def test_monokuma_no_cond(self):
mono = model.Monokuma(self.vocab_size,
self.dim,
self.num_layers,
self.dropout,
is_conditional=False,
bidirectional_encoder=False,
use_bridge=False,
use_attention=False)
avgCE = nn.CrossEntropyLoss(ignore_index=0)
mono.train()
subblock = self.tgt[0:2, :]
golds = self.tgt[1:3, :].view(-1)
mono.zero_grad()
output, attns = mono(subblock, src=None, lengths=None, start=True)
self.assertEqual(attns, None)
loss = avgCE(output, golds)
loss.backward()
nn.utils.clip_grad_norm_(mono.parameters(), 0.25)
for p in mono.parameters():
p.data.add_(-10, p.grad.data)
subblock = self.tgt[2:3, :]
golds = self.tgt[3:, :].view(-1)
output, attns = mono(subblock, src=None, lengths=None, start=False)
self.assertEqual(attns, None)
loss = avgCE(output, golds)
loss.backward()
nn.utils.clip_grad_norm_(mono.parameters(), 0.25)
for p in mono.parameters():
p.data.add_(-10, p.grad.data)
def test_epoch_continuous_data(self):
mono = model.Monokuma(len(self.dat_cont.i2w),
self.dim,
self.num_layers,
self.dropout,
is_conditional=False,
bidirectional_encoder=False,
use_bridge=False,
use_attention=False)
junko = control.Junko(mono, self.lr, self.bptt, self.interval)
val_loss, _, epoch_time = junko.epoch_continuous_data(self.dat_cont, 0)
def test_epoch_translation_data_simple(self):
mono = model.Monokuma(len(self.dat_tran.i2w),
self.dim,
self.num_layers,
self.dropout,
is_conditional=True,
bidirectional_encoder=False,
use_bridge=False,
use_attention=False)
junko = control.Junko(mono, self.lr, self.bptt, self.interval)
val_loss, _, epoch_time = junko.epoch_translation_data(self.dat_tran,
0,
shuffle=False)
def test_monokuma_attn(self):
mono = model.Monokuma(self.vocab_size,
self.dim,
self.num_layers,
self.dropout,
is_conditional=True,
bidirectional_encoder=False,
use_bridge=False,
use_attention=True)
avgCE = nn.CrossEntropyLoss(ignore_index=0)
mono.train()
# src
# 1 6 9
# 2 7 10
# 3 8 0
# 4 0 0
# 5 0 0
#
#
# subblock
# 9 3 1 golds
# 8 2 100 8 2 100
# 7 100 0
subblock = self.tgt[0:2, :]
golds = self.tgt[1:3, :].view(-1)
mono.zero_grad()
output, attns = mono(subblock,
src=self.src,
lengths=self.lengths,
start=True)
loss = avgCE(output, golds)
loss.backward()
nn.utils.clip_grad_norm_(mono.parameters(), 0.25)
for p in mono.parameters():
p.data.add_(-10, p.grad.data)
# src
# 1 6 9
# 2 7 10
# 3 8 0
# 4 0 0
# 5 0 0
#
#
# subblock
# 7 100 0 golds
# # 100 0 0
subblock = self.tgt[2:3, :]
golds = self.tgt[3:, :].view(-1)
output, attns = mono(subblock,
src=self.src,
lengths=self.lengths,
start=False)
loss = avgCE(output, golds)
loss.backward()
nn.utils.clip_grad_norm_(mono.parameters(), 0.25)
for p in mono.parameters():
p.data.add_(-10, p.grad.data)