size or config.buffer_shift: # first convet to realtime batches
assert src.size()[1] == 1
left_padding = Variable(
torch.zeros(config.buffer_size,
src.size()[1],
src.size()[-1]).cuda())
src = torch.cat((left_padding, src), dim=0)
split_idx = 0
src_new = Variable(
torch.zeros(config.buffer_size + config.buffer_shift,
mix_speech_len / config.buffer_shift + 1,
src.size()[-1]).cuda())
batch_counter = 0
while True:
print 'split_idx at:', split_idx
split_len = config.buffer_size + config.buffer_shift # the len of every split
if split_idx + split_len > src.size(
)[0]: # if pass the right length
print 'Need to add right padding with len:', (
split_idx + split_len) - src.size()[0]
right_padding = Variable(
torch.zeros((split_idx + split_len) - src.size()[0],
src.size()[1],
src.size()[-1]).cuda())
src = torch.cat((src, right_padding), dim=0)
src_split = src[split_idx:(split_idx + split_len)]
src_new[:, batch_counter] = src_split
break
src_split = src[split_idx:(split_idx + split_len)]
src_new[:, batch_counter] = src_split
split_idx += config.buffer_shift
batch_counter += 1
assert batch_counter + 1 == src_new.size()[1]
src_len[0] = config.buffer_shift + config.buffer_size
src_len = src_len.expand(1, src_new.size()[1])
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_new,
src_len,
dict_spk2idx,
tgt,
beam_size=config.beam_size)
else:
samples, alignment, hiddens, predicted_masks = model.beam_sample(
src_new,
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 occurs here************:', info
continue
if config.top1:
predicted_masks = torch.cat([predicted_masks, 1 - predicted_masks],
1)
if config.buffer_size and config.buffer_shift: # then recover the whole maps
# masks:[7,topk,buffer_size+buffer_shift,fre]
masks_recover = Variable(
torch.zeros(1, predicted_masks.size(1), mix_speech_len,
speech_fre).cuda())
recover_idx = 0
for batch_counter in range(predicted_masks.size(0)):
if not batch_counter == predicted_masks.size(0) - 1:
masks_recover[:, :, recover_idx:recover_idx +
config.buffer_shift] = predicted_masks[
batch_counter, :,
-1 * config.buffer_shift:]
else: # the last shift
assert mix_speech_len - recover_idx == config.buffer_shift - right_padding.size(
0)
masks_recover[:, :, recover_idx:] = predicted_masks[
batch_counter, :,
-1 * config.buffer_shift:(-1 * right_padding.size(0))]
recover_idx += config.buffer_shift
predicted_masks = masks_recover
src = Variable(torch.from_numpy(eval_data['mix_feas'])).transpose(
0, 1).cuda()
# '''
# 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,Var)
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 batch_idx <= (500 / config.batch_size
): #only the former batches counts the SDR
# x_input_map_multi=x_input_map_multi[:,:,:config.buffer_shift]
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_outputwaddd')
del predicted_maps, predicted_masks, x_input_map_multi
SDR_SUM = np.append(SDR_SUM, bss_test.cal('batch_outputwaddd/'))
print 'SDR_aver_now:', SDR_SUM.mean()
lera.log({'SDR sample': SDR_SUM.mean()})
# raw_input('Press any key to continue......')
elif batch_idx == (500 / config.batch_size) + 1 and SDR_SUM.mean(
) > best_SDR: #only record the best SDR once.
print 'Best SDR from {}---->{}'.format(best_SDR, SDR_SUM.mean())
best_SDR = SDR_SUM.mean()
# save_model(log_path+'checkpoint_bestSDR{}.pt'.format(best_SDR))
# '''
candidate += [
convertToLabels(dict_idx2spk, s, dict_spk2idx['<EOS>'])
for s in samples
]
# source += raw_src
reference += raw_tgt
print 'samples:', samples
print 'can:{}, \nref:{}'.format(candidate[-1 * config.batch_size:],
reference[-1 * config.batch_size:])
alignments += [align for align in alignment]
batch_idx += 1
def 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 train(epoch):
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])
lera.log({
'lr': scheduler.get_lr()[0],
})
if opt.model == 'gated':
model.current_epoch = epoch
global e, updates, total_loss, start_time, report_total, report_correct, total_loss_sgm, total_loss_ss
train_data_gen = prepare_data('once', 'train')
while True:
print '\n'
train_data = train_data_gen.next()
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']
]
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 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) # 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 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.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 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()
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_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, step):
#lera.log('epoch', epoch)
epoch += 1
for input, _ in DataLoader(datasets[dataset], batch_size=batch_size, pin_memory=use_cuda, num_workers=2, shuffle=True, drop_last=True):
if use_cuda:
input = input.cuda()
step += 1
ze = enc(V(input))
index = min_dist(V(ze.data), embeddings)
sz = index.size()
zq = (embeddings[index.view(-1)] # [batch_size * x * x, D] containing vectors from embeddings
.view(sz[0], sz[1], sz[2], D) # [batch_size, x, x, D]
.permute(0, 3, 1, 2)) # [batch_size, D, x, x]
emb_loss = (zq - V(ze.data)).pow(2).sum(1).mean() + 1e-2 * embeddings.pow(2).mean()
# detach zq so it won't backprop to embeddings with recon loss
zq = V(zq.data, requires_grad=True)
output = dec(zq)
commit_loss = beta * (ze - V(zq.data)).pow(2).sum(1).mean()
recon_loss = F.mse_loss(output, V(input))
optimizer.zero_grad()
commit_loss.backward(retain_graph=True)
emb_loss.backward()
recon_loss.backward()
# pass data term gradient from decoder to encoder
ze.backward(zq.grad)
optimizer.step()
emb_count[index.data.view(-1)] = 1
emb_count.sub_(0.01).clamp_(min=0)
unique_embeddings = emb_count.gt(0).sum()
sensitivity.add_(emb_loss.data[0] * (K - unique_embeddings) / K)
sensitivity[emb_count.gt(0)] = 0
lera.log({
'recon_loss': recon_loss.data[0],
'commit_loss': commit_loss.data[0],
'unique_embeddings': emb_count.gt(0).sum(),
}, console=True)
# make comparison image
if lera.every(seconds=60):
input = input.cpu()[0:8,:,:,:]
w = input.size(-1)
output = output.data.cpu()[0:8,:,:,:]
result = (torch.stack([input, output]) # [2, 8, 3, w, w]
.transpose(0, 1).contiguous() # [8, 2, 3, w, w]
.view(4, 4, 3, w, w) # [4, 4, 3, w, w]
.permute(0, 3, 1, 4, 2).contiguous() # [4, w, 4, w, 3]
.view(w * 4, w * 4, 3)) # [w * 4, w * 4, 3]
lera.log_image('reconstruction', result.numpy(), clip=(0, 1))
# continue training
if step < total_steps:
train(epoch, step)
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.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) == 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, alignment, hiddens, predicted_masks = model.module.beam_sample(
src, src_len, dict_spk2idx, tgt, beam_size=config.beam_size)
else:
samples, predicted_masks = model.beam_sample(
src, src_len, dict_spk2idx, tgt, beam_size=config.beam_size)
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 = 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_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))
def eval(epoch,test_or_valid='valid'):
# config.batch_size=1
global updates,model
model.eval()
# print '\n\n测试的时候请设置config里的batch_size为1！！！please set the batch_size as 1'
reference, candidate, source, alignments = [], [], [], []
e = epoch
print(('Test or valid:', test_or_valid))
eval_data_gen = prepare_data('once', test_or_valid, config.MIN_MIX, config.MAX_MIX)
SDR_SUM = np.array([])
SDRi_SUM = np.array([])
batch_idx = 0
global best_SDR, Var
# for iii in range(2000):
while True:
print(('-' * 30))
eval_data = next(eval_data_gen)
if eval_data == False:
print(('SDR_aver_eval_epoch:', SDR_SUM.mean()))
print(('SDRi_aver_eval_epoch:', SDRi_SUM.mean()))
break # 如果这个epoch的生成器没有数据了，直接进入下一个epoch
src = Variable(torch.from_numpy(eval_data['mix_feas']))
# raw_tgt = [sorted(spk.keys()) for spk in eval_data['multi_spk_fea_list']]
raw_tgt= eval_data['batch_order']
feas_tgt = models.rank_feas(raw_tgt, eval_data['multi_spk_fea_list']) # 这里是目标的图谱
top_k = len(raw_tgt[0])
# 要保证底下这几个都是longTensor(长整数）
# tgt = Variable(torch.from_numpy(np.array([[0]+[dict_spk2idx[spk] for spk in spks]+[dict_spk2idx['<EOS>']] for spks in raw_tgt],dtype=np.int))).transpose(0,1) #转换成数字，然后前后加开始和结束符号。
tgt = Variable(torch.from_numpy(np.array([[0,1,2,102] for __ in range(config.batch_size)], dtype=np.int))).transpose(0, 1) # 转换成数字，然后前后加开始和结束符号。
src_len = Variable(torch.LongTensor(config.batch_size).zero_() + mix_speech_len).unsqueeze(0)
tgt_len = Variable(torch.LongTensor([len(one_spk) for one_spk in eval_data['multi_spk_fea_list']])).unsqueeze(0)
# tgt_len = Variable(torch.LongTensor(config.batch_size).zero_()+len(eval_data['multi_spk_fea_list'][0])).unsqueeze(0)
if config.WFM:
siz = src.size() # bs,T,F
assert len(siz) == 3
# topk_max = config.MAX_MIX # 最多可能的topk个数
topk_max = 2 # 最多可能的topk个数
x_input_map_multi = torch.unsqueeze(src, 1).expand(siz[0], topk_max, siz[1], siz[2]).contiguous().view(-1, siz[1], siz[ 2]) # bs,topk,T,F
feas_tgt_tmp = feas_tgt.view(siz[0], -1, siz[1], siz[2])
feas_tgt_square = feas_tgt_tmp * feas_tgt_tmp
feas_tgt_sum_square = torch.sum(feas_tgt_square, dim=1, keepdim=True).expand(siz[0], topk_max, siz[1], siz[2])
WFM_mask = feas_tgt_square / (feas_tgt_sum_square + 1e-15)
feas_tgt = x_input_map_multi.view(siz[0], -1, siz[1], siz[2]).data * WFM_mask # bs,topk,T,F
feas_tgt = feas_tgt.view(-1, siz[1], siz[2]) # bs*topk,T,F
WFM_mask = WFM_mask.cuda()
del x_input_map_multi
elif config.PSM:
siz = src.size() # bs,T,F
assert len(siz) == 3
# topk_max = config.MAX_MIX # 最多可能的topk个数
topk_max = 2 # 最多可能的topk个数
x_input_map_multi = torch.unsqueeze(src, 1).expand(siz[0], topk_max, siz[1], siz[2]).contiguous() # bs,topk,T,F
feas_tgt_tmp = feas_tgt.view(siz[0], -1, siz[1], siz[2])
IRM=feas_tgt_tmp/(x_input_map_multi+1e-15)
angle_tgt=models.rank_feas(raw_tgt, eval_data['multi_spk_angle_list']).view(siz[0],-1,siz[1],siz[2])
angle_mix=Variable(torch.from_numpy(np.array(eval_data['mix_angle']))).unsqueeze(1).expand(siz[0], topk_max, siz[1], siz[2]).contiguous()
ang=np.cos(angle_mix-angle_tgt)
ang=np.clip(ang,0,None)
# feas_tgt = x_input_map_multi *np.clip(IRM.numpy()*ang,0,1) # bs,topk,T,F
# feas_tgt = x_input_map_multi *IRM*ang # bs,topk,T,F
feas_tgt = feas_tgt.view(siz[0],-1,siz[1],siz[2])*ang # bs,topk,T,F
feas_tgt = feas_tgt.view(-1, siz[1], siz[2]) # bs*topk,T,F
del x_input_map_multi
if use_cuda:
src = src.cuda().transpose(0, 1)
tgt = tgt.cuda()
src_len = src_len.cuda()
tgt_len = tgt_len.cuda()
feas_tgt = feas_tgt.cuda()
if config.WFM:
WFM_mask = WFM_mask.cuda()
predicted_masks, enc_attn_list = model(src, src_len, tgt, tgt_len,
dict_spk2idx) # 这里的outputs就是hidden_outputs，还没有进行最后分类的隐层，可以直接用
print('predicted mask size:', predicted_masks.size(),'should be topk,bs,T,F') # topk,bs,T,F
# try:
# '''
# expand the raw mixed-features to topk_max channel.
src = src.transpose(0, 1)
siz = src.size()
assert len(siz) == 3
# if samples[0][-1] != dict_spk2idx['<EOS>']:
# print '*'*40+'\nThe model is far from good. End the evaluation.\n'+'*'*40
# break
topk_max = config.MAX_MIX
x_input_map_multi = torch.unsqueeze(src, 1).expand(siz[0], topk_max, siz[1], siz[2])
predicted_masks=predicted_masks.transpose(0, 1)
# if config.WFM:
# feas_tgt = x_input_map_multi.data * WFM_mask
# 注意,bs是第二维
assert predicted_masks.shape == x_input_map_multi.shape
assert predicted_masks.size(0) == config.batch_size
if 1 and len(opt.gpus) > 1:
ss_loss,best_pmt = model.module.separation_pit_loss(x_input_map_multi, predicted_masks, feas_tgt, )
else:
ss_loss,best_pmt = model.separation_pit_loss(x_input_map_multi, predicted_masks, feas_tgt)
print(('loss for ss,this batch:', ss_loss.cpu().item()))
print('best perms for this batch:', best_pmt)
lera.log({
'ss_loss_' + test_or_valid: ss_loss.cpu().item(),
})
writer.add_scalars('scalar/loss',{'ss_loss_'+test_or_valid:ss_loss.cpu().item()},updates+batch_idx)
del ss_loss
if batch_idx>10:
break
if False: #this part is to test the checkpoints sequencially.
batch_idx += 1
if batch_idx%100==0:
updates=updates+1000
opt.restore='/data1/shijing_data/2020-02-14-04:58:17/Transformer_PIT_{}.pt'.format(updates)
print('loading checkpoint...\n', opt.restore)
checkpoints = torch.load(opt.restore)
model.module.load_state_dict(checkpoints['model'])
break
continue
# '''''
if 0 and batch_idx <= (500 / config.batch_size): # only the former batches counts the SDR
predicted_maps = predicted_masks * x_input_map_multi
predicted_maps = predicted_maps.view(-1,mix_speech_len,speech_fre)
# predicted_maps=Variable(feas_tgt)
utils.bss_eval2(config, predicted_maps, eval_data['multi_spk_fea_list'], raw_tgt, eval_data,
dst='batch_output_test')
# utils.bss_eval(config, predicted_maps, eval_data['multi_spk_fea_list'], raw_tgt, eval_data,
# dst='batch_output_test')
del predicted_maps, predicted_masks, x_input_map_multi
try:
sdr_aver_batch, sdri_aver_batch= bss_test.cal('batch_output_test/')
SDR_SUM = np.append(SDR_SUM, sdr_aver_batch)
SDRi_SUM = np.append(SDRi_SUM, sdri_aver_batch)
except(AssertionError):
print('Errors in calculating the SDR')
print(('SDR_aver_now:', SDR_SUM.mean()))
print(('SRi_aver_now:', SDRi_SUM.mean()))
lera.log({'SDR sample'+test_or_valid: SDR_SUM.mean()})
lera.log({'SDRi sample'+test_or_valid: SDRi_SUM.mean()})
writer.add_scalars('scalar/loss',{'SDR_sample_'+test_or_valid:sdr_aver_batch},updates)
# raw_input('Press any key to continue......')
elif batch_idx == (200 / config.batch_size) + 1 and SDR_SUM.mean() > best_SDR: # only record the best SDR once.
print(('Best SDR from {}---->{}'.format(best_SDR, SDR_SUM.mean())))
best_SDR = SDR_SUM.mean()
# save_model(log_path+'checkpoint_bestSDR{}.pt'.format(best_SDR))
# '''
# candidate += [convertToLabels(dict_idx2spk, s, dict_spk2idx['<EOS>']) for s in samples]
# source += raw_src
# reference += raw_tgt
# print(('samples:', samples))
# print(('can:{}, \nref:{}'.format(candidate[-1 * config.batch_size:], reference[-1 * config.batch_size:])))
# alignments += [align for align in alignment]
batch_idx += 1
result = utils.eval_metrics(reference, candidate, dict_spk2idx, log_path)
print(('hamming_loss: %.8f | micro_f1: %.4f |recall: %.4f | precision: %.4f'
% (result['hamming_loss'], result['micro_f1'], result['micro_recall'], result['micro_precision'], )))
def train(epoch):
global e, updates, total_loss, start_time, report_total,report_correct, total_loss_sgm, total_loss_ss
e = epoch
model.train()
SDR_SUM = np.array([])
SDRi_SUM = np.array([])
if updates<=config.warmup: #如果不在warm阶段就正常规划
pass
elif config.schedule and scheduler.get_lr()[0]>4e-5:
scheduler.step()
print(("Decaying learning rate to %g" % scheduler.get_lr()[0],updates))
lera.log({
'lr': [group['lr'] for group in optim.optimizer.param_groups][0],
})
if opt.model == 'gated':
model.current_epoch = epoch
train_data_gen = prepare_data('once', 'train')
while True:
if updates <= config.warmup: # 如果在warm就开始warmup
tmp_lr = config.learning_rate * min(max(updates,1)** (-0.5),
max(updates,1) * (config.warmup ** (-1.5)))
for param_group in optim.optimizer.param_groups:
param_group['lr'] = tmp_lr
scheduler.base_lrs=list([group['lr'] for group in optim.optimizer.param_groups])
if updates%100==0: #记录一下
print(updates)
print("Warmup learning rate to %g" % tmp_lr)
lera.log({
'lr': [group['lr'] for group in optim.optimizer.param_groups][0],
})
train_data = next(train_data_gen)
if train_data == False:
print(('SDR_aver_epoch:', SDR_SUM.mean()))
print(('SDRi_aver_epoch:', SDRi_SUM.mean()))
break # 如果这个epoch的生成器没有数据了，直接进入下一个epoch
src = Variable(torch.from_numpy(train_data['mix_feas']))
# raw_tgt = [spk.keys() for spk in train_data['multi_spk_fea_list']]
# raw_tgt = [sorted(spk.keys()) for spk in train_data['multi_spk_fea_list']]
raw_tgt=train_data['batch_order']
feas_tgt = models.rank_feas(raw_tgt, train_data['multi_spk_fea_list']) # 这里是目标的图谱,bs*Topk,len,fre
# 要保证底下这几个都是longTensor(长整数）
tgt_max_len = config.MAX_MIX + 2 # with bos and eos.
tgt = Variable(torch.from_numpy(np.array(
[[0] + [dict_spk2idx[spk] for spk in spks] + (tgt_max_len - len(spks) - 1) * [dict_spk2idx['<EOS>']] for
spks in raw_tgt], dtype=np.int))).transpose(0, 1) # 转换成数字，然后前后加开始和结束符号。
# tgt = Variable(torch.from_numpy(np.array([[0,1,2,102] for __ in range(config.batch_size)], dtype=np.int))).transpose(0, 1) # 转换成数字，然后前后加开始和结束符号。
src_len = Variable(torch.LongTensor(config.batch_size).zero_() + mix_speech_len).unsqueeze(0)
tgt_len = Variable(
torch.LongTensor([len(one_spk) for one_spk in train_data['multi_spk_fea_list']])).unsqueeze(0)
if config.WFM:
siz = src.size() # bs,T,F
assert len(siz) == 3
# topk_max = config.MAX_MIX # 最多可能的topk个数
topk_max = 2 # 最多可能的topk个数
x_input_map_multi = torch.unsqueeze(src, 1).expand(siz[0], topk_max, siz[1], siz[2]).contiguous().view(-1, siz[1], siz[ 2]) # bs,topk,T,F
feas_tgt_tmp = feas_tgt.view(siz[0], -1, siz[1], siz[2])
feas_tgt_square = feas_tgt_tmp * feas_tgt_tmp
feas_tgt_sum_square = torch.sum(feas_tgt_square, dim=1, keepdim=True).expand(siz[0], topk_max, siz[1], siz[2])
WFM_mask = feas_tgt_square / (feas_tgt_sum_square + 1e-15)
feas_tgt = x_input_map_multi.view(siz[0], -1, siz[1], siz[2]).data * WFM_mask # bs,topk,T,F
feas_tgt = feas_tgt.view(-1, siz[1], siz[2]) # bs*topk,T,F
WFM_mask = WFM_mask.cuda()
del x_input_map_multi
elif config.PSM:
siz = src.size() # bs,T,F
assert len(siz) == 3
# topk_max = config.MAX_MIX # 最多可能的topk个数
topk_max = 2 # 最多可能的topk个数
x_input_map_multi = torch.unsqueeze(src, 1).expand(siz[0], topk_max, siz[1], siz[2]).contiguous() # bs,topk,T,F
feas_tgt_tmp = feas_tgt.view(siz[0], -1, siz[1], siz[2])
IRM=feas_tgt_tmp/(x_input_map_multi+1e-15)
angle_tgt=models.rank_feas(raw_tgt, train_data['multi_spk_angle_list']).view(siz[0],-1,siz[1],siz[2]) # bs,topk,T,F
angle_mix=Variable(torch.from_numpy(np.array(train_data['mix_angle']))).unsqueeze(1).expand(siz[0], topk_max, siz[1], siz[2]).contiguous()
ang=np.cos(angle_mix-angle_tgt)
ang=np.clip(ang,0,None)
# feas_tgt = x_input_map_multi *np.clip(IRM.numpy()*ang,0,1) # bs,topk,T,F
# feas_tgt = x_input_map_multi *IRM*ang # bs,topk,T,F
feas_tgt = feas_tgt.view(siz[0],-1,siz[1],siz[2])*ang # bs,topk,T,F
feas_tgt = feas_tgt.view(-1, siz[1], siz[2]) # bs*topk,T,F
del x_input_map_multi
elif config.frame_mask:
siz = src.size() # bs,T,F
assert len(siz) == 3
# topk_max = config.MAX_MIX # 最多可能的topk个数
topk_max = 2 # 最多可能的topk个数
x_input_map_multi = torch.unsqueeze(src, 1).expand(siz[0], topk_max, siz[1], siz[2]).contiguous() # bs,topk,T,F
feas_tgt_tmp = feas_tgt.view(siz[0], -1, siz[1], siz[2])
feas_tgt_time=torch.sum(feas_tgt_tmp,3).transpose(1,2) #bs,T,topk
for v1 in feas_tgt_time:
for v2 in v1:
if v2[0]>v2[1]:
v2[0]=1
v2[1]=0
else:
v2[0]=0
v2[1]=1
frame_mask=feas_tgt_time.transpose(1,2).unsqueeze(-1) #bs,topk,t,1
feas_tgt=x_input_map_multi*frame_mask
feas_tgt = feas_tgt.view(-1, siz[1], siz[2]) # bs*topk,T,F
if use_cuda:
src = src.cuda().transpose(0, 1)
tgt = tgt.cuda()
src_len = src_len.cuda()
tgt_len = tgt_len.cuda()
feas_tgt = feas_tgt.cuda()
model.zero_grad()
if config.use_center_loss:
center_loss.zero_grad()
# aim_list 就是找到有正经说话人的地方的标号
aim_list = (tgt[1:-1].transpose(0, 1).contiguous().view(-1) != dict_spk2idx['<EOS>']).nonzero().squeeze()
aim_list = aim_list.data.cpu().numpy()
multi_mask, enc_attn_list = model(src, src_len, tgt, tgt_len,
dict_spk2idx) # 这里的outputs就是hidden_outputs，还没有进行最后分类的隐层，可以直接用
print('mask size:', multi_mask.size()) # topk,bs,T,F
# print('mask:', multi_mask[0,0,:3:3]) # topk,bs,T,F
# writer.add_histogram('global gamma',gamma, updates)
src = src.transpose(0, 1)
# expand the raw mixed-features to topk_max channel.
siz = src.size()
assert len(siz) == 3
topk_max = config.MAX_MIX # 最多可能的topk个数
x_input_map_multi = torch.unsqueeze(src, 1).expand(siz[0], topk_max, siz[1], siz[2]).contiguous()#.view(-1, siz[1], siz[2])
# x_input_map_multi = x_input_map_multi[aim_list]
# x_input_map_multi = x_input_map_multi.transpose(0, 1) #topk,bs,T,F
multi_mask = multi_mask.transpose(0, 1)
# if config.WFM:
# feas_tgt = x_input_map_multi.data * WFM_mask
# 注意,bs是第二维
assert multi_mask.shape == x_input_map_multi.shape
assert multi_mask.size(0) == config.batch_size
if 1 and len(opt.gpus) > 1: #先ss获取Perm
ss_loss, best_pmt = model.module.separation_pit_loss(x_input_map_multi, multi_mask, feas_tgt)
else:
ss_loss, best_pmt = model.separation_pit_loss(x_input_map_multi, multi_mask, feas_tgt)
print('loss for SS,this batch:', ss_loss.cpu().item())
print('best perms for this batch:', best_pmt)
writer.add_scalars('scalar/loss',{'ss_loss':ss_loss.cpu().item()},updates)
loss = ss_loss
loss.backward()
total_loss_ss += ss_loss.cpu().item()
lera.log({
'ss_loss': ss_loss.cpu().item(),
})
if updates>3 and updates % config.eval_interval in [0, 1, 2, 3, 4, 5, 6, 7, 8, 9,]:
assert multi_mask.shape==x_input_map_multi.shape
assert multi_mask.size(0)==config.batch_size
predicted_maps = (multi_mask * x_input_map_multi).view(siz[0]*topk_max,siz[1],siz[2])
# predicted_maps=Variable(feas_tgt)
# utils.bss_eval(config, predicted_maps, train_data['multi_spk_fea_list'], raw_tgt, train_data, dst=log_path+'batch_output/')
utils.bss_eval2(config, predicted_maps, train_data['multi_spk_fea_list'], raw_tgt, train_data, dst=log_path+'batch_output')
del predicted_maps, multi_mask, x_input_map_multi
sdr_aver_batch, sdri_aver_batch= bss_test.cal(log_path+'batch_output/')
lera.log({'SDR sample': sdr_aver_batch})
lera.log({'SDRi sample': sdri_aver_batch})
writer.add_scalars('scalar/loss',{'SDR_sample':sdr_aver_batch,'SDRi_sample':sdri_aver_batch},updates)
SDR_SUM = np.append(SDR_SUM, sdr_aver_batch)
SDRi_SUM = np.append(SDRi_SUM, sdri_aver_batch)
print(('SDR_aver_now:', SDR_SUM.mean()))
print(('SDRi_aver_now:', SDRi_SUM.mean()))
# Heatmap here
# n_layer个 (head*bs) x lq x dk
'''
import matplotlib.pyplot as plt
ax = plt.gca()
ax.invert_yaxis()
raw_src=models.rank_feas(raw_tgt, train_data['multi_spk_fea_list'])
att_idx=1
att = enc_attn_list[-1].view(config.trans_n_head,config.batch_size,mix_speech_len,mix_speech_len).data.cpu().numpy()[:,att_idx]
for head in range(config.trans_n_head):
xx=att[head]
plt.matshow(xx, cmap=plt.cm.hot, vmin=0,vmax=0.05)
plt.colorbar()
plt.savefig(log_path+'batch_output/'+'head_{}.png'.format(head))
plt.matshow(raw_src[att_idx*2+0].transpose(0,1), cmap=plt.cm.hot, vmin=0,vmax=2)
plt.colorbar()
plt.savefig(log_path+'batch_output/'+'source0.png')
plt.matshow(raw_src[att_idx*2+1].transpose(0,1), cmap=plt.cm.hot, vmin=0,vmax=2)
plt.colorbar()
plt.savefig(log_path+'batch_output/'+'source1.png')
1/0
'''
total_loss += loss.cpu().item()
optim.step()
updates += 1
if updates % 30 == 0:
logging(
"time: %6.3f, epoch: %3d, updates: %8d, train loss this batch: %6.3f,ss loss: %6.6f\n"
% (time.time() - start_time, epoch, updates, loss , total_loss_ss / 30.0))
total_loss_sgm, total_loss_ss = 0, 0
# continue
if 0 and updates % config.eval_interval == 0 and epoch > 3: #建议至少跑几个epoch再进行测试，否则模型还没学到东西，会有很多问题。
logging("time: %6.3f, epoch: %3d, updates: %8d, train loss: %6.5f\n"
% (time.time() - start_time, epoch, updates, total_loss/config.eval_interval))
print(('evaluating after %d updates...\r' % updates))
eval(epoch,'valid') # eval的时候batch_size会变成1
eval(epoch,'test') # eval的时候batch_size会变成1
model.train()
total_loss = 0
start_time = 0
report_total = 0
report_correct = 0
if 1 and updates % config.save_interval == 1:
save_model(log_path + 'Transformer_PIT_{}.pt'.format(updates))
def eval(epoch, 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 train(epoch):
e = epoch
model.train()
SDR_SUM = np.array([])
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:
try:
train_data = train_data_gen.next()
if train_data == False:
print 'SDR_aver_epoch:', SDR_SUM.mean()
break #如果这个epoch的生成器没有数据了，直接进入下一个epoch
src = Variable(torch.from_numpy(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,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 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()
# 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 = 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()
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_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 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
# dis_loss model
if config.is_dis:
dis_loss = models.loss.dis_loss(config, topk_max, 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],
'loss:': loss.data[0],
})
if (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_outputjaa')
del predicted_maps, multi_mask, x_input_map_multi
# raw_input('wait to continue......')
sdr_aver_batch = bss_test.cal('batch_outputjaa/')
lera.log({'SDR sample': sdr_aver_batch})
SDR_SUM = np.append(SDR_SUM, sdr_aver_batch)
print 'SDR_aver_now:', SDR_SUM.mean()
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])
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
except RuntimeError, eeee:
print 'Erros here eeee: ', eeee
continue
except Exception, dddd:
print '\n\n\nRare errors: ', dddd
continue
def eval(epoch):
model.eval()
reference, candidate, source, alignments = [], [], [], []
e = epoch
test_or_valid = 'test'
print 'Test or valid:', test_or_valid
eval_data_gen = prepare_data_aim('once', test_or_valid, config.MIN_MIX,
config.MAX_MIX)
# for raw_src, src, src_len, raw_tgt, tgt, tgt_len in validloader:
SDR_SUM = np.array([])
batch_idx = 0
global best_SDR, Var
while True:
# for ___ in range(2):
print '-' * 30
eval_data = eval_data_gen.next()
if eval_data == False:
print 'SDR_aver_eval_epoch:', SDR_SUM.mean()
break #如果这个epoch的生成器没有数据了，直接进入下一个epoch
src = Variable(torch.from_numpy(eval_data['mix_feas']))
raw_tgt = [
sorted(spk.keys()) for spk in eval_data['multi_spk_fea_list']
]
feas_tgt = models.rank_feas(raw_tgt,
eval_data['multi_spk_fea_list']) #这里是目标的图谱
top_k = len(raw_tgt[0])
# 要保证底下这几个都是longTensor(长整数）
# tgt = Variable(torch.from_numpy(np.array([[0]+[dict_spk2idx[spk] for spk in spks]+[dict_spk2idx['<EOS>']] for spks in raw_tgt],dtype=np.int))).transpose(0,1) #转换成数字，然后前后加开始和结束符号。
tgt = Variable(torch.ones(
top_k + 2, config.batch_size)) # 这里随便给一个tgt，为了测试阶段tgt的名字无所谓其实。
src_len = Variable(
torch.LongTensor(config.batch_size).zero_() +
mix_speech_len).unsqueeze(0)
tgt_len = Variable(
torch.LongTensor([
len(one_spk) for one_spk in eval_data['multi_spk_fea_list']
])).unsqueeze(0)
# tgt_len = Variable(torch.LongTensor(config.batch_size).zero_()+len(eval_data['multi_spk_fea_list'][0])).unsqueeze(0)
if config.WFM:
tmp_size = feas_tgt.size()
assert len(tmp_size) == 4
feas_tgt_sum = torch.sum(feas_tgt, dim=1, keepdim=True)
feas_tgt_sum_square = (feas_tgt_sum *
feas_tgt_sum).expand(tmp_size)
feas_tgt_square = feas_tgt * feas_tgt
WFM_mask = feas_tgt_square / feas_tgt_sum_square
if use_cuda:
src = src.cuda().transpose(0, 1)
tgt = tgt.cuda()
src_len = src_len.cuda()
tgt_len = tgt_len.cuda()
feas_tgt = feas_tgt.cuda()
if config.WFM:
WFM_mask = WFM_mask.cuda()
try:
if 1 and len(opt.gpus) > 1:
# samples, alignment = model.module.sample(src, src_len)
samples, alignment, hiddens, predicted_masks = model.module.beam_sample(
src,
src_len,
dict_spk2idx,
tgt,
beam_size=config.beam_size)
else:
samples, alignment, hiddens, predicted_masks = model.beam_sample(
src,
src_len,
dict_spk2idx,
tgt,
beam_size=config.beam_size)
# samples, alignment, hiddens, predicted_masks = model.beam_sample(src, src_len, dict_spk2idx, tgt, beam_size=config.beam_size)
except TabError, info:
print '**************Error occurs here************:', info
continue
if config.top1:
predicted_masks = torch.cat([predicted_masks, 1 - predicted_masks],
1)
# '''
# expand the raw mixed-features to topk_max channel.
src = src.transpose(0, 1)
siz = src.size()
assert len(siz) == 3
topk_max = feas_tgt.size()[1]
assert samples[0][-1] == dict_spk2idx['<EOS>']
topk_max = len(samples[0]) - 1
x_input_map_multi = torch.unsqueeze(src,
1).expand(siz[0], topk_max, siz[1],
siz[2])
if config.WFM:
feas_tgt = x_input_map_multi.data * WFM_mask
if test_or_valid == 'valid':
if 1 and len(opt.gpus) > 1:
ss_loss = model.module.separation_loss(x_input_map_multi,
predicted_masks,
feas_tgt, Var)
else:
ss_loss = model.separation_loss(x_input_map_multi,
predicted_masks, feas_tgt)
print 'loss for ss,this batch:', ss_loss.data[0]
lera.log({
'ss_loss_' + test_or_valid: ss_loss.data[0],
})
del ss_loss, hiddens
# '''''
if batch_idx <= (500 / config.batch_size
): #only the former batches counts the SDR
predicted_maps = predicted_masks * x_input_map_multi
# predicted_maps=Variable(feas_tgt)
utils.bss_eval2(config,
predicted_maps,
eval_data['multi_spk_fea_list'],
raw_tgt,
eval_data,
dst='batch_outputjaa')
del predicted_maps, predicted_masks, x_input_map_multi
SDR_SUM = np.append(SDR_SUM, bss_test.cal('batch_outputjaa/'))
print 'SDR_aver_now:', SDR_SUM.mean()
lera.log({'SDR sample': SDR_SUM.mean()})
# raw_input('Press any key to continue......')
elif batch_idx == (500 / config.batch_size) + 1 and SDR_SUM.mean(
) > best_SDR: #only record the best SDR once.
print 'Best SDR from {}---->{}'.format(best_SDR, SDR_SUM.mean())
best_SDR = SDR_SUM.mean()
# save_model(log_path+'checkpoint_bestSDR{}.pt'.format(best_SDR))
# '''
candidate += [
convertToLabels(dict_idx2spk, s, dict_spk2idx['<EOS>'])
for s in samples
]
# source += raw_src
reference += raw_tgt
print 'samples:', samples
print 'can:{}, \nref:{}'.format(candidate[-1 * config.batch_size:],
reference[-1 * config.batch_size:])
alignments += [align for align in alignment]
batch_idx += 1
def 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')
train_data_gen = musdb.DB(root="~/MUSDB18/",
subsets='train',
split='train')
train_data_gen = batch_generator(
list(train_data_gen),
config.batch_size,
)
# while 1:
# mix,ref=next(train_data_gen)
# import soundfile as sf
# sf.write('mix.wav',mix[0,0],44100)
# sf.write('vocal.wav',ref[0,0,0],44100)
# sf.write('drum.wav',ref[0,1,0],44100)
# sf.write('bass.wav',ref[0,2,0],44100)
# sf.write('other.wav',ref[0,3,0],44100)
# pass
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
padded_mixture, mixture_lengths, padded_source = train_data
# source:bs,2channel,T target:bs,4(vocals,drums,bass,other),2channel,T
padded_mixture = torch.from_numpy(padded_mixture).float()
topk_this_batch = padded_source.shape[1]
mixture_lengths = torch.from_numpy(mixture_lengths)
padded_source = torch.from_numpy(padded_source).float()
# 要保证底下这几个都是longTensor(长整数）
if use_cuda:
padded_mixture = padded_mixture.cuda().transpose(0, 1)
mixture_lengths = mixture_lengths.cuda()
padded_source = padded_source.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 0 and loss < -5:
import soundfile as sf
idx_in_batch = 0
sf.write(
str(idx_in_batch) + '_mix.wav',
padded_mixture.transpose(
0, 1).data.cpu().numpy()[idx_in_batch].transpose(), 44100)
sf.write(
str(idx_in_batch) + '_ref_vocal.wav',
padded_source.data.cpu().numpy()[idx_in_batch, 0].transpose(),
44100)
sf.write(
str(idx_in_batch) + '_ref_drum.wav',
padded_source.data.cpu().numpy()[idx_in_batch, 1].transpose(),
44100)
sf.write(
str(idx_in_batch) + '_ref_bass.wav',
padded_source.data.cpu().numpy()[idx_in_batch, 2].transpose(),
44100)
sf.write(
str(idx_in_batch) + '_ref_other.wav',
padded_source.data.cpu().numpy()[idx_in_batch, 3].transpose(),
44100)
model.zero_grad()
outputs, pred, spks_ordre_list, multi_mask, y_map = model(
None,
None,
None,
None,
dict_spk2idx,
None,
mix_wav=padded_mixture,
clean_wavs=padded_source.transpose(
0, 1)) # 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())
print('y map size:', y_map.size())
# print('spk order:', spks_ordre_list) # bs,topk
# writer.add_histogram('global gamma',gamma, updates)
multi_mask = multi_mask.transpose(0, 1)
y_map = y_map.transpose(0, 1)
spks_ordre_list = spks_ordre_list.transpose(0, 1)
# expand the raw mixed-features to topk_max channel.
topk_max = topk_this_batch # 最多可能的topk个数
if config.greddy_tf and config.add_last_silence:
multi_mask, silence_channel = torch.split(multi_mask,
[topk_this_batch, 1],
dim=1)
silence_channel = silence_channel[:, 0]
assert len(padded_source.shape) == 3
# padded_source = torch.cat([padded_source,torch.zeros(padded_source.size(0),1,padded_source.size(2))],1)
if 1 and len(opt.gpus) > 1:
ss_loss_silence = model.module.silence_loss(silence_channel)
else:
ss_loss_silence = model.silence_loss(silence_channel)
print('loss for SS silence,this batch:',
ss_loss_silence.cpu().item())
writer.add_scalars(
'scalar/loss',
{'ss_loss_silence': ss_loss_silence.cpu().item()}, updates)
lera.log({'ss_loss_silence': ss_loss_silence.cpu().item()})
if torch.isnan(ss_loss_silence):
ss_loss_silence = 0
if config.use_tas:
# print('source',padded_source)
# print('est', multi_mask)
if 1 and len(opt.gpus) > 1:
ss_loss = model.module.separation_tas_sdr_order_loss(
padded_mixture.transpose(0, 1), multi_mask, y_map,
mixture_lengths)
else:
ss_loss = model.separation_tas_sdr_order_loss(
padded_mixture, multi_mask, y_map, 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)
print('greddy perms for this batch:',
[ii for ii in spks_ordre_list.data.cpu().numpy()])
writer.add_scalars('scalar/loss', {'ss_loss': ss_loss.cpu().item()},
updates)
loss = ss_loss
if config.add_last_silence:
loss = loss + 0.1 * ss_loss_silence
loss.backward()
# print 'totallllllllllll loss:',loss
total_loss_ss += ss_loss.cpu().item()
lera.log({
'ss_loss_' + str(topk_this_batch): ss_loss.cpu().item(),
'loss:': loss.cpu().item(),
'pre_min': multi_mask.data.cpu().numpy().min(),
'pre_max': multi_mask.data.cpu().numpy().max(),
})
if 1 or loss < -5:
import soundfile as sf
idx_in_batch = 0
y0 = multi_mask.data.cpu().numpy()[idx_in_batch, 0]
y1 = multi_mask.data.cpu().numpy()[idx_in_batch, 1]
y2 = multi_mask.data.cpu().numpy()[idx_in_batch, 2]
y3 = multi_mask.data.cpu().numpy()[idx_in_batch, 3]
# sf.write(str(idx_in_batch)+'_pre_0.wav',multi_mask.data.cpu().numpy()[idx_in_batch,0].transpose(),44100)
# sf.write(str(idx_in_batch)+'_pre_1.wav',multi_mask.data.cpu().numpy()[idx_in_batch,1].transpose(),44100)
# sf.write(str(idx_in_batch)+'_pre_2.wav',multi_mask.data.cpu().numpy()[idx_in_batch,2].transpose(),44100)
# sf.write(str(idx_in_batch)+'_pre_3.wav',multi_mask.data.cpu().numpy()[idx_in_batch,3].transpose(),44100)
print('y0 range:', y0.min(), y0.max())
print('y1 range:', y1.min(), y1.max())
print('y2 range:', y2.min(), y2.max())
print('y3 range:', y3.min(), y3.max())
# input('wait')
print('*' * 50)
if 0 and updates > 10 and updates % config.eval_interval in [
0, 1, 2, 3, 4, 5
]:
utils.bss_eval_tas(config,
multi_mask,
train_data['multi_spk_fea_list'],
raw_tgt,
train_data,
dst=log_path + '/batch_output1')
sdr_aver_batch, sdri_aver_batch = bss_test.cal(log_path +
'/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()
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, 0,
total_loss_sgm / 30.0, total_loss_ss / 30.0, 0))
# 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, 0))
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 + 'TDAAv4_conditional_{}.pt'.format(updates))
def train(args):
try:
os.makedirs(args.save_img_path)
except OSError:
pass
try:
os.makedirs(args.weight_path)
except OSError:
pass
lera.log_hyperparams(
{
"title": "hw2",
"batch_size": args.bs,
"epochs": args.epochs,
"g_lr": args.g_lr,
"d_lr": args.d_lr,
"z_size": args.z_size,
}
)
# dataset
dataloader = data_loader(
args.data_path, args.imgsize, args.bs, shuffle=True
)
# model
generator = Generator(args.bs, args.imgsize, z_dim=args.z_size).cuda()
discriminator = Discriminator(args.bs, args.imgsize).cuda()
if args.pre_epochs != 0:
generator.load_state_dict(
torch.load(
join(f"{args.weight_path}", f"generator_{args.pre_epochs}.pth")
)
)
discriminator.load_state_dict(
torch.load(
join(
f"{args.weight_path}",
f"discriminator_{args.pre_epochs}.pth",
)
)
)
# optimizer
g_optimizer = torch.optim.Adam(
filter(lambda p: p.requires_grad, generator.parameters()), lr=args.g_lr
)
d_optimizer = torch.optim.SGD(
filter(lambda p: p.requires_grad, discriminator.parameters()),
lr=args.d_lr,
)
# validate noise
fixed_noise = torch.randn(9, args.z_size)
fixed_noise = torch.tensor(fixed_noise).cuda()
# train
for epoch in range(args.pre_epochs, args.epochs):
for i, data in enumerate(dataloader):
discriminator.train()
generator.train()
# train discriminator
if i % 5 == 0:
d_optimizer.zero_grad()
real_img = torch.tensor(data[0]).cuda() * 2 - 1 # (-1, 1)
d__real, _, _ = discriminator(real_img)
z = torch.randn(args.bs, args.z_size)
z = torch.tensor(z).cuda()
fake_img, _, _ = generator(z)
d_fake, _, _ = discriminator(fake_img)
# hinge loss
d_loss_real = torch.nn.ReLU()(1.0 - d__real).mean()
d_loss_fake = torch.nn.ReLU()(1.0 + d_fake).mean()
d_loss = d_loss_real + d_loss_fake
d_loss.backward()
d_optimizer.step()
# train generator
g_optimizer.zero_grad()
z = torch.randn(args.bs, args.z_size)
z = torch.tensor(z).cuda()
fake_img, _, _ = generator(z)
g_fake, _, _ = discriminator(fake_img)
# hinge loss
g_loss = -g_fake.mean()
g_loss.backward()
g_optimizer.step()
if i % 100 == 0:
lera.log({"d_loss": d_loss.item(), "g_loss": g_loss.item()})
print(
"[epoch:%4d/%4d %3d/%3d] \t d_loss: %0.6f \t g_loss: %0.6f"
% (
epoch + 1,
args.epochs,
i,
len(dataloader),
d_loss.item(),
g_loss.item(),
)
)
if i % 300 == 0:
validate(
generator, i, epoch, args.save_img_path, fixed_noise
)
torch.save(
discriminator.state_dict(),
f"./{args.weight_path}/discriminator_{epoch+1}.pth",
)
torch.save(
generator.state_dict(),
f"./{args.weight_path}/generator_{epoch+1}.pth",
)
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, 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 == (5000 / 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
def eval(epoch):
# config.batch_size=1
model.eval()
print '\n\n测试的时候请设置config里的batch_size为1！！！please set the batch_size as 1'
reference, candidate, source, alignments = [], [], [], []
e = epoch
test_or_valid = 'test'
print('Test or valid:', test_or_valid)
eval_data_gen = prepare_data('once', test_or_valid, config.MIN_MIX,
config.MAX_MIX)
SDR_SUM = np.array([])
SDRi_SUM = np.array([])
batch_idx = 0
global best_SDR, Var
while True:
print('-' * 30)
eval_data = eval_data_gen.next()
if eval_data == False:
print('SDR_aver_eval_epoch:', SDR_SUM.mean())
print('SDRi_aver_eval_epoch:', SDRi_SUM.mean())
break # 如果这个epoch的生成器没有数据了，直接进入下一个epoch
src = Variable(torch.from_numpy(eval_data['mix_feas']))
raw_tgt = [
sorted(spk.keys()) for spk in eval_data['multi_spk_fea_list']
]
feas_tgt = models.rank_feas(
raw_tgt, eval_data['multi_spk_fea_list']) # 这里是目标的图谱
top_k = len(raw_tgt[0])
# 要保证底下这几个都是longTensor(长整数）
# tgt = Variable(torch.from_numpy(np.array([[0]+[dict_spk2idx[spk] for spk in spks]+[dict_spk2idx['<EOS>']] for spks in raw_tgt],dtype=np.int))).transpose(0,1) #转换成数字，然后前后加开始和结束符号。
tgt = Variable(torch.ones(
top_k + 2, config.batch_size)) # 这里随便给一个tgt，为了测试阶段tgt的名字无所谓其实。
src_len = Variable(
torch.LongTensor(config.batch_size).zero_() +
mix_speech_len).unsqueeze(0)
tgt_len = Variable(
torch.LongTensor([
len(one_spk) for one_spk in eval_data['multi_spk_fea_list']
])).unsqueeze(0)
# tgt_len = Variable(torch.LongTensor(config.batch_size).zero_()+len(eval_data['multi_spk_fea_list'][0])).unsqueeze(0)
if config.WFM:
tmp_size = feas_tgt.size()
assert len(tmp_size) == 4
feas_tgt_sum = torch.sum(feas_tgt, dim=1, keepdim=True)
feas_tgt_sum_square = (feas_tgt_sum *
feas_tgt_sum).expand(tmp_size)
feas_tgt_square = feas_tgt * feas_tgt
WFM_mask = feas_tgt_square / feas_tgt_sum_square
if use_cuda:
src = src.cuda().transpose(0, 1)
tgt = tgt.cuda()
src_len = src_len.cuda()
tgt_len = tgt_len.cuda()
feas_tgt = feas_tgt.cuda()
if config.WFM:
WFM_mask = WFM_mask.cuda()
if 1 and len(opt.gpus) > 1:
samples, alignment, hiddens, predicted_masks = model.module.beam_sample(
src, src_len, dict_spk2idx, tgt, beam_size=config.beam_size)
else:
samples, alignment, hiddens, predicted_masks = model.beam_sample(
src, src_len, dict_spk2idx, tgt, beam_size=config.beam_size)
# '''
# expand the raw mixed-features to topk_max channel.
src = src.transpose(0, 1)
siz = src.size()
assert len(siz) == 3
# if samples[0][-1] != dict_spk2idx['<EOS>']:
# print '*'*40+'\nThe model is far from good. End the evaluation.\n'+'*'*40
# break
topk_max = len(samples[0]) - 1
x_input_map_multi = torch.unsqueeze(src,
1).expand(siz[0], topk_max, siz[1],
siz[2])
if config.WFM:
feas_tgt = x_input_map_multi.data * WFM_mask
if 0 and test_or_valid == 'valid':
if 1 and len(opt.gpus) > 1:
ss_loss = model.module.separation_loss(
x_input_map_multi,
predicted_masks,
feas_tgt,
)
else:
ss_loss = model.separation_loss(x_input_map_multi,
predicted_masks, feas_tgt)
print('loss for ss,this batch:', ss_loss.cpu().item())
lera.log({
'ss_loss_' + test_or_valid: ss_loss.cpu().item(),
})
del ss_loss, hiddens
# '''''
if batch_idx <= (500 / config.batch_size
): # only the former batches counts the SDR
predicted_maps = predicted_masks * x_input_map_multi
# predicted_maps=Variable(feas_tgt)
utils.bss_eval2(config,
predicted_maps,
eval_data['multi_spk_fea_list'],
raw_tgt,
eval_data,
dst='batch_output')
del predicted_maps, predicted_masks, x_input_map_multi
try:
SDR_SUM, SDRi_SUM = np.append(SDR_SUM,
bss_test.cal('batch_output/'))
except AssertionError, wrong_info:
print 'Errors in calculating the SDR', wrong_info
print('SDR_aver_now:', SDR_SUM.mean())
print('SDRi_aver_now:', SDRi_SUM.mean())
lera.log({'SDR sample' + test_or_valid: SDR_SUM.mean()})
lera.log({'SDRi sample' + test_or_valid: SDRi_SUM.mean()})
# raw_input('Press any key to continue......')
elif batch_idx == (500 / config.batch_size) + 1 and SDR_SUM.mean(
) > best_SDR: # only record the best SDR once.
print('Best SDR from {}---->{}'.format(best_SDR, SDR_SUM.mean()))
best_SDR = SDR_SUM.mean()
def 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):
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 train(config):
print('Random seed: %d' % int(config.seed))
torch.manual_seed(config.seed)
torch.backends.cudnn.benchmark = True
dset = config.dataset
if dset == 'modelnet10' or dset == 'modelnet40':
dataset = ClsDataset(root=config.root, npoints=config.npoints, train=True)
test_dataset = ClsDataset(root=config.root, npoints=config.npoints, train=False)
else:
raise NotImplementedError('Dataset not supported.')
print('Selected %s' % dset)
dataloader = torch.utils.data.DataLoader(dataset, batch_size=config.batchsize, shuffle=True,
num_workers=config.workers)
test_dataloader = torch.utils.data.DataLoader(test_dataset, batch_size=config.batchsize, shuffle=True,
num_workers=config.workers)
num_classes = dataset.num_classes
print('number of classes: %d' % num_classes)
print('train set size: %d | test set size: %d' % (len(dataset), len(test_dataset)))
try:
os.makedirs(config.outf)
except:
pass
blue = lambda x: '\033[94m' + x + '\033[0m'
yellow = lambda x: '\033[93m' + x + '\033[0m'
red = lambda x: '\033[91m' + x + '\033[0m'
classifier = PointNetCls(k=num_classes)
if config.model != '':
classifier.load_state_dict(torch.load(config.model))
optimizer = optim.SGD(classifier.parameters(), lr=config.lr, momentum=config.momentum)
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
classifier.to(device)
if config.mgpu:
classifier = torch.nn.DataParallel(classifier, device_ids=config.gpuids)
num_batch = len(dataset) / config.batchsize
lera.log_hyperparams({
'title': dset,
'batchsize': config.batchsize,
'epochs': config.nepochs,
'npoints': config.npoints,
'optimizer': 'SGD',
'lr': config.lr,
})
for epoch in range(config.nepochs):
train_acc_epoch, test_acc_epoch = [], []
for i, data in enumerate(dataloader):
points, labels = data
points = points.transpose(2, 1)
labels = labels[:, 0]
points, labels = points.to(device), labels.to(device)
optimizer.zero_grad()
classifier = classifier.train()
pred, _ = classifier(points)
pred = pred.view(-1, num_classes)
# print(pred.size(), labels.size())
loss = F.nll_loss(pred, labels)
loss.backward()
optimizer.step()
pred_choice = pred.data.max(1)[1]
correct = pred_choice.eq(labels.data).cpu().sum()
train_acc = correct.item() / float(config.batchsize)
print('epoch %d: %d/%d | train loss: %f | train acc: %f' % (epoch+1, i+1, num_batch+1, loss.item(), train_acc))
train_acc_epoch.append(train_acc)
lera.log({
'train loss': loss.item(),
'train acc': train_acc
})
if (i+1) % 10 == 0:
j, data = next(enumerate(test_dataloader, 0))
points, labels = data
points = points.transpose(2, 1)
labels = labels[:, 0]
points, labels = points.to(device), labels.to(device)
classifier = classifier.eval()
with torch.no_grad():
pred, _ = classifier(points)
pred = pred.view(-1, num_classes)
loss = F.nll_loss(pred, labels)
pred_choice = pred.data.max(1)[1]
correct = pred_choice.eq(labels.data).cpu().sum()
test_acc = correct.item() / float(config.batchsize)
print(blue('epoch %d: %d/%d | test loss: %f | test acc: %f') % (epoch+1, i+1, num_batch+1, loss.item(), test_acc))
test_acc_epoch.append(test_acc)
lera.log({
'test loss': loss.item(),
'test acc': test_acc
})
print(yellow('epoch %d | mean train acc: %f') % (epoch+1, np.mean(train_acc_epoch)))
print(red('epoch %d | mean test acc: %f') % (epoch+1, np.mean(test_acc_epoch)))
lera.log({
'train acc epoch': np.mean(train_acc_epoch),
'test acc epoch': np.mean(test_acc_epoch)})
torch.save(classifier.state_dict(), '%s/%s_model_%d.pth' % (config.outf, config.dataset, epoch))
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, 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'],
)))