def test_mask_t(self):
batsize = 5
lstm = q.LSTMCell(9, 10)
x_t = torch.randn(batsize, 9)
mask_t = torch.tensor([1, 1, 0, 1, 0])
c_tm1 = torch.randn(1, 10)
h_tm1 = torch.randn(1, 10)
lstm.c_0 = torch.tensor(c_tm1)
lstm.y_0 = torch.tensor(h_tm1)
y_t = lstm(x_t, mask_t=mask_t)
self.assertEqual((batsize, 10), y_t.detach().numpy().shape)
self.assertTrue(
np.allclose(h_tm1[0].detach().numpy(), y_t[2].detach().numpy()))
self.assertFalse(
np.allclose(h_tm1[0].detach().numpy(), y_t[1].detach().numpy()))
self.assertTrue(
np.allclose(h_tm1[0].detach().numpy(), y_t[4].detach().numpy()))
self.assertTrue(
np.allclose(h_tm1[0].detach().numpy(),
lstm.y_tm1[2].detach().numpy()))
self.assertFalse(
np.allclose(h_tm1[0].detach().numpy(),
lstm.y_tm1[1].detach().numpy()))
self.assertTrue(
np.allclose(h_tm1[0].detach().numpy(),
lstm.y_tm1[4].detach().numpy()))
def test_dropout_rec(self):
batsize = 5
lstm = q.LSTMCell(9, 10, dropout_rec=0.5)
x_t = torch.randn(batsize, 9)
c_tm1 = torch.randn(1, 10)
y_tm1 = torch.randn(1, 10)
lstm.c_0 = torch.tensor(c_tm1)
lstm.y_0 = torch.tensor(y_tm1)
y_t = lstm(x_t)
self.assertEqual((5, 10), y_t.detach().numpy().shape)
self.assertEqual(lstm.training, True)
lstm.train(False)
self.assertEqual(lstm.training, False)
q.batch_reset(lstm)
pred1 = lstm(x_t)
q.batch_reset(lstm)
pred2 = lstm(x_t)
self.assertTrue(
np.allclose(pred1.detach().numpy(),
pred2.detach().numpy()))
lstm.train(True)
self.assertEqual(lstm.training, True)
q.batch_reset(lstm)
pred1 = lstm(x_t)
q.batch_reset(lstm)
pred2 = lstm(x_t)
self.assertFalse(
np.allclose(pred1.detach().numpy(),
pred2.detach().numpy()))
def test_zoneout(self):
batsize = 5
q.LSTMCell.debug = False
lstm = q.LSTMCell(9, 10, zoneout=0.5)
x_t = Variable(torch.FloatTensor(np.random.random((batsize, 9))))
h_tm1 = Variable(torch.FloatTensor(np.random.random((batsize, 10))))
c_tm1 = Variable(torch.FloatTensor(np.random.random((batsize, 10))))
lstm.set_init_states(c_tm1, h_tm1)
y_t = lstm(x_t)
self.assertEqual((5, 10), y_t.data.numpy().shape)
self.assertEqual(lstm.training, True)
lstm.train(mode=False)
self.assertEqual(lstm.training, False)
lstm.reset_state()
pred1 = lstm(x_t)
lstm.reset_state()
pred2 = lstm(x_t)
# must be equal in prediction mode
self.assertTrue(np.allclose(pred1.data.numpy(), pred2.data.numpy()))
lstm.train(mode=True)
self.assertEqual(lstm.training, True)
lstm.reset_state()
pred1 = lstm(x_t)
lstm.reset_state()
pred2 = lstm(x_t)
# must not be equal in prediction mode
self.assertFalse(np.allclose(pred1.data.numpy(), pred2.data.numpy()))
def test_lstm_layer_shapes(self):
batsize = 5
seqlen = 7
q.LSTMCell.debug = False
lstm = q.LSTMCell(9, 10)
lstm = lstm.to_layer().return_all()
x = Variable(torch.FloatTensor(np.random.random((batsize, seqlen, 9))))
y = lstm(x)
self.assertEqual((batsize, seqlen, 10), y.data.numpy().shape)
def test_lstm_shapes_non_cudnn(self):
batsize = 5
lstm = q.LSTMCell(9, 10, use_cudnn_cell=False)
x_t = Variable(torch.FloatTensor(np.random.random((batsize, 9))))
h_tm1 = Variable(torch.FloatTensor(np.random.random((batsize, 10))))
c_tm1 = Variable(torch.FloatTensor(np.random.random((batsize, 10))))
lstm.set_init_states(c_tm1, h_tm1)
y_t = lstm(x_t)
self.assertEqual((5, 10), y_t.data.numpy().shape)
def test_lstm_shapes(self):
batsize = 5
lstm = q.LSTMCell(9, 10)
x_t = torch.randn(batsize, 9)
c_tm1 = torch.randn(1, 10)
y_tm1 = torch.randn(1, 10)
lstm.c_0 = torch.tensor(c_tm1)
lstm.y_0 = torch.tensor(y_tm1)
y_t = lstm(x_t)
self.assertEqual((5, 10), y_t.detach().numpy().shape)
self.assertTrue(
np.allclose(lstm.y_tm1.detach().numpy(),
y_t.detach().numpy()))
q.batch_reset(lstm)
def test_it(self):
x = np.random.randint(0, 100, (1000, 7))
y_inp = x[:, :-1]
y_out = x[:, 1:]
wD = dict((chr(xi), xi) for xi in range(100))
ctx = torch.randn(1000, 8, 30)
decoder_emb = q.WordEmb(20, worddic=wD)
decoder_lstm = q.LSTMCell(20, 30)
decoder_att = q.DotAttention()
decoder_out = q.WordLinout(60, worddic=wD)
decoder_cell = q.DecoderCell(decoder_emb, decoder_lstm, decoder_att,
None, decoder_out)
decoder_tf = q.TFDecoder(decoder_cell)
y = decoder_tf(torch.tensor(x), ctx=ctx)
self.assertTrue(y.size(), (1000, 7, 100))
# endregion
def test_it(self):
D = "<MASK> [RED] NT(START) NT(a) T(b) NT(c) T(d) T(e) NT(f) T(g) T(h) T(i)"
D = dict(zip(D.split(), range(len(D.split()))))
tok2act = {
k: (2 if k == "[RED]" else 1 if k[:2] == "NT" else 0)
for k in D
}
class CustomCombiner(StackCellCombiner):
def forward(self, _x, mask):
ret = (_x * mask.unsqueeze(-1).float()).sum(
1) / mask.float().sum(1).unsqueeze(-1).clamp_min(1e-6)
ret = ret.detach() # TODO: for grad debugging
return ret
class CustomWordLinout(q.WordLinout):
def update(self, _):
pass
class Tok2Act(torch.nn.Module):
def __init__(self, t2a, D):
super(Tok2Act, self).__init__()
self.D = D
t2a_ = torch.zeros(max(D.values()) + 1).long()
for k, v in t2a.items():
t2a_[D[k]] = v
self.register_buffer("t2a", t2a_)
def forward(self, _x):
return self.t2a[_x]
embdim = 4
coredim = 5
emb = q.WordEmb(embdim, worddic=D)
core = q.LSTMCell(embdim, coredim, dropout_rec=.1)
# combiner = BasicCombiner(embdim)
combiner = CustomCombiner()
att = BasicAttention()
out = CustomWordLinout(coredim * 2, worddic=D)
tok2act = Tok2Act(tok2act, D)
cell = StackCell(emb=emb,
tok2act=tok2act,
core=core,
combiner=combiner,
att=att,
out=out)
ctx = torch.randn(2, 6, coredim)
cell.save_ctx(ctx)
ex1 = "NT(START) NT(a) T(b) NT(c) T(d) T(e) [RED] NT(f) T(g) T(h) [RED] T(i) [RED]"
ex2 = "NT(START) NT(a) NT(c) T(d) T(e) [RED] [RED]"
x1 = [D[exi] for exi in ex1.split()] + [0]
x2 = [D[exi] for exi in ex2.split()]
x2 = x2 + [0] * (len(x1) - len(x2))
x = torch.tensor([x1, x2])
cell._debug_embs = torch.nn.Parameter(torch.zeros(2, len(x1), embdim))
ys = []
for i in range(len(x[0])):
y = cell(x[:, i])
ys.append(y)
# print(cell._debug_embs)
print(cell._debug_embs.size())
l = ys[11][0].sum()
l.backward()
print(cell._debug_embs.grad)
print(cell._stack)
def run_seq2seq_(
lr=0.001,
batsize=32,
evalbatsize=256,
epochs=100,
warmup=5,
embdim=50,
encdim=100,
numlayers=2,
dropout=.0,
wreg=1e-6,
cuda=False,
gpu=0,
):
settings = locals().copy()
device = torch.device("cpu") if not cuda else torch.device("cuda", gpu)
tt = q.ticktock("script")
tt.msg("running seq2seq on LC-QuAD")
tt.tick("loading data")
xsm, ysm, teststart, tok2act = load_data()
_tok2act = {ysm.RD[k]: v for k, v in tok2act.items()}
print("Some examples:")
for i in range(5):
print(
f"{xsm[i]}\n ->{ysm[i]}\n -> {Node.from_transitions(' '.join(ysm[i].split()[1:]), _tok2act)}"
)
print("Non-leaf tokens:")
print({ysm.RD[k]: v for k, v in tok2act.items() if v > 0})
devstart = teststart - 500
trainds = torch.utils.data.TensorDataset(
torch.tensor(xsm.matrix[:devstart]).long(),
torch.tensor(ysm.matrix[:devstart, :-1]).long(),
torch.tensor(ysm.matrix[:devstart, 1:]).long())
valds = torch.utils.data.TensorDataset(
torch.tensor(xsm.matrix[devstart:teststart]).long(),
torch.tensor(ysm.matrix[devstart:teststart, :-1]).long(),
torch.tensor(ysm.matrix[devstart:teststart, 1:]).long())
testds = torch.utils.data.TensorDataset(
torch.tensor(xsm.matrix[teststart:]).long(),
torch.tensor(ysm.matrix[teststart:, :-1]).long(),
torch.tensor(ysm.matrix[teststart:, 1:]).long())
tt.msg(
f"Data splits: train: {len(trainds)}, valid: {len(valds)}, test: {len(testds)}"
)
tloader = torch.utils.data.DataLoader(trainds,
batch_size=batsize,
shuffle=True)
vloader = torch.utils.data.DataLoader(valds,
batch_size=evalbatsize,
shuffle=False)
xloader = torch.utils.data.DataLoader(testds,
batch_size=evalbatsize,
shuffle=False)
tt.tock("data loaded")
# model
enclayers, declayers = numlayers, numlayers
decdim = encdim
xemb = q.WordEmb(embdim, worddic=xsm.D)
yemb = q.WordEmb(embdim, worddic=ysm.D)
encdims = [embdim] + [encdim // 2] * enclayers
xenc = q.LSTMEncoder(embdim,
*encdims[1:],
bidir=True,
dropout_in_shared=dropout)
decdims = [embdim] + [decdim] * declayers
dec_core = torch.nn.Sequential(*[
q.LSTMCell(decdims[i - 1],
decdims[i],
dropout_in=dropout,
dropout_rec=dropout) for i in range(1, len(decdims))
])
yout = q.WordLinout(encdim + decdim, worddic=ysm.D)
dec_cell = semparse.rnn.LuongCell(emb=yemb,
core=dec_core,
out=yout,
dropout=dropout)
decoder = q.TFDecoder(dec_cell)
testdecoder = q.FreeDecoder(dec_cell, maxtime=100)
m = Seq2Seq(xemb, xenc, decoder)
testm = Seq2Seq(xemb, xenc, testdecoder, test=True)
# test model
tt.tick("running a batch")
test_y = m(*iter(tloader).next()[:-1])
q.batch_reset(m)
test_y = testm(*iter(vloader).next()[:-1])
q.batch_reset(m)
tt.tock(f"ran a batch: {test_y.size()}")
optim = torch.optim.Adam(m.parameters(), lr=lr, weight_decay=wreg)
tlosses = [
q.CELoss(mode="logits", ignore_index=0),
q.Accuracy(ignore_index=0),
q.SeqAccuracy(ignore_index=0)
]
xlosses = [
q.CELoss(mode="logits", ignore_index=0),
q.Accuracy(ignore_index=0),
q.SeqAccuracy(ignore_index=0)
]
tlosses = [q.LossWrapper(l) for l in tlosses]
vlosses = [q.LossWrapper(l) for l in xlosses]
xlosses = [q.LossWrapper(l) for l in xlosses]
trainloop = partial(q.train_epoch,
model=m,
dataloader=tloader,
optim=optim,
losses=tlosses,
device=device)
devloop = partial(q.test_epoch,
model=testm,
dataloader=vloader,
losses=vlosses,
device=device)
testloop = partial(q.test_epoch,
model=testm,
dataloader=xloader,
losses=xlosses,
device=device)
lrplateau = q.util.ReduceLROnPlateau(optim,
mode="max",
factor=.1,
patience=3,
cooldown=1,
warmup=warmup,
threshold=0.,
verbose=True,
eps=1e-9)
on_after_valid = [lambda: lrplateau.step(vlosses[1].get_epoch_error())]
_devloop = partial(devloop, on_end=on_after_valid)
stoptrain = [lambda: all([pg["lr"] <= 1e-7 for pg in optim.param_groups])]
tt.tick("training")
q.run_training(trainloop,
_devloop,
max_epochs=epochs,
check_stop=stoptrain)
tt.tock("done training")
tt.tick("testing")
testres = testloop()
print(testres)
settings["testres"] = testres
tt.tock("tested")
devres = devloop()
print(devres, vlosses[0].get_epoch_error())
return vlosses[1].get_epoch_error()
