def run(lr=0.001):
x = np.random.random((1000, 5)).astype("float32")
y = np.random.randint(0, 5, (1000, )).astype("int64")
trainloader = q.dataload(x[:800], y[:800], batch_size=100)
validloader = q.dataload(x[800:], y[800:], batch_size=100)
m = torch.nn.Sequential(torch.nn.Linear(5, 100), torch.nn.Linear(100, 5))
m[1].weight.requires_grad = False
losses = q.lossarray(torch.nn.CrossEntropyLoss())
params = m.parameters()
for param in params:
print(param.requires_grad)
init_val = m[1].weight.detach().numpy()
optim = torch.optim.Adam(q.params_of(m), lr=lr)
trainer = q.trainer(m).on(trainloader).loss(losses).optimizer(
optim).epochs(100)
# for b, (i, e) in trainer.inf_batches():
# print(i, e)
validator = q.tester(m).on(validloader).loss(losses)
q.train(trainer, validator).run()
new_val = m[1].weight.detach().numpy()
print(np.linalg.norm(new_val - init_val))
def run(lr=20.,
dropout=0.2,
dropconnect=0.2,
gradnorm=0.25,
epochs=25,
embdim=200,
encdim=200,
numlayers=2,
seqlen=35,
batsize=20,
eval_batsize=10,
cuda=False,
gpu=0,
test=False):
tt = q.ticktock("script")
device = torch.device("cpu")
if cuda:
device = torch.device("cuda", gpu)
tt.tick("loading data")
train_batches, valid_batches, test_batches, D = \
load_data(batsize=batsize, eval_batsize=eval_batsize, seqlen=seqlen)
tt.tock("data loaded")
print("{} batches in train".format(len(train_batches)))
tt.tick("creating model")
dims = [embdim] + ([encdim] * numlayers)
m = RNNLayer_LM(*dims, worddic=D, dropout=dropout)
if test:
for i, batch in enumerate(train_batches):
y = m(batch[0])
if i > 5:
break
print(y.size())
loss = q.SeqKLLoss(time_average=True, size_average=True, mode="logits")
ppl_loss = q.SeqPPL_Loss(time_average=True,
size_average=True,
mode="logits")
optim = torch.optim.SGD(q.params_of(m), lr=lr)
gradclip = q.ClipGradNorm(gradnorm)
trainer = q.trainer(m).on(train_batches).loss(loss).optimizer(
optim).device(device).hook(m).hook(gradclip)
tester = q.tester(m).on(valid_batches).loss(
loss, ppl_loss).device(device).hook(m)
tt.tock("created model")
tt.tick("training")
q.train(trainer, tester).run(epochs=epochs)
tt.tock("trained")
def run_classify(lr=0.001,
seqlen=6,
numex=500,
epochs=25,
batsize=10,
test=True,
cuda=False,
gpu=0):
device = torch.device("cpu")
if cuda:
device = torch.device("cuda", gpu)
# region construct data
colors = "red blue green magenta cyan orange yellow grey salmon pink purple teal".split(
)
D = dict(zip(colors, range(len(colors))))
inpseqs = []
targets = []
for i in range(numex):
inpseq = list(np.random.choice(colors, seqlen, replace=False))
target = np.random.choice(range(len(inpseq)), 1)[0]
target_class = D[inpseq[target]]
inpseq[target] = "${}$".format(inpseq[target])
inpseqs.append("".join(inpseq))
targets.append(target_class)
sm = q.StringMatrix()
sm.tokenize = lambda x: list(x)
for inpseq in inpseqs:
sm.add(inpseq)
sm.finalize()
print(sm[0])
print(sm.D)
targets = np.asarray(targets)
data = q.dataload(sm.matrix[:-100], targets[:-100], batch_size=batsize)
valid_data = q.dataload(sm.matrix[-100:],
targets[-100:],
batch_size=batsize)
# endregion
# region model
embdim = 20
enc2inpdim = 45
encdim = 20
outdim = 20
emb = q.WordEmb(embdim, worddic=sm.D) # sm dictionary (characters)
out = q.WordLinout(outdim, worddic=D) # target dictionary
# encoders:
enc1 = q.RNNEncoder(embdim, encdim, bidir=True)
enc2 = q.RNNCellEncoder(enc2inpdim, outdim // 2, bidir=True)
# model
class Model(torch.nn.Module):
def __init__(self, dim, _emb, _out, _enc1, _enc2, **kw):
super(Model, self).__init__(**kw)
self.dim, self.emb, self.out, self.enc1, self.enc2 = dim, _emb, _out, _enc1, _enc2
self.score = torch.nn.Sequential(
torch.nn.Linear(dim, 1, bias=False), torch.nn.Sigmoid())
self.emb_expander = ExpandVecs(embdim, enc2inpdim, 2)
self.enc_expander = ExpandVecs(encdim * 2, enc2inpdim, 2)
def forward(self, x, with_att=False):
# embed and encode
xemb, xmask = self.emb(x)
xenc = self.enc1(xemb, mask=xmask)
# compute attention
xatt = self.score(xenc).squeeze(
2) * xmask.float()[:, :xenc.size(1)]
# encode again
_xemb = self.emb_expander(xemb[:, :xenc.size(1)])
_xenc = self.enc_expander(xenc)
_, xenc2 = self.enc2(_xemb,
gate=xatt,
mask=xmask[:, :xenc.size(1)],
ret_states=True)
scores = self.out(xenc2.view(xenc.size(0), -1))
if with_att:
return scores, xatt
else:
return scores
model = Model(40, emb, out, enc1, enc2)
# endregion
# region test
if test:
inps = torch.tensor(sm.matrix[0:2])
outs = model(inps)
# endregion
# region train
optimizer = torch.optim.Adam(q.params_of(model), lr=lr)
trainer = q.trainer(model).on(data).loss(torch.nn.CrossEntropyLoss(), q.Accuracy())\
.optimizer(optimizer).hook(q.ClipGradNorm(5.)).device(device)
validator = q.tester(model).on(valid_data).loss(
q.Accuracy()).device(device)
q.train(trainer, validator).run(epochs=epochs)
# endregion
# region check attention #TODO
# feed a batch
inpd = torch.tensor(sm.matrix[400:410])
outd, att = model(inpd, with_att=True)
outd = torch.max(outd, 1)[1].cpu().detach().numpy()
inpd = inpd.cpu().detach().numpy()
att = att.cpu().detach().numpy()
rD = {v: k for k, v in sm.D.items()}
roD = {v: k for k, v in D.items()}
for i in range(len(att)):
inpdi = " ".join([rD[x] for x in inpd[i]])
outdi = roD[outd[i]]
print("input: {}\nattention: {}\nprediction: {}".format(
inpdi, " ".join(["{:.1f}".format(x) for x in att[i]]), outdi))
def run(
lr=0.0001,
batsize=64,
epochs=100000,
lamda=10,
disciters=5,
burnin=-1,
validinter=1000,
devinter=100,
cuda=False,
gpu=0,
z_dim=128,
test=False,
dim_d=128,
dim_g=128,
):
settings = locals().copy()
logger = q.log.Logger(prefix="wgan_resnet_cifar")
logger.save_settings(**settings)
print("started")
burnin = disciters if burnin == -1 else burnin
if test:
validinter = 10
burnin = 1
batsize = 2
devinter = 1
tt = q.ticktock("script")
device = torch.device("cpu") if not cuda else torch.device("cuda", gpu)
tt.tick("creating networks")
gen = OldGenerator(z_dim, dim_g).to(device)
crit = OldDiscriminator(dim_d).to(device)
tt.tock("created networks")
# test
# z = torch.randn(3, z_dim).to(device)
# x = gen(z)
# s = crit(x)
# data
# load cifar
tt.tick("loading data")
traincifar, testcifar = load_cifar_dataset(train=True), load_cifar_dataset(
train=False)
print(len(traincifar))
gen_data_d = q.gan.gauss_dataset(z_dim, len(traincifar))
disc_data = q.datacat([traincifar, gen_data_d], 1)
gen_data = q.gan.gauss_dataset(z_dim)
gen_data_valid = q.gan.gauss_dataset(z_dim, 50000)
disc_data = q.dataload(disc_data, batch_size=batsize, shuffle=True)
gen_data = q.dataload(gen_data, batch_size=batsize, shuffle=True)
gen_data_valid = q.dataload(gen_data_valid,
batch_size=batsize,
shuffle=False)
validcifar_loader = q.dataload(testcifar,
batch_size=batsize,
shuffle=False)
dev_data_gauss = q.gan.gauss_dataset(z_dim, len(testcifar))
dev_disc_data = q.datacat([testcifar, dev_data_gauss], 1)
dev_disc_data = q.dataload(dev_disc_data,
batch_size=batsize,
shuffle=False)
# q.embed()
tt.tock("loaded data")
disc_model = q.gan.WGAN(crit, gen, lamda=lamda).disc_train()
gen_model = q.gan.WGAN(crit, gen, lamda=lamda).gen_train()
disc_optim = torch.optim.Adam(q.params_of(crit), lr=lr, betas=(0.5, 0.9))
gen_optim = torch.optim.Adam(q.params_of(gen), lr=lr, betas=(0.5, 0.9))
disc_bt = UnquantizeTransform()
disc_trainer = q.trainer(disc_model).on(disc_data).optimizer(disc_optim).loss(3).device(device)\
.set_batch_transformer(lambda a, b: (disc_bt(a), b))
gen_trainer = q.trainer(gen_model).on(gen_data).optimizer(gen_optim).loss(
1).device(device)
fidandis = q.gan.FIDandIS(device=device)
if not test:
fidandis.set_real_stats_with(validcifar_loader)
saver = q.gan.GenDataSaver(logger, "saved.npz")
generator_validator = q.gan.GeneratorValidator(gen, [fidandis, saver],
gen_data_valid,
device=device,
logger=logger,
validinter=validinter)
train_validator = q.tester(disc_model).on(dev_disc_data).loss(3).device(device)\
.set_batch_transformer(lambda a, b: (disc_bt(a), b))
train_validator.validinter = devinter
tt.tick("training")
gan_trainer = q.gan.GANTrainer(disc_trainer,
gen_trainer,
validators=(generator_validator,
train_validator),
lr_decay=True)
gan_trainer.run(epochs, disciters=disciters, geniters=1, burnin=burnin)
tt.tock("trained")
def run(lr=0.01,
epochs=10,
batsize=64,
momentum=0.5,
cuda=False,
gpu=0,
seed=1):
settings = locals().copy()
logger = q.Logger(prefix="mnist")
logger.save_settings(**settings)
torch.manual_seed(seed)
if cuda:
torch.cuda.set_device(gpu)
torch.cuda.manual_seed(seed)
kwargs = {}
train_loader = torch.utils.data.DataLoader(datasets.MNIST(
'../../datasets/mnist',
train=True,
download=True,
transform=transforms.Compose([
transforms.ToTensor(),
transforms.Normalize((0.1307, ), (0.3081, ))
])),
batch_size=batsize,
shuffle=True,
**kwargs)
test_loader = torch.utils.data.DataLoader(datasets.MNIST(
'../../datasets/mnist',
train=False,
transform=transforms.Compose([
transforms.ToTensor(),
transforms.Normalize((0.1307, ), (0.3081, ))
])),
batch_size=batsize,
shuffle=False,
**kwargs)
class Net(nn.Module):
def __init__(self):
super(Net, self).__init__()
self.conv1 = nn.Conv2d(1, 10, kernel_size=5)
self.conv2 = nn.Conv2d(10, 20, kernel_size=5)
self.conv2_drop = nn.Dropout2d()
self.fc1 = nn.Linear(320, 50)
self.fc2 = nn.Linear(50, 10)
def forward(self, x):
x = F.relu(F.max_pool2d(self.conv1(x), 2))
x = F.relu(F.max_pool2d(self.conv2_drop(self.conv2(x)), 2))
x = x.view(-1, 320)
x = F.relu(self.fc1(x))
x = F.dropout(x, training=self.training)
x = self.fc2(x)
return F.log_softmax(x, dim=1)
model = Net()
optim = torch.optim.SGD(model.parameters(), lr=lr, momentum=momentum)
trainer = q.trainer(model).on(train_loader)\
.loss(torch.nn.NLLLoss(), q.Accuracy())\
.optimizer(optim).cuda(cuda)
validator = q.tester(model).on(test_loader)\
.loss(torch.nn.NLLLoss(), q.Accuracy())\
.cuda(cuda)
logger.loglosses(trainer, "train.losses")
logger.loglosses(validator, "valid.losses")
q.train(trainer, validator).run(epochs)
def run(lr=0.0001,
batsize=64,
epochs=100000,
lamda=10,
disciters=5,
burnin=-1,
validinter=1000,
devinter=100,
cuda=False,
gpu=0,
z_dim=128,
test=False,
dim_d=128,
dim_g=128,
vggversion=13,
vgglayer=9,
vggvanilla=False, # if True, makes trainable feature transform
extralayers=False, # adds a couple extra res blocks to generator to match added VGG
pixelpenalty=False, # if True, uses penalty based on pixel-wise interpolate
inceptionpath="/data/lukovnik/",
normalwgan=False,
):
# vggvanilla=True and pixelpenalty=True makes a normal WGAN
settings = locals().copy()
logger = q.log.Logger(prefix="wgan_resnet_cifar_feat")
logger.save_settings(**settings)
burnin = disciters if burnin == -1 else burnin
if test:
validinter=10
burnin=1
batsize=2
devinter = 1
tt = q.ticktock("script")
device = torch.device("cpu") if not cuda else torch.device("cuda", gpu)
tt.tick("creating networks")
if not normalwgan:
print("doing wgan-feat")
gen = OldGenerator(z_dim, dim_g, extra_layers=extralayers).to(device)
inpd = get_vgg_outdim(vggversion, vgglayer)
crit = ReducedDiscriminator(inpd, dim_d).to(device)
subvgg = SubVGG(vggversion, vgglayer, pretrained=not vggvanilla)
else:
print("doing normal wgan")
gen = OldGenerator(z_dim, dim_g, extra_layers=False).to(device)
crit = OldDiscriminator(dim_d).to(device)
subvgg = None
tt.tock("created networks")
# test
# z = torch.randn(3, z_dim).to(device)
# x = gen(z)
# s = crit(x)
# data
# load cifar
tt.tick("loading data")
traincifar, testcifar = load_cifar_dataset(train=True), load_cifar_dataset(train=False)
print(len(traincifar), len(testcifar))
gen_data_d = q.gan.gauss_dataset(z_dim, len(traincifar))
disc_data = q.datacat([traincifar, gen_data_d], 1)
gen_data = q.gan.gauss_dataset(z_dim)
gen_data_valid = q.gan.gauss_dataset(z_dim, 50000)
swd_gen_data = q.gan.gauss_dataset(z_dim, 10000)
swd_real_data = []
swd_shape = traincifar[0].size()
for i in range(10000):
swd_real_data.append(testcifar[i])
swd_real_data = torch.stack(swd_real_data, 0)
disc_data = q.dataload(disc_data, batch_size=batsize, shuffle=True)
gen_data = q.dataload(gen_data, batch_size=batsize, shuffle=True)
gen_data_valid = q.dataload(gen_data_valid, batch_size=batsize, shuffle=False)
validcifar_loader = q.dataload(testcifar, batch_size=batsize, shuffle=False)
swd_batsize = 64
swd_gen_data = q.dataload(swd_gen_data, batch_size=swd_batsize, shuffle=False)
swd_real_data = q.dataload(swd_real_data, batch_size=swd_batsize, shuffle=False)
dev_data_gauss = q.gan.gauss_dataset(z_dim, len(testcifar))
dev_disc_data = q.datacat([testcifar, dev_data_gauss], 1)
dev_disc_data = q.dataload(dev_disc_data, batch_size=batsize, shuffle=False)
# q.embed()
tt.tock("loaded data")
if not normalwgan:
disc_model = q.gan.WGAN_F(crit, gen, subvgg, lamda=lamda, pixel_penalty=pixelpenalty).disc_train()
gen_model = q.gan.WGAN_F(crit, gen, subvgg, lamda=lamda, pixel_penalty=pixelpenalty).gen_train()
else:
disc_model = q.gan.WGAN(crit, gen, lamda=lamda).disc_train()
gen_model = q.gan.WGAN(crit, gen, lamda=lamda).gen_train()
disc_params = q.params_of(crit)
if vggvanilla and not normalwgan:
disc_params += q.params_of(subvgg)
disc_optim = torch.optim.Adam(disc_params, lr=lr, betas=(0.5, 0.9))
gen_optim = torch.optim.Adam(q.params_of(gen), lr=lr, betas=(0.5, 0.9))
disc_bt = UnquantizeTransform()
disc_trainer = q.trainer(disc_model).on(disc_data).optimizer(disc_optim).loss(3).device(device)\
.set_batch_transformer(lambda a, b: (disc_bt(a), b))
gen_trainer = q.trainer(gen_model).on(gen_data).optimizer(gen_optim).loss(1).device(device)
# fidandis = q.gan.FIDandIS(device=device)
tfis = q.gan.tfIS(inception_path=inceptionpath, gpu=gpu)
# if not test:
# fidandis.set_real_stats_with(validcifar_loader)
saver = q.gan.GenDataSaver(logger, "saved.npz")
generator_validator = q.gan.GeneratorValidator(gen, [tfis, saver], gen_data_valid, device=device,
logger=logger, validinter=validinter)
train_validator = q.tester(disc_model).on(dev_disc_data).loss(3).device(device)\
.set_batch_transformer(lambda a, b: (disc_bt(a), b))
train_validator.validinter = devinter
tt.tick("initializing SWD")
swd = q.gan.SlicedWassersteinDistance(swd_shape)
swd.prepare_reals(swd_real_data)
tt.tock("SWD initialized")
swd_validator = q.gan.GeneratorValidator(gen, [swd], swd_gen_data, device=device,
logger=logger, validinter=validinter, name="swd")
tt.tick("training")
gan_trainer = q.gan.GANTrainer(disc_trainer, gen_trainer,
validators=(generator_validator, train_validator, swd_validator),
lr_decay=True)
gan_trainer.run(epochs, disciters=disciters, geniters=1, burnin=burnin)
tt.tock("trained")