class PoemWAE(nn.Module):
def __init__(self, config, api, PAD_token=0, pretrain_weight=None):
super(PoemWAE, self).__init__()
self.vocab = api.vocab
self.vocab_size = len(self.vocab)
self.rev_vocab = api.rev_vocab
self.go_id = self.rev_vocab["<s>"]
self.eos_id = self.rev_vocab["</s>"]
self.maxlen = config.maxlen
self.clip = config.clip
self.lambda_gp = config.lambda_gp
self.lr_gan_g = config.lr_gan_g
self.lr_gan_d = config.lr_gan_d
self.n_d_loss = config.n_d_loss
self.temp = config.temp
self.init_w = config.init_weight
self.embedder = nn.Embedding(self.vocab_size,
config.emb_size,
padding_idx=PAD_token)
if pretrain_weight is not None:
self.embedder.weight.data.copy_(torch.from_numpy(pretrain_weight))
# 用同一个seq_encoder来编码标题和前后两句话
self.seq_encoder = Encoder(self.embedder, config.emb_size,
config.n_hidden, True, config.n_layers,
config.noise_radius)
# 由于Poem这里context是title和last sentence双向GRU编码后的直接cat，4*hidden
# 注意如果使用Poemwar_gmp则使用子类中的prior_net，即混合高斯分布的一个先验分布
self.prior_net = Variation(config.n_hidden * 4,
config.z_size,
dropout_rate=config.dropout,
init_weight=self.init_w) # p(e|c)
# 注意这儿原来是给Dialog那个任务用的，3*hidden
# Poem数据集上，将title和上一句，另外加上x都分别用双向GRU编码并cat，因此是6*hidden
self.post_net = Variation(config.n_hidden * 6,
config.z_size,
dropout_rate=config.dropout,
init_weight=self.init_w)
self.post_generator = nn.Sequential(
nn.Linear(config.z_size, config.z_size),
nn.BatchNorm1d(config.z_size, eps=1e-05, momentum=0.1), nn.ReLU(),
nn.Linear(config.z_size, config.z_size),
nn.BatchNorm1d(config.z_size, eps=1e-05, momentum=0.1), nn.ReLU(),
nn.Linear(config.z_size, config.z_size))
self.post_generator.apply(self.init_weights)
self.prior_generator = nn.Sequential(
nn.Linear(config.z_size, config.z_size),
nn.BatchNorm1d(config.z_size, eps=1e-05, momentum=0.1), nn.ReLU(),
nn.Linear(config.z_size, config.z_size),
nn.BatchNorm1d(config.z_size, eps=1e-05, momentum=0.1), nn.ReLU(),
nn.Linear(config.z_size, config.z_size))
self.prior_generator.apply(self.init_weights)
self.init_decoder_hidden = nn.Sequential(
nn.Linear(config.n_hidden * 4 + config.z_size,
config.n_hidden * 4),
nn.BatchNorm1d(config.n_hidden * 4, eps=1e-05, momentum=0.1),
nn.ReLU())
# 由于Poem这里context是title和last sentence双向GRU编码后的直接cat，因此hidden_size变为z_size + 4*hidden
# 修改：decoder的hidden_size还设为n_hidden, init_hidden使用一个MLP将cat变换为n_hidden
self.decoder = Decoder(self.embedder,
config.emb_size,
config.n_hidden * 4,
self.vocab_size,
n_layers=1)
self.discriminator = nn.Sequential(
# 因为Poem的cat两个双向编码，这里改为4*n_hidden + z_size
nn.Linear(config.n_hidden * 4 + config.z_size,
config.n_hidden * 2),
nn.BatchNorm1d(config.n_hidden * 2, eps=1e-05, momentum=0.1),
nn.LeakyReLU(0.2),
nn.Linear(config.n_hidden * 2, config.n_hidden * 2),
nn.BatchNorm1d(config.n_hidden * 2, eps=1e-05, momentum=0.1),
nn.LeakyReLU(0.2),
nn.Linear(config.n_hidden * 2, 1),
)
self.discriminator.apply(self.init_weights)
# optimizer 定义，分别对应三个模块的训练，注意！三个模块的optimizer不相同
# self.optimizer_AE = optim.SGD(list(self.seq_encoder.parameters())
self.optimizer_AE = optim.SGD(
list(self.seq_encoder.parameters()) +
list(self.post_net.parameters()) +
list(self.post_generator.parameters()) +
list(self.init_decoder_hidden.parameters()) +
list(self.decoder.parameters()),
lr=config.lr_ae)
self.optimizer_G = optim.RMSprop(
list(self.post_net.parameters()) +
list(self.post_generator.parameters()) +
list(self.prior_net.parameters()) +
list(self.prior_generator.parameters()),
lr=self.lr_gan_g)
self.optimizer_D = optim.RMSprop(self.discriminator.parameters(),
lr=self.lr_gan_d)
self.lr_scheduler_AE = optim.lr_scheduler.StepLR(self.optimizer_AE,
step_size=10,
gamma=0.8)
self.criterion_ce = nn.CrossEntropyLoss()
def init_weights(self, m):
if isinstance(m, nn.Linear):
m.weight.data.uniform_(-self.init_w, self.init_w)
# nn.init.kaiming_normal_(m.weight.data)
# nn.init.kaiming_uniform_(m.weight.data)
m.bias.data.fill_(0)
# x: (batch, 2*n_hidden)
# c: (batch, 2*2*n_hidden)
def sample_code_post(self, x, c):
z, _, _ = self.post_net(torch.cat((x, c),
1)) # 输入：(batch, 3*2*n_hidden)
z = self.post_generator(z)
return z
def sample_code_prior_sentiment(self, c, align):
choice_statistic = self.prior_net(c, align) # e: (batch, z_size)
return choice_statistic
def sample_code_prior(self, c):
z, _, _ = self.prior_net(c) # e: (batch, z_size)
z = self.prior_generator(z) # z: (batch, z_size)
return z
# 输入 title, context, target, target_lens.
# c由title和context encode之后的hidden相concat而成
def train_AE(self, title, context, target, target_lens):
self.seq_encoder.train()
self.decoder.train()
# import pdb
# pdb.set_trace()
# (batch, 2 * hidden_size)
title_last_hidden, _ = self.seq_encoder(title)
context_last_hidden, _ = self.seq_encoder(context)
# (batch, 2 * hidden_size)
x, _ = self.seq_encoder(target[:, 1:], target_lens - 1)
# context_embedding
c = torch.cat((title_last_hidden, context_last_hidden),
1) # (batch, 2 * hidden_size * 2)
z = self.sample_code_post(x, c) # (batch, z_size)
# 标准的autoencoder的decode，decoder初态为x, c的cat，将target错位输入
# output: (batch, len, vocab_size) len是9，即7+标点+</s>
output = self.decoder(self.init_decoder_hidden(torch.cat((z, c), 1)),
None, target[:, :-1], target_lens - 1)
flattened_output = output.view(-1, self.vocab_size)
dec_target = target[:, 1:].contiguous().view(-1)
mask = dec_target.gt(0) # 即判断target的token中是否有0（pad项）
masked_target = dec_target.masked_select(mask) # 选出非pad项
output_mask = mask.unsqueeze(1).expand(
mask.size(0), self.vocab_size) # [(batch_sz * seq_len) x n_tokens]
masked_output = flattened_output.masked_select(output_mask).view(
-1, self.vocab_size)
self.optimizer_AE.zero_grad()
loss = self.criterion_ce(masked_output / self.temp, masked_target)
loss.backward()
torch.nn.utils.clip_grad_norm_(
list(self.seq_encoder.parameters()) +
list(self.decoder.parameters()), self.clip)
self.optimizer_AE.step()
return [('train_loss_AE', loss.item())]
# G是来缩短W距离的，可以类比VAE里面的缩小KL散度项
def train_G(self,
title,
context,
target,
target_lens,
sentiment_mask=None,
mask_type=None):
self.seq_encoder.eval()
self.optimizer_G.zero_grad()
for p in self.discriminator.parameters():
p.requires_grad = False
title_last_hidden, _ = self.seq_encoder(title)
context_last_hidden, _ = self.seq_encoder(context)
c = torch.cat((title_last_hidden, context_last_hidden),
1) # (batch, 2 * hidden_size * 2)
# -----------------posterior samples ---------------------------
x, _ = self.seq_encoder(target[:, 1:], target_lens - 1)
z_post = self.sample_code_post(
x.detach(), c.detach()) # 去掉梯度，防止梯度向encoder的传播 (batch, z_size)
errG_post = torch.mean(
self.discriminator(torch.cat(
(z_post, c.detach()),
1))) * self.n_d_loss # (batch, z_size + 4 * hidden)
errG_post.backward(minus_one)
# ----------------- prior samples ---------------------------
prior_z = self.sample_code_prior(c.detach())
errG_prior = torch.mean(
self.discriminator(torch.cat(
(prior_z, c.detach()), 1))) * self.n_d_loss
# import pdb
# pdb.set_trace()
errG_prior.backward(one)
self.optimizer_G.step()
for p in self.discriminator.parameters():
p.requires_grad = True
costG = errG_prior - errG_post
return [('train_loss_G', costG.item())]
# D是用来拟合W距离，loss下降说明拟合度变好，增大gradient_penalty一定程度上可以提高拟合度
# n_iters_n越大，D训练的次数越多，对应的拟合度也越好
def train_D(self, title, context, target, target_lens):
self.seq_encoder.eval()
self.discriminator.train()
self.optimizer_D.zero_grad()
batch_size = context.size(0)
title_last_hidden, _ = self.seq_encoder(title)
context_last_hidden, _ = self.seq_encoder(context)
c = torch.cat((title_last_hidden, context_last_hidden),
1) # (batch, 2, hidden_size * 2)
x, _ = self.seq_encoder(target[:, 1:], target_lens - 1)
post_z = self.sample_code_post(x, c)
errD_post = torch.mean(
self.discriminator(torch.cat(
(post_z.detach(), c.detach()), 1))) * self.n_d_loss
errD_post.backward(one)
prior_z = self.sample_code_prior(c)
errD_prior = torch.mean(
self.discriminator(torch.cat(
(prior_z.detach(), c.detach()), 1))) * self.n_d_loss
errD_prior.backward(minus_one)
# import pdb
# pdb.set_trace()
alpha = to_tensor(torch.rand(batch_size, 1))
alpha = alpha.expand(prior_z.size())
interpolates = alpha * prior_z.data + ((1 - alpha) * post_z.data)
interpolates = Variable(interpolates, requires_grad=True)
d_input = torch.cat((interpolates, c.detach()), 1)
disc_interpolates = torch.mean(self.discriminator(d_input))
gradients = torch.autograd.grad(
outputs=disc_interpolates,
inputs=interpolates,
grad_outputs=to_tensor(torch.ones(disc_interpolates.size())),
create_graph=True,
retain_graph=True,
only_inputs=True)[0]
gradient_penalty = (
(gradients.contiguous().view(gradients.size(0), -1).norm(2, dim=1)
- 1)**2).mean() * self.lambda_gp
gradient_penalty.backward()
self.optimizer_D.step()
costD = -(errD_prior - errD_post) + gradient_penalty
return [('train_loss_D', costD.item())]
def valid(self, title, context, target, target_lens, sentiment_mask=None):
self.seq_encoder.eval()
self.discriminator.eval()
self.decoder.eval()
title_last_hidden, _ = self.seq_encoder(title)
context_last_hidden, _ = self.seq_encoder(context)
c = torch.cat((title_last_hidden, context_last_hidden),
1) # (batch, 2 * hidden_size * 2)
x, _ = self.seq_encoder(target[:, 1:], target_lens - 1)
post_z = self.sample_code_post(x, c)
prior_z = self.sample_code_prior(c)
errD_post = torch.mean(self.discriminator(torch.cat((post_z, c), 1)))
errD_prior = torch.mean(self.discriminator(torch.cat((prior_z, c), 1)))
costD = -(errD_prior - errD_post)
costG = -costD
dec_target = target[:, 1:].contiguous().view(-1) # (batch_size * len)
mask = dec_target.gt(0) # 即判断target的token中是否有0（pad项）
masked_target = dec_target.masked_select(mask) # 选出非pad项
output_mask = mask.unsqueeze(1).expand(mask.size(0), self.vocab_size)
output = self.decoder(
self.init_decoder_hidden(torch.cat((post_z, c), 1)), None,
target[:, :-1], (target_lens - 1))
flattened_output = output.view(-1, self.vocab_size)
masked_output = flattened_output.masked_select(output_mask).view(
-1, self.vocab_size)
lossAE = self.criterion_ce(masked_output / self.temp, masked_target)
return [('valid_loss_AE', lossAE.item()),
('valid_loss_G', costG.item()), ('valid_loss_D', costD.item())]
# 正如论文中说的，测试生成的时候，从先验网络中拿到噪声，用G生成prior_z（即代码中的sample_code_prior(c)）
# 然后decoder将prior_z和c的cat当做输入，decode出这句诗（这和论文里面不太一样，论文里面只把prior_z当做输入）
# batch_size是1，一次测一句
# title 即标题
# context 上一句
def test(self, title_tensor, title_words, headers):
self.seq_encoder.eval()
self.discriminator.eval()
self.decoder.eval()
# tem初始化为[2,3,0,0,0,0,0,0,0]
tem = [[2, 3] + [0] * (self.maxlen - 2)]
pred_poems = []
title_tokens = [
self.vocab[e] for e in title_words[0].tolist()
if e not in [0, self.eos_id, self.go_id]
]
pred_poems.append(title_tokens)
for sent_id in range(4):
tem = to_tensor(np.array(tem))
context = tem
# vec_context = np.zeros((batch_size, self.maxlen), dtype=np.int64)
# for b_id in range(batch_size):
# vec_context[b_id, :] = np.array(context[b_id])
# context = to_tensor(vec_context)
title_last_hidden, _ = self.seq_encoder(
title_tensor) # （batch=1, 2*hidden）
if sent_id == 0:
context_last_hidden, _ = self.seq_encoder(
title_tensor) # (batch=1, 2*hidden)
else:
context_last_hidden, _ = self.seq_encoder(
context) # (batch=1, 2*hidden)
c = torch.cat((title_last_hidden, context_last_hidden),
1) # (batch, 4*hidden_size)
# 由于一次只有一首诗，batch_size = 1，因此不必repeat
prior_z = self.sample_code_prior(c)
# decode_words 是完整的一句诗
decode_words = self.decoder.testing(
init_hidden=self.init_decoder_hidden(torch.cat((prior_z, c),
1)),
maxlen=self.maxlen,
go_id=self.go_id,
mode="greedy",
header=headers[sent_id])
decode_words = decode_words[0].tolist()
# import pdb
# pdb.set_trace()
if len(decode_words) > self.maxlen:
tem = [decode_words[0:self.maxlen]]
else:
tem = [[0] * (self.maxlen - len(decode_words)) + decode_words]
pred_tokens = [
self.vocab[e] for e in decode_words[:-1]
if e != self.eos_id and e != 0
]
pred_poems.append(pred_tokens)
gen = ''
for line in pred_poems:
true_str = " ".join(line)
gen = gen + true_str + '\n'
return gen
def sample(self, title, context, repeat, go_id, end_id):
self.seq_encoder.eval()
self.decoder.eval()
title_last_hidden, _ = self.seq_encoder(title)
context_last_hidden, _ = self.seq_encoder(context)
c = torch.cat((title_last_hidden, context_last_hidden),
1) # (batch, 2 * hidden_size * 2)
c_repeated = c.expand(
repeat, -1) # 注意，我们输入的batch_size是1，这里复制repeat遍，为了后面的BLEU计算
prior_z = self.sample_code_prior(
c_repeated) # c_repeated: (batch_size=repeat, 4*hidden_size)
# (batch, max_len, 1) (batch_size, 1)
sample_words, sample_lens = self.decoder.sampling(
self.init_decoder_hidden(torch.cat((prior_z, c_repeated), 1)),
self.maxlen, go_id, end_id, "greedy")
return sample_words, sample_lens
class DeepAPI(nn.Module):
''' model. '''
def __init__(self, config, vocab_size):
super(DeepAPI, self).__init__()
self.vocab_size = vocab_size
self.maxlen = config['maxlen']
self.clip = config['clip']
self.temp = config['temp']
self.desc_embedder = nn.Embedding(vocab_size,
config['emb_size'],
padding_idx=PAD_ID)
self.api_embedder = nn.Embedding(vocab_size,
config['emb_size'],
padding_idx=PAD_ID)
# utter encoder: encode response to vector
self.encoder = Encoder(self.desc_embedder, config['emb_size'],
config['n_hidden'], True, config['n_layers'],
config['noise_radius'])
self.decoder = Decoder(self.api_embedder, config['emb_size'],
config['n_hidden'] * 2, vocab_size,
config['use_attention'], 1,
config['dropout']) # utter decoder: P(x|c,z)
self.optimizer = optim.Adadelta(list(self.encoder.parameters()) +
list(self.decoder.parameters()),
lr=config['lr_ae'],
rho=0.95)
self.criterion_ce = nn.CrossEntropyLoss()
def forward(self, descs, desc_lens, apiseqs, api_lens):
c, hids = self.encoder(descs, desc_lens)
output, _ = self.decoder(c, hids, None, apiseqs[:, :-1],
(api_lens - 1))
# decode from z, c # output: [batch x seq_len x n_tokens]
output = output.view(-1, self.vocab_size) # [batch*seq_len x n_tokens]
dec_target = apiseqs[:, 1:].contiguous().view(-1)
mask = dec_target.gt(0) # [(batch_sz*seq_len)]
masked_target = dec_target.masked_select(mask) #
output_mask = mask.unsqueeze(1).expand(
mask.size(0), self.vocab_size) # [(batch_sz*seq_len) x n_tokens]
masked_output = output.masked_select(output_mask).view(
-1, self.vocab_size)
loss = self.criterion_ce(masked_output / self.temp, masked_target)
return loss
def train_AE(self, descs, desc_lens, apiseqs, api_lens):
self.encoder.train()
self.decoder.train()
loss = self.forward(descs, desc_lens, apiseqs, api_lens)
self.optimizer.zero_grad()
loss.backward()
# `clip_grad_norm` to prevent exploding gradient in RNNs / LSTMs
torch.nn.utils.clip_grad_norm_(
list(self.encoder.parameters()) + list(self.decoder.parameters()),
self.clip)
self.optimizer.step()
return {'train_loss': loss.item()}
def valid(self, descs, desc_lens, apiseqs, api_lens):
self.encoder.eval()
self.decoder.eval()
loss = self.forward(descs, desc_lens, apiseqs, api_lens)
return {'valid_loss': loss.item()}
def sample(self, descs, desc_lens, n_samples, mode='beamsearch'):
self.encoder.eval()
self.decoder.eval()
c, hids = self.encoder(descs, desc_lens)
if mode == 'beamsearch':
sample_words, sample_lens, _ = self.decoder.beam_decode(
c, hids, None, 12, self.maxlen, n_samples)
#[batch_size x n_samples x seq_len]
sample_words, sample_lens = sample_words[0], sample_lens[0]
else:
sample_words, sample_lens = self.decoder.sampling(
c, hids, None, n_samples, self.maxlen, mode)
return sample_words, sample_lens
def adjust_lr(self):
#self.lr_scheduler_AE.step()
return None
class DA_RNN:
def __init__(self, X_dim, Y_dim, encoder_hidden_size=64, decoder_hidden_size=64,
linear_dropout=0, T=10, learning_rate=1e-5, batch_size=128, decay_rate=0.95):
self.T = T
self.decay_rate = decay_rate
self.batch_size = batch_size
self.X_dim = X_dim
self.Y_dim = Y_dim
self.encoder = Encoder(X_dim, encoder_hidden_size, T, linear_dropout).to(device)
self.decoder = Decoder(encoder_hidden_size, decoder_hidden_size, T, linear_dropout, Y_dim).to(device)
self.encoder_optim = torch.optim.Adam(params=self.encoder.parameters(), lr=learning_rate)
self.decoder_optim = torch.optim.Adam(params=self.decoder.parameters(), lr=learning_rate)
self.loss_func = torch.nn.MSELoss()
def adjust_learning_rate(self):
for enc_params, dec_params in zip(self.encoder_optim.param_groups, self.decoder_optim.param_groups):
enc_params['lr'] = enc_params['lr'] * self.decay_rate
dec_params['lr'] = dec_params['lr'] * self.decay_rate
def ToTrainingBatches(self, X, Y, shuffle_slice=True):
X_batches = []
Y_batches = []
N = X.shape[0]
batch_num = math.ceil((N-self.T)/self.batch_size)
i = self.T-1
for b in range(batch_num):
# number of output = N - T + 1
# N is length, i is an index
_batch_size = self.batch_size if N-i >= self.batch_size else N-i
X_batch = np.empty((_batch_size, self.T, self.X_dim))
Y_batch = np.empty((_batch_size, self.Y_dim))
for b_idx in range(_batch_size):
# print(N, i, i-self.T+1, i+1)
# print(X[i-self.T+1:i+1].shape)
X_batch[b_idx, :, :] = X[i-self.T+1:i+1]
Y_batch[b_idx, :] = Y[i]
i += 1
X_batches.append(X_batch)
Y_batches.append(Y_batch)
# TODO: zero padding
# print(X.shape[0], np.sum([_.shape[0] for _ in X_batches]))
if shuffle_slice:
return shuffle(X_batches, Y_batches)
else:
return X_batches, Y_batches
def ToTestingBatch(self, X):
N = X.shape[0]
X_batch = np.empty((N-self.T+1, self.T, self.X_dim))
i = self.T-1
b_idx = 0
while i < N:
X_batch[b_idx, :, :] = X[i-self.T+1:i+1]
i += 1
b_idx += 1
# TODO: zero padding
return X_batch
def train(self, X_train, Y_train, X_val, Y_val, epochs):
if len(Y_train.shape) == 1:
Y_train = Y_train[:, np.newaxis]
if len(Y_val.shape) == 1:
Y_val = Y_val[:, np.newaxis]
assert len(X_train) == len(Y_train)
assert len(X_val) == len(Y_val)
epoch_loss_hist = []
iter_loss_hist = []
N = X_train.shape[0]
for _e in range(epochs):
X_train_batches, Y_train_batches = self.ToTrainingBatches(X_train, Y_train)
for X_train_batch, Y_train_batch in zip(X_train_batches, Y_train_batches):
X_train_loss = self.train_iter(X_train_batch, Y_train_batch)
iter_loss_hist.append(np.mean(X_train_loss))
# decay learning rate
# if _e % 20 == 0:
# self.adjust_learning_rate()
epoch_loss_hist.append(iter_loss_hist[-len(X_train_batches):])
if _e % 2 == 0:
print("Epoch: {}\t".format(_e), end="")
Y_val_pred = self.predict(X_val, on_train=True)
Y_val_loss = self.loss_func(Y_val_pred, toTorch(Y_val[-(N-self.T+1):]))
print("train_loss: {:.4f} val_loss: {:.4f}".format(X_train_loss, Y_val_loss))
return epoch_loss_hist, iter_loss_hist
def train_iter(self, X, Y):
self.encoder.train(), self.decoder.train()
self.encoder_optim.zero_grad(), self.decoder_optim.zero_grad()
_, X_encoded = self.encoder(toTorch(X))
Y_pred = self.decoder(X_encoded)
loss = self.loss_func(Y_pred, toTorch(Y))
loss.backward()
self.encoder_optim.step()
self.decoder_optim.step()
return loss.item()
def predict(self, X, on_train=False):
self.encoder.eval(), self.decoder.eval()
X_batch = self.ToTestingBatch(X)
_, X_encoded = self.encoder(toTorch(X_batch))
Y_pred = self.decoder(X_encoded)
if on_train == False:
Y_pred = Y_pred.cpu().detach().numpy()
return Y_pred
class CVAE(nn.Module):
def __init__(self, config, api, PAD_token=0):
super(CVAE, self).__init__()
self.vocab = api.vocab
self.vocab_size = len(self.vocab)
self.rev_vocab = api.rev_vocab
self.go_id = self.rev_vocab["<s>"]
self.eos_id = self.rev_vocab["</s>"]
self.maxlen = config.maxlen
self.clip = config.clip
self.temp = config.temp
self.full_kl_step = config.full_kl_step
self.z_size = config.z_size
self.init_w = config.init_weight
self.softmax = nn.Softmax(dim=1)
self.embedder = nn.Embedding(self.vocab_size,
config.emb_size,
padding_idx=PAD_token)
# 对title, 每一句诗做编码
self.seq_encoder = Encoder(embedder=self.embedder,
input_size=config.emb_size,
hidden_size=config.n_hidden,
bidirectional=True,
n_layers=config.n_layers,
noise_radius=config.noise_radius)
# 先验网络的输入是 标题encode结果 + 上一句诗过encoder的结果 + 上一句情感过encoder的结果
self.prior_net = Variation(config.n_hidden * 4,
config.z_size,
dropout_rate=config.dropout,
init_weight=self.init_w)
# 后验网络，再加上x的2*hidden
# self.post_net = Variation(config.n_hidden * 6, config.z_size*2)
self.post_net = Variation(config.n_hidden * 6,
config.z_size,
dropout_rate=config.dropout,
init_weight=self.init_w)
# 词包loss的MLP
self.bow_project = nn.Sequential(
nn.Linear(config.n_hidden * 4 + config.z_size, 400),
nn.LeakyReLU(), nn.Dropout(config.dropout),
nn.Linear(400, self.vocab_size))
self.init_decoder_hidden = nn.Sequential(
nn.Linear(config.n_hidden * 4 + config.z_size, config.n_hidden),
nn.BatchNorm1d(config.n_hidden, eps=1e-05, momentum=0.1),
nn.LeakyReLU())
# self.post_generator = nn.Sequential(
# nn.Linear(config.z_size, config.z_size),
# nn.BatchNorm1d(config.z_size, eps=1e-05, momentum=0.1),
# nn.LeakyReLU(),
# nn.Linear(config.z_size, config.z_size),
# nn.BatchNorm1d(config.z_size, eps=1e-05, momentum=0.1),
# nn.LeakyReLU(),
# nn.Linear(config.z_size, config.z_size)
# )
# self.post_generator.apply(self.init_weights)
# self.prior_generator = nn.Sequential(
# nn.Linear(config.z_size, config.z_size),
# nn.BatchNorm1d(config.z_size, eps=1e-05, momentum=0.1),
# nn.ReLU(),
# nn.Dropout(config.dropout),
# nn.Linear(config.z_size, config.z_size),
# nn.BatchNorm1d(config.z_size, eps=1e-05, momentum=0.1),
# nn.ReLU(),
# nn.Dropout(config.dropout),
# nn.Linear(config.z_size, config.z_size)
# )
# self.prior_generator.apply(self.init_weights)
self.init_decoder_hidden.apply(self.init_weights)
self.bow_project.apply(self.init_weights)
self.post_net.apply(self.init_weights)
self.decoder = Decoder(embedder=self.embedder,
input_size=config.emb_size,
hidden_size=config.n_hidden,
vocab_size=self.vocab_size,
n_layers=1)
# self.optimizer_lead = optim.Adam(list(self.seq_encoder.parameters())\
# + list(self.prior_net.parameters()), lr=config.lr_lead)
self.optimizer_AE = optim.Adam(list(self.seq_encoder.parameters())\
+ list(self.prior_net.parameters())\
# + list(self.prior_generator.parameters())
+ list(self.post_net.parameters())\
# + list(self.post_generator.parameters())
+ list(self.bow_project.parameters())\
+ list(self.init_decoder_hidden.parameters())\
+ list(self.decoder.parameters()), lr=config.lr_ae)
# self.lr_scheduler_AE = optim.lr_scheduler.StepLR(self.optimizer_AE, step_size=10, gamma=0.6)
self.criterion_ce = nn.CrossEntropyLoss()
self.softmax = nn.Softmax(dim=1)
self.criterion_sent_lead = nn.CrossEntropyLoss()
def set_full_kl_step(self, kl_full_step):
self.full_kl_step = kl_full_step
def force_change_lr(self, new_init_lr_ae):
self.optimizer_AE = optim.Adam(list(self.seq_encoder.parameters()) \
+ list(self.prior_net.parameters()) \
# + list(self.prior_generator.parameters())
+ list(self.post_net.parameters()) \
# + list(self.post_generator.parameters())
+ list(self.bow_project.parameters()) \
+ list(self.init_decoder_hidden.parameters()) \
+ list(self.decoder.parameters()), lr=new_init_lr_ae)
def init_weights(self, m):
if isinstance(m, nn.Linear):
m.weight.data.uniform_(-self.init_w, self.init_w)
m.bias.data.fill_(0)
def sample_code_post(self, x, c):
# import pdb
# pdb.set_trace()
# mulogsigma = self.post_net(torch.cat((x, c), dim=1))
# mu, logsigma = torch.chunk(mulogsigma, chunks=2, dim=1)
# batch_size = c.size(0)
# std = torch.exp(0.5 * logsigma)
# epsilon = to_tensor(torch.randn([batch_size, self.z_size]))
# z = epsilon * std + mu
z, mu, logsigma = self.post_net(torch.cat(
(x, c), 1)) # 输入：(batch, 3*2*n_hidden)
# z = self.post_generator(z)
return z, mu, logsigma
def sample_code_prior(self, c, sentiment_mask=None, mask_type=None):
return self.prior_net(c,
sentiment_mask=sentiment_mask,
mask_type=mask_type)
# # input: (batch, 3)
# # target: (batch, 3)
# def criterion_sent_lead(self, input, target):
# softmax_res = self.softmax(input)
# negative_log_softmax_res = -torch.log(softmax_res + 1e-10) # (batch, 3)
# cross_entropy_loss = torch.sum(negative_log_softmax_res * target, dim=1)
# avg_cross_entropy = torch.mean(cross_entropy_loss)
# return avg_cross_entropy
# sentiment_lead: (batch, 3)
def train_AE(self,
global_t,
title,
context,
target,
target_lens,
sentiment_mask=None,
sentiment_lead=None):
self.seq_encoder.train()
self.decoder.train()
# batch_size = title.size(0)
# 每一句的情感用第二个分类器来预测，输入当前的m_hidden，输出分类结果
title_last_hidden, _ = self.seq_encoder(title)
context_last_hidden, _ = self.seq_encoder(context)
# import pdb
# pdb.set_trace()
x, _ = self.seq_encoder(target[:, 1:], target_lens - 1)
condition_prior = torch.cat((title_last_hidden, context_last_hidden),
dim=1)
z_prior, prior_mu, prior_logvar, pi, pi_final = self.sample_code_prior(
condition_prior, sentiment_mask=sentiment_mask)
z_post, post_mu, post_logvar = self.sample_code_post(
x, condition_prior)
# import pdb
# pdb.set_trace()
if sentiment_lead is not None:
self.sent_lead_loss = self.criterion_sent_lead(
input=pi, target=sentiment_lead)
else:
self.sent_lead_loss = 0
# if sentiment_lead is not None:
# self.optimizer_lead.zero_grad()
# self.sent_lead_loss.backward()
# self.optimizer_lead.step()
# return [('lead_loss', self.sent_lead_loss.item())], global_t
final_info = torch.cat((z_post, condition_prior), dim=1)
# reconstruct_loss
# import pdb
# pdb.set_trace()
output = self.decoder(init_hidden=self.init_decoder_hidden(final_info),
context=None,
inputs=target[:, :-1])
flattened_output = output.view(-1, self.vocab_size)
# flattened_output = self.softmax(flattened_output) + 1e-10
# flattened_output = torch.log(flattened_output)
dec_target = target[:, 1:].contiguous().view(-1)
mask = dec_target.gt(0) # 即判断target的token中是否有0（pad项）
masked_target = dec_target.masked_select(mask) # 选出非pad项
output_mask = mask.unsqueeze(1).expand(
mask.size(0), self.vocab_size) # [(batch_sz * seq_len) x n_tokens]
masked_output = flattened_output.masked_select(output_mask).view(
-1, self.vocab_size)
self.rc_loss = self.criterion_ce(masked_output / self.temp,
masked_target)
# kl散度
kld = gaussian_kld(post_mu, post_logvar, prior_mu, prior_logvar)
self.avg_kld = torch.mean(kld)
self.kl_weights = min(global_t / self.full_kl_step, 1.0) # 退火
self.kl_loss = self.kl_weights * self.avg_kld
# avg_bow_loss
self.bow_logits = self.bow_project(final_info)
# 说白了就是把target所有词的预测loss求个和
labels = target[:, 1:]
label_mask = torch.sign(labels).detach().float()
# 取符号变成正数，从而通过最小化来optimize
# soft_result = self.softmax(self.bow_logits) + 1e-10
# bow_loss = -torch.log(soft_result).gather(1, labels) * label_mask
bow_loss = -F.log_softmax(self.bow_logits, dim=1).gather(
1, labels) * label_mask
sum_bow_loss = torch.sum(bow_loss, 1)
self.avg_bow_loss = torch.mean(sum_bow_loss)
self.aug_elbo_loss = self.avg_bow_loss + self.kl_loss + self.rc_loss
self.total_loss = self.aug_elbo_loss + self.sent_lead_loss
# 变相增加标注集的学习率
if sentiment_mask is not None:
self.total_loss = self.total_loss * 13.33
self.optimizer_AE.zero_grad()
self.total_loss.backward()
self.optimizer_AE.step()
avg_total_loss = self.total_loss.item()
avg_lead_loss = 0 if sentiment_lead is None else self.sent_lead_loss.item(
)
avg_aug_elbo_loss = self.aug_elbo_loss.item()
avg_kl_loss = self.kl_loss.item()
avg_rc_loss = self.rc_loss.data.item()
avg_bow_loss = self.avg_bow_loss.item()
global_t += 1
return [('avg_total_loss', avg_total_loss),
('avg_lead_loss', avg_lead_loss),
('avg_aug_elbo_loss', avg_aug_elbo_loss),
('avg_kl_loss', avg_kl_loss), ('avg_rc_loss', avg_rc_loss),
('avg_bow_loss', avg_bow_loss),
('kl_weight', self.kl_weights)], global_t
def valid_AE(self,
global_t,
title,
context,
target,
target_lens,
sentiment_mask=None,
sentiment_lead=None):
self.seq_encoder.eval()
self.decoder.eval()
title_last_hidden, _ = self.seq_encoder(title)
context_last_hidden, _ = self.seq_encoder(context)
# import pdb
# pdb.set_trace()
x, _ = self.seq_encoder(target[:, 1:], target_lens - 1)
condition_prior = torch.cat((title_last_hidden, context_last_hidden),
dim=1)
z_prior, prior_mu, prior_logvar, pi, pi_final = self.sample_code_prior(
condition_prior, sentiment_mask=sentiment_mask)
z_post, post_mu, post_logvar = self.sample_code_post(
x, condition_prior)
if sentiment_lead is not None:
self.sent_lead_loss = self.criterion_sent_lead(
input=pi, target=sentiment_lead)
else:
self.sent_lead_loss = 0
# if sentiment_lead is not None:
# return [('valid_lead_loss', self.sent_lead_loss.item())], global_t
final_info = torch.cat((z_post, condition_prior), dim=1)
output = self.decoder(init_hidden=self.init_decoder_hidden(final_info),
context=None,
inputs=target[:, :-1])
flattened_output = output.view(-1, self.vocab_size)
# flattened_output = self.softmax(flattened_output) + 1e-10
# flattened_output = torch.log(flattened_output)
dec_target = target[:, 1:].contiguous().view(-1)
mask = dec_target.gt(0) # 即判断target的token中是否有0（pad项）
masked_target = dec_target.masked_select(mask) # 选出非pad项
output_mask = mask.unsqueeze(1).expand(
mask.size(0), self.vocab_size) # [(batch_sz * seq_len) x n_tokens]
masked_output = flattened_output.masked_select(output_mask).view(
-1, self.vocab_size)
self.rc_loss = self.criterion_ce(masked_output / self.temp,
masked_target)
# kl散度
kld = gaussian_kld(post_mu, post_logvar, prior_mu, prior_logvar)
self.avg_kld = torch.mean(kld)
self.kl_weights = min(global_t / self.full_kl_step, 1.0) # 退火
self.kl_loss = self.kl_weights * self.avg_kld
# avg_bow_loss
self.bow_logits = self.bow_project(final_info)
# 说白了就是把target所有词的预测loss求个和
labels = target[:, 1:]
label_mask = torch.sign(labels).detach().float()
bow_loss = -F.log_softmax(self.bow_logits, dim=1).gather(
1, labels) * label_mask
sum_bow_loss = torch.sum(bow_loss, 1)
self.avg_bow_loss = torch.mean(sum_bow_loss)
self.aug_elbo_loss = self.avg_bow_loss + self.kl_loss + self.rc_loss
avg_aug_elbo_loss = self.aug_elbo_loss.item()
avg_kl_loss = self.kl_loss.item()
avg_rc_loss = self.rc_loss.data.item()
avg_bow_loss = self.avg_bow_loss.item()
avg_lead_loss = 0 if sentiment_lead is None else self.sent_lead_loss.item(
)
return [('valid_lead_loss', avg_lead_loss),
('valid_aug_elbo_loss', avg_aug_elbo_loss),
('valid_kl_loss', avg_kl_loss), ('valid_rc_loss', avg_rc_loss),
('valid_bow_loss', avg_bow_loss)], global_t
# batch_size = 1 只输入了一个标题
# test的时候，只有先验，没有后验，更没有所谓的kl散度
def test(self, title_tensor, title_words, mask_type=None):
self.seq_encoder.eval()
self.decoder.eval()
assert title_tensor.size(0) == 1
tem = [[2, 3] + [0] * (self.maxlen - 2)]
pred_poems = []
# 过滤掉标题中的<s> </s> 0,只为了打印
title_tokens = [
self.vocab[e] for e in title_words[0].tolist()
if e not in [0, self.eos_id, self.go_id]
]
pred_poems.append(title_tokens)
for i in range(4):
tem = to_tensor(np.array(tem))
context = tem
if i == 0:
context_last_hidden, _ = self.seq_encoder(title_tensor)
else:
context_last_hidden, _ = self.seq_encoder(context)
title_last_hidden, _ = self.seq_encoder(title_tensor)
condition_prior = torch.cat(
(title_last_hidden, context_last_hidden), dim=1)
z_prior, prior_mu, prior_logvar, _, _ = self.sample_code_prior(
condition_prior, mask_type=mask_type)
final_info = torch.cat((z_prior, condition_prior), 1)
decode_words = self.decoder.testing(
init_hidden=self.init_decoder_hidden(final_info),
maxlen=self.maxlen,
go_id=self.go_id,
mode="greedy")
decode_words = decode_words[0].tolist()
if len(decode_words) >= self.maxlen:
tem = [decode_words[0:self.maxlen]]
else:
tem = [[0] * (self.maxlen - len(decode_words)) + decode_words]
pred_tokens = [
self.vocab[e] for e in decode_words[:-1]
if e != self.eos_id and e != 0 and e != self.go_id
]
pred_poems.append(pred_tokens)
gen = ""
for line in pred_poems:
cur_line = " ".join(line)
gen = gen + cur_line + '\n'
return gen
def sample(self, title, context, repeat, go_id, end_id):
self.seq_encoder.eval()
self.decoder.eval()
assert title.size(0) == 1
title_last_hidden, _ = self.seq_encoder(title)
context_last_hidden, _ = self.seq_encoder(context)
condition_prior = torch.cat((title_last_hidden, context_last_hidden),
1)
condition_prior_repeat = condition_prior.expand(repeat, -1)
z_prior_repeat, _, _, _, _ = self.sample_code_prior(
condition_prior_repeat)
final_info = torch.cat((z_prior_repeat, condition_prior_repeat), dim=1)
sample_words, sample_lens = self.decoder.sampling(
init_hidden=self.init_decoder_hidden(final_info),
maxlen=self.maxlen,
go_id=self.go_id,
eos_id=self.eos_id,
mode="greedy")
return sample_words, sample_lens
class Seq2Seq(nn.Module):
def __init__(self, config, api, pad_token=0):
super(Seq2Seq, self).__init__()
self.vocab = api.vocab
self.vocab_size = len(self.vocab)
self.rev_vocab = api.rev_vocab
self.go_id = self.rev_vocab["<s>"]
self.eos_id = self.rev_vocab["</s>"]
self.maxlen = config.maxlen
self.embedder = nn.Embedding(self.vocab_size, config.emb_size, padding_idx=pad_token)
self.encoder = Encoder(self.embedder, config.emb_size, config.n_hidden,
True, config.n_layers, config.noise_radius)
self.decoder = AttnDecoder(config=config, embedder=self.embedder, vocab_size=self.vocab_size)
self.criterion = nn.NLLLoss(reduction='none')
self.optimizer = optim.Adam(list(self.encoder.parameters())
+ list(self.decoder.parameters()),
lr=config.lr_s2s)
self.lr_scheduler_AE = optim.lr_scheduler.StepLR(self.optimizer, step_size=10, gamma=0.6)
# 将title和诗句都限制成10个字，不够了pad，超了截取
# 每次输入进来一个batch的context和target
def train_model(self, context, target, target_lens):
self.encoder.train()
self.decoder.train()
self.optimizer.zero_grad()
# (batch, 2 * n_hidden), (batch, len, 2*n_hidden)
encoder_last_hidden, encoder_output = self.encoder(context)
batch_size = encoder_last_hidden.size(0)
hidden_size = encoder_last_hidden.size(1) // 2
# (1, batch, n_hidden)
last_hidden = encoder_last_hidden.view(batch_size, 2, -1)[:, -1, :].squeeze().unsqueeze(0)
# (batch, len, n_hidden)
encoder_output = encoder_output.view(batch_size, -1, 2, hidden_size)[:, :, -1]
decoder_input = target[:, :-1] # (batch, 9)
decoder_target = target[:, 1:] # (batch, 9
step_losses = []
for i in range(self.maxlen - 1):
decoded_result, last_hidden = \
self.decoder(decoder_input=decoder_input[:, i], init_hidden=last_hidden, encoder_output=encoder_output)
step_loss = self.criterion(decoded_result, decoder_target[:, i])
step_losses.append(step_loss)
stack_loss = torch.stack(step_losses, 1) # (batch, maxlen)
sum_loss = torch.sum(stack_loss, 1) # 对每一行求和
avg_loss_batch = sum_loss / (target_lens.float() - 1) # 对每一行先求平均, decode时候每一行是9个字符
loss = torch.mean(avg_loss_batch) # 再对所有行一起求和
loss.backward()
self.optimizer.step()
return [('train_loss', loss.item())]
def valid(self, context, target, target_lens):
self.encoder.eval()
self.decoder.eval()
encoder_last_hidden, encoder_output = self.encoder(context)
batch_size = encoder_last_hidden.size(0)
hidden_size = encoder_last_hidden.size(1) // 2
# (1, batch, n_hidden)
last_hidden = encoder_last_hidden.view(batch_size, 2, -1)[:, -1, :].squeeze().unsqueeze(0)
# (batch, len, n_hidden)
encoder_output = encoder_output.view(batch_size, -1, 2, hidden_size)[:, :, -1]
decoder_input = target[:, :-1] # (batch, 9)
decoder_target = target[:, 1:] # (batch, 9
step_losses = []
for i in range(self.maxlen - 1):
decoded_result, last_hidden = \
self.decoder(decoder_input=decoder_input[:, i], init_hidden=last_hidden, encoder_output=encoder_output)
step_loss = self.criterion(decoded_result, decoder_target[:, i])
step_losses.append(step_loss)
stack_loss = torch.stack(step_losses, 1) # (batch, maxlen)
sum_loss = torch.sum(stack_loss, 1) # 对每一行求和
avg_loss_batch = sum_loss / (target_lens.float() - 1) # 对每一行先求平均, decode时候每一行是9个字符
loss = torch.mean(avg_loss_batch) # 再对所有行一起求和
return [('valid_loss', loss.item())]
def test(self, title, title_list, batch_size):
self.encoder.eval()
self.decoder.eval()
assert title.size(0) == 1
tem = title[0][0: self.maxlen].unsqueeze(0)
pred_poems = []
title_tokens = [self.vocab[e] for e in title_list[0].tolist() if e not in [0, self.eos_id, self.go_id]]
pred_poems.append(title_tokens)
for sent_id in range(4):
context = tem
if type(context) is list:
vec_context = np.zeros((batch_size, self.maxlen), dtype=np.int64)
for b_id in range(batch_size):
vec_context[b_id, :] = np.array(context[b_id])
context = to_tensor(vec_context)
encoder_last_hidden, encoder_output = self.encoder(context)
batch_size = encoder_last_hidden.size(0)
hidden_size = encoder_last_hidden.size(1) // 2
# (1, 1, n_hidden)
last_hidden = encoder_last_hidden.view(batch_size, 2, -1)[:, -1, :].unsqueeze(0)
# (batch, len, n_hidden)
encoder_output = encoder_output.view(batch_size, -1, 2, hidden_size)[:, :, -1]
# decode_words 是完整的一句诗
decode_words = self.decoder.testing(init_hidden=last_hidden, encoder_output=encoder_output,
maxlen=self.maxlen, go_id=self.go_id, mode="greedy")
decode_words = decode_words[0].tolist()
# import pdb
# pdb.set_trace()
if len(decode_words) > self.maxlen:
tem = [decode_words[0: self.maxlen]]
else:
tem = [[0] * (self.maxlen - len(decode_words)) + decode_words]
pred_tokens = [self.vocab[e] for e in decode_words[:-1] if e != self.eos_id and e != 0]
pred_poems.append(pred_tokens)
gen = ''
for line in pred_poems:
true_str = " ".join(line)
gen = gen + true_str + '\n'
return gen
def sample(self, title, context, repeat, go_id, end_id):
self.encoder.eval()
self.decoder.eval()
encoder_last_hidden, encoder_output = self.encoder(context)
batch_size = encoder_last_hidden.size(0)
hidden_size = encoder_last_hidden.size(1) // 2
# (1, batch, n_hidden)
last_hidden = encoder_last_hidden.view(batch_size, 2, -1)[:, -1].unsqueeze(0)
# (batch, len, n_hidden)
encoder_output = encoder_output.view(batch_size, -1, 2, hidden_size)[:, :, -1]
last_hidden = last_hidden.expand(1, repeat, hidden_size)
encoder_output = encoder_output.expand(repeat, -1, hidden_size)
sample_words, sample_lens = self.decoder.sampling(last_hidden, encoder_output, self.maxlen,
go_id, end_id, "greedy")
return sample_words, sample_lens
def adjust_lr(self):
self.lr_scheduler_AE.step()