hidden_size=[256, 64, 16, 4, 2])
# push to cuda if cudnn is available
if (torch.backends.cudnn.version() != None and USE_CUDA == True):
encoder_train = encoder_train.cuda()
# print the initialized architectures
now = datetime.utcnow().strftime("%Y%m%d-%H:%M:%S")
print('[LOG {}] encoder-generator architecture:\n\n{}\n'.format(
now, encoder_train))
# Decoder network instantiation
# init training network classes / architectures
decoder_train = Decoder(output_size=ori_subset_transformed.shape[1],
hidden_size=[2, 4, 16, 64, 256])
# push to cuda if cudnn is available
if (torch.backends.cudnn.version() != None) and (USE_CUDA == True):
decoder_train = decoder_train.cuda()
# print the initialized architectures
now = datetime.utcnow().strftime("%Y%m%d-%H:%M:%S")
print('[LOG {}] decoder architecture:\n\n{}\n'.format(now, decoder_train))
# Discriminator network
# init training network classes / architectures
discriminator_train = Discriminator(input_size=2,
hidden_size=[256, 16, 4, 2],
output_size=1)
# push to cuda if cudnn is available
if (torch.backends.cudnn.version() != None) and (USE_CUDA == True):
discriminator_train = discriminator_train.cuda()
# print the initialized architectures
now = datetime.utcnow().strftime("%Y%m%d-%H:%M:%S")
print('[LOG {}] discriminator architecture:\n\n{}\n'.format(
args.num_symbols, args.min_count, args.lower,
args.enable_cuda)
if args.enc_type.lower() == "transformer":
encoder = TransformerEncoder(args.dim, corpus.vocab_size_e,
corpus.max_pos, enable_cuda)
else:
encoder = Encoder(args.dim, corpus.vocab_size_e, corpus.max_pos,
args.enc_type, enable_cuda)
valid = corpus.load_data(args.english_valid, args.french_valid)
eos = corpus.dict_f.word2index["</s>"]
decoder = Decoder(args.dim, corpus.vocab_size_f, eos,
corpus.longest_english, args.dec_type, args.attention)
if enable_cuda:
encoder.cuda()
decoder.cuda()
# Train
losses, bleus = train(corpus, valid, encoder, decoder, args.lr,
args.epochs, args.batch_size, enable_cuda,
args.tf_ratio, args.max_length)
# Plot losses and save figure
plt.figure(figsize=(15, 10))
plt.plot([i for i in range(len(losses))], losses)
plt.scatter([i for i in range(len(losses))], losses)
plt.savefig("loss_enc={}_dec={}_att={}.png".format(args.enc_type,
args.dec_type,
args.attention))
plt.figure(figsize=(15, 10))
class Mem2SeqRunner(ExperimentRunnerBase):
def __init__(self, args):
super(Mem2SeqRunner, self).__init__(args)
# Model parameters
self.gru_size = 128
self.emb_size = 128
#TODO: Try hops 4 with task 3
self.hops = 3
self.dropout = 0.2
self.encoder = Encoder(self.hops, self.nwords, self.gru_size)
self.decoder = Decoder(self.emb_size, self.hops, self.gru_size,
self.nwords)
self.optim_enc = torch.optim.Adam(self.encoder.parameters(), lr=0.001)
self.optim_dec = torch.optim.Adam(self.decoder.parameters(), lr=0.001)
if self.loss_weighting:
self.optim_loss_weights = torch.optim.Adam([self.loss_weights],
lr=0.0001)
self.scheduler = lr_scheduler.ReduceLROnPlateau(self.optim_dec,
mode='max',
factor=0.5,
patience=1,
min_lr=0.0001,
verbose=True)
if self.use_cuda:
self.cross_entropy = self.cross_entropy.cuda()
self.encoder = self.encoder.cuda()
self.decoder = self.decoder.cuda()
if self.loss_weighting:
self.loss_weights = self.loss_weights.cuda()
def train_batch_wrapper(self, batch, new_epoch, clip_grads):
context = batch[0].transpose(0, 1)
responses = batch[1].transpose(0, 1)
index = batch[2].transpose(0, 1)
sentinel = batch[3].transpose(0, 1)
context_lengths = batch[4]
target_lengths = batch[5]
return self.train_batch(context, responses, index, sentinel, new_epoch,
context_lengths, target_lengths, clip_grads)
def train_batch(self, context, responses, index, sentinel, new_epoch,
context_lengths, target_lengths, clip_grads):
# (TODO): remove transpose
if new_epoch: # (TODO): Change this part
self.loss = 0
self.ploss = 0
self.vloss = 0
self.n = 1
context = context.type(self.TYPE)
responses = responses.type(self.TYPE)
index = index.type(self.TYPE)
sentinel = sentinel.type(self.TYPE)
self.optim_enc.zero_grad()
self.optim_dec.zero_grad()
if self.loss_weighting:
self.optim_loss_weights.zero_grad()
h = self.encoder(context.transpose(0, 1))
self.decoder.load_memory(context.transpose(0, 1))
y = torch.from_numpy(np.array([2] * context.size(1),
dtype=int)).type(self.TYPE)
y_len = 0
h = h.unsqueeze(0)
output_vocab = torch.zeros(max(target_lengths), context.size(1),
self.nwords)
output_ptr = torch.zeros(max(target_lengths), context.size(1),
context.size(0))
if self.use_cuda:
output_vocab = output_vocab.cuda()
output_ptr = output_ptr.cuda()
while y_len < responses.size(0): # TODO: Add EOS condition
p_ptr, p_vocab, h = self.decoder(context, y, h)
output_vocab[y_len] = p_vocab
output_ptr[y_len] = p_ptr
#TODO: Add teqacher forcing ratio
y = responses[y_len].type(self.TYPE)
y_len += 1
# print(loss)
mask_v = torch.ones(output_vocab.size())
mask_p = torch.ones(output_ptr.size())
if self.use_cuda:
mask_p = mask_p.cuda()
mask_v = mask_v.cuda()
for i in range(responses.size(1)):
mask_v[target_lengths[i]:, i, :] = 0
mask_p[target_lengths[i]:, i, :] = 0
loss_v = self.cross_entropy(
output_vocab.contiguous().view(-1, self.nwords),
responses.contiguous().view(-1))
loss_ptr = self.cross_entropy(
output_ptr.contiguous().view(-1, context.size(0)),
index.contiguous().view(-1))
if self.loss_weighting:
loss = loss_ptr/(2*self.loss_weights[0]*self.loss_weights[0]) + loss_v/(2*self.loss_weights[1]*self.loss_weights[1]) + \
torch.log(self.loss_weights[0] * self.loss_weights[1])
loss_ptr = loss_ptr / (2 * self.loss_weights[0] *
self.loss_weights[0])
loss_v = loss_v / (2 * self.loss_weights[1] * self.loss_weights[1])
else:
loss = loss_ptr + loss_v
loss.backward()
ec = torch.nn.utils.clip_grad_norm_(self.encoder.parameters(), 10.0)
dc = torch.nn.utils.clip_grad_norm_(self.decoder.parameters(), 10.0)
self.optim_enc.step()
self.optim_dec.step()
if self.loss_weighting:
self.optim_loss_weights.step()
self.loss += loss.item()
self.vloss += loss_v.item()
self.ploss += loss_ptr.item()
return loss.item(), loss_v.item(), loss_ptr.item()
def evaluate_batch(self,
batch_size,
input_batches,
input_lengths,
target_batches,
target_lengths,
target_index,
target_gate,
src_plain,
profile_memory=None):
# Set to not-training mode to disable dropout
self.encoder.train(False)
self.decoder.train(False)
# Run words through encoder
decoder_hidden = self.encoder(input_batches.transpose(0,
1)).unsqueeze(0)
self.decoder.load_memory(input_batches.transpose(0, 1))
# Prepare input and output variables
decoder_input = Variable(torch.LongTensor([2] * batch_size))
decoded_words = []
all_decoder_outputs_vocab = Variable(
torch.zeros(max(target_lengths), batch_size, self.nwords))
all_decoder_outputs_ptr = Variable(
torch.zeros(max(target_lengths), batch_size,
input_batches.size(0)))
# all_decoder_outputs_gate = Variable(torch.zeros(self.max_r, batch_size))
# Move new Variables to CUDA
if self.use_cuda:
all_decoder_outputs_vocab = all_decoder_outputs_vocab.cuda()
all_decoder_outputs_ptr = all_decoder_outputs_ptr.cuda()
# all_decoder_outputs_gate = all_decoder_outputs_gate.cuda()
decoder_input = decoder_input.cuda()
p = []
for elm in src_plain:
elm_temp = [word_triple[0] for word_triple in elm]
p.append(elm_temp)
self.from_whichs = []
acc_gate, acc_ptr, acc_vac = 0.0, 0.0, 0.0
# Run through decoder one time step at a time
for t in range(max(target_lengths)):
decoder_ptr, decoder_vacab, decoder_hidden = self.decoder(
input_batches, decoder_input, decoder_hidden)
all_decoder_outputs_vocab[t] = decoder_vacab
topv, topvi = decoder_vacab.data.topk(1)
all_decoder_outputs_ptr[t] = decoder_ptr
topp, toppi = decoder_ptr.data.topk(1)
top_ptr_i = torch.gather(input_batches[:, :, 0], 0,
Variable(toppi.view(1,
-1))).transpose(0, 1)
next_in = [
top_ptr_i[i].item() if
(toppi[i].item() < input_lengths[i] - 1) else topvi[i].item()
for i in range(batch_size)
]
# if next_in in self.kb_entry.keys():
# ptr_distr.append([next_in, decoder_vacab.data])
decoder_input = Variable(
torch.LongTensor(next_in)) # Chosen word is next input
if self.use_cuda: decoder_input = decoder_input.cuda()
temp = []
from_which = []
for i in range(batch_size):
if (toppi[i].item() < len(p[i]) - 1):
temp.append(p[i][toppi[i].item()])
from_which.append('p')
else:
if target_index[t][i] != toppi[i].item():
self.incorrect_sentinel += 1
ind = topvi[i].item()
if ind == 3:
temp.append('<eos>')
else:
temp.append(self.i2w[ind])
from_which.append('v')
decoded_words.append(temp)
self.from_whichs.append(from_which)
self.from_whichs = np.array(self.from_whichs)
loss_v = self.cross_entropy(
all_decoder_outputs_vocab.contiguous().view(-1, self.nwords),
target_batches.contiguous().view(-1))
loss_ptr = self.cross_entropy(
all_decoder_outputs_ptr.contiguous().view(-1,
input_batches.size(0)),
target_index.contiguous().view(-1))
if self.loss_weighting:
loss = loss_ptr/(2*self.loss_weights[0]*self.loss_weights[0]) + loss_v/(2*self.loss_weights[1]*self.loss_weights[1]) + \
torch.log(self.loss_weights[0] * self.loss_weights[1])
else:
loss = loss_ptr + loss_v
self.loss += loss.item()
self.vloss += loss_v.item()
self.ploss += loss_ptr.item()
self.n += 1
# Set back to training mode
self.encoder.train(True)
self.decoder.train(True)
return decoded_words, self.from_whichs # , acc_ptr, acc_vac
def save_models(self, path):
torch.save(self.encoder.state_dict(),
os.path.join(path, 'encoder.pth'))
torch.save(self.decoder.state_dict(),
os.path.join(path, 'decoder.pth'))
def load_models(self, path: str = '.'):
self.encoder.load_state_dict(
torch.load(os.path.join(path, 'encoder.pth')))
self.decoder.load_state_dict(
torch.load(os.path.join(path, 'decoder.pth')))
decoder_bytes = urllib.request.urlopen(decoder_model_name)
# Load tensor from io.BytesIO object
encoder_buffer = io.BytesIO(encoder_bytes.read())
decoder_buffer = io.BytesIO(decoder_bytes.read())
# init training network classes / architectures
encoder_eval = Encoder(input_size=ori_subset_transformed.shape[1],
hidden_size=[256, 64, 16, 4, 2])
decoder_eval = Decoder(output_size=ori_subset_transformed.shape[1],
hidden_size=[2, 4, 16, 64, 256])
# push to cuda if cudnn is available
if (torch.backends.cudnn.version() != None) and (USE_CUDA == True):
encoder_eval = encoder_eval.cuda()
decoder_eval = decoder_eval.cuda()
# load trained models
# since the model was trained on a gpu and will be restored in a cpu we need to provide: map_location = 'cpu'
encoder_eval.load_state_dict(torch.load(encoder_buffer, map_location='cpu'))
decoder_eval.load_state_dict(torch.load(decoder_buffer, map_location='cpu'))
## specify a dataloader that provides the ability to evaluate the journal entrie in an "unshuffled" batch-wise manner:
# convert pre-processed data to pytorch tensor
torch_dataset = torch.from_numpy(ori_subset_transformed.values).float()
# convert to pytorch tensor - none cuda enabled
dataloader_eval = DataLoader(torch_dataset,
batch_size=mini_batch_size,
shuffle=False,
num_workers=0)
momentum_rate = float(cfg.get('params', 'momentum_rate'))
encoder = Encoder(wv_size, e_hidden_size)
decoder = Decoder(2 * e_hidden_size, wv_size, d_hidden_size, len(ch_index),
d_linear_size)
try:
encoder.load_state_dict(torch.load('model/encoder.params.pkl'))
decoder.load_state_dict(torch.load('model/decoder.params.pkl'))
print('load')
except:
pass
if use_cuda:
encoder = encoder.cuda()
decoder = decoder.cuda()
criterion = nn.NLLLoss()
encoder_optimizer = optim.SGD(encoder.parameters(),
lr=encoder_lr,
momentum=momentum_rate)
decoder_optimizer = optim.SGD(decoder.parameters(),
lr=decoder_lr,
momentum=momentum_rate)
#encoder_optimizer = optim.Adam(encoder.parameters(), lr=0.001)
#decoder_optimizer = optim.Adam(decoder.parameters(), lr=0.001)
def train(keywords, sentences):
poem_type = len(sentences[0]) - 1
