def test_dual_chain_rewrite():
"""Runs regular chain gradient, makes sure memory usage makes sense."""
tf.reset_default_graph()
tf_dev = tf.device('/cpu:0')
tf_dev.__enter__()
n = 5
nodes1 = make_chain_tanh_constant(n, "a")
nodes2 = make_chain_tanh_constant(n, "b")
a0,b0 = nodes1[0], nodes2[0]
a, b = nodes1[-1], nodes2[-1]
grad = memory_saving_gradients.gradients([a+b], [a0, b0],
checkpoints=[nodes1[2], nodes2[2]])
sess = create_session()
sessrun(tf.global_variables_initializer())
sessrun([grad[0].op, grad[1].op])
peak_memory = cpu_peak()
# normal usage comes from 2*n nodes + default ygrad node + 2 gradient nodes
# here we save two 2 units of memory by dropping 2 activations (a1/b1) temporarily
# also, this moves "peak memory" scenario lower down the chain
# where the final addition node activations are no longer needed (another -1)
expected_peak = (2*(n-1)+1)*10**6
util.report_memory(peak_memory, expected_peak)
# since two independent chains, some variability in node scheduling
# allow 1MB slack
if not REMOVE_ASSERTS:
assert (peak_memory - expected_peak) < 4.1e6, "Difference too large."
def test_chain_rewrite(linearize=False):
"""Take chain of length 5, save 2 nodes, make sure 2 units of RAM is
saved."""
tf.reset_default_graph()
tf_dev = tf.device('/cpu:0')
tf_dev.__enter__()
n = 5
a0, a1, a2, a3, a4 = make_chain_tanh(n)
grad = memory_saving_gradients.gradients([a4], [a0], checkpoints=[a1,a3])[0]
expected_peak = (n+1-2)*10**6 # subtract 2 since we recompute 2
sess = create_session()
sessrun(tf.global_variables_initializer())
sessrun(grad.op)
if linearize:
linearize_lib.linearize()
peak_memory = cpu_peak()
util.report_memory(peak_memory, expected_peak)
if not REMOVE_ASSERTS:
assert (peak_memory - expected_peak) < 1e6+10000, "Difference too large."
def test_resnet_rewrite(linearize=False):
tf.reset_default_graph()
tf_dev = tf.device('/cpu:0')
tf_dev.__enter__()
n = 6
nodes = make_resnet(n)
a0 = nodes[0]
a = nodes[-1]
checkpoints = [nodes[3], nodes[5]] # ['a03_add:0', 'a05_add:0']
grad = memory_saving_gradients.gradients([a], [a0], checkpoints=[nodes[2]])[0]
if linearize:
added = linearize_lib.linearize(grad.op)
sess = create_session()
sessrun(tf.global_variables_initializer())
sessrun(grad.op)
peak_memory = cpu_peak()
# 1 for activation of each tanh node + 1 for initial backprop node
# + 1 temporary memory for computing the adds,
# -1 for discarding, then recomputing a1_tanh
expected_peak = (n-1)*10**6
util.report_memory(peak_memory, expected_peak)
if not REMOVE_ASSERTS:
assert (peak_memory - expected_peak) < 1.1*10**6, "Difference too large."
def test_resnet_rewrite(linearize=False):
tf.reset_default_graph()
tf_dev = tf.device('/cpu:0')
tf_dev.__enter__()
n = 6
nodes = make_resnet(n)
a0 = nodes[0]
a = nodes[-1]
checkpoints = [nodes[3], nodes[5]] # ['a03_add:0', 'a05_add:0']
grad = memory_saving_gradients.gradients([a], [a0],
checkpoints=[nodes[2]])[0]
if linearize:
added = linearize_lib.linearize(grad.op)
sess = create_session()
sessrun(tf.global_variables_initializer())
sessrun(grad.op)
peak_memory = cpu_peak()
# 1 for activation of each tanh node + 1 for initial backprop node
# + 1 temporary memory for computing the adds,
# -1 for discarding, then recomputing a1_tanh
expected_peak = (n - 1) * 10**6
util.report_memory(peak_memory, expected_peak)
if not REMOVE_ASSERTS:
assert (peak_memory -
expected_peak) < 1.1 * 10**6, "Difference too large."
def test_chain_rewrite(linearize=False):
"""Take chain of length 5, save 2 nodes, make sure 2 units of RAM is
saved."""
tf.reset_default_graph()
tf_dev = tf.device('/cpu:0')
tf_dev.__enter__()
n = 5
a0, a1, a2, a3, a4 = make_chain_tanh(n)
grad = memory_saving_gradients.gradients([a4], [a0], checkpoints=[a1,
a3])[0]
expected_peak = (n + 1 - 2) * 10**6 # subtract 2 since we recompute 2
sess = create_session()
sessrun(tf.global_variables_initializer())
sessrun(grad.op)
if linearize:
linearize_lib.linearize()
peak_memory = cpu_peak()
util.report_memory(peak_memory, expected_peak)
if not REMOVE_ASSERTS:
assert (peak_memory -
expected_peak) < 1e6 + 10000, "Difference too large."
def test_dual_chain_rewrite():
"""Runs regular chain gradient, makes sure memory usage makes sense."""
tf.reset_default_graph()
tf_dev = tf.device('/cpu:0')
tf_dev.__enter__()
n = 5
nodes1 = make_chain_tanh_constant(n, "a")
nodes2 = make_chain_tanh_constant(n, "b")
a0, b0 = nodes1[0], nodes2[0]
a, b = nodes1[-1], nodes2[-1]
grad = memory_saving_gradients.gradients(
[a + b], [a0, b0], checkpoints=[nodes1[2], nodes2[2]])
sess = create_session()
sessrun(tf.global_variables_initializer())
sessrun([grad[0].op, grad[1].op])
peak_memory = cpu_peak()
# normal usage comes from 2*n nodes + default ygrad node + 2 gradient nodes
# here we save two 2 units of memory by dropping 2 activations (a1/b1) temporarily
# also, this moves "peak memory" scenario lower down the chain
# where the final addition node activations are no longer needed (another -1)
expected_peak = (2 * (n - 1) + 1) * 10**6
util.report_memory(peak_memory, expected_peak)
# since two independent chains, some variability in node scheduling
# allow 1MB slack
if not REMOVE_ASSERTS:
assert (peak_memory - expected_peak) < 4.1e6, "Difference too large."
def test_chain_rewrite_save_last():
"""Take chain of length 5, save last node. This saved no memory, and is
and edge case that should raise exception by rewriter."""
tf.reset_default_graph()
tf_dev = tf.device('/cpu:0')
tf_dev.__enter__()
n = 5
a0, a1, a2, a3, a4 = make_chain_tanh(n)
try:
grad = memory_saving_gradients.gradients([a4], [a0], checkpoints=[a4])[0]
except Exception:
return
else:
if not REMOVE_ASSERTS:
assert "Should've been 'no checkpoints nodes found' exception"
def test_chain_rewrite_save_last():
"""Take chain of length 5, save last node. This saved no memory, and is
and edge case that should raise exception by rewriter."""
tf.reset_default_graph()
tf_dev = tf.device('/cpu:0')
tf_dev.__enter__()
n = 5
a0, a1, a2, a3, a4 = make_chain_tanh(n)
try:
grad = memory_saving_gradients.gradients([a4], [a0], checkpoints=[a4])[0]
except Exception:
return
else:
if not REMOVE_ASSERTS:
assert "Should've been 'no checkpoints nodes found' exception"
def test_chain_rewrite_save_first():
"""Take chain of length 5, save first node."""
tf.reset_default_graph()
tf_dev = tf.device('/cpu:0')
tf_dev.__enter__()
n = 5
a0, a1, a2, a3, a4 = make_chain_tanh_constant(n)
grad = memory_saving_gradients.gradients([a4], [a0], checkpoints=[a1, a3])[0]
expected_peak = (n+1-2)*10**6
sess = create_session()
sessrun(tf.global_variables_initializer())
sessrun(grad.op)
peak_memory = cpu_peak()
util.report_memory(peak_memory, expected_peak)
if not REMOVE_ASSERTS:
assert (peak_memory - expected_peak) < 1.1e6, "Difference too large."
def test_chain_rewrite_save_one_before_last():
"""Take chain of length 5, save first node."""
tf.reset_default_graph()
tf_dev = tf.device('/cpu:0')
tf_dev.__enter__()
n = 5
a0, a1, a2, a3, a4 = make_chain_tanh_constant(n)
grad = memory_saving_gradients.gradients([a4], [a0], checkpoints=[a2])[0]
expected_peak = (n + 1 - 2) * 10**6
sess = create_session()
sessrun(tf.global_variables_initializer())
sessrun(grad.op)
peak_memory = cpu_peak()
util.report_memory(peak_memory, expected_peak)
if not REMOVE_ASSERTS:
assert (peak_memory - expected_peak) < 1.1e6, "Difference too large."
def gradients_collection(ys, xs, grad_ys=None, **kwargs):
return memory_saving_gradients.gradients(ys,
xs,
grad_ys,
checkpoints='collection',
**kwargs)
def main():
args = parser.parse_args()
enc = encoder.get_encoder(args.model_name)
hparams = model.default_hparams()
with open(os.path.join('models', args.model_name, 'hparams.json')) as f:
hparams.override_from_dict(json.load(f))
if args.sample_length > hparams.n_ctx:
raise ValueError(
"Can't get samples longer than window size: %s" % hparams.n_ctx)
if args.model_name == '774M':
args.memory_saving_gradients = True
if args.optimizer == 'adam':
args.only_train_transformer_layers = True
config = tf.ConfigProto()
config.gpu_options.allow_growth = True
config.graph_options.rewrite_options.layout_optimizer = rewriter_config_pb2.RewriterConfig.OFF
with tf.Session(config=config) as sess:
context = tf.placeholder(tf.int32, [args.batch_size, None])
context_in = randomize(context, hparams, args.noise)
output = model.model(hparams=hparams, X=context_in)
loss = tf.reduce_mean(
tf.nn.sparse_softmax_cross_entropy_with_logits(
labels=context[:, 1:], logits=output['logits'][:, :-1]))
tf_sample = sample.sample_sequence(
hparams=hparams,
length=args.sample_length,
context=context,
batch_size=args.batch_size,
temperature=1.0,
top_k=args.top_k,
top_p=args.top_p)
all_vars = [v for v in tf.trainable_variables() if 'model' in v.name]
#this line is to hopefully reduce memory usage (found on Twitter: https://twitter.com/BasedBlue/status/1169601983046672385?s=20)
edgeindex = -1 * args.layers_to_train
train_vars = all_vars[edgeindex:]
print("Training", args.layers_to_train, "raw layers out of", len(all_vars))
train_vars = [v for v in train_vars if '/h' in v.name] if args.only_train_transformer_layers else train_vars
print("Training", len(train_vars), "net layers out of", len(all_vars))
if args.optimizer == 'adam':
opt = tf.train.AdamOptimizer(learning_rate=args.learning_rate)
elif args.optimizer == 'sgd':
opt = tf.train.GradientDescentOptimizer(learning_rate=args.learning_rate)
elif args.optimizer == 'adafactor':
opt = AdafactorOptimizer(learning_rate=args.learning_rate)
else:
exit('Bad optimizer:', args.optimizer)
if args.accumulate_gradients > 1:
if args.memory_saving_gradients:
exit("Memory saving gradients are not implemented for gradient accumulation yet.")
opt = AccumulatingOptimizer(
opt=opt,
var_list=train_vars)
opt_reset = opt.reset()
opt_compute = opt.compute_gradients(loss)
opt_apply = opt.apply_gradients()
summary_loss = tf.summary.scalar('loss', opt_apply)
else:
if args.memory_saving_gradients:
opt_grads = memory_saving_gradients.gradients(loss, train_vars)
else:
opt_grads = tf.gradients(loss, train_vars)
opt_grads = list(zip(opt_grads, train_vars))
opt_apply = opt.apply_gradients(opt_grads)
summary_loss = tf.summary.scalar('loss', loss)
summary_lr = tf.summary.scalar('learning_rate', args.learning_rate)
summaries = tf.summary.merge([summary_lr, summary_loss])
summary_log = tf.summary.FileWriter(
os.path.join(CHECKPOINT_DIR, args.run_name))
saver = tf.train.Saver(
var_list=all_vars,
max_to_keep=5,
keep_checkpoint_every_n_hours=2)
sess.run(tf.global_variables_initializer())
if args.restore_from == 'latest':
ckpt = tf.train.latest_checkpoint(
os.path.join(CHECKPOINT_DIR, args.run_name))
if ckpt is None:
# Get fresh GPT weights if new run.
ckpt = tf.train.latest_checkpoint(
os.path.join('models', args.model_name))
elif args.restore_from == 'fresh':
ckpt = tf.train.latest_checkpoint(
os.path.join('models', args.model_name))
else:
ckpt = tf.train.latest_checkpoint(args.restore_from)
print('Loading checkpoint', ckpt)
saver.restore(sess, ckpt)
print('Loading dataset...')
chunks = load_dataset(enc, args.dataset, args.combine, encoding=args.encoding)
data_sampler = Sampler(chunks)
if args.val_every > 0:
if args.val_dataset:
val_chunks = load_dataset(enc, args.val_dataset, args.combine, encoding=args.encoding)
else:
val_chunks = chunks
print('dataset has', data_sampler.total_size, 'tokens')
print('Training...')
if args.val_every > 0:
# Sample from validation set once with fixed seed to make
# it deterministic during training as well as across runs.
val_data_sampler = Sampler(val_chunks, seed=1)
val_batches = [[val_data_sampler.sample(1024) for _ in range(args.val_batch_size)]
for _ in range(args.val_batch_count)]
counter = 1
counter_path = os.path.join(CHECKPOINT_DIR, args.run_name, 'counter')
if os.path.exists(counter_path):
# Load the step number if we're resuming a run
# Add 1 so we don't immediately try to save again
with open(counter_path, 'r') as fp:
counter = int(fp.read()) + 1
def save():
maketree(os.path.join(CHECKPOINT_DIR, args.run_name))
print(
'Saving',
os.path.join(CHECKPOINT_DIR, args.run_name,
'model-{}').format(counter))
saver.save(
sess,
os.path.join(CHECKPOINT_DIR, args.run_name, 'model'),
global_step=counter)
with open(counter_path, 'w') as fp:
fp.write(str(counter) + '\n')
def generate_samples():
print('Generating samples...')
context_tokens = data_sampler.sample(1)
all_text = []
index = 0
while index < args.sample_num:
out = sess.run(
tf_sample,
feed_dict={context: args.batch_size * [context_tokens]})
for i in range(min(args.sample_num - index, args.batch_size)):
text = enc.decode(out[i])
text = '======== SAMPLE {} ========\n{}\n'.format(
index + 1, text)
all_text.append(text)
index += 1
print(text)
maketree(os.path.join(SAMPLE_DIR, args.run_name))
with open(
os.path.join(SAMPLE_DIR, args.run_name,
'samples-{}').format(counter), 'w', encoding=args.encoding) as fp:
fp.write('\n'.join(all_text))
def sample_batch():
ret = [data_sampler.sample(1024) for _ in range(args.batch_size)]
# print (enc.decode(ret[0]))
return ret
avg_loss = (0.0, 0.0)
bval_loss = (0.0, 0.0)
start_time = time.time()
best_val_loss = 99
missed_val_checkpoints = 0
try:
while counter < args.stop_after:
if counter % args.sample_every == 0:
generate_samples()
if args.accumulate_gradients > 1:
sess.run(opt_reset)
for _ in range(args.accumulate_gradients):
sess.run(
opt_compute, feed_dict={context: sample_batch()})
(v_loss, v_summary) = sess.run((opt_apply, summaries))
else:
(_, v_loss, v_summary) = sess.run(
(opt_apply, loss, summaries),
feed_dict={context: sample_batch()})
summary_log.add_summary(v_summary, counter)
avg_loss = (avg_loss[0] * 0.98 + v_loss,
avg_loss[1] * 0.98 + 1.0)
print(
'[{counter} | {time:2.2f}] loss={loss:2.2f} avg={avg:2.2f}'
.format(
counter=counter,
time=time.time() - start_time,
loss=v_loss,
avg=avg_loss[0] / avg_loss[1]))
if args.val_every > 0 and counter % args.val_every == 0:
valbatch = [val_data_sampler.sample(1024) for _ in range(args.batch_size)]
valacc = sess.run(loss, feed_dict={context: valbatch})
bval_loss = (bval_loss[0] * 0.9 + valacc, bval_loss[1] * 0.9 + 1.0)
av_val_loss = bval_loss[0] / bval_loss[1]
av_train_loss = avg_loss[0] / avg_loss[1]
print(
'[{counter} | {time:2.2f}] VAL_loss={loss:2.4f} VAL_avg={avg:2.4f} best={best:2.4f}'
.format(
counter=counter,
time=time.time() - start_time,
loss=valacc,
avg=av_val_loss,
best=best_val_loss))
if counter >= args.save_every and counter % args.save_every == 0: # check for validation checkpoints every save_every iterations.
if av_val_loss < best_val_loss and av_val_loss > av_train_loss: # got a good one from validation, save a checkpoint (every save_every) -- but don't save before val loss goes above train loss
save()
best_val_loss = av_val_loss
missed_val_checkpoints = 0
else: # missed a validation checkpoint. tolerate like 10 of these.
if av_val_loss > av_train_loss: # don't count a missed checkpoint while val loss is under training loss
missed_val_checkpoints += 1
if missed_val_checkpoints > 19: # missed too many save opportunities, stop training
counter = args.stop_after + 1
print('stopping training due to val loss not improving.')
counter += 1
except KeyboardInterrupt:
print('interrupted')
def grads(ys, xs, grad_ys=None, **kwargs):
return memory_saving_gradients.gradients(ys, xs, grad_ys,
checkpoints='speed', **kwargs)
def abstract_model_xy(sess, hps, feeds, train_iterator, test_iterator, data_init, lr, f_loss):
# == Create class with static fields and methods
class m(object):
pass
m.sess = sess
m.feeds = feeds
m.lr = lr
# === Loss and optimizer
loss_train, stats_train = f_loss(train_iterator, True)
all_params = tf.trainable_variables()
if hps.gradient_checkpointing == 1:
from memory_saving_gradients import gradients
gs = gradients(loss_train, all_params)
else:
gs = tf.gradients(loss_train, all_params)
optimizer = {'adam': optim.adam, 'adamax': optim.adamax,
'adam2': optim.adam2}[hps.optimizer]
train_op, polyak_swap_op, ema = optimizer(
all_params, gs, alpha=lr, hps=hps)
if hps.direct_iterator:
m.train = lambda _lr: sess.run([train_op, stats_train], {lr: _lr})[1]
else:
def _train(_lr):
_x, _y = train_iterator()
return sess.run([train_op, stats_train], {feeds['x']: _x,
feeds['y']: _y, lr: _lr})[1]
m.train = _train
m.polyak_swap = lambda: sess.run(polyak_swap_op)
# === Testing
loss_test, stats_test = f_loss(test_iterator, False, reuse=True)
if hps.direct_iterator:
m.test = lambda: sess.run(stats_test)
else:
def _test():
_x, _y = test_iterator()
return sess.run(stats_test, {feeds['x']: _x,
feeds['y']: _y})
m.test = _test
# === Saving and restoring
saver = tf.train.Saver()
saver_ema = tf.train.Saver(ema.variables_to_restore())
m.save_ema = lambda path: saver_ema.save(
sess, path, write_meta_graph=False)
m.save = lambda path: saver.save(sess, path, write_meta_graph=False)
m.restore = lambda path: saver.restore(sess, path)
# === Initialize the parameters
if hps.restore_path != '':
m.restore(hps.restore_path)
else:
with Z.arg_scope([Z.get_variable_ddi, Z.actnorm], init=True):
results_init = f_loss(None, True, reuse=True)
sess.run(tf.global_variables_initializer())
sess.run(results_init, {feeds['x']: data_init['x'],
feeds['y']: data_init['y']})
sess.run(hvd.broadcast_global_variables(0))
return m
def model_fn(features, labels, mode, params):
del labels
cfg = params['cfg']
model = models.model(cfg)
y = features['y']
if mode == tf.estimator.ModeKeys.PREDICT:
###########
# PREDICT #
###########
predictions = {'generated_images': model.sample(y, temp=0.75)}
return tpu_estimator.TPUEstimatorSpec(mode=mode,
predictions=predictions)
is_training = (mode == tf.estimator.ModeKeys.TRAIN)
real_images = features['real_images']
f_loss, eps = model.f_loss(real_images, y)
if mode == tf.estimator.ModeKeys.TRAIN:
#########
# TRAIN #
#########
f_loss = tf.reduce_mean(f_loss)
with tf.variable_scope('Regularization'):
for v in tf.trainable_variables():
if 'invconv' in v.name:
det = tf.matrix_determinant(v * tf.transpose(v))
f_loss += tf.square(det - 1)
if cfg.use_l2_regularization:
for v in tf.trainable_variables():
if 'actnorm' not in v.name:
f_loss += cfg.l2_regularization_factor * tf.nn.l2_loss(
v)
if not cfg.use_tpu and cfg.report_histograms:
for v in tf.trainable_variables():
tf.summary.histogram(v.name.replace(':', '_'), v)
global_step = tf.train.get_or_create_global_step()
rate = tf.minimum(tf.cast(global_step, tf.float32) / 2000.0, 1.0)
#lr = int(real_images.get_shape()[0]) * cfg.lr
lr = cfg.lr * rate
#from AMSGrad import AMSGrad
optimizer = tf.train.AdamOptimizer(learning_rate=lr,
beta1=cfg.beta1,
epsilon=cfg.adam_eps)
tf.summary.scalar('lr', lr)
if cfg.use_tpu:
optimizer = tpu_optimizer.CrossShardOptimizer(optimizer)
with tf.control_dependencies(tf.get_collection(
tf.GraphKeys.UPDATE_OPS)):
with tf.variable_scope('TrainOps'):
if cfg.memory_saving_gradients:
from memory_saving_gradients import gradients
gs = gradients(f_loss, tf.trainable_variables())
else:
gs = tf.gradients(f_loss, tf.trainable_variables())
if cfg.use_gradient_clipping:
gs = [tf.clip_by_value(g, -100., 100.) for g in gs]
grads_and_vars = list(zip(gs, tf.trainable_variables()))
train_op = optimizer.apply_gradients(grads_and_vars)
increment_step = tf.assign_add(
tf.train.get_or_create_global_step(), 1)
joint_op = tf.group([train_op, increment_step])
return tpu_estimator.TPUEstimatorSpec(mode=mode,
loss=f_loss,
train_op=joint_op)
elif mode == tf.estimator.ModeKeys.EVAL:
########
# EVAL #
########
def _eval_metric_fn(f_loss):
return {'f_loss': tf.metrics.mean(f_loss)}
return tpu_estimator.TPUEstimatorSpec(mode=mode,
loss=tf.reduce_mean(f_loss),
eval_metrics=(_eval_metric_fn,
[f_loss]))
raise ValueError('Invalid mode provided to model_fn')
def train_main(dataset,
model_name='1250M',
seed=None,
msg=True,
batch_size=16,
learning_rate=0.00002,
sample_length=512,
sample_num=1,
sample_every=100,
run_name='run1',
restore_from='latest',
save_every=1000,
combine=50000):
enc = encoder.get_encoder(model_name)
hparams = model.default_hparams()
with open(os.path.join('models', model_name, 'hparams.json')) as f:
hparams.override_from_dict(json.load(f))
print('n_ctx: ', hparams.n_ctx, 'n_head: ', hparams.n_head, 'n_embd: ',
hparams.n_embd, 'n_layer: ', hparams.n_layer)
if sample_length is None:
sample_length = hparams.n_ctx
elif sample_length > hparams.n_ctx:
raise ValueError("Can't get samples longer than window size: %s" %
hparams.n_ctx)
# TF config
config = tf.ConfigProto()
#device_map = { 0:2, 0:3, 1:2, 1:3 }
#config.gpu_options.visible_device_list = str(device_map[hvd.rank()])
config.gpu_options.visible_device_list = str(hvd.local_rank())
config.gpu_options.allow_growth = True
global_step = tf.Variable(0, trainable=False)
with tf.Session(config=config) as sess:
context = tf.placeholder(tf.int32, [batch_size, None])
np.random.seed(seed)
tf.set_random_seed(seed)
output = model.model(hparams=hparams, X=context)
loss = tf.reduce_mean(
tf.nn.sparse_softmax_cross_entropy_with_logits(
labels=context[:, 1:], logits=output['logits'][:, :-1]))
tf_sample = sample.sample_sequence(hparams=hparams,
length=sample_length,
context=context,
batch_size=batch_size,
temperature=0.9,
top_k=40)
#global_step = tf.Variable(0, trainable=False)
counter = 1
train_vars = [v for v in tf.trainable_variables() if 'model' in v.name]
#opt = tf.train.AdamOptimizer(learning_rate=learning_rate)
# l4rz 11/10/2019
decayed_lr = tf.train.exponential_decay(learning_rate,
global_step,
200,
0.999,
staircase=True)
opt = tf.train.AdamOptimizer(decayed_lr)
#opt = tf.train.GradientDescentOptimizer(decayed_lr)
opt = hvd.DistributedOptimizer(opt)
# this is original horovod
#train_op = opt.minimize(loss, var_list=train_vars)
# this is ours
if (msg):
print('Using memory saving gradients')
opt_grads = memory_saving_gradients.gradients(loss, train_vars)
opt_grads = list(zip(opt_grads, train_vars))
train_op = opt.apply_gradients(opt_grads, global_step=global_step)
else:
print('Not using memory saving gradients')
#train_op = opt.minimize(loss, var_list=train_vars)
# l4rz 11/10
train_op = opt.minimize(loss,
var_list=train_vars,
global_step=global_step)
# [1,2]<stderr>:TypeError: apply_gradients() missing 1 required positional argument: 'grads_and_vars'
#summary_loss = tf.summary.scalar('loss', train_op)
#_, lv = sess.run((train_op, loss), feed_dict={context: batch})
# Horovod: broadcast initial variable states from rank 0 to all other processes.
# This is necessary to ensure consistent initialization of all workers when
# training is started with random weights or restored from a checkpoint.
print('Running hvd.broadcast_global_variables')
bcast = hvd.broadcast_global_variables(0)
print('Done')
saver = tf.train.Saver(var_list=train_vars,
max_to_keep=5,
keep_checkpoint_every_n_hours=2)
print('Running global_variables_initializer')
sess.run(tf.global_variables_initializer())
print('Done')
if restore_from == 'latest':
ckpt = tf.train.latest_checkpoint(
os.path.join(CHECKPOINT_DIR, run_name))
if ckpt is None:
# Get fresh GPT weights if new run.
ckpt = tf.train.latest_checkpoint(
os.path.join('models', model_name))
elif restore_from == 'fresh':
ckpt = tf.train.latest_checkpoint(
os.path.join('models', model_name))
# comment out when running for 1st time
else:
ckpt = tf.train.latest_checkpoint(restore_from)
print(str(hvd.local_rank()), 'Loading checkpoint', ckpt)
saver.restore(sess, ckpt)
# uncomment when running for first time INIT THE MODEL
#print('tf.global_variables_initializer()')
#sess.run(tf.global_variables_initializer())
bcast.run()
print(str(hvd.local_rank()), 'Loading dataset...')
chunks = load_dataset(enc, dataset, combine)
data_sampler = Sampler(chunks)
print(str(hvd.local_rank()), 'dataset has', data_sampler.total_size,
'tokens')
print(str(hvd.local_rank()), 'Training...')
counter = 1
if os.path.exists(os.path.join(CHECKPOINT_DIR, run_name, 'counter')):
# Load the step number if we're resuming a run
# Add 1 so we don't immediately try to save again
with open(os.path.join(CHECKPOINT_DIR, run_name, 'counter'),
'r') as fp:
counter = int(fp.read()) + 1
def save():
maketree(os.path.join(CHECKPOINT_DIR, run_name))
print(
'Saving',
os.path.join(CHECKPOINT_DIR, run_name,
'model-{}').format(counter))
saver.save(sess,
os.path.join(CHECKPOINT_DIR, run_name, 'model'),
global_step=counter)
with open(os.path.join(CHECKPOINT_DIR, run_name, 'counter'),
'w') as fp:
fp.write(str(counter) + '\n')
def generate_samples():
context_tokens = data_sampler.sample(1)
all_text = []
index = 0
while index < sample_num:
out = sess.run(
tf_sample,
feed_dict={context: batch_size * [context_tokens]})
for i in range(min(sample_num - index, batch_size)):
text = enc.decode(out[i])
text = '======== SAMPLE {} ========\n{}\n'.format(
index + 1, text)
all_text.append(text)
index += 1
print(text)
maketree(os.path.join(SAMPLE_DIR, run_name))
with open(
os.path.join(SAMPLE_DIR, run_name,
'samples-{}').format(counter), 'w') as fp:
fp.write('\n'.join(all_text))
avg_loss = (0.0, 0.0)
start_time = time.time()
try:
while True:
batch = [data_sampler.sample(1024) for _ in range(batch_size)]
_, lv = sess.run((train_op, loss), feed_dict={context: batch})
avg_loss = (avg_loss[0] * 0.99 + lv, avg_loss[1] * 0.99 + 1.0)
if hvd.rank() == 0:
if counter % save_every == 0:
save()
if counter % sample_every == 0:
generate_samples()
print(
'[{counter} | {time:2.2f}] loss={loss:2.4f} avg={avg:2.4f} lr={lr:.2e}'
.format(counter=counter,
time=time.time() - start_time,
loss=lv,
avg=avg_loss[0] / avg_loss[1],
lr=decayed_lr.eval()))
counter += 1
except KeyboardInterrupt:
print('interrupted')
if hvd.rank() == 0:
save()
def main():
args = parser.parse_args()
enc = encoder.get_encoder(args.model_name)
hparams = model.default_hparams()
with open(os.path.join('models', args.model_name, 'hparams.json')) as f:
hparams.override_from_dict(json.load(f))
if args.sample_length > hparams.n_ctx:
raise ValueError("Can't get samples longer than window size: %s" %
hparams.n_ctx)
if args.model_name == '355M':
args.memory_saving_gradients = True
if args.optimizer == 'adam':
args.only_train_transformer_layers = True
config = tf.ConfigProto()
config.gpu_options.allow_growth = True
config.graph_options.rewrite_options.layout_optimizer = rewriter_config_pb2.RewriterConfig.OFF
with tf.Session(config=config) as sess:
context = tf.placeholder(tf.int32, [args.batch_size, None])
context_in = randomize(context, hparams, args.noise)
output = model.model(hparams=hparams, X=context_in)
loss = tf.reduce_mean(
tf.nn.sparse_softmax_cross_entropy_with_logits(
labels=context[:, 1:], logits=output['logits'][:, :-1]))
if args.val_every > 0:
val_context = tf.placeholder(tf.int32, [args.val_batch_size, None])
val_output = model.model(hparams=hparams, X=val_context)
val_loss = tf.reduce_mean(
tf.nn.sparse_softmax_cross_entropy_with_logits(
labels=val_context[:, 1:],
logits=val_output['logits'][:, :-1]))
val_loss_summary = tf.summary.scalar('val_loss', val_loss)
tf_sample = sample.sample_sequence(hparams=hparams,
length=args.sample_length,
context=context,
batch_size=args.batch_size,
temperature=1.0,
top_k=args.top_k,
top_p=args.top_p)
all_vars = [v for v in tf.trainable_variables() if 'model' in v.name]
train_vars = [v for v in all_vars if '/h' in v.name
] if args.only_train_transformer_layers else all_vars
if args.optimizer == 'adam':
opt = tf.train.AdamOptimizer(learning_rate=args.learning_rate)
elif args.optimizer == 'sgd':
opt = tf.train.GradientDescentOptimizer(
learning_rate=args.learning_rate)
else:
exit('Bad optimizer:', args.optimizer)
if args.accumulate_gradients > 1:
if args.memory_saving_gradients:
exit(
"Memory saving gradients are not implemented for gradient accumulation yet."
)
opt = AccumulatingOptimizer(opt=opt, var_list=train_vars)
opt_reset = opt.reset()
opt_compute = opt.compute_gradients(loss)
opt_apply = opt.apply_gradients()
summary_loss = tf.summary.scalar('loss', opt_apply)
else:
if args.memory_saving_gradients:
opt_grads = memory_saving_gradients.gradients(loss, train_vars)
else:
opt_grads = tf.gradients(loss, train_vars)
opt_grads = list(zip(opt_grads, train_vars))
opt_apply = opt.apply_gradients(opt_grads)
summary_loss = tf.summary.scalar('loss', loss)
summary_lr = tf.summary.scalar('learning_rate', args.learning_rate)
summaries = tf.summary.merge([summary_lr, summary_loss])
summary_log = tf.summary.FileWriter(
os.path.join(CHECKPOINT_DIR, args.run_name))
saver = tf.train.Saver(var_list=all_vars,
max_to_keep=5,
keep_checkpoint_every_n_hours=2)
sess.run(tf.global_variables_initializer())
if args.restore_from == 'latest':
ckpt = tf.train.latest_checkpoint(
os.path.join(CHECKPOINT_DIR, args.run_name))
if ckpt is None:
# Get fresh GPT weights if new run.
ckpt = tf.train.latest_checkpoint(
os.path.join('models', args.model_name))
elif args.restore_from == 'fresh':
ckpt = tf.train.latest_checkpoint(
os.path.join('models', args.model_name))
else:
ckpt = tf.train.latest_checkpoint(args.restore_from)
print('Loading checkpoint', ckpt)
saver.restore(sess, ckpt)
print('Loading dataset...')
chunks = load_dataset(enc,
args.dataset,
args.combine,
encoding=args.encoding)
data_sampler = Sampler(chunks)
if args.val_every > 0:
if args.val_dataset:
val_chunks = load_dataset(enc,
args.val_dataset,
args.combine,
encoding=args.encoding)
else:
val_chunks = chunks
print('dataset has', data_sampler.total_size, 'tokens')
print('Training...')
if args.val_every > 0:
# Sample from validation set once with fixed seed to make
# it deterministic during training as well as across runs.
val_data_sampler = Sampler(val_chunks, seed=1)
val_batches = [[
val_data_sampler.sample(1024)
for _ in range(args.val_batch_size)
] for _ in range(args.val_batch_count)]
counter = 1
counter_path = os.path.join(CHECKPOINT_DIR, args.run_name, 'counter')
if os.path.exists(counter_path):
# Load the step number if we're resuming a run
# Add 1 so we don't immediately try to save again
with open(counter_path, 'r') as fp:
counter = int(fp.read()) + 1
def save():
maketree(os.path.join(CHECKPOINT_DIR, args.run_name))
print(
'Saving',
os.path.join(CHECKPOINT_DIR, args.run_name,
'model-{}').format(counter))
saver.save(sess,
os.path.join(CHECKPOINT_DIR, args.run_name, 'model'),
global_step=counter)
with open(counter_path, 'w') as fp:
fp.write(str(counter) + '\n')
def generate_samples():
print('Generating samples...')
context_tokens = data_sampler.sample(1)
all_text = []
index = 0
while index < args.sample_num:
out = sess.run(
tf_sample,
feed_dict={context: args.batch_size * [context_tokens]})
for i in range(min(args.sample_num - index, args.batch_size)):
text = enc.decode(out[i])
text = '======== SAMPLE {} ========\n{}\n'.format(
index + 1, text)
all_text.append(text)
index += 1
print(text.encode('utf8'))
maketree(os.path.join(SAMPLE_DIR, args.run_name))
with open(os.path.join(SAMPLE_DIR, args.run_name,
'samples-{}').format(counter),
'w',
encoding=args.encoding) as fp:
fp.write('\n'.join(all_text))
def validation():
print('Calculating validation loss...')
losses = []
for batch in tqdm.tqdm(val_batches):
losses.append(
sess.run(val_loss, feed_dict={val_context: batch}))
v_val_loss = np.mean(losses)
v_summary = sess.run(val_loss_summary,
feed_dict={val_loss: v_val_loss})
summary_log.add_summary(v_summary, counter)
summary_log.flush()
print('[{counter} | {time:2.2f}] validation loss = {loss:2.2f}'.
format(counter=counter,
time=time.time() - start_time,
loss=v_val_loss))
def sample_batch():
return [data_sampler.sample(1024) for _ in range(args.batch_size)]
avg_loss = (0.0, 0.0)
start_time = time.time()
try:
while True:
if counter % args.save_every == 0:
save()
if counter % args.sample_every == 0:
generate_samples()
if args.val_every > 0 and (counter % args.val_every == 0
or counter == 1):
validation()
if args.accumulate_gradients > 1:
sess.run(opt_reset)
for _ in range(args.accumulate_gradients):
sess.run(opt_compute,
feed_dict={context: sample_batch()})
(v_loss, v_summary) = sess.run((opt_apply, summaries))
else:
(_, v_loss, v_summary) = sess.run(
(opt_apply, loss, summaries),
feed_dict={context: sample_batch()})
summary_log.add_summary(v_summary, counter)
avg_loss = (avg_loss[0] * 0.99 + v_loss,
avg_loss[1] * 0.99 + 1.0)
print(
'[{counter} | {time:2.2f}] loss={loss:2.2f} avg={avg:2.2f}'
.format(counter=counter,
time=time.time() - start_time,
loss=v_loss,
avg=avg_loss[0] / avg_loss[1]))
counter += 1
except KeyboardInterrupt:
print('interrupted')
save()
def gradients_auto(ys, xs, grad_ys=None, **kwargs):
return memory_saving_gradients.gradients(ys,
xs,
grad_ys,
checkpoints='memory',
**kwargs)
def main():
args = parser.parse_args()
try:
logdir = os.path.join(CHECKPOINT_DIR, args.run_name)
with open('logdir.txt', 'w') as z:
z.write(logdir)
except:
pass
enc = get_encoder(model_name)
hparams = model.default_hparams()
with open(os.path.join(model_name, 'hparams.json')) as f:
hparams.override_from_dict(json.load(f))
if args.sample_length > hparams.n_ctx:
raise ValueError("Can't get samples longer than window size: %s" %
hparams.n_ctx)
if args.model_name == '345M':
args.memory_saving_gradients = True
args.only_train_transformer_layers = True
config = tf.ConfigProto()
config.gpu_options.allow_growth = True
config.graph_options.rewrite_options.layout_optimizer = rewriter_config_pb2.RewriterConfig.OFF
with tf.Session(config=config) as sess:
context = tf.placeholder(tf.int32, [args.batch_size, None])
output = model.model(hparams=hparams, X=context)
loss = tf.reduce_mean(
tf.nn.sparse_softmax_cross_entropy_with_logits(
labels=context[:, 1:], logits=output['logits'][:, :-1]))
if args.val_every > 0:
val_context = tf.placeholder(tf.int32, [args.val_batch_size, None])
val_output = model.model(hparams=hparams, X=val_context)
val_loss = tf.reduce_mean(
tf.nn.sparse_softmax_cross_entropy_with_logits(
labels=val_context[:, 1:],
logits=val_output['logits'][:, :-1]))
val_loss_summary = tf.summary.scalar('val_loss', val_loss)
tf_sample = sample.sample_sequence(hparams=hparams,
length=args.sample_length,
context=context,
batch_size=args.batch_size,
temperature=1.0,
top_k=40)
all_vars = [v for v in tf.trainable_variables() if 'model' in v.name]
train_vars = [v for v in all_vars if '/h' in v.name
] if args.only_train_transformer_layers else all_vars
if args.accumulate_gradients > 1:
if args.memory_saving_gradients:
exit(
"Memory saving gradients are not implemented for gradient accumulation yet."
)
opt = AccumulatingOptimizer(
opt=tf.train.AdamOptimizer(learning_rate=args.learning_rate),
var_list=train_vars)
opt_reset = opt.reset()
opt_compute = opt.compute_gradients(loss)
opt_apply = opt.apply_gradients()
summary_loss = tf.summary.scalar('loss', opt_apply)
else:
opt = tf.train.AdamOptimizer(learning_rate=args.learning_rate)
if args.memory_saving_gradients:
opt_grads = memory_saving_gradients.gradients(loss, train_vars)
else:
opt_grads = tf.gradients(loss, train_vars)
opt_grads = list(zip(opt_grads, train_vars))
opt_apply = opt.apply_gradients(opt_grads)
summary_loss = tf.summary.scalar('loss', loss)
summary_log = tf.summary.FileWriter(
os.path.join(CHECKPOINT_DIR, args.run_name))
saver = tf.train.Saver(var_list=all_vars,
max_to_keep=5,
keep_checkpoint_every_n_hours=2)
sess.run(tf.global_variables_initializer())
if args.restore_from == 'latest':
ckpt = tf.train.latest_checkpoint(
os.path.join(CHECKPOINT_DIR, args.run_name))
if ckpt is None:
# Get fresh GPT weights if new run.
ckpt = tf.train.latest_checkpoint(os.path.join(model_name))
elif args.restore_from == 'fresh':
ckpt = tf.train.latest_checkpoint(os.path.join(model_name))
else:
ckpt = tf.train.latest_checkpoint(args.restore_from)
print('Loading checkpoint', ckpt)
saver.restore(sess, ckpt)
#print('Loading dataset...')
#chunks = load_dataset(enc, args.dataset, args.combine)
#data_sampler = Sampler(chunks)
print('Loading train dataset...')
from_name, ques_name, to_name = name_parts(args.dataset)
trn_chunks_from = load_dataset(
enc, from_name, args.combine) #if args.dataset else chunks
#trn_chunks_ques = load_dataset(enc, ques_name, args.combine) if args.dataset else chunks
trn_chunks_to = load_dataset(
enc, to_name, args.combine) #if args.dataset else chunks
skip_delimeter = True
char = '\t'
trn_data_sampler_from = SamplerVal(trn_chunks_from,
enc,
char=char,
skip_delimeter=skip_delimeter)
#trn_data_sampler_ques = SamplerVal(trn_chunks_ques, enc, char=char, skip_delimeter=skip_delimeter)
trn_data_sampler_to = SamplerVal(trn_chunks_to,
enc,
char=char,
skip_delimeter=skip_delimeter)
len_v = 0
data_sampler = []
for i in range(trn_data_sampler_from.total_size):
v = (
#enc.encode('\nQ: ') +
trn_data_sampler_from.get(i) +
#enc.encode('. \nA: ') +
trn_data_sampler_to.get(i) # +
#enc.encode('. ')
)
v = v[:HIDDEN_SIZE - 1]
len_v += len(v)
#data_sampler.extend(v) ##
data_sampler.append(v)
pass
if len_v < HIDDEN_SIZE:
mult = HIDDEN_SIZE // len_v + 1
for i in range(mult):
x = data_sampler[:]
data_sampler.extend(x)
data_sampler = Sampler([np.array(data_sampler)])
#if not args.train_special and len_v >= HIDDEN_SIZE:
# data_sampler = Sampler([np.array(data_sampler)])
if args.val_every > 0 and False:
val_chunks = load_dataset(
enc, args.val_dataset,
args.combine) if args.val_dataset else chunks
if not isinstance(data_sampler, list):
print('dataset has', data_sampler.total_size, 'tokens')
print('Training...')
if args.val_every > 0:
# Sample from validation set once with fixed seed to make
# it deterministic during training as well as across runs.
val_data_sampler = Sampler(val_chunks, seed=1)
val_batches = [[
val_data_sampler.sample(1024)
for _ in range(args.val_batch_size)
] for _ in range(args.val_batch_count)]
counter = 1
counter_path = os.path.join(CHECKPOINT_DIR, args.run_name, 'counter')
if os.path.exists(counter_path):
# Load the step number if we're resuming a run
# Add 1 so we don't immediately try to save again
with open(counter_path, 'r') as fp:
counter = int(fp.read()) + 1
def save():
maketree(os.path.join(CHECKPOINT_DIR, args.run_name))
print(
'Saving',
os.path.join(CHECKPOINT_DIR, args.run_name,
'model-{}').format(counter))
saver.save(sess,
os.path.join(CHECKPOINT_DIR, args.run_name, 'model'),
global_step=counter)
with open(counter_path, 'w') as fp:
fp.write(str(counter) + '\n')
#print(model_name, 'mn')
GPT2_DIR_X = model_name
cd = CHECKPOINT_DIR + "/" + args.run_name
if not os.path.isfile(cd + '/' + 'encoder.json'):
os.system("cp " + GPT2_DIR_X + '/' + 'encoder.json ' + cd +
'/.')
os.system('cp ' + GPT2_DIR_X + "/" + 'vocab.bpe ' + cd + '/.')
def generate_samples():
print('Generating samples...')
#context_tokens = data_sampler.sample(1)
#context_tokens = data_sampler[0]
context_tokens = trn_data_sampler_from.get(
random.randint(0, trn_data_sampler_from.total_size))
#print(enc.decode(context_tokens), len(context_tokens))
#print(args.batch_size * [context_tokens])
all_text = []
index = 0
while index < args.sample_num:
out = sess.run(
tf_sample,
feed_dict={context: args.batch_size * [context_tokens]})
for i in range(min(args.sample_num - index, args.batch_size)):
text = enc.decode(out[i])
text = '======== SAMPLE {} ========\n{}\n'.format(
index + 1, text)
all_text.append(text)
index += 1
print(text)
maketree(os.path.join(SAMPLE_DIR, args.run_name))
with open(
os.path.join(SAMPLE_DIR, args.run_name,
'samples-{}').format(counter), 'w') as fp:
fp.write('\n'.join(all_text))
def validation():
print('Calculating validation loss...')
losses = []
for batch in tqdm.tqdm(val_batches):
losses.append(
sess.run(val_loss, feed_dict={val_context: batch}))
v_val_loss = np.mean(losses)
v_summary = sess.run(val_loss_summary,
feed_dict={val_loss: v_val_loss})
summary_log.add_summary(v_summary, counter)
summary_log.flush()
print('[{counter} | {time:2.2f}] validation loss = {loss:2.2f}'.
format(counter=counter,
time=time.time() - start_time,
loss=v_val_loss))
def sample_batch():
#z = [data_sampler.sample(1024) for _ in range(args.batch_size)]
#print(len(data_sampler))
#print(len(data_sampler[0]))
z = [data_sampler[random.randint(0, args.batch_size)]]
#print(enc.decode(z[0]))
#print(z[1],'\n1' ,z[2],'\n2' ,z[3] ,len(data_sampler[0]))
#exit()
return z
avg_loss = (0.0, 0.0)
start_time = time.time()
try:
while counter != args.stop_after:
if counter % args.save_every == 0:
save()
if counter % args.sample_every == 0:
generate_samples()
pass
if args.val_every > 0 and (counter % args.val_every == 0
or counter == 1):
validation()
if args.accumulate_gradients > 1:
sess.run(opt_reset)
for _ in range(args.accumulate_gradients):
sess.run(opt_compute,
feed_dict={context: sample_batch()})
(v_loss, v_summary) = sess.run((opt_apply, summary_loss))
else:
(_, v_loss, v_summary) = sess.run(
(opt_apply, loss, summary_loss),
feed_dict={context: sample_batch()})
summary_log.add_summary(v_summary, counter)
avg_loss = (avg_loss[0] * 0.99 + v_loss,
avg_loss[1] * 0.99 + 1.0)
print(
'[{counter} | {time:2.2f}] loss={loss:2.2f} avg={avg:2.2f}'
.format(counter=counter,
time=time.time() - start_time,
loss=v_loss,
avg=avg_loss[0] / avg_loss[1]))
counter += 1
except KeyboardInterrupt:
print('\ninterrupted')
finally:
save()
def abstract_model_xy(sess, hps, feeds, train_iterator, test_iterator,
data_init, lr, f_loss):
# == Create class with static fields and methods
class m(object):
pass
m.sess = sess
m.feeds = feeds
m.lr = lr
# === Loss and optimizer
loss_train, stats_train = f_loss(train_iterator, True)
all_params = tf.trainable_variables()
if hps.gradient_checkpointing == 1:
from memory_saving_gradients import gradients
gs = gradients(loss_train, all_params)
else:
gs = tf.gradients(loss_train, all_params)
optimizer = {
'adam': optim.adam,
'adamax': optim.adamax,
'adam2': optim.adam2
}[hps.optimizer]
train_op, polyak_swap_op, ema = optimizer(all_params,
gs,
alpha=lr,
hps=hps)
if hps.direct_iterator:
m.train = lambda _lr: sess.run([train_op, stats_train], {lr: _lr})[1]
else:
def _train(_lr):
_x, _y = train_iterator()
return sess.run([train_op, stats_train], {
feeds['x']: _x,
feeds['y']: _y,
lr: _lr
})[1]
m.train = _train
m.polyak_swap = lambda: sess.run(polyak_swap_op)
# === Testing
loss_test, stats_test = f_loss(test_iterator, False, reuse=True)
if hps.direct_iterator:
m.test = lambda: sess.run(stats_test)
else:
def _test():
_x, _y = test_iterator()
return sess.run(stats_test, {feeds['x']: _x, feeds['y']: _y})
m.test = _test
# === Saving and restoring
saver = tf.train.Saver()
saver_ema = tf.train.Saver(ema.variables_to_restore())
m.save_ema = lambda path: saver_ema.save(
sess, path, write_meta_graph=False)
m.save = lambda path: saver.save(sess, path, write_meta_graph=False)
m.restore = lambda path: saver.restore(sess, path)
# === Initialize the parameters
if hps.restore_path != '':
m.restore(hps.restore_path)
else:
with Z.arg_scope([Z.get_variable_ddi, Z.actnorm], init=True):
results_init = f_loss(None, True, reuse=True)
sess.run(tf.global_variables_initializer())
sess.run(results_init, {
feeds['x']: data_init['x'],
feeds['y']: data_init['y']
})
# sess.run(hvd.broadcast_global_variables(0))
return m
def train_main(dataset,
valset,
model_name='774M',
seed=None,
batch_size=1,
batch_length=1024,
sample_length=1023,
sample_num=1,
sample_every=100,
run_name='run1',
restore_from='latest',
stop_after=None,
learning_rate=0.001,
beta1=0.9,
beta2=0.999,
epsilon=1e-08,
save_every=1000,
layers_to_train=144):
enc = encoder.get_encoder(model_name)
hparams = model.default_hparams()
with open(os.path.join('models', model_name, 'hparams.json')) as f:
hparams.override_from_dict(json.load(f))
if sample_length is None:
sample_length = hparams.n_ctx // 2
elif sample_length > hparams.n_ctx:
raise ValueError("Can't get samples longer than window size: %s" %
hparams.n_ctx)
config = tf.ConfigProto()
config.gpu_options.allow_growth = True
config.graph_options.rewrite_options.layout_optimizer = rewriter_config_pb2.RewriterConfig.OFF
with tf.Session(config=config) as sess:
context = tf.placeholder(tf.int32, [batch_size, None])
np.random.seed(seed)
tf.set_random_seed(seed)
output = model.model(hparams=hparams, X=context)
loss = tf.reduce_mean(
tf.nn.sparse_softmax_cross_entropy_with_logits(
labels=context[:, 1:], logits=output['logits'][:, :-1]))
tf_sample = sample.sample_sequence(hparams=hparams,
length=sample_length,
context=context,
batch_size=batch_size,
temperature=1.0,
top_k=40)
all_vars = [v for v in tf.trainable_variables() if 'model' in v.name]
#this line is to hopefully reduce memory usage (found on Twitter: https://twitter.com/BasedBlue/status/1169601983046672385?s=20)
train_vars = all_vars[-layers_to_train:]
print("Training", layers_to_train, "layers out of", len(all_vars))
decay_rate = adafactor_decay_rate_adam(beta2)
opt = AdafactorOptimizer(learning_rate=learning_rate,
decay_rate=decay_rate,
beta1=beta1,
name="Adafactor")
opt_grads = memory_saving_gradients.gradients(loss, train_vars)
opt_grads = list(zip(opt_grads, train_vars))
opt_apply = opt.apply_gradients(opt_grads)
summary_loss = tf.summary.scalar('loss', loss)
saver = tf.train.Saver(var_list=all_vars,
max_to_keep=5,
keep_checkpoint_every_n_hours=2)
sess.run(tf.global_variables_initializer())
if restore_from == 'latest':
ckpt = tf.train.latest_checkpoint(
os.path.join(CHECKPOINT_DIR, run_name))
if ckpt is None:
# Get fresh GPT weights if new run.
ckpt = tf.train.latest_checkpoint(
os.path.join('models', model_name))
elif restore_from == 'fresh':
ckpt = tf.train.latest_checkpoint(
os.path.join('models', model_name))
else:
ckpt = tf.train.latest_checkpoint(restore_from)
print('Loading checkpoint', ckpt)
saver.restore(sess, ckpt)
print('Loading dataset...')
chunks = load_dataset(enc, dataset)
data_sampler = Sampler(chunks)
print('dataset has', data_sampler.total_size, 'tokens')
print('Training...')
print('Loading valset...')
val_chunks = load_dataset(enc, valset)
val_data_sampler = Sampler(val_chunks)
print('valset has', val_data_sampler.total_size, 'tokens')
print('Training...')
counter = 1
if os.path.exists(os.path.join(CHECKPOINT_DIR, run_name, 'counter')):
# Load the step number if we're resuming a run
# Add 1 so we don't immediately try to save again
with open(os.path.join(CHECKPOINT_DIR, run_name, 'counter'),
'r') as fp:
counter = int(fp.read()) + 1
def save():
maketree(os.path.join(CHECKPOINT_DIR, run_name))
print(
'Saving',
os.path.join(CHECKPOINT_DIR, run_name,
'model-{}').format(counter))
saver.save(sess,
os.path.join(CHECKPOINT_DIR, run_name, 'model'),
global_step=counter)
with open(os.path.join(CHECKPOINT_DIR, run_name, 'counter'),
'w') as fp:
fp.write(str(counter) + '\n')
def generate_samples():
context_tokens = data_sampler.sample(1)
all_text = []
index = 0
while index < sample_num:
out = sess.run(
tf_sample,
feed_dict={context: batch_size * [context_tokens]})
for i in range(min(sample_num - index, batch_size)):
text = enc.decode(out[i])
text = '======== SAMPLE {} ========\n{}\n'.format(
index + 1, text)
all_text.append(text)
index += 1
print(text)
maketree(os.path.join(SAMPLE_DIR, run_name))
with open(
os.path.join(SAMPLE_DIR, run_name,
'samples-{}').format(counter), 'w') as fp:
fp.write('\n'.join(all_text))
avg_loss = (0.0, 0.0)
val_loss = (0.0, 0.0)
start_time = time.time()
best_val_loss = 99
missed_val_checkpoints = 0
try:
while counter < stop_after:
#if counter % save_every == 0:
# save()
if counter % sample_every == 0:
generate_samples()
batch = [
data_sampler.sample(batch_length)
for _ in range(batch_size)
]
_, lv = sess.run((opt_apply, loss), feed_dict={context: batch})
avg_loss = (avg_loss[0] * 0.99 + lv, avg_loss[1] * 0.99 + 1.0)
print(
'[{counter} | {time:2.2f}] loss={loss:2.4f} avg={avg:2.4f}'
.format(counter=counter,
time=time.time() - start_time,
loss=lv,
avg=avg_loss[0] / avg_loss[1]))
if counter % 5 == 0:
valbatch = [
val_data_sampler.sample(batch_length)
for _ in range(batch_size)
]
valacc = sess.run(loss, feed_dict={context: valbatch})
val_loss = (val_loss[0] * 0.99 + valacc,
val_loss[1] * 0.99 + 1.0)
av_val_loss = val_loss[0] / val_loss[1]
print(
'[{counter} | {time:2.2f}] VAL_loss={loss:2.4f} VAL_avg={avg:2.4f} best={best:2.4f}'
.format(counter=counter,
time=time.time() - start_time,
loss=valacc,
avg=av_val_loss,
best=best_val_loss))
if counter >= save_every and counter % save_every == 0: # check for validation checkpoints every save_every iterations.
if av_val_loss < best_val_loss: # got a good one from validation, save a checkpoint (every save_every)
save()
best_val_loss = av_val_loss
missed_val_checkpoints = 0
else: # missed a validation checkpoint. tolerate like 10 of these.
missed_val_checkpoints += 1
if missed_val_checkpoints > 9: # missed too many save opportunities, stop training
counter = stop_after + 1
counter += 1
except KeyboardInterrupt:
print('interrupted')
def gradients_collection(ys, xs, grad_ys=None, **kwargs):
return memory_saving_gradients.gradients(ys, xs, grad_ys,
checkpoints='collection', **kwargs)
def compute_gradients(optimizer,
loss,
var_list=None,
gate_gradients=Optimizer.GATE_OP,
aggregation_method=None,
colocate_gradients_with_ops=False,
grad_loss=None):
if callable(loss):
from tensorflow.python.eager import backprop
with backprop.GradientTape() as tape:
if var_list is not None:
tape.watch(var_list)
loss_value = loss()
# Scale loss if using a "mean" loss reduction and multiple towers.
# Have to be careful to call distribute_lib.get_loss_reduction()
# *after* loss() is evaluated, so we know what loss reduction it uses.
# TODO(josh11b): Test that we handle weight decay in a reasonable way.
if (distribute_lib.get_loss_reduction() ==
variable_scope.VariableAggregation.MEAN):
num_towers = distribution_strategy_context.get_distribution_strategy(
).num_towers
if num_towers > 1:
loss_value *= (1. / num_towers)
if var_list is None:
var_list = tape.watched_variables()
# TODO(jhseu): Figure out why GradientTape's gradients don't require loss
# to be executed.
with ops.control_dependencies([loss_value]):
grads = tape.gradient(loss_value, var_list, grad_loss)
return list(zip(grads, var_list))
# Non-callable/Tensor loss case
if context.executing_eagerly():
raise RuntimeError(
"`loss` passed to Optimizer.compute_gradients should "
"be a function when eager execution is enabled.")
# Scale loss if using a "mean" loss reduction and multiple towers.
if (distribute_lib.get_loss_reduction() ==
variable_scope.VariableAggregation.MEAN):
num_towers = distribution_strategy_context.get_distribution_strategy(
).num_towers
if num_towers > 1:
loss *= (1. / num_towers)
if gate_gradients not in [
Optimizer.GATE_NONE, Optimizer.GATE_OP,
Optimizer.GATE_GRAPH
]:
raise ValueError(
"gate_gradients must be one of: Optimizer.GATE_NONE, "
"Optimizer.GATE_OP, Optimizer.GATE_GRAPH. Not %s" %
gate_gradients)
optimizer._assert_valid_dtypes([loss])
if grad_loss is not None:
optimizer._assert_valid_dtypes([grad_loss])
if var_list is None:
var_list = (variables.trainable_variables() +
ops.get_collection(
ops.GraphKeys.TRAINABLE_RESOURCE_VARIABLES))
else:
var_list = nest.flatten(var_list)
# pylint: disable=protected-access
var_list += ops.get_collection(
ops.GraphKeys._STREAMING_MODEL_PORTS)
# pylint: enable=protected-access
from tensorflow.python.training.optimizer import _get_processor
processors = [_get_processor(v) for v in var_list]
if not var_list:
raise ValueError("No variables to optimize.")
var_refs = [p.target() for p in processors]
# original gradients computation
# grads = tf.gradients(
# loss, var_refs, grad_ys=grad_loss,
# gate_gradients=(gate_gradients == Optimizer.GATE_OP),
# aggregation_method=aggregation_method,
# colocate_gradients_with_ops=colocate_gradients_with_ops)
# using gradient check-pointing
from memory_saving_gradients import gradients
# setting outputs of different networks
tensors_to_checkpoint = self.get_tensors_to_checkpoint()
# just specifying memory as parameter fails
grads = gradients(
loss,
var_refs,
grad_ys=grad_loss,
gate_gradients=(gate_gradients == Optimizer.GATE_OP),
aggregation_method=aggregation_method,
colocate_gradients_with_ops=colocate_gradients_with_ops,
checkpoints='speed')
if gate_gradients == Optimizer.GATE_GRAPH:
grads = control_flow_ops.tuple(grads)
grads_and_vars = list(zip(grads, var_list))
optimizer._assert_valid_dtypes([
v for g, v in grads_and_vars
if g is not None and v.dtype != dtypes.resource
])
return grads_and_vars
def create_optimizer(loss, init_lr, num_train_steps, num_warmup_steps, use_tpu, prune_config_flag):
"""Creates an optimizer training op."""
global_step = tf.train.get_or_create_global_step()
learning_rate = tf.constant(value=init_lr, shape=[], dtype=tf.float32)
# Implements linear decay of the learning rate.
learning_rate = tf.train.polynomial_decay(
learning_rate,
global_step,
num_train_steps,
end_learning_rate=0.0,
power=1.0,
cycle=False)
# Implements linear warmup. I.e., if global_step < num_warmup_steps, the
# learning rate will be `global_step/num_warmup_steps * init_lr`.
if num_warmup_steps:
global_steps_int = tf.cast(global_step, tf.int32)
warmup_steps_int = tf.constant(num_warmup_steps, dtype=tf.int32)
global_steps_float = tf.cast(global_steps_int, tf.float32)
warmup_steps_float = tf.cast(warmup_steps_int, tf.float32)
warmup_percent_done = global_steps_float / warmup_steps_float
warmup_learning_rate = init_lr * warmup_percent_done
is_warmup = tf.cast(global_steps_int < warmup_steps_int, tf.float32)
learning_rate = (
(1.0 - is_warmup) * learning_rate + is_warmup * warmup_learning_rate)
# It is recommended that you use this optimizer for fine tuning, since this
# is how the model was trained (note that the Adam m/v variables are NOT
# loaded from init_checkpoint.)
optimizer = AdamWeightDecayOptimizer(
learning_rate=learning_rate,
weight_decay_rate=0.01,
beta_1=0.9,
beta_2=0.999,
epsilon=1e-6,
exclude_from_weight_decay=["LayerNorm", "layer_norm", "bias"])
if use_tpu:
optimizer = tf.contrib.tpu.CrossShardOptimizer(optimizer)
# memory_saving_gradients.DEBUG_LOGGING = True
tvars = tf.trainable_variables()
if os.getenv('DISABLE_GRAD_CHECKPOINT'):
grads = tf.gradients(loss, tvars)
else:
grads = memory_saving_gradients.gradients(loss, tvars, checkpoints='memory')
# This is how the model was pre-trained.
(grads, _) = tf.clip_by_global_norm(grads, clip_norm=1.0)
train_op = optimizer.apply_gradients(
zip(grads, tvars), global_step=global_step)
# Pruning mask update ops
if prune_config_flag:
tf.logging.info(f'Pruning with configs {prune_config_flag}')
prune_config = get_pruning_hparams().parse(prune_config_flag)
prune = Pruning(prune_config, global_step=global_step)
mask_update_op = prune.conditional_mask_update_op()
prune.add_pruning_summaries()
else:
tf.logging.info('No pruning config provided, skipping pruning')
mask_update_op = tf.no_op()
# Normally the global step update is done inside of `apply_gradients`.
# However, `AdamWeightDecayOptimizer` doesn't do this. But if you use
# a different optimizer, you should probably take this line out.
new_global_step = global_step + 1
train_op = tf.group(train_op, mask_update_op, [global_step.assign(new_global_step)])
return train_op
def main():
enc = encoder.get_encoder(args.model_name)
hparams = model.default_hparams()
hparams.batch_size=args.batch_size
hparams.seq_len=args.seq_len
##data_path
args.train_data_path=args.data_dir+args.dataset+'/train.txt'
args.eval_data_path=args.data_dir+args.dataset+'/dev.txt'
args.test_data_path=args.data_dir+args.dataset+'/test.txt'
args.eval_data_path=args.test_data_path ###Test mode only!
args.gpt_save_path=args.gpt_save_dir+args.dataset+'/'
args.dis_save_path=args.dis_save_dir+args.dataset+'/'
args.gpt_sample_dir2=args.gpt_sample_dir+args.dataset+'/'
args.dis_sample_dir2=args.dis_sample_dir+args.dataset+'/'
args.log_path=args.log_dir+args.dataset+'/'
maketree(args.gpt_save_dir)
maketree(args.dis_save_dir)
maketree(args.gpt_save_path)
maketree(args.dis_save_path)
maketree(args.gpt_sample_dir)
maketree(args.dis_sample_dir)
maketree(args.gpt_sample_dir2)
maketree(args.dis_sample_dir2)
maketree(args.log_dir)
maketree(args.log_path)
with open(os.path.join('models', args.model_name, 'hparams.json')) as f:
hparams.override_from_dict(json.load(f))
if args.sample_length > hparams.n_ctx:
raise ValueError(
"Can't get samples longer than window size: %s" % hparams.n_ctx)
if args.model_name == '345M':
args.memory_saving_gradients = True
if args.optimizer == 'adam':
args.only_train_transformer_layers = True
config = tf.ConfigProto()
config.gpu_options.allow_growth = True
config.graph_options.rewrite_options.layout_optimizer = rewriter_config_pb2.RewriterConfig.OFF
with tf.Session(config=config) as sess:
scope_discri='distri'
def get_dis_logit_and_prob_single_step(context, scope):
with tf.variable_scope(scope, reuse=tf.AUTO_REUSE):
context=tf.reshape(context, [-1, args.seq_len])
emb=tf.get_variable(name='emb', initializer=tf.random.normal([hparams.n_vocab, 32], 0, 0.02))
context_emb=tf.nn.embedding_lookup(emb, context)
logit=dis(context_emb, scope=scope_discri)
prob=tf.sigmoid(logit+1e-7)
return logit, prob
def get_dis_logit_and_prob(context, context_len, scope):
##Pay attention to context_len here. temporary changes!!!!!!!!!!!!!!!!!!!
context_mask=(1-tf.sequence_mask(context_len-1, args.seq_len-1, dtype=tf.float32))*1e3
context_mask2=tf.sequence_mask(context_len-1, args.seq_len-1, dtype=tf.float32)
ones=tf.ones(shape=[tf.shape(context_len)[0], args.seq_len], dtype=tf.int32)*enc.encoder['<|endoftext|>']
input_tensor_list=[]
for i in range(1, args.seq_len):
input_tensor_list.append(tf.concat([context[:, :i+1], ones[:,i+1:]], axis=1))
input_tensor=tf.concat(input_tensor_list, axis=0)
log_prob, _=get_dis_logit_and_prob_single_step(input_tensor, scope=scope)
log_prob=tf.transpose(tf.reshape(log_prob, [args.seq_len-1, -1]))
log_prob+=tf.cast(context_mask, tf.float32)
log_prob_min=tf.reduce_min(log_prob, axis=1)
prob_min=tf.exp(log_prob_min)
return log_prob_min, prob_min, log_prob
##Build discriminator
def build_dis_layer(scope):
context_pos_discri = tf.placeholder(tf.int32, [None, args.seq_len])
context_pos_discri_len = tf.placeholder(tf.int32, [None])
context_neg_discri = tf.placeholder(tf.int32, [None, args.seq_len])
context_neg_discri_len = tf.placeholder(tf.int32, [None])
label_pos_discri=tf.ones([tf.shape(context_pos_discri_len)[0]], dtype=tf.float32)
label_neg_discri=tf.zeros([tf.shape(context_neg_discri_len)[0]], dtype=tf.float32)
logit_pos_discri, prob_pos_discri, mask=get_dis_logit_and_prob(context_pos_discri, context_pos_discri_len, scope=scope)
logit_neg_discri, _, _=get_dis_logit_and_prob(context_neg_discri, context_neg_discri_len, scope=scope)
loss_pre_pos_discri=tf.nn.sigmoid_cross_entropy_with_logits(labels=label_pos_discri, logits=logit_pos_discri)
loss_pos_discri=tf.reduce_mean(loss_pre_pos_discri)
loss_pre_neg_discri=tf.nn.sigmoid_cross_entropy_with_logits(labels=label_neg_discri, logits=logit_neg_discri)
loss_neg_discri=tf.reduce_mean(loss_pre_neg_discri)
loss_discri=(loss_pos_discri*args.pos_loss_weight+loss_neg_discri)/(1+args.pos_loss_weight)
train_var_list_discri=[x for x in tf.global_variables() if scope in x.name]
train_op_discri=tf.train.AdamOptimizer().minimize(loss_discri, var_list=train_var_list_discri)
var_list_discri=[x for x in tf.global_variables() if scope in x.name]
initializer_discri=tf.variables_initializer(var_list_discri)
saver_discri=tf.train.Saver(var_list=var_list_discri, max_to_keep=1)
print('discri: {} build succeed!'.format(scope))
return context_pos_discri,context_pos_discri_len, context_neg_discri,context_neg_discri_len, loss_pos_discri, loss_neg_discri, loss_discri, train_op_discri, initializer_discri, saver_discri, prob_pos_discri, mask, logit_pos_discri
class dis_class:
def __init__(self, layer_num=1, scope=scope_discri):
self.model=[]
self.dis=np.zeros([layer_num], dtype=np.float32)
print(layer_num)
for i in range(layer_num):
layer={'scope': scope+str(i)}
layer['context_pos_discri'],layer['context_pos_discri_len'], layer['context_neg_discri'],layer['context_neg_discri_len'], layer['loss_pos_discri'], layer['loss_neg_discri'], layer['loss_discri'], layer['train_op_discri'], layer['initializer_discri'], layer['saver_discri'], layer['prob_pos_discri'], layer['mask'], layer['logit_pos_discri'] = build_dis_layer(scope+str(i))
self.model.append(layer)
def prob(self, context, context_len, layer=-1):
if layer==-1:
layer=len(self.model)
prob_final=tf.ones(tf.shape(context)[0], dtype=tf.float32)
for i in range(layer):
item=self.model[i]
scope=item['scope']
_, prob, _=get_dis_logit_and_prob(context, context_len, scope=scope)
prob_final*=prob
return prob_final
def log_prob_step(self, context, layer=-1):
if layer==-1:
layer=len(self.model)
prob_final=tf.ones(tf.shape(context)[0], dtype=tf.float32)
log_prob_list=[]
for i in range(layer):
item=self.model[i]
scope=item['scope']
log_prob, prob=get_dis_logit_and_prob_single_step(context, scope=scope)
log_prob_list.append(tf.expand_dims(log_prob, 1))
log_prob_final=tf.concat(log_prob_list, axis=1)
return log_prob_final
Dis=dis_class(layer_num=args.layer_num)
context = tf.placeholder(tf.int32, [None, None])
context_len=tf.placeholder(tf.int32, [None])
context_mask=tf.sequence_mask(context_len-1, args.seq_len-1, dtype=tf.float32)
context_in=context
output = model.model(hparams=hparams, X=context_in)
loss_tensor = tf.nn.sparse_softmax_cross_entropy_with_logits(labels=context[:, 1:], logits=output['logits'][:, :-1])*context_mask
loss=tf.reduce_sum(loss_tensor, axis=1)/(tf.reduce_sum(context_mask, axis=1)+1e-7)
loss_sen=tf.reduce_sum(loss)
loss=tf.reduce_mean(loss)
if args.val_every > 0:
def transform_np(x, lift=args.exponential_param):
x=x-0.5
x=x+np.abs(x)
return lift*x**2
def transform(x, lift=args.exponential_param):
x=x-0.5
x=x+tf.abs(x)
return lift*x**2
val_context = tf.placeholder(tf.int32, [args.val_batch_size, args.seq_len])
val_context_len=tf.placeholder(tf.int32, [args.batch_size])
NLL_bias=tf.placeholder(tf.float32, [])
val_context_mask=tf.sequence_mask(val_context_len-1, args.seq_len-1, dtype=tf.float32)
val_output = model.model(hparams=hparams, X=val_context)
val_loss_tensor =tf.nn.sparse_softmax_cross_entropy_with_logits(labels=val_context[:, 1:], logits=val_output['logits'][:, :-1])*val_context_mask
val_context_prob_cut=Dis.prob(val_context, val_context_len)
val_NLL_cut=tf.log(val_context_prob_cut+1e-7)
val_loss=tf.reduce_sum(val_loss_tensor, axis=1)/(tf.reduce_sum(val_context_mask, axis=1)+1e-7)
val_loss_cut=(tf.reduce_sum(val_loss_tensor, axis=1)+NLL_bias)/(tf.reduce_sum(val_context_mask, axis=1)+1e-7)-val_NLL_cut/tf.cast(val_context_len, tf.float32)
val_loss_sum=tf.reduce_sum(val_loss_tensor, axis=1)
val_loss_cut_sum=(tf.reduce_sum(val_loss_tensor, axis=1)+NLL_bias)-val_NLL_cut
val_loss_mean=tf.reduce_mean(val_loss)
val_loss_cut_mean=tf.reduce_mean(val_loss_cut)
val_loss_summary = tf.summary.scalar('val_loss', val_loss_mean)
tf_sample = sample.sample_sequence(
hparams=hparams,
length=args.seq_len,
context=context,
batch_size=args.batch_size,
temperature=1.0,
top_k=args.top_k,
top_p=args.top_p,
start_token=enc.encoder['<|endoftext|>'])
start_token=enc.encoder['<|endoftext|>']
all_vars = [v for v in tf.trainable_variables() if 'model' in v.name]
train_vars = [v for v in all_vars if '/h' in v.name] if args.only_train_transformer_layers else all_vars
if args.optimizer == 'adam':
opt = tf.train.AdamOptimizer(learning_rate=args.learning_rate)
elif args.optimizer == 'sgd':
opt = tf.train.GradientDescentOptimizer(learning_rate=args.learning_rate)
else:
exit('Bad optimizer:', args.optimizer)
if args.accumulate_gradients > 1:
if args.memory_saving_gradients:
exit("Memory saving gradients are not implemented for gradient accumulation yet.")
opt = AccumulatingOptimizer(
opt=opt,
var_list=train_vars)
opt_reset = opt.reset()
opt_compute = opt.compute_gradients(loss)
opt_apply = opt.apply_gradients()
summary_loss = tf.summary.scalar('loss', opt_apply)
else:
if args.memory_saving_gradients:
opt_grads = memory_saving_gradients.gradients(loss, train_vars)
else:
opt_grads = tf.gradients(loss, train_vars)
opt_grads = list(zip(opt_grads, train_vars))
opt_apply = opt.apply_gradients(opt_grads)
summary_loss = tf.summary.scalar('loss', loss)
summary_lr = tf.summary.scalar('learning_rate', args.learning_rate)
summaries = tf.summary.merge([summary_lr, summary_loss])
summary_log = tf.summary.FileWriter(
os.path.join(CHECKPOINT_DIR, args.run_name))
saver = tf.train.Saver(var_list=all_vars, max_to_keep=1)
sess.run(tf.global_variables_initializer())
if args.restore_from == 'latest':
ckpt = tf.train.latest_checkpoint(
os.path.join(CHECKPOINT_DIR, args.run_name))
if ckpt is None:
# Get fresh GPT weights if new run.
ckpt = tf.train.latest_checkpoint(
os.path.join('models', args.model_name))
elif args.restore_from == 'fresh':
ckpt = tf.train.latest_checkpoint(
os.path.join('models', args.model_name))
else:
ckpt = tf.train.latest_checkpoint(args.restore_from)
print('Loading checkpoint', ckpt)
saver.restore(sess, ckpt)
print('Loading dataset...')
data_list, data_len = load_dataset(enc, args.train_data_path, args.seq_len)
data_sampler = Sampler(data_list, data_len )
if args.val_every > 0:
val_data_list, val_data_len = load_dataset(enc, args.eval_data_path, args.seq_len)
print('dataset has', data_sampler.total_size, 'tokens')
print('Training...')
if args.val_every > 0:
# Sample from validation set once with fixed seed to make
# it deterministic during training as well as across runs.
val_data_sampler = Sampler(val_data_list, val_data_len, seed=1)
val_batches = [val_data_sampler.sample(args.batch_size) for _ in range(args.val_batch_count)]
counter = 0
counter_path = os.path.join(CHECKPOINT_DIR, args.run_name, 'counter')
if os.path.exists(counter_path):
# Load the step number if we're resuming a run
# Add 1 so we don't immediately try to save again
with open(counter_path, 'r') as fp:
counter = int(fp.read()) + 1
def save():
maketree(os.path.join(CHECKPOINT_DIR, args.run_name))
print(
'Saving',
os.path.join(CHECKPOINT_DIR, args.run_name,
'model-{}').format(counter))
saver.save(
sess,
os.path.join(CHECKPOINT_DIR, args.run_name, 'model'),
global_step=counter)
with open(counter_path, 'w') as fp:
fp.write(str(counter) + '\n')
def train_step_discri(layer_id=0, mask_train_epoch=0):
pos_samples, pos_samples_len=data_sampler.sample(args.batch_size)
neg_samples=generate_negative_sample(layer_id=layer_id)
neg_samples_len=get_array_len(neg_samples)
_, loss=sess.run([Dis.model[layer_id]['train_op_discri'], Dis.model[layer_id]['loss_discri']], feed_dict={Dis.model[layer_id]['context_pos_discri']: pos_samples,Dis.model[layer_id]['context_pos_discri_len']: pos_samples_len, Dis.model[layer_id]['context_neg_discri']: neg_samples, Dis.model[layer_id]['context_neg_discri_len']: neg_samples_len})
return loss
def generate_negative_samples(layer_id, generate_num=args.batch_size):
result_list=[]
generate_num_now=0
samples_mem=[]
while generate_num_now<generate_num:
t=time.time()
sample_id=generate_negative_sample(layer_id=layer_id)
samples=[]
t1=time.time()
selected_id_list=np.arange(len(sample_id))
t2=time.time()
result_list.append(sample_id[selected_id_list])
generate_num_now+=len(selected_id_list)
return np.concatenate(result_list, axis=0)[:generate_num]
def get_array_len(sample_array):
lens=[]
for item in sample_array:
for i in range(1, len(item)):
if item[i]==enc.encoder['<|endoftext|>']:
break
lens.append(i)
return np.array(lens).astype(np.int32)
def generate_discri_sample3(layer_id=-1, sample_size=10000, save_path='/mnt/cephfs_new_wj/mlnlp/miaoning/Experiment/gpt-2-sep/samples/discri/sample2.txt'):
samples=[]
while len(samples)<sample_size:
sample_id=generate_negative_sample(layer_id)
for i in range(len(sample_id)):
sample_tem=enc.decode(sample_id[i]).split('<|endoftext|>')[1].split('\n')[0]
samples.append(sample_tem)
print(len(samples))
with open(save_path, 'w') as g:
g.write('\n'.join(samples))
def eval_discri_NLL(layer_id=0):
losses_pos=[]
losses_neg=[]
for batch in tqdm.tqdm(val_batches):
pos_samples, pos_samples_len=batch
neg_samples=generate_negative_sample(layer_id=layer_id)
neg_samples_len=get_array_len(neg_samples)
loss_pos, mask=sess.run([Dis.model[layer_id]['loss_pos_discri'], Dis.model[layer_id]['mask']], feed_dict={Dis.model[layer_id]['context_pos_discri']: pos_samples, Dis.model[layer_id]['context_pos_discri_len']: pos_samples_len})
#print(mask)
loss_neg=sess.run(Dis.model[layer_id]['loss_neg_discri'], feed_dict={Dis.model[layer_id]['context_neg_discri']: neg_samples, Dis.model[layer_id]['context_neg_discri_len']: neg_samples_len})
losses_pos.append(loss_pos)
losses_neg.append(loss_neg)
return np.mean(losses_pos), np.mean(losses_neg)
def get_discri_quantile(layer_id=0, quantile=0.85):
logits_list=[]
for batch in tqdm.tqdm(val_batches):
pos_samples, pos_samples_len=batch
logits, mask=sess.run([Dis.model[layer_id]['logit_pos_discri'], Dis.model[layer_id]['mask']], feed_dict={Dis.model[layer_id]['context_pos_discri']: pos_samples, Dis.model[layer_id]['context_pos_discri_len']: pos_samples_len})
print(np.min(mask, axis=1)[:10])
print(logits[:10])
with open('mask.pkl', 'wb') as g:
pkl.dump(mask, g)
logits_list.extend(list(logits))
break
with open('logits.pkl', 'wb') as g:
pkl.dump(sorted(logits_list), g)
#print(sorted(logits_list))
print('finish')
return sorted(logits_list)[int(len(logits_list)*(1-quantile))]
def train_discri(train_step, eval_every, train_layer_list=list(range(len(Dis.model)))):
#sess.run(initializer_discri)
print('Start Discri training')
train_losses=[]
for layer_id in train_layer_list:
flag=0
for epoch in range(train_step):
if epoch % eval_every==0:
train_losses=np.mean(train_losses)
train_losses=[]
eval_NLL_pos, eval_NLL_neg=eval_discri_NLL(layer_id)
eval_loss=(eval_NLL_pos*args.pos_loss_weight+eval_NLL_neg)/(args.pos_loss_weight+1)
print('layer_id:{} discri eval loss:{}'.format(layer_id, eval_loss))
print('layer_id:{} discri NLL pos: {}, discri NLL neg: {}'.format(layer_id, eval_NLL_pos, eval_NLL_neg))
print(epoch)
if epoch==0:
eval_loss_old=eval_loss
else:
print(eval_loss, eval_loss_old)
if eval_loss<eval_loss_old:
eval_loss_old=eval_loss
save_path=args.dis_save_path+str(layer_id)+'/'
if not os.path.isdir(save_path):
os.mkdir(save_path)
Dis.model[layer_id]['saver_discri'].save(sess, save_path+'a')
print('model discri saved!')
flag=0
else:
if epoch>=200:
flag+=1
if flag>=4:
break
train_loss=train_step_discri(layer_id)
print('layer_id:{} discri train loss:{}'.format(layer_id, train_loss))
train_losses.append(train_loss)
return eval_loss_old
tf_sample_0 = sample_link.sample_sequence(
hparams=hparams,
length=args.seq_len,
context=context,
batch_size=args.batch_size,
temperature=1.0,
top_k=args.top_k,
top_p=args.top_p,
start_token=enc.encoder['<|endoftext|>'])
tf_sample_dict={}
def generate_negative_sample(layer_id=0):
##output the filtered result of layer layer_id-1
if layer_id==0:
tf_sample=tf_sample_0
sample = data_sampler.sample(args.batch_size)[0][:,0:1]
out = sess.run(
tf_sample,
feed_dict={context: sample})[:,:args.seq_len]
for i in range(len(out)):
flag=0
for j in range(len(out[i])):
if flag==2:
out[i][j]=start_token
continue
if out[i][j]==start_token:
flag+=1
return out
else:
if layer_id==-1:
layer_id=len(Dis.model)
if layer_id in tf_sample_dict:
tf_sample=tf_sample_dict[layer_id]
else:
tf_sample = sample_link.sample_sequence_ISMC_threshold(
Dis=Dis,
layer=layer_id,
hparams=hparams,
length=args.seq_len,
context=context,
batch_size=args.batch_size,
temperature=1.0,
top_k=args.top_k,
top_p=args.top_p,
start_token=enc.encoder['<|endoftext|>'])
tf_sample_dict[layer_id]=tf_sample
sample = data_sampler.sample(args.batch_size)[0][:,0:1]
out = sess.run(
tf_sample,
feed_dict={context: sample})[:,:args.seq_len]
for i in range(len(out)):
flag=0
for j in range(len(out[i])):
if flag==2:
out[i][j]=start_token
continue
if out[i][j]==start_token:
flag+=1
return out
def validation():
print('Calculating validation loss...')
start_time=time.time()
losses = []
rates=[]
for batch in tqdm.tqdm(val_batches):
losses.append(sess.run(val_loss_mean, feed_dict={val_context: batch[0], val_context_len: batch[1]}))
v_val_loss = np.mean(losses)
v_summary = sess.run(val_loss_summary, feed_dict={val_loss_mean: v_val_loss})
summary_log.add_summary(v_summary, counter)
summary_log.flush()
print(
'[{counter} | {time:2.2f}] validation loss = {loss:2.2f}'
.format(
counter=counter,
time=time.time() - start_time,
loss=v_val_loss))
return v_val_loss
def validation_cut(NLL_bias_0=0):
print('Calculating validation loss...')
losses = []
rates=[]
for batch in tqdm.tqdm(val_batches):
losses.append(sess.run(val_loss_cut_mean, feed_dict={val_context: batch[0], val_context_len: batch[1], NLL_bias:NLL_bias_0}))
v_val_loss = np.mean(losses)
print(
'[{counter} | {time:2.2f}] validation cut loss = {loss:2.2f}'
.format(
counter=counter,
time=time.time() - start_time,
loss=v_val_loss))
return v_val_loss
def sample_batch():
return [data_sampler.sample(1024) for _ in range(args.batch_size)]
def train_gpt():
val_loss_old=10000.0
avg_loss = (0.0, 0.0)
start_time = time.time()
counter=0
while True:
#pretraining
if counter % args.save_every == 0:
pass
#save()
if counter % args.sample_every == 0:
pass
#generate_samples()
if args.val_every > 0 and (counter % args.val_every == 0 or counter == 1):
val_loss_1=validation()
print(str(counter //args.val_every))
if val_loss_1>=val_loss_old:
print('pre-training ends!')
break
else:
val_loss_old=val_loss_1
saver.save(sess, args.gpt_save_path+'a')
print('save succeed!')
if args.accumulate_gradients > 1:
sess.run(opt_reset)
for _ in range(args.accumulate_gradients):
batch, batch_len=data_sampler.sample(args.batch_size)
sess.run(
opt_compute, feed_dict={context: batch, context_len:batch_len})
(v_loss, v_summary) = sess.run((opt_apply, summaries))
else:
batch, batch_len=data_sampler.sample(args.batch_size)
(_, v_loss, v_summary) = sess.run(
(opt_apply, loss, summaries),
feed_dict={context: batch, context_len:batch_len})
summary_log.add_summary(v_summary, counter)
avg_loss = (avg_loss[0] * 0.9 + v_loss,
avg_loss[1] * 0.9 + 1.0)
print(
'[{counter} | {time:2.2f}] loss={loss:2.2f} avg={avg:2.2f}'
.format(
counter=counter,
time=time.time() - start_time,
loss=v_loss,
avg=avg_loss[0] / avg_loss[1]))
counter += 1
class log_writer:
def __init__(self, path):
self.path=path
with open(path, 'w') as g:
g.write('')
def __call__(self, string, verbose=False):
with open(self.path, 'a') as g:
g.write(string+'\n')
if verbose:
print(string)
try:
if args.finetune:
#Finetune GPT-2
train_gpt()
if True:
#Restore Finetuned model
save_path=tf.train.latest_checkpoint(args.gpt_save_path)
saver.restore(sess, save_path)
print('Load gpt2 succeeded!')
if args.evaluate_finetune:
#Evaluate finetuning baseline
print(validation())
if args.evaluate_finetune:
#Calculate reverse-ppl for finetuning baseline
sample_path=args.gpt_sample_dir2+'sample.txt'
generate_discri_sample3(layer_id=0, sample_size=3000, save_path=sample_path)
rev_ppl=train.file_f(train_data_path=sample_path, val_data_path=args.eval_data_path)
Log_writer=log_writer(args.log_path+'finetune')
Log_writer('finetuning_rev_ppl: {}'.format(rev_ppl), verbose=True)
##Begin tailoring
if True:
Log_writer=log_writer(args.log_path+'discri')
for layer in range(args.layer_num):
print(layer)
if args.train_tailor:
#Train ratio estimator
train_discri(500, 10, [layer])
if True:
#Restore ratio estimator
for layer_id in range(layer+1):
save_path=args.dis_save_path+str(layer_id)+'/'
print(save_path)
save_path=tf.train.latest_checkpoint(save_path)
print(save_path)
Dis.model[layer_id]['saver_discri'].restore(sess, save_path)
if False:
#Save quantile for analysis
with open(args.dis_sample_dir2+'quantile.pkl', 'rb') as f:
pkl.load(f)
print('Load dis model succeeded!')
if True:
if layer==0:
quantile=0.85
else:
quantile=0.9
Dis.dis[layer]=get_discri_quantile(layer, quantile)
with open(args.dis_sample_dir2+'quantile.pkl', 'wb') as g:
pkl.dump(Dis.dis, g)
print(Dis.dis)
if args.evaluate_tailor:
#Generate sample for ERS and calculate reverse-ppl
sample_path=args.dis_sample_dir2+'_sample_layer_'+str(layer)
generate_discri_sample3(layer_id=layer+1, sample_size=3000, save_path=sample_path)
rev_ppl=train.file_f(train_data_path=sample_path, val_data_path=args.eval_data_path)
Log_writer('layer: {}, dis_rev_ppl: {}'.format(layer, rev_ppl), verbose=True)
except KeyboardInterrupt:
print('interrupted')
def finetune(sess,
dataset,
steps=-1,
model_name='124M',
model_dir='models',
combine=50000,
batch_size=1,
learning_rate=0.0001,
accumulate_gradients=5,
restore_from='latest',
run_name='run1',
checkpoint_dir='checkpoint',
sample_every=100,
sample_length=1023,
sample_num=1,
multi_gpu=False,
save_every=1000,
print_every=1,
max_checkpoints=1,
use_memory_saving_gradients=False,
only_train_transformer_layers=False,
optimizer='adam',
overwrite=False,
val_dataset=None,
val_batch_size=2,
val_batch_count=40,
val_every=0):
"""Finetunes the model on the given dataset.
Adapted from https://github.com/nshepperd/gpt-2/blob/finetuning/train.py.
See that file for parameter definitions.
"""
# assert model_name not in ['774M', '1558M'] or multi_gpu, "Currently, a modern single GPU cannot finetune the 774M GPT-2 model or larger."
SAMPLE_DIR = 'samples'
checkpoint_path = os.path.join(checkpoint_dir, run_name)
def maketree(path):
try:
os.makedirs(path)
except:
pass
maketree(checkpoint_path)
files = [f for f in os.listdir(checkpoint_path)]
for file in ['hparams.json', 'encoder.json', 'vocab.bpe']:
try:
shutil.copyfile(os.path.join(model_dir, model_name, file),
os.path.join(checkpoint_path, file))
except FileNotFoundError as fnf_error:
print(
"You need to download the GPT-2 model first via download_gpt2()"
)
raise (fnf_error)
enc = encoder.get_encoder(checkpoint_path)
hparams = model.default_hparams()
with open(os.path.join(checkpoint_path, 'hparams.json')) as f:
hparams.override_from_dict(json.load(f))
if sample_length > hparams.n_ctx:
raise ValueError("Can't get samples longer than window size: %s" %
hparams.n_ctx)
if model_name not in ['117M', '124M']:
use_memory_saving_gradients = True
only_train_transformer_layers = True
accumulate_gradients = 1
context = tf.compat.v1.placeholder(tf.int32, [batch_size, None])
gpus = []
if multi_gpu:
gpus = get_available_gpus()
output = model.model(hparams=hparams, X=context, gpus=gpus)
loss = tf.reduce_mean(
input_tensor=tf.nn.sparse_softmax_cross_entropy_with_logits(
labels=context[:, 1:], logits=output['logits'][:, :-1]))
# validation code
if val_every > 0:
val_context = tf.placeholder(tf.int32, [val_batch_size, None])
val_output = model.model(hparams=hparams, X=val_context,
reuse=True) # added reuse=True
val_loss = tf.reduce_mean(
tf.nn.sparse_softmax_cross_entropy_with_logits(
labels=val_context[:,
1:], logits=val_output['logits'][:, :-1]))
val_loss_summary = tf.summary.scalar('val_loss', val_loss)
tf_sample = sample.sample_sequence(hparams=hparams,
length=sample_length,
context=context,
batch_size=batch_size,
temperature=1.0,
top_k=40)
all_vars = [
v for v in tf.compat.v1.trainable_variables() if 'model' in v.name
]
train_vars = [v for v in all_vars if '/h' in v.name
] if only_train_transformer_layers else all_vars
if optimizer == 'adam':
opt = tf.compat.v1.train.AdamOptimizer(learning_rate=learning_rate)
elif optimizer == 'sgd':
opt = tf.compat.v1.train.GradientDescentOptimizer(
learning_rate=learning_rate)
if accumulate_gradients > 1:
if use_memory_saving_gradients:
exit(
"Memory saving gradients are not implemented for gradient accumulation yet."
)
opt = AccumulatingOptimizer(opt=opt, var_list=train_vars)
opt_reset = opt.reset()
opt_compute = opt.compute_gradients(loss)
opt_apply = opt.apply_gradients()
summary_loss = tf.compat.v1.summary.scalar('loss', opt_apply)
else:
if use_memory_saving_gradients:
opt_grads = memory_saving_gradients.gradients(loss, train_vars)
else:
opt_grads = tf.gradients(ys=loss, xs=train_vars)
opt_grads = list(zip(opt_grads, train_vars))
opt_apply = opt.apply_gradients(opt_grads)
summary_loss = tf.compat.v1.summary.scalar('loss', loss)
summary_log = tf.compat.v1.summary.FileWriter(checkpoint_path)
saver = tf.compat.v1.train.Saver(var_list=all_vars,
max_to_keep=max_checkpoints)
sess.run(tf.compat.v1.global_variables_initializer())
if restore_from == 'latest':
ckpt = tf.train.latest_checkpoint(checkpoint_path)
if ckpt is None:
# Get fresh GPT weights if new run.
ckpt = tf.train.latest_checkpoint(
os.path.join(model_dir, model_name))
elif restore_from == 'fresh':
ckpt = tf.train.latest_checkpoint(os.path.join(model_dir, model_name))
else:
ckpt = tf.train.latest_checkpoint(restore_from)
print('Loading checkpoint', ckpt)
saver.restore(sess, ckpt)
print('Loading dataset...')
chunks = load_dataset(enc, dataset, combine)
data_sampler = Sampler(chunks)
# validation code
if val_every > 0:
if val_dataset:
val_chunks = load_dataset(enc, val_dataset, combine)
else:
val_chunks = chunks
print('dataset has', data_sampler.total_size, 'tokens')
print('Training...')
# validation code
if val_every > 0:
# Sample from validation set once with fixed seed to make
# it deterministic during training as well as across runs.
val_data_sampler = Sampler(val_chunks, seed=1)
val_batches = [[
val_data_sampler.sample(1024) for _ in range(val_batch_size)
] for _ in range(val_batch_count)]
counter = 1
counter_path = os.path.join(checkpoint_path, 'counter')
if os.path.exists(counter_path) and restore_from == 'latest':
# Load the step number if we're resuming a run
# Add 1 so we don't immediately try to save again
with open(counter_path, 'r') as fp:
counter = int(fp.read()) + 1
counter_base = counter
def save():
maketree(checkpoint_path)
print('Saving',
os.path.join(checkpoint_path, 'model-{}').format(counter - 1))
saver.save(sess,
os.path.join(checkpoint_path, 'model'),
global_step=counter - 1)
with open(counter_path, 'w') as fp:
fp.write(str(counter - 1) + '\n')
def generate_samples():
context_tokens = data_sampler.sample(1)
all_text = []
index = 0
while index < sample_num:
out = sess.run(tf_sample,
feed_dict={context: batch_size * [context_tokens]})
for i in range(min(sample_num - index, batch_size)):
text = enc.decode(out[i])
text = '======== SAMPLE {} ========\n{}\n'.format(
index + 1, text)
all_text.append(text)
index += 1
print(text)
maketree(os.path.join(SAMPLE_DIR, run_name))
with open(
os.path.join(SAMPLE_DIR, run_name,
'samples-{}').format(counter), 'w') as fp:
fp.write('\n'.join(all_text))
# validation code
def validation():
print('Calculating validation loss...')
losses = []
for batch in tqdm(val_batches):
losses.append(sess.run(val_loss, feed_dict={val_context: batch}))
v_val_loss = np.mean(losses)
v_summary = sess.run(val_loss_summary,
feed_dict={val_loss: v_val_loss})
summary_log.add_summary(v_summary, counter)
summary_log.flush()
print('[{counter} | {time:2.2f}] validation loss = {loss:2.2f}'.format(
counter=counter, time=time.time() - start_time, loss=v_val_loss))
return v_val_loss
def sample_batch():
return [data_sampler.sample(1024) for _ in range(batch_size)]
if overwrite and restore_from == 'latest':
for file in files:
if file.startswith('model') or file.startswith('events'):
os.remove(os.path.join(checkpoint_path, file))
save()
avg_loss = (0.0, 0.0)
start_time = time.time()
#Trying out a change to finetune that saves only when validation loss decreases
if steps:
steps = int(steps)
try:
while True:
if steps > 0 and counter == (counter_base + steps):
#save()
return
# if (counter - 1) % save_every == 0 and counter > 1:
# save()
if (counter - 1) % sample_every == 0 and counter > 1:
generate_samples()
# validation code
if val_every > 0 and counter == 1:
v_val_loss = validation()
save()
elif val_every > 0 and counter == counter_base:
v_val_loss = validation()
elif val_every > 0 and (counter % val_every == 0):
new_v_val_loss = validation()
if new_v_val_loss < v_val_loss:
v_val_loss = new_v_val_loss
save()
if accumulate_gradients > 1:
sess.run(opt_reset)
for _ in range(accumulate_gradients):
sess.run(opt_compute, feed_dict={context: sample_batch()})
(v_loss, v_summary) = sess.run((opt_apply, summary_loss))
else:
(_, v_loss, v_summary) = sess.run(
(opt_apply, loss, summary_loss),
feed_dict={context: sample_batch()})
summary_log.add_summary(v_summary, counter)
if (counter % print_every == 0) or counter == 1:
avg_loss = (avg_loss[0] * 0.99 + v_loss,
avg_loss[1] * 0.99 + 1.0)
print(
'[{counter} | {time:2.2f}] loss={loss:2.2f} avg={avg:2.2f}'
.format(counter=counter,
time=time.time() - start_time,
loss=v_loss,
avg=avg_loss[0] / avg_loss[1]))
counter += 1
except KeyboardInterrupt:
print('interrupted')
save()
def train(dataset,
model_in_path,
model_out_path,
model_name='117M',
steps=1000,
combine=50000,
batch_size=1,
learning_rate=0.00002,
accumulate_gradients=1,
memory_saving_gradients=False,
only_train_transformer_layers=False,
optimizer='adam',
noise=0.0,
top_k=40,
top_p=0.0,
restore_from='latest',
sample_every=100,
sample_length=1023,
sample_num=1,
save_every=1000,
val_dataset=None):
# Reset the TF computation graph
tf.reset_default_graph()
# Get the checkpoint and sample directories
#checkpoint_dir = os.path.dirname(model_path)
#sample_dir = checkpoint_dir
#run_name = os.path.basename(model_path)
# Load the encoder
enc = get_encoder(model_in_path)
hparams = model.default_hparams()
with open(os.path.join(model_in_path, 'hparams.json')) as f:
hparams.override_from_dict(json.load(f))
if sample_length > hparams.n_ctx:
raise ValueError("Can't get samples longer than window size: %s" %
hparams.n_ctx)
# Size matters
if model_name == '345M':
memory_saving_gradients = True
if optimizer == 'adam':
only_train_transformer_layers = True
# Configure TF
config = tf.ConfigProto()
config.gpu_options.allow_growth = True
config.graph_options.rewrite_options.layout_optimizer = rewriter_config_pb2.RewriterConfig.OFF
# Start the session
with tf.Session(config=config) as sess:
context = tf.placeholder(tf.int32, [batch_size, None])
context_in = randomize(context, hparams, noise)
output = model.model(hparams=hparams, X=context_in)
loss = tf.reduce_mean(
tf.nn.sparse_softmax_cross_entropy_with_logits(
labels=context[:, 1:], logits=output['logits'][:, :-1]))
tf_sample = sample.sample_sequence(hparams=hparams,
length=sample_length,
context=context,
batch_size=batch_size,
temperature=1.0,
top_k=top_k,
top_p=top_p)
all_vars = [v for v in tf.trainable_variables() if 'model' in v.name]
train_vars = [v for v in all_vars if '/h' in v.name
] if only_train_transformer_layers else all_vars
if optimizer == 'adam':
opt = tf.train.AdamOptimizer(learning_rate=learning_rate)
elif optimizer == 'sgd':
opt = tf.train.GradientDescentOptimizer(
learning_rate=learning_rate)
else:
exit('Bad optimizer:', optimizer)
if accumulate_gradients > 1:
if memory_saving_gradients:
exit(
"Memory saving gradients are not implemented for gradient accumulation yet."
)
opt = AccumulatingOptimizer(opt=opt, var_list=train_vars)
opt_reset = opt.reset()
opt_compute = opt.compute_gradients(loss)
opt_apply = opt.apply_gradients()
summary_loss = tf.summary.scalar('loss', opt_apply)
else:
if memory_saving_gradients:
opt_grads = memory_saving_gradients.gradients(loss, train_vars)
else:
opt_grads = tf.gradients(loss, train_vars)
opt_grads = list(zip(opt_grads, train_vars))
opt_apply = opt.apply_gradients(opt_grads)
summary_loss = tf.summary.scalar('loss', loss)
summary_lr = tf.summary.scalar('learning_rate', learning_rate)
summaries = tf.summary.merge([summary_lr, summary_loss])
summary_log = tf.summary.FileWriter(
#os.path.join(checkpoint_dir, run_name)
model_out_path)
saver = tf.train.Saver(var_list=all_vars, max_to_keep=1)
sess.run(tf.global_variables_initializer())
if restore_from == 'latest':
ckpt = tf.train.latest_checkpoint(
#os.path.join(checkpoint_dir, run_name)
model_in_path)
if ckpt is None:
# Get fresh GPT weights if new run.
ckpt = tf.train.latest_checkpoint(
model_in_path) #os.path.join('models', model_name))
elif restore_from == 'fresh':
ckpt = tf.train.latest_checkpoint(
model_in_path) #os.path.join('models', model_name))
else:
ckpt = tf.train.latest_checkpoint(restore_from)
print('Loading checkpoint', ckpt)
saver.restore(sess, ckpt)
print('Loading dataset...')
chunks = load_dataset(enc, dataset, combine)
data_sampler = Sampler(chunks)
print('dataset has', data_sampler.total_size, 'tokens')
print('Training...')
counter = 1
counter_path = os.path.join(
model_in_path,
'counter') #os.path.join(checkpoint_dir, run_name, 'counter')
if restore_from == 'latest' and os.path.exists(counter_path):
# Load the step number if we're resuming a run
# Add 1 so we don't immediately try to save again
with open(counter_path, 'r') as fp:
counter = int(fp.read()) + 1
def save():
#maketree(os.path.join(checkpoint_dir, run_name))
maketree(model_out_path)
print(
'Saving',
#os.path.join(checkpoint_dir, run_name, 'model-{}').format(counter)
os.path.join(model_out_path, 'model-{}').format(counter))
saver.save(
sess,
#os.path.join(checkpoint_dir, run_name, 'model'),
os.path.join(model_out_path, 'model'),
global_step=counter)
with open(os.path.join(model_out_path, 'counter'), 'w') as fp:
fp.write(str(counter) + '\n')
def generate_samples():
print('Generating samples...')
context_tokens = data_sampler.sample(1)
all_text = []
index = 0
while index < sample_num:
out = sess.run(
tf_sample,
feed_dict={context: batch_size * [context_tokens]})
for i in range(min(sample_num - index, batch_size)):
text = enc.decode(out[i])
text = '======== SAMPLE {} ========\n{}\n'.format(
index + 1, text)
all_text.append(text)
index += 1
print(text)
#maketree(os.path.join(sample_dir, run_name))
maketree(model_out_path)
with open(
os.path.join(model_out_path, 'samples-{}').format(counter),
'w') as fp:
fp.write('\n'.join(all_text))
def sample_batch():
return [data_sampler.sample(1024) for _ in range(batch_size)]
avg_loss = (0.0, 0.0)
start_time = time.time()
stop = steps + counter
try:
while counter < stop + 1:
if counter % save_every == 0:
save()
'''
if counter % sample_every == 0:
generate_samples()
'''
if accumulate_gradients > 1:
sess.run(opt_reset)
for _ in range(accumulate_gradients):
sess.run(opt_compute,
feed_dict={context: sample_batch()})
(v_loss, v_summary) = sess.run((opt_apply, summaries))
else:
(_, v_loss, v_summary) = sess.run(
(opt_apply, loss, summaries),
feed_dict={context: sample_batch()})
summary_log.add_summary(v_summary, counter)
avg_loss = (avg_loss[0] * 0.99 + v_loss,
avg_loss[1] * 0.99 + 1.0)
print(
'[{counter} | {time:2.2f}] loss={loss:2.2f} avg={avg:2.2f}'
.format(counter=counter,
time=time.time() - start_time,
loss=v_loss,
avg=avg_loss[0] / avg_loss[1]))
counter += 1
print('done!')
save()
except KeyboardInterrupt:
print('interrupted')
save()
def main():
# initialize data loaders for train/test splits
if args.data_set == 'imagenet' and args.class_conditional:
raise("We currently don't have labels for the small imagenet data set")
if args.data_set == 'cifar':
import data.cifar10_data as cifar10_data
DataLoader = cifar10_data.DataLoader
elif args.data_set == 'imagenet':
import data.imagenet_data as imagenet_data
DataLoader = imagenet_data.DataLoader
else:
raise("unsupported dataset")
train_data = DataLoader(args.data_dir, 'train', args.batch_size * args.nr_gpu, rng=rng, shuffle=True, return_labels=args.class_conditional)
test_data = DataLoader(args.data_dir, 'test', args.batch_size * args.nr_gpu, shuffle=False, return_labels=args.class_conditional)
obs_shape = train_data.get_observation_size() # e.g. a tuple (32,32,3)
assert len(obs_shape) == 3, 'assumed right now'
# data place holders
x_init = tf.placeholder(tf.float32, shape=(args.init_batch_size,) + obs_shape)
xs = [tf.placeholder(tf.float32, shape=(args.batch_size, ) + obs_shape) for i in range(args.nr_gpu)]
# if the model is class-conditional we'll set up label placeholders + one-hot encodings 'h' to condition on
if args.class_conditional:
num_labels = train_data.get_num_labels()
y_init = tf.placeholder(tf.int32, shape=(args.init_batch_size,))
h_init = tf.one_hot(y_init, num_labels)
y_sample = np.split(np.mod(np.arange(args.batch_size*args.nr_gpu), num_labels), args.nr_gpu)
h_sample = [tf.one_hot(tf.Variable(y_sample[i], trainable=False), num_labels) for i in range(args.nr_gpu)]
ys = [tf.placeholder(tf.int32, shape=(args.batch_size,)) for i in range(args.nr_gpu)]
hs = [tf.one_hot(ys[i], num_labels) for i in range(args.nr_gpu)]
else:
h_init = None
h_sample = [None] * args.nr_gpu
hs = h_sample
# create the model
model_opt = { 'nr_resnet': args.nr_resnet, 'nr_filters': args.nr_filters, 'nr_logistic_mix': args.nr_logistic_mix, 'resnet_nonlinearity': args.resnet_nonlinearity}
model = tf.make_template('model', model_spec)
# run once for data dependent initialization of parameters
data_dependent_init = model(x_init, h_init, init=True, dropout_p=args.dropout_p, **model_opt)
# keep track of moving average
all_params = tf.trainable_variables()
ema = tf.train.ExponentialMovingAverage(decay=args.polyak_decay)
maintain_averages_op = tf.group(ema.apply(all_params))
ema_params = [ema.average(p) for p in all_params]
# get loss gradients over multiple GPUs + sampling
grads = []
loss_gen = []
loss_gen_test = []
new_x_gen = []
for i in range(args.nr_gpu):
with tf.device('/gpu:%d' % i):
if args.graph_cloning and i>0:
# already defined the graph once, use it again via template rather than redefining again
in_ = [xs[i]] + tf.global_variables()
res = gpu_template.apply(in_)
loss_train, loss_test, sx = res[:3]
grad = res[3:]
loss_gen.append(loss_train)
loss_gen_test.append(loss_test)
new_x_gen.append(sx)
grads.append(grad)
else:
# train
out = model(xs[i], hs[i], ema=None, dropout_p=args.dropout_p, **model_opt)
loss_gen.append(nn.discretized_mix_logistic_loss(tf.stop_gradient(xs[i]), out))
# gradients
grads.append(gradients(loss_gen[i], all_params))
# test
out = model(xs[i], hs[i], ema=ema, dropout_p=0., **model_opt)
loss_gen_test.append(nn.discretized_mix_logistic_loss(xs[i], out))
# sample
out = model(xs[i], h_sample[i], ema=ema, dropout_p=0, **model_opt)
new_x_gen.append(nn.sample_from_discretized_mix_logistic(out, args.nr_logistic_mix))
if args.graph_cloning:
in_ = [xs[0]] + tf.global_variables()
out_ = [loss_gen[0], loss_gen_test[0], new_x_gen[0]] + grads[0]
gpu_template = GraphTemplate(in_, outputs=out_)
# add losses and gradients together and get training updates
tf_lr = tf.placeholder(tf.float32, shape=[])
with tf.device('/gpu:0'):
for i in range(1,args.nr_gpu):
loss_gen[0] += loss_gen[i]
loss_gen_test[0] += loss_gen_test[i]
for j in range(len(grads[0])):
grads[0][j] += grads[i][j]
# training op
optimizer = tf.group(nn.adam_updates(all_params, grads[0], lr=tf_lr, mom1=0.95, mom2=0.9995), maintain_averages_op)
# convert loss to bits/dim
bits_per_dim = loss_gen[0]/(args.nr_gpu*np.log(2.)*np.prod(obs_shape)*args.batch_size)
bits_per_dim_test = loss_gen_test[0]/(args.nr_gpu*np.log(2.)*np.prod(obs_shape)*args.batch_size)
# sample from the model
def sample_from_model(sess):
x_gen = [np.zeros((args.batch_size,) + obs_shape, dtype=np.float32) for i in range(args.nr_gpu)]
for yi in range(obs_shape[0]):
for xi in range(obs_shape[1]):
new_x_gen_np = sess.run(new_x_gen, {xs[i]: x_gen[i] for i in range(args.nr_gpu)})
for i in range(args.nr_gpu):
x_gen[i][:,yi,xi,:] = new_x_gen_np[i][:,yi,xi,:]
return np.concatenate(x_gen, axis=0)
# turn numpy inputs into feed_dict for use with tensorflow
def make_feed_dict(data, init=False):
if type(data) is tuple:
x,y = data
else:
x = data
y = None
x = np.cast[np.float32]((x - 127.5) / 127.5) # input to pixelCNN is scaled from uint8 [0,255] to float in range [-1,1]
if init:
feed_dict = {x_init: x}
if y is not None:
feed_dict.update({y_init: y})
else:
x = np.split(x, args.nr_gpu)
feed_dict = {xs[i]: x[i] for i in range(args.nr_gpu)}
if y is not None:
y = np.split(y, args.nr_gpu)
feed_dict.update({ys[i]: y[i] for i in range(args.nr_gpu)})
return feed_dict
# //////////// perform training //////////////
if not os.path.exists(args.save_dir):
os.makedirs(args.save_dir)
print('starting training')
test_bpd = []
lr = args.learning_rate
saver = tf.train.Saver()
with tf.Session() as sess:
for epoch in range(args.max_epochs):
begin = time.time()
# init
if epoch == 0:
feed_dict = make_feed_dict(train_data.next(args.init_batch_size), init=True) # manually retrieve exactly init_batch_size examples
train_data.reset() # rewind the iterator back to 0 to do one full epoch
print('initializing the model...')
sess.run(tf.global_variables_initializer())
sess.run(data_dependent_init, feed_dict)
# train for one epoch
train_losses = []
counter = 0
for d in train_data:
counter+=1
feed_dict = make_feed_dict(d)
# forward/backward/update model on each gpu
lr *= args.lr_decay
feed_dict.update({ tf_lr: lr })
l,_ = sess.run([bits_per_dim, optimizer], feed_dict)
print(counter, l)
train_losses.append(l)
if counter>50:
if l>6.5:
assert False, "Test failed, expected loss 6.28537 at iteration 50"
else:
print("Test passed, loss %f (expected %f)"%(l, 6.28537))
sys.exit()
train_loss_gen = np.mean(train_losses)
# compute likelihood over test data
test_losses = []
for d in test_data:
feed_dict = make_feed_dict(d)
l = sess.run(bits_per_dim_test, feed_dict)
test_losses.append(l)
test_loss_gen = np.mean(test_losses)
test_bpd.append(test_loss_gen)
# log progress to console
print("Iteration %d, time = %ds, train bits_per_dim = %.4f, test bits_per_dim = %.4f" % (epoch, time.time()-begin, train_loss_gen, test_loss_gen))
sys.stdout.flush()
if epoch % args.save_interval == 0:
# generate samples from the model
sample_x = []
for i in range(args.num_samples):
sample_x.append(sample_from_model(sess))
sample_x = np.concatenate(sample_x,axis=0)
#img_tile = plotting.img_tile(sample_x[:100], aspect_ratio=1.0, border_color=1.0, stretch=True)
#img = plotting.plot_img(img_tile, title=args.data_set + ' samples')
#plotting.plt.savefig(os.path.join(args.save_dir,'%s_sample%d.png' % (args.data_set, epoch)))
#plotting.plt.close('all')
np.savez(os.path.join(args.save_dir,'%s_sample%d.npz' % (args.data_set, epoch)), sample_x)
# save params
saver.save(sess, args.save_dir + '/params_' + args.data_set + '.ckpt')
np.savez(args.save_dir + '/test_bpd_' + args.data_set + '.npz', test_bpd=np.array(test_bpd))
def main():
args = parser.parse_args()
enc = encoder.get_encoder(args.model_name)
hparams = model.default_hparams()
hparams.res_dropout = args.dropout
hparams.attn_dropout = args.dropout
epsilon = -1e10
if args.dtype == 'float32':
hparams.dtype = tf.float32
elif args.dtype == 'float16':
hparams.dtype = tf.float16
epsilon = -65500
elif args.dtype == 'bfloat16':
hparams.dtype = tf.bfloat16
epsilon = -65500
else:
print('Unknown dtype', args.dtype)
if args.float16:
hparams.dtype = tf.bfloat16
epsilon = -65500
with open(os.path.join('models', args.model_name, 'hparams.json')) as f:
hparams.override_from_dict(json.load(f))
if args.n_ctx >= 0:
hparams.n_ctx=args.n_ctx
if args.n_embd >= 0:
hparams.n_embd=args.n_embd
if args.n_head >= 0:
hparams.n_head=args.n_head
if args.n_layer >= 0:
hparams.n_layer=args.n_layer
if args.sample_length < 0:
args.sample_length = hparams.n_ctx - 1
if args.sample_length > hparams.n_ctx:
raise ValueError(
"Can't get samples longer than window size: %s" % hparams.n_ctx)
if args.sample_ctx < 0:
args.sample_ctx = hparams.n_ctx
if args.model_name == '345M':
args.memory_saving_gradients = True
if args.optimizer == 'adam':
args.only_train_transformer_layers = True
config = tf.ConfigProto()
if args.allow_growth:
config.gpu_options.allow_growth = True
if args.disable_layout_optimizer:
config.graph_options.rewrite_options.layout_optimizer = rewriter_config_pb2.RewriterConfig.OFF
with tflex.Session(config=config, init_tpu=args.init_tpu) as sess:
context = tf.placeholder(tf.int32, [args.batch_size, None])
context_in = randomize(context, hparams, args.noise)
output = model.model(hparams=hparams, X=context_in)
loss = tf.reduce_mean(
tf.nn.sparse_softmax_cross_entropy_with_logits(
labels=context[:, 1:], logits=output['logits'][:, :-1]))
if args.val_every > 0:
val_context = tf.placeholder(tf.int32, [args.val_batch_size, None])
val_output = model.model(hparams=hparams, X=val_context)
val_loss = tf.reduce_mean(
tf.nn.sparse_softmax_cross_entropy_with_logits(
labels=val_context[:, 1:], logits=val_output['logits'][:, :-1]))
val_loss_summary = tf.summary.scalar('val_loss', val_loss)
tf_sample = sample.sample_sequence(
hparams=hparams,
length=args.sample_length,
context=context,
batch_size=args.batch_size,
temperature=1.0,
top_k=args.top_k,
top_p=args.top_p,
epsilon=epsilon)
all_vars = [v for v in tf.trainable_variables() if 'model' in v.name]
train_vars = [v for v in all_vars if '/h' in v.name] if args.only_train_transformer_layers else all_vars
parameter_count = sum([np.prod(v.shape.as_list()) for v in train_vars])
print("This model is using %d parameters (%.2fM)" % (parameter_count, parameter_count/(1024.0*1024.0)))
with tf.variable_scope(tf.get_variable_scope().name, reuse=tf.AUTO_REUSE):
global_step = tflex.get_variable('global_step') or tf.get_variable('global_step', shape=(), dtype=tf.int32, trainable=False)
current_step = args.learning_rate_initial_step
global_step.load(current_step, session=sess)
if args.learning_rate_cos:
lr = tflex_sgdr.sgdr_decay_with_warmup(args.learning_rate, global_step,
warmup_steps=args.learning_rate_warmup, initial_period_steps=args.learning_rate_period, learning_rate_min=args.learning_rate_min)
else:
lr = tflex.get_variable('learn_rate') or tf.get_variable('learn_rate', shape=(), dtype=tf.float32, trainable=False)
lr.load(args.learning_rate, session=sess)
def update_lr(rate=None, step=None):
if not args.learning_rate_cos:
if step is None:
step = global_step.eval(session=sess)
if rate is None:
rate = args.learning_rate
if callable(rate):
rate = rate(step)
lr.load(rate, session=sess)
return lr.eval(session=sess)
@tflex.register_command
def set_learning_rate():
print("Current learn rate: %0.8f" % update_lr())
print("New learn rate?")
rate = input('')
if not rate:
print("Empty input; not changing anything.")
else:
try:
rate = float(rate)
except:
print("Invalid input; must be a float")
print("Setting learn rate to %0.8f" % rate)
args.learning_rate = rate
if args.optimizer == 'adam':
opt = tf.train.AdamOptimizer(learning_rate=lr)
elif args.optimizer == 'sgd':
opt = tf.train.GradientDescentOptimizer(learning_rate=lr)
elif args.optimizer == 'ada':
import tensor2tensor.utils.optimize
from tensor2tensor.utils import hparam
import tensor2tensor.models.research
from tensor2tensor.utils import registry
ada_hparams = registry.hparams('afx_mimic_adam')
ada_hparams.optimizer_adafactor_beta1 = 0.0
ada_hparams.optimizer_adafactor_factored = True
opt = tensor2tensor.utils.optimize.adafactor(learning_rate=lr, hparams=ada_hparams)
else:
exit('Bad optimizer:', args.optimizer)
#if tpu_addr:
# # https://pulsejet.github.io/blog/posts/tpu-without-estimator/
# from tensorflow.contrib.tpu.python.tpu import tpu_function
# tpu_function.get_tpu_context().set_number_of_shards(8)
# opt = tf.contrib.tpu.CrossShardOptimizer(opt)
if args.accumulate_gradients > 1:
if args.memory_saving_gradients:
exit("Memory saving gradients are not implemented for gradient accumulation yet.")
opt = AccumulatingOptimizer(
opt=opt,
var_list=train_vars)
opt_reset = opt.reset()
opt_compute = opt.compute_gradients(loss)
opt_apply = opt.apply_gradients()
summary_loss = tf.summary.scalar('loss', opt_apply)
else:
if args.memory_saving_gradients:
opt_grads = memory_saving_gradients.gradients(loss, train_vars)
else:
opt_grads = tf.gradients(loss, train_vars)
opt_grads = list(zip(opt_grads, train_vars))
opt_apply = opt.apply_gradients(opt_grads)
summary_loss = tf.summary.scalar('loss', loss)
summary_lr = tf.summary.scalar('learning_rate', lr)
summaries = tf.summary.merge([summary_lr, summary_loss])
summary_log = tf.summary.FileWriter(
os.path.join(CHECKPOINT_DIR, args.run_name))
if args.save_graph:
summary_log.add_graph(tf.get_default_graph())
saver = tflex.Saver(
var_list=all_vars,
max_to_keep=args.max_to_keep,
keep_checkpoint_every_n_hours=100000,
reshape=args.truncate_weights)
sess.run(tf.global_variables_initializer())
if args.restore_from == 'latest':
ckpt = tflex.latest_checkpoint(
os.path.join(CHECKPOINT_DIR, args.run_name))
if ckpt is None:
# Get fresh GPT weights if new run.
ckpt = tflex.latest_checkpoint(
os.path.join('models', args.model_name))
elif args.restore_from == 'fresh':
ckpt = tflex.latest_checkpoint(
os.path.join('models', args.model_name))
else:
ckpt = tflex.latest_checkpoint(args.restore_from)
print('Loading snapshot %s...' % ckpt)
t0 = time.time()
if not args.fresh_model:
saver.restore(sess, ckpt)
t1 = time.time()
print('Loaded in %f seconds' % (t1 - t0))
def make_sampler(dataset, enc, seed, combine):
if os.path.isdir(dataset) or dataset.endswith('.npz'):
chunks = load_dataset(enc, dataset, combine)
data_sampler = Sampler(chunks, seed=seed)
print('dataset has', data_sampler.total_size, 'tokens', len(chunks), 'chunks')
else:
data_sampler = TextSampler(dataset, enc, seed=seed)
return data_sampler
print('Loading dataset...')
seed = None if args.seed < 0 else args.seed
data_sampler = make_sampler(dataset=args.dataset, enc=enc, seed=seed, combine=args.combine)
if args.val_every > 0:
# Sample from validation set once with fixed seed to make
# it deterministic during training as well as across runs.
val_dataset = args.val_dataset if args.val_dataset else args.dataset
val_data_sampler = make_sampler(dataset=val_dataset, enc=enc, seed=1, combine=args.combine)
val_batches = [[val_data_sampler.sample(hparams.n_ctx) for _ in range(args.val_batch_size)]
for _ in range(args.val_batch_count)]
print('Training...')
counter = 1
counter_path = os.path.join(CHECKPOINT_DIR, args.run_name, 'counter')
if os.path.exists(counter_path):
# Load the step number if we're resuming a run
# Add 1 so we don't immediately try to save again
with open(counter_path, 'r') as fp:
counter = int(fp.read()) + 1
@tflex.register_command
def get_tarfile_name(checkpoint_folder):
"""Converts a folder path into a filename for a .tar archive"""
tarfile_name = checkpoint_folder.replace(os.path.sep, '_') + '.tar'
return tarfile_name
def copy_checkpoint_to_gdrive(run_name='run1', copy_folder=False):
"""Copies the checkpoint folder to a mounted Google Drive."""
#is_mounted()
checkpoint_folder = os.path.join('checkpoint', run_name)
if copy_folder:
shutil.copytree(checkpoint_folder, "/content/drive/My Drive/" + checkpoint_folder)
else:
file_path = get_tarfile_name(checkpoint_folder)
# Reference: https://stackoverflow.com/a/17081026
with tarfile.open(file_path, 'w') as tar:
tar.add(checkpoint_folder)
shutil.copyfile(file_path, "/content/drive/My Drive/" + file_path)
@tflex.register_command
def save():
maketree(os.path.join(CHECKPOINT_DIR, args.run_name))
print(
'Saving',
os.path.join(CHECKPOINT_DIR, args.run_name,
'model-{}').format(counter))
t0 = time.time()
saver.save(
sess,
os.path.join(CHECKPOINT_DIR, args.run_name, 'model'),
global_step=counter)
t1 = time.time()
print('Saved in %f seconds' % (t1 - t0))
with open(counter_path, 'w') as fp:
fp.write(str(counter) + '\n')
#copy_checkpoint_to_gdrive()
@tflex.register_command
def generate_samples():
print('Generating samples...')
context_tokens = data_sampler.sample(1)
all_text = []
index = 0
while index < args.sample_num:
out = sess.run(
tf_sample,
feed_dict={context: args.batch_size * [context_tokens]})
for i in range(min(args.sample_num - index, args.batch_size)):
text = enc.decode(out[i])
text = '======== SAMPLE {} ========\n{}\n'.format(
index + 1, text)
print(text)
all_text.append(text)
index += 1
maketree(os.path.join(SAMPLE_DIR, args.run_name))
with open(
os.path.join(SAMPLE_DIR, args.run_name,
'samples-{}').format(counter), 'w') as fp:
fp.write('\n'.join(all_text))
@tflex.register_command
def validation():
if args.val_every <= 0:
return
print('Calculating validation loss...')
losses = []
for batch in tqdm.tqdm(val_batches):
losses.append(sess.run(val_loss, feed_dict={val_context: batch}))
v_val_loss = np.mean(losses)
v_summary = sess.run(val_loss_summary, feed_dict={val_loss: v_val_loss})
summary_log.add_summary(v_summary, counter)
summary_log.flush()
print(
'{stamp} [{counter} | {time:2.4f}] validation loss = {loss:2.4f}'
.format(
stamp=timestamp(),
counter=counter,
time=time.time() - start_time,
loss=v_val_loss))
start_time = time.time()
def elapsed():
return time.time() - start_time
def say(msg):
print('{stamp} [{counter} | {time:2.4f}] {msg}'.format(counter=counter, time=elapsed(), msg=msg, stamp=timestamp()))
def sample_batch():
#return [data_sampler.sample(args.sample_ctx) for _ in range(args.batch_size)]
#say('Sampling batch...')
r = []
times = []
for _ in range(args.batch_size):
start = time.time()
sample = data_sampler.sample(args.sample_ctx)
end = time.time()
elapsed = (end - start)
r += [sample]
times += [elapsed]
total = sum(times)
avg = total / len(times)
#say('Sampled %d batches in %.4f seconds (avg per batch: %.4f)' % (args.batch_size, total, avg))
return r
prev_time = time.time()
avg_loss = (0.0, 0.0)
if args.debug_before_training:
import pdb
pdb.set_trace()
last_saved_time = elapsed()
while True:
try:
now = elapsed()
if args.save_time > 0 and (((now - last_saved_time) / 60.0) >= args.save_time):
save()
last_saved_time = now
elif args.save_every > 0 and (counter % args.save_every == 0):
save()
if counter % args.sample_every == 0:
generate_samples()
if args.val_every > 0 and (counter % args.val_every == 0 or counter == 1):
validation()
v_rate = update_lr()
if args.accumulate_gradients > 1:
#say('Running opt_reset...')
sess.run(opt_reset)
for _ in range(args.accumulate_gradients):
batch = sample_batch()
say('Running opt_compute...')
sess.run(opt_compute, feed_dict={context: batch})
say('Running opt_apply...')
(v_loss, v_summary) = sess.run((opt_apply, summaries))
else:
batch = sample_batch()
say('Running opt_apply...')
(_, v_loss, v_summary) = sess.run(
(opt_apply, loss, summaries),
feed_dict={context: batch})
if args.float16:
v_loss = tf.to_float(v_loss).eval()
summary_log.add_summary(v_summary, counter)
summary_log.flush()
avg_loss = (avg_loss[0] * 0.99 + v_loss,
avg_loss[1] * 0.99 + 1.0)
now = time.time()
print('{stamp} [{counter} | {time:2.4f} | {delta:2.2f}s | {ops:2.6f}tokens/s] loss={loss:2.4f} avg={avg:2.4f} rate={rate:0.7f} step={step}'
.format(
stamp=timestamp(),
counter=counter,
time=now - start_time,
delta=now - prev_time,
ops=args.sample_ctx * args.batch_size / (now - prev_time),
rate=v_rate,
loss=v_loss,
avg=avg_loss[0] / avg_loss[1],
step=current_step,
))
counter += 1
current_step += 1
global_step.load(current_step, session=sess)
tflex.check_commands_with_args(
session=sess,
stamp=timestamp(),
counter=counter,
time=now - start_time,
delta=now - prev_time,
ops=args.batch_size / (now - prev_time),
rate=v_rate,
loss=v_loss,
avg=avg_loss[0] / avg_loss[1],
avg_loss=avg_loss,
step=current_step,
train_vars=train_vars,
all_vars=all_vars,
args=args,
data_sampler=data_sampler,
ckpt=ckpt,
saver=saver,
)
if tflex.should_quit():
break
prev_time = now
if args.debug_print_all_vars:
print('all variables:')
print('name/shape/parameter_count')
param_count = 0
for x in tf.all_variables():
shape = x.shape.as_list()
count = np.prod(shape)
print(x.name, shape, count)
param_count += count
print('Total parameters:', param_count)
args.debug_print_all_vars = False
if args.debug_print_trainable_vars:
print('trainable variables:')
print('name/shape/parameter_count')
param_count = 0
for x in tf.trainable_variables():
shape = x.shape.as_list()
count = np.prod(shape)
print(x.name, shape, count)
param_count += count
print('Total parameters:', param_count)
args.debug_print_trainable_vars = False
except KeyboardInterrupt:
print('interrupted')
if args.save_on_ctrlc:
save()
if args.debug_on_ctrlc:
import pdb
pdb.set_trace()
else:
break
def gradients_auto(ys, xs, grad_ys=None, **kwargs):
return memory_saving_gradients.gradients(ys, xs, grad_ys,
checkpoints='memory', **kwargs)
def compute_gradients(loss, var_list, checkpoints='collection'):
# https://github.com/openai/gradient-checkpointing
from memory_saving_gradients import gradients
grads = gradients(loss, var_list, checkpoints=checkpoints)
grads_vars = list(zip(grads, var_list))
return grads_vars
def main():
args = parser.parse_args()
enc = get_encoder(model_name)
hparams = model.default_hparams()
with open(os.path.join(model_name, 'hparams.json')) as f:
hparams.override_from_dict(json.load(f))
if args.sample_length > hparams.n_ctx:
raise ValueError("Can't get samples longer than window size: %s" %
hparams.n_ctx)
if args.model_name == '345M':
args.memory_saving_gradients = True
args.only_train_transformer_layers = True
config = tf.ConfigProto()
config.gpu_options.allow_growth = True
config.graph_options.rewrite_options.layout_optimizer = rewriter_config_pb2.RewriterConfig.OFF
acc_total = 0
acc_over_time = []
loss_avg_over_time = []
if args.val_every > 0:
# val_context = tf.placeholder(tf.int32, [args.val_batch_size, None])
val_context = tf.placeholder(np.int32, [1, None])
val_output = model.model(hparams=hparams, X=val_context)
val_loss = tf.reduce_mean(
tf.nn.sparse_softmax_cross_entropy_with_logits(
labels=val_context[:,
1:], logits=val_output['logits'][:, :-1]))
val_loss_summary = tf.summary.scalar('val_loss', val_loss)
tf_sample_val = sample.sample_sequence(
hparams=hparams,
length=1, #args.sample_length,
context=val_context,
batch_size=1, #args.batch_size,
temperature=10.001,
top_k=1)
with tf.Session(config=config) as sess:
context = tf.placeholder(tf.int32, [args.batch_size, None])
output = model.model(hparams=hparams, X=context)
loss = tf.reduce_mean(
tf.nn.sparse_softmax_cross_entropy_with_logits(
labels=context[:, 1:], logits=output['logits'][:, :-1]))
tf_sample = sample.sample_sequence(hparams=hparams,
length=args.sample_length,
context=context,
batch_size=args.batch_size,
temperature=1.0,
top_k=40)
all_vars = [v for v in tf.trainable_variables() if 'model' in v.name]
train_vars = [v for v in all_vars if '/h' in v.name
] if args.only_train_transformer_layers else all_vars
if args.accumulate_gradients > 1:
if args.memory_saving_gradients:
exit(
"Memory saving gradients are not implemented for gradient accumulation yet."
)
opt = AccumulatingOptimizer(
opt=tf.train.AdamOptimizer(learning_rate=args.learning_rate),
var_list=train_vars)
opt_reset = opt.reset()
opt_compute = opt.compute_gradients(loss)
opt_apply = opt.apply_gradients()
summary_loss = tf.summary.scalar('loss', opt_apply)
else:
opt = tf.train.AdamOptimizer(learning_rate=args.learning_rate)
if args.memory_saving_gradients:
opt_grads = memory_saving_gradients.gradients(loss, train_vars)
else:
opt_grads = tf.gradients(loss, train_vars)
opt_grads = list(zip(opt_grads, train_vars))
opt_apply = opt.apply_gradients(opt_grads)
summary_loss = tf.summary.scalar('loss', loss)
summary_log = tf.summary.FileWriter(
os.path.join(CHECKPOINT_DIR, args.run_name))
saver = tf.train.Saver(var_list=all_vars,
max_to_keep=5,
keep_checkpoint_every_n_hours=2)
sess.run(tf.global_variables_initializer())
if args.restore_from == 'latest':
ckpt = tf.train.latest_checkpoint(
os.path.join(CHECKPOINT_DIR, args.run_name))
if ckpt is None:
# Get fresh GPT weights if new run.
ckpt = tf.train.latest_checkpoint(os.path.join(model_name))
elif args.restore_from == 'fresh':
ckpt = tf.train.latest_checkpoint(os.path.join(model_name))
else:
ckpt = tf.train.latest_checkpoint(args.restore_from)
print('Loading checkpoint', ckpt)
saver.restore(sess, ckpt)
print('Loading train dataset...')
from_name, ques_name, to_name = name_parts(
args.dataset) #'../data/train.from')
trn_chunks_from = load_dataset(
enc, from_name, args.combine) if args.val_dataset else chunks
trn_chunks_ques = load_dataset(
enc, ques_name, args.combine) if args.val_dataset else chunks
trn_chunks_to = load_dataset(
enc, to_name, args.combine) if args.val_dataset else chunks
skip_delimeter = True
trn_data_sampler_from = SamplerVal(trn_chunks_from,
enc,
skip_delimeter=skip_delimeter)
trn_data_sampler_ques = SamplerVal(trn_chunks_ques,
enc,
skip_delimeter=skip_delimeter)
trn_data_sampler_to = SamplerVal(trn_chunks_to,
enc,
skip_delimeter=skip_delimeter)
data_sampler = []
for i in range(trn_data_sampler_from.total_size):
v = (
trn_data_sampler_from.get(i) + trn_data_sampler_ques.get(i) +
enc.encode('. ') + trn_data_sampler_to.get(i) # +
#enc.encode('<|endoftext|>')
)
# v += [enc.encode(' ')[0] for _ in range(HIDDEN_SIZE - len(v) )]
if len(v) >= HIDDEN_SIZE - GENERATE_SIZE:
continue
v = v[:HIDDEN_SIZE - 1]
data_sampler.append(v)
pass
#chunks = load_dataset(enc, args.dataset, args.combine)
if not args.train_special:
data_sampler = Sampler([np.array(data_sampler)])
if args.val_every > 0:
print('Loading validation dataset...')
#val_chunks = load_dataset(enc, args.val_dataset, args.combine) if args.val_dataset else chunks
from_name, ques_name, to_name = name_parts(args.val_dataset)
val_chunks_from = load_dataset(
enc, from_name, args.combine) if args.val_dataset else chunks
val_chunks_ques = load_dataset(
enc, ques_name, args.combine) if args.val_dataset else chunks
val_chunks_to = load_dataset(
enc, to_name, args.combine) if args.val_dataset else chunks
if not args.train_special:
print('train dataset has', data_sampler.total_size, 'tokens')
else:
print('train dataset has', len(data_sampler), 'tokens')
print('Training...')
if args.val_every > 0:
val_data_sampler_from = SamplerVal(val_chunks_from, enc)
val_data_sampler_ques = SamplerVal(val_chunks_ques, enc)
val_data_sampler_to = SamplerVal(val_chunks_to, enc)
if args.val_batch_count == -1:
args.val_batch_count = val_data_sampler_from.total_size
val_batches = []
for i in range(args.val_batch_count):
v = (val_data_sampler_from.get(i) +
val_data_sampler_ques.get(i) + enc.encode('. ')
) #+ val_data_sampler_to.get(i)
#v += [enc.encode(' ')[0] for _ in range(HIDDEN_SIZE - len(v) )]
if len(v) >= HIDDEN_SIZE - GENERATE_SIZE:
continue
v = v[:HIDDEN_SIZE]
val_batches.append(v)
pass
print('val dataset has', len(val_batches), 'tokens')
counter = 1
counter_path = os.path.join(CHECKPOINT_DIR, args.run_name, 'counter')
if os.path.exists(counter_path):
# Load the step number if we're resuming a run
# Add 1 so we don't immediately try to save again
with open(counter_path, 'r') as fp:
counter = int(fp.read()) + 1
txt_file_path = os.path.join(CHECKPOINT_DIR, args.run_name,
args.run_name + '.summary.txt')
def save_summary(message=None):
if message is None:
txt = ''
fmt = '{valid:2.2f}'
if not os.path.exists(txt_file_path):
a = vars(args)
txt += 'Summary for ' + args.run_name + '\n'
txt += str(datetime.datetime.now()) + '\n\n'
txt += json.dumps(a) + '\n'
txt += '-----\n'
pass
txt += str(datetime.datetime.now()) + '\n'
txt += 'acc: ' + ', '.join(
[fmt.format(valid=i) for i in acc_over_time]) + '\n'
txt += 'loss: ' + ', '.join(
[fmt.format(valid=i) for i in loss_avg_over_time]) + '\n'
txt += 'counter: ' + str(counter) + '\n'
txt += 'time elapsed: ' + str(time.time() - start_time) + '\n'
txt += '-----\n'
else:
txt = message
print(txt)
with open(txt_file_path, 'a') as f:
f.write(txt + '\n')
def save():
if args.test:
return
maketree(os.path.join(CHECKPOINT_DIR, args.run_name))
print(
'Saving',
os.path.join(CHECKPOINT_DIR, args.run_name,
'model-{}').format(counter))
saver.save(sess,
os.path.join(CHECKPOINT_DIR, args.run_name, 'model'),
global_step=counter)
with open(counter_path, 'w') as fp:
fp.write(str(counter) + '\n')
'''
def generate_samples():
print('Generating samples...')
context_tokens = data_sampler.sample(1)
all_text = []
index = 0
while index < args.sample_num:
out = sess.run(
tf_sample,
feed_dict={context: args.batch_size * [context_tokens]})
for i in range(min(args.sample_num - index, args.batch_size)):
text = enc.decode(out[i])
text = '======== SAMPLE {} ========\n{}\n'.format(
index + 1, text)
all_text.append(text)
index += 1
print(text)
maketree(os.path.join(SAMPLE_DIR, args.run_name))
with open(
os.path.join(SAMPLE_DIR, args.run_name,
'samples-{}').format(counter), 'w') as fp:
fp.write('\n'.join(all_text))
'''
def print_status(word=None,
acc_total_in=0,
size=0,
v_loss_in=0.0,
shorten=False):
v_loss = v_loss_in
acc_out = 0
acc_total = 0
loss_out = 0.0
if word is None:
word = 'progress'
if acc_total_in != 0 and size != 0:
acc_out = acc_total_in / size * 100
acc_total = size
pass
if avg_loss[1] == 0.0 or avg_loss[0] == 0.0:
loss_out = 0.0
v_loss = 0.0
pass
elif not np.isnan(
avg_loss[0]) or True: # and not np.isnan(avg_loss[1]):
loss_out = avg_loss[0] / avg_loss[1]
print(word + ' [' + args.run_name + ']' +
' [{counter} | {time:2.2f}] loss={loss:2.2f} avg={avg:2.2f}'.
format(counter=counter,
time=time.time() - start_time,
loss=v_loss,
avg=loss_out),
'acc=' + str(acc_out),
end=' ')
print('total=' + str(acc_total), end=' ')
if len(acc_over_time) > 0 and not shorten:
print('last-acc=' + str(acc_over_time[-1]))
else:
print()
pass
def sample_batch(counter=0, randomize=False, pad_start=False):
#print(enc.encode('<|endoftext|>'), 'eot')
#print(data_sampler.sample(1024))
if not args.train_special:
return [
data_sampler.sample(HIDDEN_SIZE)[0]
for _ in range(args.batch_size)
]
else:
num = 0
z = []
while (len(z) > HIDDEN_SIZE or len(z) == 0) and num <= 5:
if randomize:
r = random.randint(1, 4)
else:
r = 0
#print('train special', r)
if pad_start:
pad = HIDDEN_SIZE - (len(data_sampler[counter]) - r)
else:
pad = 0
if randomize:
z = [[enc.encode(' ')[0]
for _ in range(pad)] + data_sampler[counter][:-r]
for _ in range(args.batch_size)]
if not randomize:
z = [data_sampler[counter]]
#print(enc.decode(z[0]))
num += 1
if num == 5:
z = z[len(z) - HIDDEN_SIZE:]
print('cannot get sample_batch')
break
return z
def validation_by_sample():
print('Generating validation...')
global acc_total
if args.val_with_loss:
losses = []
for batch in tqdm.tqdm(val_batches):
batch = np.reshape(batch, [1, -1])
v = sess.run(val_loss, feed_dict={val_context: batch})
#print(v, 'v')
losses.append(v)
v_val_loss = np.mean(losses)
v_summary = sess.run(val_loss_summary,
feed_dict={val_loss: v_val_loss})
summary_log.add_summary(v_summary, counter)
summary_log.flush()
print(
'[{counter} | {time:2.2f}] validation loss = {loss:2.2f}'.
format(counter=counter,
time=time.time() - start_time,
loss=v_val_loss))
acc_total = 0
generated = 0
for _ in range(len(val_batches)):
val_batches_in = val_batches[generated]
val_batches_in = val_batches_in[:1024]
context_tokens = np.reshape(val_batches_in, [1, -1])
#print(val_batches_in)
text_in = enc.decode(val_batches_in)
#print(text_in)
#print(context_tokens, 'ct1')
for x in range(GENERATE_SIZE):
out = sess.run(tf_sample_val,
feed_dict={val_context: context_tokens})
#print(out[0][-x:])
#print(enc.decode(out[0][-x:]))
context_tokens = out
compare = enc.decode(
val_data_sampler_to.get(generated)) # + ' <|endoftext|>'
compare = ' '.join(compare.split(' '))
generated += 1
text = enc.decode(out[0])
text_returned = ''
text_original = ''
if text.startswith(text_in):
text_returned = text[len(text_in):]
#print('-',text_returned,'-')
if args.train_special:
text_original = text
text = text_returned
if text.strip().endswith('.'): ## remove trailing period
text = text.strip()[:-1]
if text.strip().endswith('<|endoftext|>'):
text = text.strip()[:-len('<|endoftext|>')]
t_vals = text.split(' ')
if '<' in t_vals[-1] or '>' in t_vals[-1]:
t_vals = t_vals[:-1]
num = 0
while t_vals[-1] == '' and num < 10:
t_vals = t_vals[:-1]
num += 1
#print(t_vals)
t_vals = [i for i in t_vals if i != '']
#print(t_vals)
text = ' '.join(t_vals)
if compare.strip().endswith('.'):
compare = compare.strip()[:-1]
if compare.strip().endswith('<|endoftext|>'):
compare = compare.strip()[:-len('<|endoftext|>')]
notification = ''
len_bar = 40
if text.strip().lower().endswith(compare.strip().lower()):
acc_total += 1
notification = 'vv CORRECT vv'
len_bar = 40 - len(notification)
elif text_returned.strip().lower().startswith(
compare.strip().lower()):
acc_total += 1
notification = 'vv CORRECT_INITIAL vv'
len_bar = 40 - len(notification)
print(notification + "=" * len_bar + " SAMPLE " +
str(generated) + " " + "=" * len_bar + notification)
if args.train_special:
print(text_original)
else:
print(text)
print_status('old values',
acc_total_in=acc_total,
size=generated)
print("=" * 80)
return acc_total
pass
avg_loss = (0.0, 0.0)
start_time = time.time()
count_success = 0
count_success_with_skips = 0
acc = 0.0
try:
if args.test:
v_loss = 0.0
dataset = re.sub('train', 'test', args.dataset)
print(dataset)
from_name, ques_name, to_name = name_parts(dataset)
test_chunks_from = load_dataset(enc, from_name, args.combine)
test_chunks_ques = load_dataset(enc, ques_name, args.combine)
test_chunks_to = load_dataset(enc, to_name, args.combine)
val_data_sampler_from = SamplerVal(test_chunks_from, enc)
val_data_sampler_ques = SamplerVal(test_chunks_ques, enc)
val_data_sampler_to = SamplerVal(test_chunks_to, enc)
if args.val_batch_count == -1:
args.val_batch_count = val_data_sampler_from.total_size
val_batches = []
for i in range(args.val_batch_count):
v = (val_data_sampler_from.get(i) +
val_data_sampler_ques.get(i) + enc.encode('. ')
) # + val_data_sampler_to.get(i)
# v += [enc.encode(' ')[0] for _ in range(HIDDEN_SIZE - len(v) )]
if len(v) >= HIDDEN_SIZE - GENERATE_SIZE:
continue
val_batches.append(v)
acc_total = validation_by_sample()
acc = acc_total / len(val_batches) * 100
print(acc, 'test accuracy')
save_summary('Accuracy with test set ' + str(acc) + '\n')
exit()
while counter != args.stop_after:
#model_summary()
if counter % args.save_every == 0:
save()
if counter % args.sample_every == 0:
#generate_samples()
pass
if args.val_every > 0 and (counter % args.val_every
== 0): # or counter == 1):
acc_total = validation_by_sample()
acc = acc_total / len(val_batches) * 100
acc_over_time.append(acc)
if avg_loss[1] > 0.0:
loss_avg_over_time.append(avg_loss[0] / avg_loss[1])
else:
loss_avg_over_time.append(0)
counter_in = counter % len(val_batches)
if args.accumulate_gradients > 1:
sess.run(opt_reset)
for _ in range(args.accumulate_gradients):
sess.run(opt_compute,
feed_dict={context: sample_batch(counter_in)})
(v_loss, v_summary) = sess.run((opt_apply, summary_loss))
else:
(_, v_loss, v_summary) = sess.run(
(opt_apply, loss, summary_loss),
feed_dict={context: sample_batch(counter_in)})
summary_log.add_summary(v_summary, counter)
#if True:
if not np.isnan(avg_loss[0]) and not np.isnan(avg_loss[1]):
avg_loss = (avg_loss[0] * 0.99 + v_loss,
avg_loss[1] * 0.99 + 1.0)
if counter % args.val_every == 1:
if float(acc) == 100.0:
#save()
print('validation accuracy 100',
time.time() - start_time)
count_success += 1
count_success_with_skips += 1
if count_success >= 2 or count_success_with_skips >= 4:
#save_summary()
exit()
else:
count_success = 0
print_status(acc_total_in=acc_total,
size=len(val_batches),
v_loss_in=v_loss,
shorten=True)
counter += 1
except KeyboardInterrupt:
print('interrupted')
finally:
save()
save_summary()
print('save weights/summary and exit.')
def abstract_model_xy(sess, hps, feeds, train_iterators, test_iterators, data_inits, lr, f_loss):
# == Create class with static fields and methods
class m(object):
pass
m.sess = sess
m.feeds = feeds
m.lr = lr
# === Loss and optimizer
if hps.joint_train:
(loss_train_A, stats_train_A, eps_flatten_A, loss_train_B, stats_train_B, eps_flatten_B) \
= f_loss(train_iterators, is_training=True)
else:
(loss_train_A, stats_train_A, loss_train_B, stats_train_B) \
= f_loss(train_iterators, is_training=True)
all_params = tf.trainable_variables()
# Get train data op
def get_train_data():
x_A, y_A = train_iterators['A']()
x_B, y_B = train_iterators['B']()
return x_A, y_A, x_B, y_B
m.get_train_data = get_train_data
# A
with tf.variable_scope('optim_A'):
params_A = [param for param in all_params if 'A/' in param.name]
if hps.gradient_checkpointing == 1:
from memory_saving_gradients import gradients
gs_A = gradients(loss_train_A, params_A)
else:
gs_A = tf.gradients(loss_train_A, params_A)
m.optimizer_A = optim.Optimizer()
train_op_A, polyak_swap_op_A, ema_A = m.optimizer_A.adamax(
params_A, gs_A, alpha=lr, hps=hps)
if hps.direct_iterator:
m.train_A = lambda _lr: sess.run(
[train_op_A, stats_train_A], {lr: _lr})[1]
else:
def _train_A(_lr, _x_A, _y_A, _x_B, _y_B):
return sess.run([train_op_A, stats_train_A], {feeds['x_A']: _x_A,
feeds['y_A']: _y_A,
feeds['x_B']: _x_B,
feeds['y_B']: _y_B,
lr: _lr})[1]
m.train_A = _train_A
m.polyak_swap_A = lambda: sess.run(polyak_swap_op_A)
# B
with tf.variable_scope('optim_B'):
params_B = [param for param in all_params if 'B/' in param.name]
if hps.gradient_checkpointing == 1:
from memory_saving_gradients import gradients
gs_B = gradients(loss_train_B, params_B)
else:
gs_B = tf.gradients(loss_train_B, params_B)
m.optimizer_B = optim.Optimizer()
train_op_B, polyak_swap_op_B, ema_B = m.optimizer_B.adamax(
params_B, gs_B, alpha=lr, hps=hps)
if hps.direct_iterator:
m.train_B = lambda _lr: sess.run(
[train_op_B, stats_train_B], {lr: _lr})[1]
else:
def _train_B(_lr, _x_A, _y_A, _x_B, _y_B):
return sess.run([train_op_B, stats_train_B], {feeds['x_A']: _x_A,
feeds['y_A']: _y_A,
feeds['x_B']: _x_B,
feeds['y_B']: _y_B,
lr: _lr})[1]
m.train_B = _train_B
m.polyak_swap_B = lambda: sess.run(polyak_swap_op_B)
def _train(_lr, _x_A, _y_A, _x_B, _y_B):
return sess.run([train_op_A, train_op_B, stats_train_A, stats_train_B],
{feeds['x_A']: _x_A, feeds['y_A']: _y_A,
feeds['x_B']: _x_B, feeds['y_B']: _y_B,
lr: _lr})[-2:]
m.train = _train
# === Testing
loss_test_A, stats_test_A, loss_test_B, stats_test_B = f_loss(
test_iterators, False, reuse=True)
if hps.direct_iterator:
m.test_A = lambda: sess.run(stats_test_A)
m.test_B = lambda: sess.run(stats_test_B)
else:
# Get test data op
def get_test_data():
x_A, y_A = test_iterators['A']()
x_B, y_B = test_iterators['B']()
return x_A, y_A, x_B, y_B
m.get_test_data = get_test_data
def _test_A(_x_A, _y_A, _x_B, _y_B):
return sess.run(stats_test_A, {feeds['x_A']: _x_A,
feeds['y_A']: _y_A,
feeds['x_B']: _x_B,
feeds['y_B']: _y_B})
def _test_B(_x_A, _y_A, _x_B, _y_B):
return sess.run(stats_test_B, {feeds['x_A']: _x_A,
feeds['y_A']: _y_A,
feeds['x_B']: _x_B,
feeds['y_B']: _y_B})
m.test_A = _test_A
m.test_B = _test_B
# === Saving and restoring
with tf.variable_scope('saver_A'):
saver_A = tf.train.Saver()
saver_ema_A = tf.train.Saver(ema_A.variables_to_restore())
m.save_ema_A = lambda path_A: saver_ema_A.save(
sess, path_A, write_meta_graph=False)
m.save_A = lambda path_A: saver_A.save(
sess, path_A, write_meta_graph=False)
m.restore_A = lambda path_A: saver_A.restore(sess, path_A)
with tf.variable_scope('saver_B'):
saver_B = tf.train.Saver()
saver_ema_B = tf.train.Saver(ema_B.variables_to_restore())
m.save_ema_B = lambda path_B: saver_ema_B.save(
sess, path_B, write_meta_graph=False)
m.save_B = lambda path_B: saver_B.save(
sess, path_B, write_meta_graph=False)
m.restore_B = lambda path_B: saver_B.restore(sess, path_B)
print("After saver")
# === Initialize the parameters
if hps.restore_path_A != '':
m.restore_A(hps.restore_path_A)
if hps.restore_path_B != '':
m.restore_B(hps.restore_path_B)
if hps.restore_path_A == '' and hps.restore_path_B == '':
with Z.arg_scope([Z.get_variable_ddi, Z.actnorm], init=True):
results_init = f_loss(None, False, reuse=True, init=True)
all_params = tf.global_variables()
params_A = [param for param in all_params if 'A/' in param.name]
params_B = [param for param in all_params if 'B/' in param.name]
sess.run(tf.variables_initializer(params_A))
sess.run(tf.variables_initializer(params_B))
feeds_dict = {feeds['x_A']: data_inits['A']['x'],
feeds['y_A']: data_inits['A']['y'],
feeds['x_B']: data_inits['B']['x'],
feeds['y_B']: data_inits['B']['y']}
sess.run(results_init, feeds_dict)
sess.run(hvd.broadcast_global_variables(0))
return m
def main():
args = parser.parse_args()
enc = encoder.get_encoder(args.model_name, models_dir=args.models_dir)
hparams = model.default_hparams()
with open(os.path.join('models', args.model_name, 'hparams.json')) as f:
hparams.override_from_dict(json.load(f))
if args.sample_length > hparams.n_ctx:
raise ValueError("Can't get samples longer than window size: %s" %
hparams.n_ctx)
with tf.Session() as sess:
# Fully static shape required to make memory accounting in
# twremat accurate.
train_context = tf.placeholder(tf.int32, [args.batch_size, 1024])
train_context_in = randomize(train_context, hparams, args.noise)
train_output = model.model(hparams=hparams, X=train_context_in)
train_loss = tf.reduce_mean(
tf.nn.sparse_softmax_cross_entropy_with_logits(
labels=train_context[:, 1:],
logits=train_output['logits'][:, :-1]))
if args.val_every > 0:
val_context = tf.placeholder(tf.int32, [args.val_batch_size, None])
val_output = model.model(hparams=hparams, X=val_context)
val_loss = tf.reduce_mean(
tf.nn.sparse_softmax_cross_entropy_with_logits(
labels=val_context[:, 1:],
logits=val_output['logits'][:, :-1]))
val_loss_summary = tf.summary.scalar('val_loss', val_loss)
sample_context = tf.placeholder(tf.int32, [args.batch_size, None])
tf_sample = sample.sample_sequence(hparams=hparams,
length=args.sample_length,
context=sample_context,
batch_size=args.batch_size,
temperature=1.0,
top_k=args.top_k,
top_p=args.top_p)
all_vars = [v for v in tf.trainable_variables() if 'model' in v.name]
train_vars = [v for v in all_vars if '/h' in v.name
] if args.only_train_transformer_layers else all_vars
if args.optimizer == 'adam':
print('Using Adam optimizer', file=sys.stderr)
opt = tf.train.AdamOptimizer(learning_rate=args.learning_rate)
elif args.optimizer == 'sgd':
print('Using SGD optimizer', file=sys.stderr)
opt = tf.train.GradientDescentOptimizer(
learning_rate=args.learning_rate)
else:
exit('Bad optimizer:', args.optimizer)
if args.memory_saving_gradients:
if tf.VERSION >= '2':
exit(
'Memory saving gradients are not supported in tensorflow 2.x'
)
import memory_saving_gradients
opt_grads = memory_saving_gradients.gradients(
train_loss, train_vars)
elif args.twremat:
import tfremat
opt_grads = tf.gradients(train_loss, train_vars)
(train_loss, opt_grads) = tfremat.tf_remat(
(train_loss, opt_grads), memlimit=args.twremat_memlimit)
else:
opt_grads = tf.gradients(train_loss, train_vars)
opt_grads = list(zip(opt_grads, train_vars))
opt_apply = opt.apply_gradients(opt_grads)
summary_loss = tf.summary.scalar('loss', train_loss)
# if args.twremat:
# import tfremat
# # Applying tfremat to opt_apply has more accurate
# # accounting but is a bit iffier since side effecting ops
# # have more restrictions for correctness. If in doubt
# # revert back to version using opt_grads above.
# (opt_apply, train_loss, summary_loss) = (
# tfremat.tf_remat((opt_apply, train_loss, summary_loss), memlimit=args.twremat_memlimit))
summary_lr = tf.summary.scalar('learning_rate', args.learning_rate)
summaries = tf.summary.merge([summary_lr, summary_loss])
summary_log = tf.summary.FileWriter(
os.path.join(CHECKPOINT_DIR, args.run_name))
saver = tf.train.Saver(var_list=all_vars,
max_to_keep=5,
keep_checkpoint_every_n_hours=2)
sess.run(tf.global_variables_initializer())
if args.restore_from == 'latest':
ckpt = tf.train.latest_checkpoint(
os.path.join(CHECKPOINT_DIR, args.run_name))
if ckpt is None:
# Get fresh GPT weights if new run.
ckpt = tf.train.latest_checkpoint(
os.path.join('models', args.model_name))
elif args.restore_from == 'fresh':
ckpt = tf.train.latest_checkpoint(
os.path.join('models', args.model_name))
else:
ckpt = tf.train.latest_checkpoint(args.restore_from)
print('Loading checkpoint', ckpt)
saver.restore(sess, ckpt)
print('Loading dataset...')
chunks = load_dataset(enc,
args.dataset,
args.combine,
encoding=args.encoding)
data_sampler = Sampler(chunks)
if args.val_every > 0:
if args.val_dataset:
val_chunks = load_dataset(enc,
args.val_dataset,
args.combine,
encoding=args.encoding)
else:
val_chunks = chunks
print('dataset has', data_sampler.total_size, 'tokens')
print('Training...')
if args.val_every > 0:
# Sample from validation set once with fixed seed to make
# it deterministic during training as well as across runs.
val_data_sampler = Sampler(val_chunks, seed=1)
val_batches = [[
val_data_sampler.sample(1024)
for _ in range(args.val_batch_size)
] for _ in range(args.val_batch_count)]
counter = 1
counter_path = os.path.join(CHECKPOINT_DIR, args.run_name, 'counter')
if os.path.exists(counter_path):
# Load the step number if we're resuming a run
# Add 1 so we don't immediately try to save again
with open(counter_path, 'r') as fp:
counter = int(fp.read()) + 1
def save():
maketree(os.path.join(CHECKPOINT_DIR, args.run_name))
print(
'Saving',
os.path.join(CHECKPOINT_DIR, args.run_name,
'model-{}').format(counter))
saver.save(sess,
os.path.join(CHECKPOINT_DIR, args.run_name, 'model'),
global_step=counter)
with open(counter_path, 'w') as fp:
fp.write(str(counter) + '\n')
def generate_samples():
print('Generating samples...')
context_tokens = data_sampler.sample(1)
all_text = []
index = 0
while index < args.sample_num:
out = sess.run(tf_sample,
feed_dict={
sample_context:
args.batch_size * [context_tokens]
})
for i in range(min(args.sample_num - index, args.batch_size)):
text = enc.decode(out[i])
text = '======== SAMPLE {} ========\n{}\n'.format(
index + 1, text)
all_text.append(text)
index += 1
print(text)
maketree(os.path.join(SAMPLE_DIR, args.run_name))
with open(os.path.join(SAMPLE_DIR, args.run_name,
'samples-{}').format(counter),
'w',
encoding=args.encoding) as fp:
fp.write('\n'.join(all_text))
def validation():
print('Calculating validation loss...')
losses = []
for batch in tqdm.tqdm(val_batches):
losses.append(
sess.run(val_loss, feed_dict={val_context: batch}))
v_val_loss = np.mean(losses)
v_summary = sess.run(val_loss_summary,
feed_dict={val_loss: v_val_loss})
summary_log.add_summary(v_summary, counter)
summary_log.flush()
print('[{counter} | {time:2.2f}] validation loss = {loss:2.2f}'.
format(counter=counter,
time=time.time() - start_time,
loss=v_val_loss))
def sample_batch():
return [data_sampler.sample(1024) for _ in range(args.batch_size)]
avg_loss = (0.0, 0.0)
start_time = time.time()
# print('Evaluating grads..')
# tf2.profiler.experimental.start('logdir')
# sess.run((opt_apply, train_loss, summaries), feed_dict={train_context: sample_batch()})
# tf2.profiler.experimental.stop()
# print('Succeeded')
# exit()
try:
while True:
if counter % args.save_every == 0:
save()
if counter % args.sample_every == 0:
generate_samples()
if args.val_every > 0 and (counter % args.val_every == 0
or counter == 1):
validation()
(_, v_loss, v_summary) = sess.run(
(opt_apply, train_loss, summaries),
feed_dict={train_context: sample_batch()})
summary_log.add_summary(v_summary, counter)
avg_loss = (avg_loss[0] * 0.99 + v_loss,
avg_loss[1] * 0.99 + 1.0)
print(
'[{counter} | {time:2.2f}] loss={loss:2.2f} avg={avg:2.2f}'
.format(counter=counter,
time=time.time() - start_time,
loss=v_loss,
avg=avg_loss[0] / avg_loss[1]))
counter += 1
except KeyboardInterrupt:
print('interrupted')
save()
def grads(ys, xs, grad_ys=None, **kwargs):
return memory_saving_gradients.gradients(ys,
xs,
grad_ys,
checkpoints='speed',
**kwargs)
def main():
args = parser.parse_args()
folder_id = get_id(args.gdir)
#xmpp = SendMsgBot(jid, password, to, "Starting GPT-2")
#xmpp.register_plugin('xep_0030') # Service Discovery
#xmpp.register_plugin('xep_0199') # XMPP Ping
#xmpp.connect()
#threading = Thread(target=xmpp.process, daemon=True).start()
download_checkpoint(folder_id)
#send_m('checkpoint downloaded')
enc = encoder.get_encoder(args.model_name)
hparams = model.default_hparams()
with open(os.path.join('models', args.model_name, 'hparams.json')) as f:
hparams.override_from_dict(json.load(f))
if args.sample_length > hparams.n_ctx:
raise ValueError(
"Can't get samples longer than window size: %s" % hparams.n_ctx)
if args.model_name == '345M':
args.memory_saving_gradients = True
# if args.optimizer == 'adam':
# args.only_train_transformer_layers = True
config = tf.ConfigProto()
config.gpu_options.allow_growth = True
config.graph_options.rewrite_options.layout_optimizer = rewriter_config_pb2.RewriterConfig.OFF
with tf.Session(config=config) as sess:
context = tf.placeholder(tf.int32, [args.batch_size, None])
context_in = randomize(context, hparams, args.noise)
output = model.model(hparams=hparams, X=context_in)
loss = tf.reduce_mean(tf.nn.sparse_softmax_cross_entropy_with_logits(labels=context[:, 1:], logits=output['logits'][:, :-1]))
if args.val_every > 0:
val_context = tf.placeholder(tf.int32, [args.val_batch_size, None])
val_output = model.model(hparams=hparams, X=val_context)
val_loss = tf.reduce_mean(
tf.nn.sparse_softmax_cross_entropy_with_logits(
labels=val_context[:, 1:], logits=val_output['logits'][:, :-1]))
val_loss_summary = tf.summary.scalar('val_loss', val_loss)
tf_sample = sample.sample_sequence(
hparams=hparams,
length=args.sample_length,
context=context,
batch_size=args.batch_size,
temperature=1.0,
top_k=args.top_k,
top_p=args.top_p)
all_vars = [v for v in tf.trainable_variables() if 'model' in v.name]
train_vars = [v for v in all_vars if '/h' in v.name] if args.only_train_transformer_layers else all_vars
if args.optimizer == 'adam':
opt = tf.train.AdamOptimizer(learning_rate=args.learning_rate)
elif args.optimizer == 'sgd':
opt = tf.train.GradientDescentOptimizer(learning_rate=args.learning_rate)
else:
exit('Bad optimizer:', args.optimizer)
if args.accumulate_gradients > 1:
if args.memory_saving_gradients:
exit("Memory saving gradients are not implemented for gradient accumulation yet.")
opt = AccumulatingOptimizer(
opt=opt,
var_list=train_vars)
opt_reset = opt.reset()
opt_compute = opt.compute_gradients(loss)
opt_apply = opt.apply_gradients()
summary_loss = tf.summary.scalar('loss', opt_apply)
else:
if args.memory_saving_gradients:
opt_grads = memory_saving_gradients.gradients(loss, train_vars)
else:
opt_grads = tf.gradients(loss, train_vars)
opt_grads = list(zip(opt_grads, train_vars))
opt_apply = opt.apply_gradients(opt_grads)
summary_loss = tf.summary.scalar('loss', loss)
summary_lr = tf.summary.scalar('learning_rate', args.learning_rate)
summaries = tf.summary.merge([summary_lr, summary_loss])
summary_log = tf.summary.FileWriter(os.path.join(CHECKPOINT_DIR, args.run_name))
saver = tf.train.Saver(var_list=all_vars, max_to_keep=5, keep_checkpoint_every_n_hours=2)
sess.run(tf.global_variables_initializer())
if args.restore_from == 'latest':
ckpt = tf.train.latest_checkpoint(os.path.join(CHECKPOINT_DIR, args.run_name))
if ckpt is None:
# Get fresh GPT weights if new run.
ckpt = tf.train.latest_checkpoint(os.path.join('models', args.model_name))
elif args.restore_from == 'fresh':
ckpt = tf.train.latest_checkpoint(os.path.join('models', args.model_name))
else:
ckpt = tf.train.latest_checkpoint(args.restore_from)
print('Loading checkpoint', ckpt)
#send_m('Loading ' + str(ckpt))
saver.restore(sess, ckpt)
print('Loading dataset...')
#send_m('Loading dataset...')
#chunks = load_dataset(enc, args.dataset, args.combine)
ds_path = f'{CHECKPOINT_DIR}//run1//{args.dataset}'
chunks = load_dataset(enc, ds_path, args.combine)
data_sampler = Sampler(chunks)
print(f'{ds_path} has', data_sampler.total_size, 'tokens')
if args.val_every > 0:
val_chunks = load_dataset(enc, args.val_dataset, args.combine) if args.val_dataset else chunks
if args.enc:
print(colored(f'Trying writing Data.npz encoded from this dataset to {args.enc}', 'red'))
np.savez_compressed(args.enc, *chunks)
upload_npz(args.enc, folder_id)
#send_m(f'{args.dataset} has ' + str(data_sampler.total_size) + ' tokens' + ' Start training...')
print('Training...')
if args.val_every > 0:
# Sample from validation set once with fixed seed to make
# it deterministic during training as well as across runs.
val_data_sampler = Sampler(val_chunks, seed=1)
val_batches = [[val_data_sampler.sample(1024) for _ in range(args.val_batch_size)]
for _ in range(args.val_batch_count)]
counter = 1
counter_path = os.path.join(CHECKPOINT_DIR, args.run_name, 'counter')
if os.path.exists(counter_path):
# Load the step number if we're resuming a run
# Add 1 so we don't immediately try to save again
with open(counter_path, 'r') as fp:
counter = int(fp.read()) + 1
def save():
maketree(os.path.join(CHECKPOINT_DIR, args.run_name))
print(
'Saving',
os.path.join(CHECKPOINT_DIR, args.run_name,
'model-{}').format(counter))
saver.save(
sess,
os.path.join(CHECKPOINT_DIR, args.run_name, 'model'),
global_step=counter)
with open(counter_path, 'w') as fp:
fp.write(str(counter) + '\n')
save_gdisk(counter, folder_id)
def generate_samples():
print('Generating samples...')
#send_m('Generating samples...')
context_tokens = data_sampler.sample(1)
all_text = []
index = 0
while index < args.sample_num:
out = sess.run(
tf_sample,
feed_dict={context: args.batch_size * [context_tokens]})
for i in range(min(args.sample_num - index, args.batch_size)):
text = enc.decode(out[i])
text = '======== SAMPLE {} ========\n{}\n'.format(
index + 1, text)
all_text.append(text)
index += 1
print(text)
#send_m(text)
maketree(os.path.join(SAMPLE_DIR, args.run_name))
with open(
os.path.join(SAMPLE_DIR, args.run_name,
'samples-{}').format(counter), 'w') as fp:
fp.write('\n'.join(all_text))
def validation():
print('Calculating validation loss...')
losses = []
for batch in tqdm.tqdm(val_batches):
losses.append(sess.run(val_loss, feed_dict={val_context: batch}))
v_val_loss = np.mean(losses)
v_summary = sess.run(val_loss_summary, feed_dict={val_loss: v_val_loss})
summary_log.add_summary(v_summary, counter)
summary_log.flush()
print(
'[{counter} | {time:2.2f}] validation loss = {loss:2.2f}'
.format(
counter=counter,
time=time.time() - start_time,
loss=v_val_loss))
def sample_batch():
return [data_sampler.sample(1024) for _ in range(args.batch_size)]
avg_loss = (0.0, 0.0)
start_time = time.time()
last_time = time.time()
cur_counter, min_loss = 1, 2.0
print(colored(f'Model >>> {args.gdir}\nLearning rate is {args.learning_rate}', 'blue'))
print(colored(f'model optimizer >>> {args.optimizer}\nRestricted to train only transformer layer={args.only_train_transformer_layers}', 'blue'))
#send_m(f'Model >>> {args.model_name}\nLearning rate is {args.learning_rate}')
try:
while True:
if counter % args.save_every == 0:
save()
if check_quota():
return() # exit train
if counter % args.sample_every == 0:
generate_samples()
if args.val_every > 0 and (counter % args.val_every == 0 or counter == 1):
validation()
if args.accumulate_gradients > 1:
sess.run(opt_reset)
for _ in range(args.accumulate_gradients):
sess.run(opt_compute, feed_dict={context: sample_batch()})
(v_loss, v_summary) = sess.run((opt_apply, summaries))
else:
(_, v_loss, v_summary) = sess.run((opt_apply, loss, summaries), feed_dict={context: sample_batch()})
summary_log.add_summary(v_summary, counter)
avg_loss = (avg_loss[0] * 0.99 + v_loss,
avg_loss[1] * 0.99 + 1.0)
a_loss = avg_loss[0] / avg_loss[1]
time_all = int((time.time() - start_time) / 60)
time_iter = time.time() - last_time
stats = f'[{counter} | {cur_counter} | {time_all}m | {time_iter:2.2f}s] loss={v_loss:2.2f} avg={a_loss:2.2f}'
if not(cur_counter % 50):
print(colored(stats, 'red' if a_loss > min_loss else 'yellow'))
if a_loss < min_loss:
min_loss = a_loss
#send_m(stats)
last_time = time.time()
counter += 1
cur_counter += 1
except Exception as e:
#send_m('Stoped ' + str(e.__class__))
print('Stoped', e.__class__)