def default_init():
self.actions_executed = 0
self.ntrain_called = 0
self.nstore_called = 0
self.experiences = []
nxt_model = src_model.deep_clone()
inshape = list(src_model.get_input().shape())
batchin = [mbatch_size] + inshape
# environment interaction
self.obs = tc.EVariable(inshape, 0, 'obs', ctx=self.ctx)
self.act_idx = tc.TenncorAPI(self.ctx).argmax(src_model.connect(self.obs))
self.act_idx.tag("recovery", "act_idx")
# training
self.src_obs = tc.EVariable(batchin, 0, 'src_obs', ctx=self.ctx)
self.nxt_obs = tc.EVariable(batchin, 0, 'nxt_obs', ctx=self.ctx)
self.src_outmask = tc.EVariable([mbatch_size] + list(src_model.shape()), 1, 'src_outmask', ctx=self.ctx)
self.nxt_outmask = tc.EVariable([mbatch_size], 1, 'nxt_outmask', ctx=self.ctx)
self.rewards = tc.EVariable([mbatch_size], 0, 'rewards', ctx=self.ctx)
self.prediction_err = tc.api.identity(tc.apply_update([src_model, nxt_model],
get_dqnupdate(update_fn, target_update_rate),
get_dqnerror(self, discount_rate), ctx=self.ctx))
self.prediction_err.tag("recovery", "prediction_err")
tc.optimize(optimize_cfg, self.ctx)
def default_init():
self.name = name
self.dataset_idx = 0
self.api = tc.TenncorAPI(self.ctx)
# prevent shuffling to allow predictable recovery
self.dataset = helper.load(self.oxfile)
self.train_inputs = train_inputs
labelled_inputs = [
self.api.identity(train_input)
for train_input in self.train_inputs
]
outputs = connect_fn(labelled_inputs, self.ctx)
if not isinstance(outputs, Iterable):
outputs = [outputs]
self.train_outputs = [
self.api.identity(train_output) for train_output in outputs
]
for i, linput in enumerate(labelled_inputs):
linput.tag('recovery', 'input_{}'.format(i))
for i, loutput in enumerate(self.train_outputs):
loutput.tag('recovery', 'output_{}'.format(i))
tc.optimize(optimize_cfg, self.ctx)
def tc_mlp_grad(matrix_dim):
import client.profile.profile as tc_prof
import tenncor as tc
n_in = matrix_dim
n_out = int(n_in / 2)
batch_size = 1
# regular mlp
brain = tc.api.layer.link([
tc.api.layer.dense([n_in], [matrix_dim]),
tc.api.layer.bind(tc.api.sigmoid),
tc.api.layer.dense([matrix_dim], [n_out]),
tc.api.layer.bind(tc.api.sigmoid),
])
invar = tc.Variable(np.zeros([batch_size, n_in], dtype=float), 'in')
# out = brain.connect(invar)
# expected_out = tc.Variable(np.zeros([batch_size, n_out], dtype=float), 'expected_out')
# err = tc.api.square(expected_out - out)
train_input = tc.Variable([batch_size, n_in], label='train_input')
train_output = tc.Variable([batch_size, n_out], label='train_output')
invar_batch = batch_generate(n_in, batch_size)
test_batch = batch_generate(n_in, batch_size)
test_batch_out = avgevry2(test_batch)
invar.assign(invar_batch)
train_input.assign(test_batch)
train_output.assign(test_batch_out)
train_err = tc.apply_update(
[brain], lambda err, leaves: tc.api.approx.sgd(
err, leaves, learning_rate=learning_rate), lambda models: tc.api.
error.sqr_diff(train_output, models[0].connect(train_input)))
tc.optimize("external/com_github_mingkaic_tenncor/cfg/optimizations.json")
tc_prof.remote_profile('localhost:8069', [train_err])
def main(args):
# default_ts = time.time()
default_ts = 0
parser = argparse.ArgumentParser(description=prog_description)
parser.add_argument('--seed',
dest='seed',
type=str2bool,
nargs='?',
const=False,
default=True,
help='Whether to seed or not (default: True)')
parser.add_argument('--seedval',
dest='seedval',
type=int,
nargs='?',
default=int(default_ts),
help='Random seed value (default: <current time>)')
parser.add_argument('--nbatch',
dest='nbatch',
type=int,
nargs='?',
default=3,
help='Batch size when training (default: 3)')
parser.add_argument('--n_train',
dest='n_train',
type=int,
nargs='?',
default=3000,
help='Number of times to train (default: 3000)')
parser.add_argument('--n_test',
dest='n_test',
type=int,
nargs='?',
default=500,
help='Number of times to test (default: 500)')
parser.add_argument('--save',
dest='save',
nargs='?',
default='',
help='Filename to save model (default: <blank>)')
parser.add_argument(
'--load',
dest='load',
nargs='?',
default='models/gd.onnx',
help='Filename to load pretrained model (default: models/gd.onnx)')
args = parser.parse_args(args)
if args.seed:
print('seeding {}'.format(args.seedval))
tc.seed(args.seedval)
np.random.seed(args.seedval)
nunits = 9
ninput = 10
noutput = int(ninput / 2)
nbatch = args.nbatch
train_input = tc.EVariable([nbatch, ninput])
train_exout = tc.EVariable([nbatch, noutput])
model = tc.api.layer.link([
tc.api.layer.dense([ninput], [nunits]),
tc.api.layer.bind(tc.api.sigmoid),
tc.api.layer.dense([nunits], [noutput]),
tc.api.layer.bind(tc.api.sigmoid),
], train_input)
train = tc.apply_update(
[model],
lambda err, leaves: tc.api.approx.sgd(err, leaves, learning_rate=0.9),
lambda models: tc.api.loss.mean_squared(train_exout, models[0].connect(
train_input)))
untrained = model.deep_clone()
trained = model.deep_clone()
try:
print('loading ' + args.load)
trained = tc.load_from_file(args.load)[0]
print('successfully loaded from ' + args.load)
except Exception as e:
print(e)
print('failed to load from "{}"'.format(args.load))
testin = tc.EVariable([ninput], label='testin')
untrained_out = untrained.connect(testin)
trained_out = model.connect(testin)
pretrained_out = trained.connect(testin)
tc.optimize("cfg/optimizations.json")
show_every_n = 500
start = time.time()
for i in range(args.n_train):
batch, batch_out = batch_generate(ninput, nbatch)
train_input.assign(batch.reshape(nbatch, ninput))
train_exout.assign(batch_out.reshape(nbatch, noutput))
err = train.get()
if i % show_every_n == show_every_n - 1:
print('training {}\ntraining error:\n{}'.format(i + 1, err))
print('training time: {} seconds'.format(time.time() - start))
untrained_err = 0
trained_err = 0
pretrained_err = 0
for i in range(args.n_test):
if i % show_every_n == show_every_n - 1:
print('testing {}'.format(i + 1))
test_batch, test_batch_out = batch_generate(ninput, 1)
testin.assign(test_batch)
untrained_data = untrained_out.get()
trained_data = trained_out.get()
pretrained_data = pretrained_out.get()
untrained_err += np.mean(abs(untrained_data - test_batch_out))
trained_err += np.mean(abs(trained_data - test_batch_out))
pretrained_err += np.mean(abs(pretrained_data - test_batch_out))
untrained_err /= args.n_test
trained_err /= args.n_test
pretrained_err /= args.n_test
print('untrained mlp error rate: {}%'.format(untrained_err * 100))
print('trained mlp error rate: {}%'.format(trained_err * 100))
print('pretrained mlp error rate: {}%'.format(pretrained_err * 100))
print('is structurally equal?:', cmp.is_equal(pretrained_out, trained_out))
print('% data equal:', cmp.percent_dataeq(pretrained_out, trained_out))
try:
print('saving')
if tc.save_to_file(args.save, [model]):
print('successfully saved to {}'.format(args.save))
except Exception as e:
print(e)
print('failed to write to "{}"'.format(args.save))
def main(args):
default_ts = time.time()
parser = argparse.ArgumentParser(description=prog_description)
parser.add_argument('--seed',
dest='seed',
type=str2bool,
nargs='?',
const=False,
default=False,
help='Whether to seed or not (default: True)')
parser.add_argument('--seedval',
dest='seedval',
type=int,
nargs='?',
default=int(default_ts),
help='Random seed value (default: <current time>)')
parser.add_argument('--n_train',
dest='n_train',
type=int,
nargs='?',
default=10001,
help='Number of times of train (default: 10001)')
parser.add_argument('--save',
dest='save',
nargs='?',
default='',
help='Filename to save model (default: <blank>)')
parser.add_argument(
'--load',
dest='load',
nargs='?',
default='models/latin_lstm.onnx',
help=
'Filename to load pretrained model (default: models/latin_lstm.onnx)')
args = parser.parse_args(args)
if args.seed:
print('seeding {}'.format(args.seedval))
tc.seed(args.seedval)
np.random.seed(args.seedval)
else:
np.random.seed(seed=0)
# Read data and setup maps for integer encoding and decoding.
data = open('models/data/latin_input.txt', 'r').read()
chars = sorted(list(
set(data))) # Sort makes model predictable (if seeded).
data_size, vocab_size = len(data), len(chars)
print('data has %d characters, %d unique.' % (data_size, vocab_size))
char_to_ix = {ch: i for i, ch in enumerate(chars)}
ix_to_char = {i: ch for i, ch in enumerate(chars)}
# Hyper parameters
h_size, o_size, N = vocab_size, vocab_size, vocab_size # Hidden size is set to vocab_size, assuming that level of abstractness is approximately proportional to vocab_size (but can be set to any other value).
seq_length = 25 # Longer sequence lengths allow for lengthier latent dependencies to be trained.
learning_rate = 1e-1
print_interval = 100
def old_winit(shape, label):
return tc.variable(np.random.uniform(-0.05, 0.05, shape.as_list()),
label)
model = tc.api.layer.link([
tc.api.layer.lstm(tc.Shape([N]),
h_size,
seq_length,
kernel_init=tc.api.init.random_uniform(-0.05, 0.05)),
tc.api.layer.dense([h_size], [o_size],
kernel_init=tc.api.init.random_uniform(-0.05,
0.05)),
tc.api.layer.bind(lambda x: tc.api.softmax(x, 0, 1)),
])
untrained_model = model.deep_clone()
pretrained_model = model.deep_clone()
try:
print('loading ' + args.load)
pretrained_model = tc.load_from_file(args.load)[0]
print('successfully loaded from ' + args.load)
except Exception as e:
print(e)
print('failed to load from "{}"'.format(args.load))
sample_inp = tc.EVariable([1, vocab_size], 0)
trained_prob = tc.api.slice(model.connect(sample_inp), 0, 1, 1)
untrained_prob = tc.api.slice(untrained_model.connect(sample_inp), 0, 1, 1)
pretrained_prob = tc.api.slice(pretrained_model.connect(sample_inp), 0, 1,
1)
train_inps = tc.EVariable([seq_length, vocab_size], 0)
train_exout = tc.EVariable([seq_length, vocab_size], 0)
train_err = tc.apply_update(
[model], lambda error, leaves: tc.api.approx.adagrad(
error, leaves, learning_rate=learning_rate, epsilon=1e-8), lambda
models: encoded_loss(train_exout, models[0].connect(train_inps)))
tc.optimize("cfg/optimizations.json")
smooth_loss = -np.log(1.0 / vocab_size) * seq_length
p = 0
start = time.time()
for i in range(args.n_train):
# Reset memory if appropriate
if p + seq_length + 1 >= len(data) or i == 0:
p = 0
# Get input and target sequence
inputs = [char_to_ix[ch] for ch in data[p:p + seq_length]]
encoded_inp = one_encode(inputs, vocab_size)
encoded_out = one_encode(
[char_to_ix[ch] for ch in data[p + 1:p + seq_length + 1]],
vocab_size)
# Occasionally sample from oldModel and print result
if i % print_interval == 0:
sample_ix = sample(sample_inp, trained_prob, inputs[0], 1000)
print('----\n%s\n----' % (''.join(ix_to_char[ix]
for ix in sample_ix)))
# Get gradients for current oldModel based on input and target sequences
train_inps.assign(encoded_inp)
train_exout.assign(encoded_out)
loss = train_err.get()
smooth_loss = smooth_loss * 0.999 + loss * 0.001
# Occasionally print loss information
if i % print_interval == 0:
print('iter %d, loss: %f, smooth loss: %f' %
(i, loss, smooth_loss))
print('batch training time: {} seconds'.format(time.time() -
start))
start = time.time()
# Prepare for next iteration
p += seq_length
untrained_sample = sample(sample_inp, untrained_prob, char_to_ix[data[0]],
1000)
trained_sample = sample(sample_inp, trained_prob, char_to_ix[data[0]],
1000)
pretrained_sample = sample(sample_inp, pretrained_prob,
char_to_ix[data[0]], 1000)
print('--untrained--\n%s\n----' % (''.join(ix_to_char[ix]
for ix in untrained_sample)))
print('--trained--\n%s\n----' % (''.join(ix_to_char[ix]
for ix in trained_sample)))
print('--pretrained--\n%s\n----' % (''.join(ix_to_char[ix]
for ix in pretrained_sample)))
try:
print('saving')
if tc.save_to_file(args.save, [model]):
print('successfully saved to {}'.format(args.save))
except Exception as e:
print(e)
print('failed to write to "{}"'.format(args.save))
tc.api.layer.bind(tc.api.softmax),
])
training_data = generate_training_data(embedding, corpus)
winput = tc.variable(np.random.rand(nwords) * 2 - 1, 'input')
woutput = tc.variable(np.random.rand(2 * window, nwords) * 2 - 1, 'output')
y_pred = model.connect(winput)
train_err = tc.apply_update([model],
lambda error, leaves: tc.api.approx.sgd(error, leaves, lr),
lambda models: tc.api.reduce_sum(tc.api.pow( \
tc.api.extend(models[0].connect(winput), [1, 2 * window]) - woutput, 2.)))
tc.optimize("cfg/optimizations.json")
# Cycle through each epoch
for i in range(epochs):
# Intialise loss to 0
loss = 0
# Cycle through each training sample
# w_t = vector for target word, w_c = vectors for context words
for w_t, w_c in training_data:
wcdata = np.array(w_c)
ydata = y_pred.get().reshape(1, nwords)
for j in range(2 * window - wcdata.shape[0]):
wcdata = np.concatenate((wcdata, ydata), 0)
winput.assign(np.array(w_t))
woutput.assign(wcdata)
def main(args):
default_ts = time.time()
parser = argparse.ArgumentParser(description=prog_description)
parser.add_argument('--seed',
dest='seed',
type=str2bool,
nargs='?',
const=False,
default=False,
help='Whether to seed or not (default: True)')
parser.add_argument('--seedval',
dest='seedval',
type=int,
nargs='?',
default=int(default_ts),
help='Random seed value (default: <current time>)')
parser.add_argument('--n_batch',
dest='n_batch',
type=int,
nargs='?',
default=100,
help='Batch size when training (default: 100)')
parser.add_argument('--n_train',
dest='n_train',
type=int,
nargs='?',
default=2000,
help='Number of times to train (default: 2000)')
parser.add_argument('--n_test',
dest='n_test',
type=int,
nargs='?',
default=5,
help='Number of times to test (default: 5)')
parser.add_argument('--save',
dest='save',
nargs='?',
default='',
help='Filename to save model (default: <blank>)')
parser.add_argument(
'--load',
dest='load',
nargs='?',
default='models/rnn.onnx',
help='Filename to load pretrained model (default: models/rnn.onnx)')
args = parser.parse_args(args)
if args.seed:
print('seeding {}'.format(args.seedval))
tc.seed(args.seedval)
np.random.seed(args.seedval)
else:
np.random.seed(seed=1)
# dataset parameters
n_train = args.n_train
n_test = args.n_test
sequence_len = 7
# training parameters
n_batch = args.n_batch
lmbd = 0.5
learning_rate = 0.05
momentum_term = 0.80
eps = 1e-6
# model parameters
nunits = 3 # Number of states in the recurrent layer
ninput = 2
noutput = 1
model = tc.api.layer.link([
tc.api.layer.dense(inshape=[ninput],
hidden_dims=[nunits],
kernel_init=weight_init),
tc.api.layer.rnn(indim=nunits,
hidden_dim=nunits,
activation=tc.api.tanh,
nseq=sequence_len,
seq_dim=2,
kernel_init=weight_init,
bias_init=tc.api.init.zeros()),
tc.api.layer.dense(inshape=[nunits],
hidden_dims=[noutput],
kernel_init=weight_init),
#tc.api.layer.dense(inshape=[ninput], hidden_dims=[nunits]),
#tc.api.layer.rnn(indim=nunits, hidden_dim=nunits,
# activation=tc.api.tanh, nseq=sequence_len, seq_dim=2),
#tc.api.layer.dense(inshape=[nunits], hidden_dims=[noutput]),
tc.api.layer.bind(tc.api.sigmoid),
])
untrained = model.deep_clone()
trained = model.deep_clone()
try:
print('loading ' + args.load)
trained = tc.load_from_file(args.load)[0]
print('successfully loaded from ' + args.load)
except Exception as e:
print(e)
print('failed to load from "{}"'.format(args.load))
train_invar = tc.EVariable([n_batch, sequence_len, ninput])
train_exout = tc.EVariable([n_batch, sequence_len, noutput])
tinput = tc.api.permute(train_invar, [0, 2, 1])
toutput = tc.api.permute(train_exout, [0, 2, 1])
train_err = tc.apply_update(
[model], lambda error, leaves: make_rms_prop(
error, leaves, learning_rate, momentum_term, lmbd, eps),
lambda models: loss(toutput, models[0].connect(tinput))
) #tc.api.reduce_mean(tc.api.loss.cross_entropy(toutput, models[0].connect(tinput))))
# create training samples
train_input, train_output = create_dataset(n_train, sequence_len)
print(f'train_input tensor shape: {train_input.shape}')
print(f'train_output tensor shape: {train_output.shape}')
test_invar = tc.EVariable([n_test, sequence_len, ninput])
tin = tc.api.permute(test_invar, [0, 2, 1])
untrained_out = tc.api.round(untrained.connect(tin))
trained_out = tc.api.round(model.connect(tin))
pretrained_out = tc.api.round(trained.connect(tin))
tc.optimize("cfg/optimizations.json")
ls_of_loss = []
start = time.time()
for i in range(5):
for j in range(n_train // n_batch):
xbatch = train_input[j:j + n_batch, :, :]
tbatch = train_output[j:j + n_batch, :, :]
train_invar.assign(xbatch)
train_exout.assign(tbatch)
# Add loss to list to plot
ls_of_loss.append(train_err.get())
print('training time: {} seconds'.format(time.time() - start))
# Plot the loss over the iterations
fig = plt.figure(figsize=(5, 3))
plt.plot(ls_of_loss, 'b-')
plt.xlabel('minibatch iteration')
plt.ylabel('$\\xi$', fontsize=15)
plt.title('Loss over backprop iteration')
plt.xlim(0, 99)
fig.subplots_adjust(bottom=0.2)
plt.show()
test_input, test_output = create_dataset(n_test, sequence_len)
test_invar.assign(test_input)
got_untrained = untrained_out.get()
got_trained = trained_out.get()
got_pretrained = pretrained_out.get()
for i in range(test_input.shape[0]):
left = test_input[i, :, 0]
right = test_input[i, :, 1]
expected = test_output[i, :, :]
yuntrained = got_untrained[:, i, :]
ytrained = got_trained[:, i, :]
ypretrained = got_pretrained[:, i, :]
left = ''.join([str(int(d)) for d in left])
left_num = int(''.join(reversed(left)), 2)
right = ''.join([str(int(d)) for d in right])
right_num = int(''.join(reversed(right)), 2)
expected = ''.join([str(int(d[0])) for d in expected])
expected_num = int(''.join(reversed(expected)), 2)
yuntrained = ''.join([str(int(d[0])) for d in yuntrained])
yuntrained_num = int(''.join(reversed(yuntrained)), 2)
ytrained = ''.join([str(int(d[0])) for d in ytrained])
ytrained_num = int(''.join(reversed(ytrained)), 2)
ypretrained = ''.join([str(int(d[0])) for d in ypretrained])
ypretrained_num = int(''.join(reversed(ypretrained)), 2)
print(f'left: {left:s} {left_num:2d}')
print(f'right: + {right:s} {right_num:2d}')
print(f' ------- --')
print(f'expected: = {expected:s} {expected_num:2d}')
print(f'untrained: = {yuntrained:s} {yuntrained_num:2d}')
print(f'trained: = {ytrained:s} {ytrained_num:2d}')
print(f'pretrained: = {ypretrained:s} {ypretrained_num:2d}')
print('')
try:
print('saving')
if tc.save_to_file(args.save, [model]):
print('successfully saved to {}'.format(args.save))
except Exception as e:
print(e)
print('failed to write to "{}"'.format(args.save))
def main(args):
default_ts = time.time()
parser = argparse.ArgumentParser(description=prog_description)
parser.add_argument('--seed',
dest='seed',
type=str2bool,
nargs='?',
const=False,
default=True,
help='Whether to seed or not (default: True)')
parser.add_argument('--seedval',
dest='seedval',
type=int,
nargs='?',
default=int(default_ts),
help='Random seed value (default: <current time>)')
parser.add_argument(
'--pre_nepoch',
dest='pretrain_epochs',
type=int,
nargs='?',
default=1000,
help='Number of epochs when pretraining (default: 1000)')
parser.add_argument('--nepochs',
dest='finetune_epochs',
type=int,
nargs='?',
default=200,
help='Number of epochs when finetuning (default: 200)')
parser.add_argument(
'--cdk',
dest='cdk',
type=int,
nargs='?',
default=1,
help='Length of the Contrastive divergence chain (default: 1)')
parser.add_argument('--save',
dest='save',
nargs='?',
default='',
help='Filename to save model (default: <blank>)')
parser.add_argument(
'--load',
dest='load',
nargs='?',
default='models/dbn.onnx',
help='Filename to load pretrained model (default: models/dbn.onnx)')
args = parser.parse_args(args)
if args.seed:
print('seeding {}'.format(args.seedval))
tc.seed(args.seedval)
np.random.seed(args.seedval)
x = np.array([[1, 1, 1, 0, 0, 0], [1, 0, 1, 0, 0, 0], [1, 1, 1, 0, 0, 0],
[0, 0, 1, 1, 1, 0], [0, 0, 1, 1, 0, 0], [0, 0, 1, 1, 1, 0],
[0, 0, 1, 1, 0, 1]])
y = np.array([[1, 0], [1, 0], [1, 0], [0, 1], [0, 1], [0, 1], [0, 0]])
# construct DBN
rbms = [
tc.api.layer.rbm(6, 3),
tc.api.layer.rbm(3, 3),
]
dense = tc.api.layer.dense([3], [2], kernel_init=tc.api.init.zeros())
softmax_dim = 0
rbm_interlace = zip([rbm.fwd() for rbm in rbms],
len(rbms) * [tc.api.layer.bind(tc.api.sigmoid)])
model = tc.api.layer.link(
[e for inters in rbm_interlace for e in inters] + [
dense,
tc.api.layer.bind(lambda x: tc.api.softmax(x, softmax_dim, 1))
])
untrained = model.deep_clone()
trained = model.deep_clone()
try:
print('loading ' + args.load)
trained = tc.load_from_file(args.load)[0]
print('successfully loaded from ' + args.load)
except Exception as e:
print(e)
print('failed to load from "{}"'.format(args.load))
trainer = tc.DBNTrainer(rbms,
dense,
softmax_dim,
x.shape[0],
pretrain_lr=0.1,
train_lr=0.1,
cdk=args.cdk)
def pretrain_log(epoch, layer):
if epoch % 100 == 0:
print('Pre-training layer {}, epoch {}, cost {}'.format(
layer, epoch, trainer.reconstruction_cost(layer)))
def finetune_log(epoch):
if epoch % 100 == 0:
print('Training epoch {}, cost {}'.format(epoch,
trainer.training_cost()))
# pre-training (TrainUnsupervisedDBN)
trainer.pretrain(x, nepochs=args.pretrain_epochs, logger=pretrain_log)
# fine-tuning (DBNSupervisedFineTuning)
trainer.finetune(x, y, nepochs=args.finetune_epochs, logger=finetune_log)
# test
x = np.array([1, 1, 0, 0, 0, 0])
var = tc.variable(x)
untrained_out = untrained.connect(var)
out = model.connect(var)
trained_out = trained.connect(var)
tc.optimize("cfg/optimizations.json")
# since x is similar to first 3 rows of x, expect results simlar to first 3 rows of y [1, 0]
print('untrained_out: ', untrained_out.get())
print('out: ', out.get())
print('trained_out: ', trained_out.get())
try:
print('saving')
if tc.save_to_file(args.save, [model]):
print('successfully saved to {}'.format(args.save))
except Exception as e:
print(e)
print('failed to write to "{}"'.format(args.save))
def main(args):
default_ts = time.time()
parser = argparse.ArgumentParser(description=prog_description)
parser.add_argument('--seed', dest='seed',
type=str2bool, nargs='?', const=False, default=False,
help='Whether to seed or not (default: True)')
parser.add_argument('--seedval', dest='seedval', type=int, nargs='?', default=int(default_ts),
help='Random seed value (default: <current time>)')
parser.add_argument('--n_train', dest='n_train', type=int, nargs='?', default=100,
help='Number of times of train (default: 100)')
parser.add_argument('--save', dest='save', nargs='?', default='',
help='Filename to save model (default: <blank>)')
parser.add_argument('--load', dest='load', nargs='?', default='models/fast_lstm.onnx',
help='Filename to load pretrained model (default: models/fast_lstm.onnx)')
args = parser.parse_args(args)
if args.seed:
print('seeding {}'.format(args.seedval))
tc.seed(args.seedval)
np.random.seed(args.seedval)
else:
np.random.seed(seed=0)
# parameters for input data dimension and lstm cell count
mem_cell_ct = 100
x_dim = 50
y_list = [-0.5, 0.2, 0.1, -0.5]
input_val_arr = [np.random.random(x_dim) for _ in y_list]
model = tc.api.layer.lstm(tc.Shape([x_dim]), mem_cell_ct, len(y_list),
kernel_init=tc.api.init.xavier_uniform(1),
bias_init=tc.api.init.xavier_uniform(1))
untrained_model = model.deep_clone()
pretrained_model = model.deep_clone()
try:
print('loading ' + args.load)
pretrained_model = tc.load_from_file(args.load)[0]
print('successfully loaded from ' + args.load)
except Exception as e:
print(e)
print('failed to load from "{}"'.format(args.load))
test_inputs = tc.variable(np.array(input_val_arr), 'test_input')
test_exout = tc.variable(np.array(y_list), 'test_exout')
untrained = tc.api.slice(untrained_model.connect(test_inputs), 0, 1, 0)
hiddens = tc.api.slice(model.connect(test_inputs), 0, 1, 0)
pretrained = tc.api.slice(pretrained_model.connect(test_inputs), 0, 1, 0)
err = tc.api.reduce_sum(loss(test_exout, hiddens))
train_err = tc.apply_update([model],
lambda error, leaves: tc.api.approx.sgd(error, leaves, learning_rate=0.1),
lambda models: loss(test_exout, models[0].connect(test_inputs)))
tc.optimize("cfg/optimizations.json")
start = time.time()
for cur_iter in range(args.n_train):
train_err.get()
print("iter {}: y_pred = {}, loss: {}".format(
cur_iter, hiddens.get().flatten(), err.get()))
print("expecting = {}".format(np.array(y_list)))
print("untrained_y_pred = {}".format(untrained.get().flatten()))
print("trained_y_pred = {}".format(hiddens.get().flatten()))
print("pretrained_y_pred = {}".format(pretrained.get().flatten()))
print('training time: {} seconds'.format(time.time() - start))
try:
print('saving')
if tc.save_to_file(args.save, [model]):
print('successfully saved to {}'.format(args.save))
except Exception as e:
print(e)
print('failed to write to "{}"'.format(args.save))