def train(opt):
data, word2ix, ix2word = get_data(opt)
data = torch.from_numpy(data)
dataloader = td.DataLoader(data,
batch_size=opt.batch_size,
shuffle=True,
num_workers=1)
model = PoetryModel(len(word2ix), 128, 256)
optimizer = torch.optim.Adam(model.parameters(), lr=opt.lr)
criterion = nn.CrossEntropyLoss()
if opt.use_gpu:
model.cuda()
criterion.cuda()
for epoch in range(opt.epoch):
for step, data_ in enumerate(dataloader):
data_ = data_.long().transpose(1, 0).contiguous()
if opt.use_gpu: data_ = data_.cuda()
optimizer.zero_grad()
input_, target = Variable(data_[:-1, :]), Variable(data_[1:, :])
output, _ = model(input_)
loss = criterion(output, target.view(-1))
loss.backward()
optimizer.step()
if (1 + step) % 10 == 0:
print("current loss", loss.data)
t.save(model.state_dict(), '%s_%s.pth' % (opt.model_prefix, epoch))
def run(self):
# 1 获取数据
data, char_to_ix, ix_to_chars = get_data(self.config)
vocab_size = len(char_to_ix)
print('样本数：%d' % len(data))
print('词典大小： %d' % vocab_size)
# 2 设置dataloader
data = torch.from_numpy(data)
data_loader = Data.DataLoader(data,
batch_size=self.config.batch_size,
shuffle=True,
num_workers=1)
# 3 创建模型
model = PoetryModel(vocab_size=vocab_size,
embedding_dim=self.config.embedding_dim,
hidden_dim=self.config.hidden_dim,
device=self.device,
layer_num=self.config.layer_num)
model.to(self.device)
# 4 创建优化器
optimizer = optim.Adam(model.parameters(),
lr=self.config.lr,
weight_decay=self.config.weight_decay)
# 5 创建损失函数,使用与logsoftmax的输出
criterion = nn.CrossEntropyLoss()
# 6.训练
self.train(data_loader, model, optimizer, criterion, char_to_ix,
ix_to_chars)
def train():
datas = np.load("tang.npz")
data = datas['data']
ix2word = datas['ix2word'].item()
word2ix = datas['word2ix'].item()
data = torch.from_numpy(data)
dataloader = DataLoader(data[:5000],
batch_size=config.batch_size,
shuffle=True,
num_workers=2)
model = PoetryModel(len(word2ix),
embedding_dim=config.embedding_dim,
hidden_dim=config.hidden_dim)
optimizer = optim.Adam(model.parameters(), lr=config.lr)
criterion = nn.CrossEntropyLoss()
model.to(config.device)
f = open('result.txt', 'w')
loss_history = []
for epoch in range(config.epoch):
start_time = time.time()
temp_loss = 0
for step, batch_data in enumerate(dataloader):
batch_data = batch_data.long().transpose(1, 0).contiguous()
optimizer.zero_grad()
trn, target = batch_data[:-1, :], batch_data[1:, :]
output, _ = model(trn)
loss = criterion(output, target.view(-1))
loss.backward()
optimizer.step()
temp_loss += loss.item()
if step % config.print_freq == 0 or step == len(dataloader) - 1:
print("Train: [{:2d}/{}] Step: {:03d}/{:03d} Loss: {} ".format(
epoch + 1, config.epoch, step,
len(dataloader) - 1, loss.item()))
loss_history.append(temp_loss / len(len(dataloader)))
elapsed_time = time.time() - start_time
print("Epoch: %d" % epoch + " " + "Loss: %d" % loss_history[-1] +
" Epoch time: " +
time.strftime("%H: %M: %S", time.gmtime(elapsed_time)))
torch.save(model.state_dict(), config.model_path)
def train():
# 获取数据
data, word2ix, ix2word = get_data()
data = torch.from_numpy(data)
dataloader = DataLoader(data,
batch_size=config.batch_size,
shuffle=True,
num_workers=1)
# 模型定义
model = PoetryModel(len(word2ix), config.embedding_dim, config.hidden_dim,
config.num_layers)
optimizer = optim.Adam(model.parameters(),
lr=config.lr,
weight_decay=config.weight_decay)
scheduler = StepLR(optimizer,
step_size=config.lr_step,
gamma=config.lr_gamma)
criterion = nn.CrossEntropyLoss()
model.to(config.device)
for epoch in range(config.epoch):
total_loss = 0
for data_ in tqdm(dataloader):
# 训练
data_ = data_.long().transpose(1, 0).contiguous()
data_ = data_.to(config.device)
optimizer.zero_grad()
input_, target = data_[:-1, :], data_[1:, :]
output, _ = model(input_)
loss = criterion(output, target.view(-1))
loss.backward()
optimizer.step()
total_loss += loss.item()
scheduler.step()
print("epoch: ", epoch, "loss: ", total_loss / len(dataloader))
torch.save(model.state_dict(),
'%s_%s.pth' % (config.model_prefix, epoch))
def train(**kwargs:dict) -> None:
for k, v in kwargs.items():
setattr(opt, k, v)
#vis = Visdom(env=opt.env)
#data get
data, word2ix, ix2word = get_data(opt)
data = t.from_numpy(data)
dataloader = t.utils.data.DataLoader(
data,
batch_size=opt.batch_size,
shuffle = True,
)
model = PoetryModel(len(word2ix), 2, 2)
optimizer = t.optim.Adam(model.parameters(), lr=opt.lr)
criterion = nn.CrossEntropyLoss()
if opt.model_path:
model.load_state_dict(t.load(opt.model_path))
if opt.user_gpu:
model.cuda()
criterion.cuda()
for epoch in range(opt.epoch):
for ii, data_ in tqdm.tqdm(enumerate(dataloader)):
data_ = data.long().transpose(1, 0).contiguous()
if opt.user_gpu : data_ = data_.cuda()
optimizer.zero_grad()
input_, target = V(data_[:-1, :]), V(data[1:, :])
ouput, _ = model(input_)
loss = criterion(ouput, target.view(-1))
loss.backward()
optimizer.step()
for i in range(len(data)):
data[i] = toList(data[i])
data[i].append("<EOP>")
# save the word dic for sample method
p.dump(word_to_ix, file('wordDic', 'w'))
# save all avaible word
# wordList = open('wordList','w')
# for w in word_to_ix:
# wordList.write(w.encode('utf-8'))
# wordList.close()
model = PoetryModel(len(word_to_ix), 256, 256)
model.cuda(
) # running on GPU,if you want to run it on CPU,delete all .cuda() usage.
optimizer = optim.RMSprop(model.parameters(), lr=0.01, weight_decay=0.0001)
criterion = nn.NLLLoss()
one_hot_var_target = {}
for w in word_to_ix:
one_hot_var_target.setdefault(w, make_one_hot_vec_target(w, word_to_ix))
epochNum = 10
TRAINSIZE = len(data)
batch = 100
def test():
v = int(TRAINSIZE / batch)
loss = 0
counts = 0
def train(**kwargs):
for k, v in kwargs.items():
setattr(opt, k, v)
vis = Visualizer(env=opt.env)
# 获取数据
data, word2ix, ix2word = get_data(opt)
data = t.from_numpy(data)
dataloader = t.utils.data.DataLoader(data,
batch_size=opt.batch_size,
shuffle=True,
num_workers=1)
# 模型定义
model = PoetryModel(len(word2ix), 128, 256)
optimizer = t.optim.Adam(model.parameters(), lr=opt.lr)
criterion = nn.CrossEntropyLoss()
if opt.model_path:
model.load_state_dict(t.load(opt.model_path))
if opt.use_gpu:
model.cuda()
criterion.cuda()
loss_meter = meter.AverageValueMeter()
for epoch in range(opt.epoch):
loss_meter.reset()
for ii, data_ in tqdm.tqdm(enumerate(dataloader)):
# 训练
data_ = data_.long().transpose(1, 0).contiguous()
if opt.use_gpu: data_ = data_.cuda()
optimizer.zero_grad()
input_, target = Variable(data_[:-1, :]), Variable(data_[1:, :])
output, _ = model(input_)
loss = criterion(output, target.view(-1))
loss.backward()
optimizer.step()
loss_meter.add(loss.data[0])
# 可视化
if (1 + ii) % opt.plot_every == 0:
if os.path.exists(opt.debug_file):
ipdb.set_trace()
vis.plot('loss', loss_meter.value()[0])
# 诗歌原文
poetrys = [[ix2word[_word] for _word in data_[:, _iii]]
for _iii in range(data_.size(1))][:16]
vis.text('</br>'.join([''.join(poetry) for poetry in poetrys]), win=u'origin_poem')
gen_poetries = []
# 分别以这几个字作为诗歌的第一个字，生成8首诗
for word in list(u'春江花月夜凉如水'):
gen_poetry = ''.join(generate(model, word, ix2word, word2ix))
gen_poetries.append(gen_poetry)
vis.text('</br>'.join([''.join(poetry) for poetry in gen_poetries]), win=u'gen_poem')
t.save(model.state_dict(), '%s_%s.pth' % (opt.model_prefix, epoch))
def train(**kwargs):
for k, v in kwargs.items():
setattr(config, k, v)
device = torch.device('cuda') if use_cuda else torch.device('cpu')
# 获取数据
data, vocab = get_data(config.filepath)
np.random.shuffle(data)
l = len(data)
dev_data = data[:l // 5 - 1]
data = data[l // 5:]
data = torch.from_numpy(data)
dev_data = torch.from_numpy(dev_data)
dataloader = D.DataLoader(data,
batch_size=config.batch_size,
shuffle=True,
num_workers=4)
dev_dataloader = D.DataLoader(dev_data,
batch_size=config.batch_size,
shuffle=True,
num_workers=4)
# 模型定义
model = PoetryModel(len(vocab.word2idx), 128, 256)
# if config.model_path:
# model.load_state_dict(torch.load(config.model_path))
model.to(device)
# SGD, SGD with momentum, Nesterov, Adagrad, Adadelta, Adam
# optimizer = torch.optim.SGD(model.parameters(), lr=config.lr)
# optimizer = torch.optim.SGD(model.parameters(), lr=config.lr, momentum=0.9)
# optimizer = torch.optim.SGD(model.parameters(), lr=config.lr, momentum=0.9, nesterov=True)
# optimizer = torch.optim.Adagrad(model.parameters(), lr=config.lr)
# optimizer = torch.optim.Adadelta(model.parameters())
optimizer = torch.optim.Adam(model.parameters(), lr=config.lr)
criterion = nn.CrossEntropyLoss()
pre_pp = 0
cnt = -1
loss_his = []
pp_his = []
for epoch in range(config.epoch):
for ii, data_ in enumerate(dataloader):
# 训练
data_ = data_.long().transpose(1, 0).contiguous()
data_ = data_.to(device)
optimizer.zero_grad()
input_, target = data_[:-1, :], data_[1:, :]
output, _ = model(input_)
loss = criterion(output, target.view(-1))
loss.backward()
optimizer.step()
print("epoch", epoch, "step", ii, "loss", loss.item())
loss_his.append(loss.item())
# 测试
if (1 + ii) % config.gen_every == 0:
# "'春江花月夜凉如水'"
word = "春"
gen_poetry = ''.join(generate(model, word, vocab))
print(gen_poetry)
if (1 + ii) % config.pp_every == 0:
pp = check_perplexity(model, dev_dataloader)
if pre_pp < pp:
cnt += 1
pre_pp = pp
print(pp.cpu().numpy())
pp_his.append(pp.cpu().numpy())
if cnt >= config.tolerance:
torch.save(model.state_dict(),
'%s_final.pth' % str(int(time.time())))
print("epoch", epoch, "step", ii, "final loss",
loss.item())
for word in ["日", "红", "山", "夜", "湖", "海", "月"]:
gen_poetry = ''.join(generate(model, word, vocab))
print(gen_poetry)
return loss_his, pp_his
if (epoch + 1) % config.save_every == 0 or epoch + 1 == config.epoch:
torch.save(model.state_dict(),
'%s_%s.pth' % (str(int(time.time())), str(epoch)))
return loss_his, pp_his
def train(**kwargs):
for k, v in kwargs.items():
setattr(opt, k, v)
vis = Visualizer(env=opt.env)
#获取数据
data, word2ix, ix2word = get_data(opt)
data = t.from_numpy(data)
dataloader = t.utils.data.DataLoader(data,
batch_size=opt.batch_size,
shuffle=True,
num_workers=1)
#模型定义
model = PoetryModel(len(word2ix), 128, 256)
optimizer = t.optim.Adam(model.parameters(), lr=opt.lr)
criterion = nn.CrossEntropyLoss()
if opt.model_path:
model.load_state_dict(t.load(opt.model_path))
if opt.use_gpu:
model.cuda()
criterion.cuda()
loss_meter = meter.AverageValueMeter()
for epoch in range(opt.epoch):
loss_meter.reset()
for li, data_ in tqdm.tqdm(enumerate(dataloader)):
#训练
data_ = data_.long().transpose(1, 0).contiguous()
if opt.use_gpu: data_ = data_.cuda()
optimizer.zero_grad()
##输入和目标错开
input_, target = Variable(data_[:-1, :]), Variable(data_[1:, :])
output, _ = model(input_)
loss = criterion(output, target.view(-1))
loss.backward()
optimizer.step()
loss_meter.add(loss.data[0])
# 可视化
if (1 + ii) % opt.plot_every == 0:
if os.path.exists(opt.debug_file):
ipdb.set_trace()
vis.plot('loss', loss_meter.value()[0])
#诗歌原文
poetrys = [[ix2word[_word] for _word in data_[:, -iii]]
for _iii in range(data_.size(1))][:16]
vis.text('</br>'.join([''.join(poetry) for poetry in poetrys]),
win=u'origin_poem')
gen_poetries = []
#分别以这几个字作为诗歌的第一个字，生成8首诗
for word in list(u'春江花月夜凉如水'):
gen_poetry = ''.join(
generate(model, word, ix2word, word2ix))
gen_poetries.append(gen_poetry)
vis.text('</br>'.join(
[''.join(poetry) for poetry in gen_poetries]),
win=u'gen_poem')
t.save(model.state_dict(), '%s_%s.pth' % (opt.model_prefix, epoch))
def train(**kwargs):
for k, v in kwargs.items():
setattr(opt, k, v)
opt.device = t.device('cuda' if t.cuda.is_available() else 'cpu')
device = opt.device
vis = Visualizer(env=opt.env)
# 获取数据
data_all = np.load(opt.pickle_path)
data = data_all['data']
word2ix = data_all['word2ix'].item()
ix2word = data_all['ix2word'].item()
data = t.from_numpy(data)
dataloader = DataLoader(data,
batch_size=opt.batch_size,
shuffle=True,
num_workers=1)
# 模型定义
model = PoetryModel(len(word2ix), 128, 256)
optimizer = t.optim.Adam(model.parameters(), lr=opt.lr)
loss_func = nn.CrossEntropyLoss()
if opt.model_path:
model.load_state_dict(
t.load(opt.model_path, map_location=t.device('cpu')))
model.to(device)
loss_avg = 0
for epoch in range(opt.epoch):
for ii, data_ in tqdm(enumerate(dataloader)):
data_ = data_.long()
data_ = data_.to(device)
optimizer.zero_grad()
input_, target = data_[:, :-1], data_[:, 1:]
output, _ = model(input_)
loss = loss_func(output, target.reshape(-1))
loss.backward()
optimizer.step()
loss_avg += loss.item()
# 可视化
if (ii + 1) % opt.plot_every == 0:
vis.plot('loss', loss_avg / opt.plot_every)
loss_avg = 0
poetrys = [[ix2word[_word] for _word in data_[i].tolist()]
for i in range(data_.shape[0])][:16]
vis.text('</br>'.join([''.join(poetry) for poetry in poetrys]),
win='origin_poem')
gen_poetries = []
for word in list('春江花月夜凉如水'):
gen_poetry = ''.join(
generate(model, word, ix2word, word2ix))
gen_poetries.append(gen_poetry)
vis.text('</br>'.join(
[''.join(poetry) for poetry in gen_poetries]),
win='gen_poem')
t.save(model.state_dict(), '%s_%s.pth' % (opt.model_prefix, epoch))
def train(**kwargs):
for k,v in kwargs.items():
setattr(opt,k,v)
vis = Visualizer(env=opt.env)
# 获取数据
data,word2ix,ix2word = get_data(opt)
data = t.from_numpy(data)#把数据类型转为tensor
dataloader = t.utils.data.DataLoader(data,#初始化Dataloader类实例
batch_size=opt.batch_size,
shuffle=True,
num_workers=1)
# 模型定义
model = PoetryModel(len(word2ix), 128, 256)#(vocab_size, embedding_dim, hidden_dim)
optimizer = t.optim.Adam(model.parameters(), lr=opt.lr)
criterion = nn.CrossEntropyLoss()#损失函数定义为交叉熵
if opt.model_path:
model.load_state_dict(t.load(opt.model_path))
if opt.use_gpu:
model.cuda()
criterion.cuda()
loss_meter = meter.AverageValueMeter()
for epoch in range(opt.epoch):
loss_meter.reset()
for ii,data_ in tqdm.tqdm(enumerate(dataloader)): #tqdm进度条工具
#取一个batch的数据
# 训练
#data_.size:(batch_size,maxlen)
data_ = data_.long().transpose(1,0).contiguous()#转置后返回一个内存连续的有相同数据的tensor
# if epoch==0 and ii ==0:
# print('size of data_ after transpose: \n',data_.size())
if opt.use_gpu: data_ = data_.cuda()
optimizer.zero_grad()#梯度清零
input_,target = Variable(data_[:-1,:]),Variable(data_[1:,:])#input_是所有句子的前maxlen-1个item的集合，
#target是所有句子的后maxlen-１个item的集合
#以＂床前明月光＂为例，输入是＂床前明月＂，要预测＂前明月光＂
output,_ = model(input_)
#Tensor.view(-1)按照第０个维度逐个元素读取将张量展开成数组
loss = criterion(output,target.view(-1))
loss.backward()
optimizer.step()
loss_meter.add(loss.data[0])
# 可视化
if (1+ii)%opt.plot_every==0:
if os.path.exists(opt.debug_file):#如果存在调试文件，
#则进入调试模式
ipdb.set_trace()
vis.plot('loss',loss_meter.value()[0])
# 诗歌原文
poetrys=[ [ix2word[_word] for _word in data_[:,_iii]] #每一个句子（诗歌）的每一个item（id）要转换成文本
for _iii in range(data_.size(1))][:16]#_iii的取值范围[,127]
vis.text('</br>'.join([''.join(poetry) for poetry in poetrys]),win=u'origin_poem')
#在visdom中输出这些句子（诗歌）中的前１６个
gen_poetries = []
# 分别以这几个字作为诗歌的第一个字，生成8首诗
for word in list(u'春江花月夜凉如水'):
gen_poetry = ''.join(generate(model,word,ix2word,word2ix))
gen_poetries.append(gen_poetry)
vis.text('</br>'.join([''.join(poetry) for poetry in gen_poetries]),win=u'gen_poem')
t.save(model.state_dict(),'%s_%s.pth' %(opt.model_prefix,epoch))
def train(**kwargs):
for k, v in kwargs.items():
setattr(opt, k, v)
vis = Visualizer(env=opt.env)
# 获取数据
data, word2ix, ix2word = get_data(opt)
data = t.from_numpy(data)
dataloader = t.utils.data.DataLoader(data, batch_size=opt.batch_size, shuffle=True, num_workers=2)
# 定义model
model = PoetryModel(len(word2ix), opt.embedding_dim, opt.hidden_dim)
# 优化器
optimizer = t.optim.Adam(model.parameters(), lr=opt.lr)
# Loss Function
criterion = nn.CrossEntropyLoss()
# 使用预训练的模型，为了可持续训练
if opt.model_path and os.path.exists(opt.model_path):
model.load_state_dict(t.load(opt.model_path))
# GPU related
if opt.use_gpu:
model = model.to(device)
criterion = criterion.to(device)
# loss 计量器
loss_meter = meter.AverageValueMeter()
# for loop
for epoch in range(opt.epoch):
loss_meter.reset()
# for : batching dataset
for i, data_ in tqdm.tqdm(enumerate(dataloader)):
# 训练
# data_
# size: [128, 125] 每次取128行，每行一首诗，长度为125
# type: Tensor
# dtype: torch.int32 应该转成long
# 这行代码信息量很大：
# 第一步：int32 to long
# 第二步：将行列互换，为了并行计算的需要
# 第三步：将数据放置在连续内存里，避免后续有些操作报错
data_ = data_.long().transpose(0, 1).contiguous()
# GPU related
if opt.use_gpu:
data_ = data_.to(device)
# 到这里 data_.dtype又变成了torch.int64
# print(data_.dtype)
# 清空梯度
optimizer.zero_grad()
# 错位训练，很容易理解
# 把前n-1行作为input，把后n-1行作为target : model的输入
# 这么做还是为了并行计算的需要
# input_ 加下划线是为了和built_in function input区分开
input_, target = data_[:-1, :], data_[1:, :]
# model的返回值 output和hidden
# 这里hidden没什么用
output, _ = model(input_)
# 计算loss
target = target.view(-1)
# 新的target.size() [15872] 124 * 128 = 15872
# output.size() [15872, 8293] 8293 是词汇量的大小
loss = criterion(output, target)
# 反向传播
loss.backward()
# optimizer梯度下降更新参数
optimizer.step()
loss_meter.add(loss.data[0])
# 可视化
if (1 + i) % opt.plot_every == 0:
if os.path.exists(opt.debug_file):
ipdb.set_trace()
vis.plot('loss', loss_meter.value()[0])
# 诗歌原文
poetrys = [[ix2word[_word.item()] for _word in data_[:, _iii]]
for _iii in range(data_.size(1))][:16]
vis.text('</br>'.join([''.join(poetry) for poetry in poetrys]), win=u'origin_poem')
gen_poetries = []
# 分别以这几个字作为诗歌的第一个字，生成8首诗
for word in list(u'春江花月夜凉如水'):
gen_poetry = ''.join(generate(model, word, ix2word, word2ix))
gen_poetries.append(gen_poetry)
vis.text('</br>'.join([''.join(poetry) for poetry in gen_poetries]), win=u'gen_poem')
# 迭代一次epoch，保存一下模型
t.save(model.state_dict(), '%s_%s.pth' % (opt.model_prefix, epoch))
def train_torch_lstm(conf, args=None):
pdata = PoemData()
pdata.read_data(conf)
pdata.get_vocab()
if conf.use_gpu:
device = torch.device('cuda')
else:
device = torch.device('cpu')
model = PoetryModel(pdata.vocab_size, conf, device)
train_data = pdata.train_data
test_data = pdata.test_data
train_data = torch.from_numpy(np.array(train_data['pad_words']))
dev_data = torch.from_numpy(np.array(test_data['pad_words']))
dataloader = DataLoader(train_data,
batch_size=conf.batch_size,
shuffle=True,
num_workers=conf.num_workers)
devloader = DataLoader(dev_data,
batch_size=conf.batch_size,
shuffle=True,
num_workers=conf.num_workers)
optimizer = Adam(model.parameters(), lr=conf.learning_rate)
criterion = nn.CrossEntropyLoss()
loss_meter = meter.AverageValueMeter()
if conf.load_best_model:
model.load_state_dict(torch.load(conf.beat_model_path))
if conf.use_gpu:
model.cuda()
criterion.cuda()
step = 0
bestppl = 1e9
early_stop_controller = 0
for epoch in range(conf.n_epochs):
losses = []
loss_meter.reset()
model.train()
for i, data in enumerate(dataloader):
data = data.long().transpose(1, 0).contiguous()
if conf.use_gpu:
data = data.cuda()
input, target = data[:-1, :], data[1:, :]
optimizer.zero_grad()
output, _ = model(input)
loss = criterion(output, target.contiguous().view(-1))
loss.backward()
optimizer.step()
losses.append(loss.item())
loss_meter.add(loss.item())
step += 1
if step % 100 == 0:
print("epoch_%d_step_%d_loss:%0.4f" %
(epoch + 1, step, loss.item()))
train_loss = float(loss_meter.value()[0])
model.eval()
for i, data in enumerate(devloader):
data = data.long().transpose(1, 0).contiguous()
if conf.use_gpu:
data = data.cuda()
input, target = data[:-1, :], data[1:, :]
output, _ = model(input)
loss = criterion(output, target.view(-1))
loss_meter.add(loss.item())
ppl = math.exp(loss_meter.value()[0])
print("epoch_%d_loss:%0.4f , ppl:%0.4f" % (epoch + 1, train_loss, ppl))
if epoch % conf.save_every == 0:
torch.save(model.state_dict(),
"{0}_{1}".format(conf.model_prefix, epoch))
fout = open("{0}out_{1}".format(conf.out_path, epoch),
'w',
encoding='utf-8')
for word in list('日红山夜湖海月'):
gen_poetry = generate_poet(model, word, pdata.vocab, conf)
fout.write("".join(gen_poetry) + '\n\n')
fout.close()
if ppl < bestppl:
bestppl = ppl
early_stop_controller = 0
torch.save(model.state_dict(),
"{0}_{1}".format(conf.best_model_path, "best_model"))
else:
early_stop_controller += 1
if early_stop_controller > conf.patience:
print("early stop.")
break
def train(Config):
torch.multiprocessing.set_sharing_strategy('file_system')
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
datas = np.load("data/chinese-poetry-master/tang .npz", allow_pickle=True)
data = datas["data"]
ix2word = datas['ix2word'].item()
word2ix = datas['word2ix'].item()
data = torch.from_numpy(data)
print(data.shape)
# #去掉空格
# t_data = data.view(-1)
# flat_data = t_data.numpy()
# no_space_data = []
# for i in flat_data:
# if (i != 8292):
# no_space_data.append(i)
# slice_size = 48
# txt = [no_space_data[i:i+slice_size] for i in range(0,len(no_space_data),slice_size)]
# txt = np.array(txt[:-1])#去掉最后一个不够48的数据
# txt = torch.from_numpy(txt).long()
# print(txt.shape)
#datas = PoemDataSet(Config.data_path, 48)
#data = datas.no_space_data#datas['data']
#ix2word = datas.ix2word#datas['ix2word'].item()
#word2ix = datas.word2ix#datas['word2ix'].item()
dataLoader = DataLoader(data,
batch_size=Config.batch_size,
shuffle=Config.shuffle,
num_workers=Config.num_workers)
model = PoetryModel(len(word2ix),
embedding_dim=Config.embedding_dim,
hidden_dim=Config.hidden_dim).to(device)
optimizer = optim.Adam(model.parameters(), lr=Config.lr)
scheduler = torch.optim.lr_scheduler.StepLR(optimizer,
step_size=10,
gamma=0.1) # 学习率调整
crierion = nn.CrossEntropyLoss()
loss_meter = meter.AverageValueMeter()
top1 = meter.AverageValueMeter()
#top1 = utils.AverageMeter()
# if Config.model_path:
# model.load_state_dict(torch.load(Config.model_path))
train_loss_list = []
train_accuracy_list = []
for epoch in range(Config.epoch):
loss_meter.reset()
top1.reset()
for ii, data_ in enumerate(
dataLoader): #tqdm.tqdm(enumerate(dataLoader)):
#inputs, labels =Variable(data_[0]), Variable(data_[1])#.to(device)
data_ = data_.long().transpose(1, 0).contiguous()
inputs, labels = Variable(data_[:-1, :]), Variable(data_[1:, :])
print(inputs.size(1))
optimizer.zero_grad()
# 上边一句，将输入的诗句错开一个字，形成训练和目标
output, _ = model(inputs)
loss = crierion(output, labels.view(-1))
loss.backward()
optimizer.step()
loss_meter.add(loss.item())
_, pred = output.topk(1)
prec1, prec2 = accuracy(output, labels, topk=(1, 2))
n = inputs.size(0)
top1.add(prec1.item())
#data = data.long().transpose(1,0).contiguous()
if (1 + ii) % Config.plot_every == 0:
if os.path.exists(Config.debug_file):
ipdb.set_trace()
# 下面是对目前模型情况的测试，诗歌原文、
#print(inputs.size(1))
#print(inputs.numpy()[:1].shape)
# poetrys = [[ix2word[_word] for _word in inputs.numpy()[:, _iii]]
# for _iii in range(inputs.size(1))][0]
# poetrys =["".join(poetry) for poetry in poetrys]
# print("origen")
#print(poetrys)
# 上面句子嵌套了两个循环，主要是将诗歌索引的前十六个字变成原文
gen_poetries = []
start = u"春江花月夜凉如水"
gen_poetry = "".join(
generate(model, start, ix2word, word2ix, Config))
# for word in list(u"春江花月夜凉如水"):
# gen_poetry = "".join(generate(model, word, ix2word, word2ix,Config))
# gen_poetries.append(gen_poetry)
# gen_poetries="</br>".join(["".join(poetry) for poetry in gen_poetries])
print("genetate")
print(gen_poetry)
# if os.path.exists(Config.tensorboard_path) == False:
# os.mkdir(Config.tensorboard_path)
# writer = SummaryWriter(Config.tensorboard_path)
# writer.add_scalar('Train/Loss', loss.item(), epoch)
# writer.add_scalar('Train/Accuracy', 100*prec1.item()/output.size(0), epoch)
#
# writer.flush()
train_loss_list.append(loss.item())
train_accuracy_list.append(100 * prec1.item() / output.size(0))
print('train %d epoch loss: %.3f acc: %.3f ' %
(epoch + 1, loss_meter.mean, 100 * top1.mean / output.size(0)))
scheduler.step()
x1 = range(0, configs.epoch)
y1 = train_loss_list
y3 = train_accuracy_list
plt.subplot(2, 1, 1)
plt.plot(x1, y1, 'o-')
plt.legend(["train_loss"])
plt.title('Loss vs. epoches')
plt.ylabel('Loss')
plt.subplot(2, 1, 2)
plt.plot(x1, y3, '.-')
plt.legend("train_accuracy")
plt.xlabel('Accuracy vs. epoches')
plt.ylabel('Accuracy')
plt.show()
plt.savefig("pw_LSTM" + "_accuracy_loss.jpg")
torch.save(model.state_dict(), "%s_%s.pth" % (Config.model_prefix, epoch))
def train(**kwargs):
for k, v in kwargs.items():
setattr(opt, k, v)
opt.device = t.device('cuda:0') if opt.use_gpu else t.device('cpu')
device = opt.device
#vis = Visualizer(env=opt.env)
# 获取数据
data, word2ix, ix2word = get_data(opt)
data = t.from_numpy(data) #[57580,125]
dataloader = t.utils.data.DataLoader(data,
batch_size=opt.batch_size,
shuffle=True,
num_workers=1)
# 模型定义
model = PoetryModel(len(word2ix), 128, 256)
optimizer = t.optim.Adam(model.parameters(), lr=opt.lr)
criterion = nn.CrossEntropyLoss()
if opt.model_path:
model.load_state_dict(t.load(opt.model_path))
model.to(device)
# AverageMeter类用来管理一些变量的更新
loss_meter = meter.AverageValueMeter()
for epoch in range(opt.epoch):
loss_meter.reset() # 在每一个epoch 都要进行reset一遍
for ii, data_ in tqdm.tqdm(enumerate(dataloader)):
# 训练
#contiguous：view只能用在contiguous的variable上。如果在view之前用了transpose, permute等，需要用contiguous()来返回一个contiguous copy。
#一种可能的解释是：
#有些tensor并不是占用一整块内存，而是由不同的数据块组成，而tensor的view()操作依赖于内存是整块的，这时只需要执行contiguous()这个函数，把tensor变成在内存中连续分布的形式。
# 也就是说使用contiguous()是为了能够使用view（）
data_ = data_.long().transpose(
1, 0).contiguous() #data_ shape:[seq_len,batch_size]
data_ = data_.to(device)
optimizer.zero_grad()
#input_ shape:[124,128] target shape:[124,128]
input_, target = data_[:-1, :], data_[1:, :]
output, _ = model(
input_
) # ouput shape:[seq_len * batch_size,vocab_size] 此处seq_len为124
loss = criterion(
output, target.view(-1)) # target需要规整成[seq_len * batch_size]
loss.backward()
optimizer.step()
# 更新loss_meter
loss_meter.add(loss.item())
'''
# 可视化
if (1 + ii) % opt.plot_every == 0:
if os.path.exists(opt.debug_file):
ipdb.set_trace()
vis.plot('loss', loss_meter.value()[0])
# 诗歌原文
poetrys = [[ix2word[_word] for _word in data_[:, _iii].tolist()]
for _iii in range(data_.shape[1])][:16]
vis.text('</br>'.join([''.join(poetry) for poetry in poetrys]), win=u'origin_poem')
gen_poetries = []
# 分别以这几个字作为诗歌的第一个字，生成8首诗
for word in list(u'春江花月夜凉如水'):
gen_poetry = ''.join(generate(model, word, ix2word, word2ix))
gen_poetries.append(gen_poetry)
vis.text('</br>'.join([''.join(poetry) for poetry in gen_poetries]), win=u'gen_poem')
'''
# 每一个epoch都打印下loss的值
print('epoch:%d, loss:%.3f' % (epoch, loss_meter.value()[0]))
#需要改进的地方是 得到模型在验证集的结果 根据结果只保存最好的模型
t.save(model.state_dict(), '%s_%s.pth' % (opt.model_prefix, epoch))
return ix2word, word2ix
for i in range(len(data)):
data[i] = toList(data[i])
data[i].append("<EOP>")
# save the word dic for sample method
p.dump(word_to_ix, file('wordDic', 'w'))
# save all avaible word
# wordList = open('wordList','w')
# for w in word_to_ix:
# wordList.write(w.encode('utf-8'))
# wordList.close()
model = PoetryModel(len(word_to_ix), 256, 256)
model.cuda(
) # running on GPU,if you want to run it on CPU,delete all .cuda() usage.
optimizer = optim.Adam(model.parameters(), lr=0.001, weight_decay=0.0001)
criterion = nn.NLLLoss()
one_hot_var_target = {}
for w in word_to_ix:
one_hot_var_target.setdefault(w, make_one_hot_vec_target(w, word_to_ix))
epochNum = 100
TRAINSIZE = len(data)
batch = 200
def test():
model.eval()
v = int(TRAINSIZE / batch)
loss = 0