def test_baseline_nnlm_init(self):
model = NNLM(ctx_size=2,
vocab_size=4,
embed_dim=10,
h_dim=4,
num_h=2,
use_dropout=True,
embed_dropout=True,
drop_probability=0.1)
print("RUN GRAPH:")
for layer in tx.layers_to_list(model.run_outputs):
print(layer.full_str())
print("=" * 60)
print("TRAIN GRAPH:")
for layer in tx.layers_to_list(model.train_outputs):
print(layer.full_str())
print("=" * 60)
print("EVAL GRAPH:")
for layer in tx.layers_to_list(model.eval_outputs):
print(layer.full_str())
print("=" * 60)
runner = tx.ModelRunner(model)
runner.log_graph("/tmp/")
def test_nce_nnlm(self):
vocab_size = 1000
embed_size = 100
nce_samples = 10
model = NNLM(ctx_size=2,
vocab_size=vocab_size,
h_activation=tx.relu,
embed_dim=embed_size,
embed_share=True,
num_h=1,
h_dim=128,
use_f_predict=True,
use_dropout=True,
drop_probability=0.75,
embed_dropout=True,
use_nce=False,
nce_samples=nce_samples)
# model.eval_tensors.append(model.train_loss_tensors[0])
runner = tx.ModelRunner(model)
options = tf.RunOptions(trace_level=tf.RunOptions.FULL_TRACE)
# options = None
runner.set_session(runtime_stats=True, run_options=options)
runner.set_log_dir("/tmp/")
runner.log_graph()
runner.config_optimizer(
tf.train.GradientDescentOptimizer(learning_rate=0.01),
gradient_op=lambda grad: tf.clip_by_norm(grad, 1.0))
# runner.config_optimizer(tf.train.AdamOptimizer(learning_rate=0.005))
data = np.array([[0, 2], [5, 7], [9, 8], [3, 4], [1, 9]])
labels = np.array([[32], [56], [12], [2], [5]])
# data = np.array([[0, 2], [5, 7], [9, 8], [3, 4], [3, 2]])
# labels = np.array([[32], [56], [12], [2], [7]])
# data = np.array([[0, 2]])
# labels = np.array([[32]])
ppl_curve = []
for i in tqdm(range(3000)):
res = runner.train(data, labels, output_loss=True)
# print(res)
# print(res)
if i % 5 == 0:
# print(res)
res = runner.eval(data, labels)
print(res)
ppl_curve.append(np.exp(res))
ppl = sns.lineplot(x=np.array(list(range(len(ppl_curve)))),
y=np.array(ppl_curve))
print(ppl_curve)
plt.show()
def run(progress=False, **kwargs):
arg_dict.from_dict(kwargs)
args = arg_dict.to_namespace()
# ======================================================================================
# Load Params, Prepare results assets
# ======================================================================================
# Experiment parameter summary
res_param_filename = os.path.join(args.out_dir,
"params_{id}.csv".format(id=args.run_id))
with open(res_param_filename, "w") as param_file:
writer = csv.DictWriter(f=param_file, fieldnames=arg_dict.keys())
writer.writeheader()
writer.writerow(arg_dict)
param_file.flush()
# make dir for model checkpoints
if args.save_model:
model_ckpt_dir = os.path.join(args.out_dir,
"model_{id}".format(id=args.run_id))
os.makedirs(model_ckpt_dir, exist_ok=True)
model_path = os.path.join(model_ckpt_dir,
"nnlm_{id}.ckpt".format(id=args.run_id))
# start perplexity file
ppl_header = ["id", "run", "epoch", "step", "lr", "dataset", "perplexity"]
ppl_fname = os.path.join(args.out_dir,
"perplexity_{id}.csv".format(id=args.run_id))
ppl_file = open(ppl_fname, "w")
ppl_writer = csv.DictWriter(f=ppl_file, fieldnames=ppl_header)
ppl_writer.writeheader()
# ======================================================================================
# Load Corpus & Vocab
# ======================================================================================
corpus = h5py.File(os.path.join(args.corpus,
"ptb_{}.hdf5".format(args.ngram_size)),
mode='r')
vocab = marisa_trie.Trie(corpus["vocabulary"])
def data_pipeline(data,
epochs=1,
batch_size=args.batch_size,
shuffle=False):
def chunk_fn(x):
return chunk_it(x, chunk_size=batch_size * 1000)
if epochs > 1:
data = repeat_apply(chunk_fn, data, epochs)
else:
data = chunk_fn(data)
if shuffle:
data = shuffle_it(data, args.shuffle_buffer_size)
data = batch_it(data, size=batch_size, padding=False)
return data
# ======================================================================================
# MODEL
# ======================================================================================
# Activation functions
if args.h_act == "relu":
h_act = tx.relu
h_init = tx.he_normal_init()
elif args.h_act == "tanh":
h_act = tx.tanh
h_init = tx.glorot_uniform()
elif args.h_act == "elu":
h_act = tx.elu
h_init = tx.he_normal_init()
# Parameter Init
if args.embed_init == "normal":
embed_init = tx.random_normal(mean=0., stddev=args.embed_init_val)
elif args.embed_init == "uniform":
embed_init = tx.random_uniform(minval=-args.embed_init_val,
maxval=args.embed_init_val)
if args.logit_init == "normal":
logit_init = tx.random_normal(mean=0., stddev=args.logit_init_val)
elif args.logit_init == "uniform":
logit_init = tx.random_uniform(minval=-args.logit_init_val,
maxval=args.logit_init_val)
f_init = None
if args.use_f_predict:
if args.f_init == "normal":
f_init = tx.random_normal(mean=0., stddev=args.f_init_val)
elif args.f_init == "uniform":
f_init = tx.random_uniform(minval=-args.f_init_val,
maxval=args.f_init_val)
model = NNLM(ctx_size=args.ngram_size - 1,
vocab_size=len(vocab),
embed_dim=args.embed_dim,
embed_init=embed_init,
embed_share=args.embed_share,
logit_init=logit_init,
h_dim=args.h_dim,
num_h=args.num_h,
h_activation=h_act,
h_init=h_init,
use_dropout=args.dropout,
drop_probability=args.keep_prob,
embed_dropout=args.embed_dropout,
l2_loss=args.l2_loss,
l2_weight=args.l2_loss_coef,
use_f_predict=args.use_f_predict,
f_init=f_init,
logit_bias=args.logit_bias)
model_runner = tx.ModelRunner(model)
# we use an InputParam because we might want to change it during training
lr_param = tx.InputParam(value=args.lr)
if args.optimizer == "sgd":
optimizer = tf.train.GradientDescentOptimizer(
learning_rate=lr_param.tensor)
elif args.optimizer == "adam":
optimizer = tf.train.AdamOptimizer(learning_rate=lr_param.tensor,
beta1=args.optimizer_beta1,
beta2=args.optimizer_beta2,
epsilon=args.optimizer_epsilon)
elif args.optimizer == "ams":
optimizer = tx.AMSGrad(learning_rate=lr_param.tensor,
beta1=args.optimizer_beta1,
beta2=args.optimizer_beta2,
epsilon=args.optimizer_epsilon)
def clip_grad_global(grads):
grads, _ = tf.clip_by_global_norm(grads, 12)
return grads
def clip_grad_local(grad):
return tf.clip_by_norm(grad, args.clip_value)
if args.clip_grads:
if args.clip_local:
clip_fn = clip_grad_local
else:
clip_fn = clip_grad_global
if args.clip_grads:
model_runner.config_optimizer(optimizer,
optimizer_params=lr_param,
gradient_op=clip_fn,
global_gradient_op=not args.clip_local)
else:
model_runner.config_optimizer(optimizer, optimizer_params=lr_param)
# ======================================================================================
# EVALUATION
# ======================================================================================
def eval_model(runner,
dataset_it,
len_dataset=None,
display_progress=False):
if display_progress:
pb = tqdm(total=len_dataset, ncols=60)
batches_processed = 0
sum_loss = 0
for batch in dataset_it:
batch = np.array(batch, dtype=np.int64)
ctx = batch[:, :-1]
target = batch[:, -1:]
mean_loss = runner.eval(ctx, target)
sum_loss += mean_loss
if display_progress:
pb.update(args.batch_size)
batches_processed += 1
if display_progress:
pb.close()
return np.exp(sum_loss / batches_processed)
def evaluation(runner: tx.ModelRunner,
pb,
cur_epoch,
step,
display_progress=False):
##pb.write("[Eval Validation Set]")
val_data = corpus["validation"]
ppl_validation = eval_model(
runner, data_pipeline(val_data, epochs=1, shuffle=False),
len(val_data), display_progress)
res_row = {
"id": args.id,
"run": args.run,
"epoch": cur_epoch,
"step": step,
"lr": lr_param.value,
"dataset": "validation",
"perplexity": ppl_validation
}
ppl_writer.writerow(res_row)
if args.eval_test:
# pb.write("[Eval Test Set]")
test_data = corpus["test"]
ppl_test = eval_model(
runner, data_pipeline(test_data, epochs=1, shuffle=False),
len(test_data), display_progress)
res_row = {
"id": args.id,
"run": args.run,
"epoch": cur_epoch,
"step": step,
"lr": lr_param.value,
"dataset": "test",
"perplexity": ppl_test
}
ppl_writer.writerow(res_row)
ppl_file.flush()
if args.eval_test:
pb.write("test. ppl = {}".format(ppl_test))
# pb.write("valid. ppl = {}".format(ppl_validation))
return ppl_validation
# ======================================================================================
# TRAINING LOOP
# ======================================================================================
print("starting TF")
# preparing evaluation steps
# I use ceil because I make sure we have padded batches at the end
epoch_step = 0
global_step = 0
current_epoch = 0
patience = 0
cfg = tf.ConfigProto()
cfg.gpu_options.allow_growth = True
sess = tf.Session(config=cfg)
model_runner.set_session(sess)
model_runner.init_vars()
training_dset = corpus["training"]
progress = tqdm(total=len(training_dset) * args.epochs, position=1)
training_data = data_pipeline(training_dset,
epochs=args.epochs,
shuffle=True)
evals = []
try:
for ngram_batch in training_data:
epoch = progress.n // len(training_dset) + 1
# Start New Epoch
if epoch != current_epoch:
current_epoch = epoch
epoch_step = 0
if progress:
progress.write("epoch: {}".format(current_epoch))
# Eval Time
if epoch_step == 0:
current_eval = evaluation(model_runner, progress, epoch,
global_step)
evals.append(current_eval)
if global_step > 0:
if args.early_stop:
if evals[-2] - evals[-1] < args.eval_threshold:
if patience >= 3:
break
patience += 1
else:
# restart patience and adjust lr
patience = 0
# lr decay only at the start of each epoch
if args.lr_decay and len(evals) > 0:
if evals[-2] - evals[-1] < args.eval_threshold:
lr_param.value = max(
lr_param.value * args.lr_decay_rate,
args.lr_decay_threshold)
if progress:
progress.write("lr decreased to {}".format(
lr_param.value))
# ================================================
# TRAIN MODEL
# ================================================
ngram_batch = np.array(ngram_batch, dtype=np.int64)
ctx_ids = ngram_batch[:, :-1]
word_ids = ngram_batch[:, -1:]
model_runner.train(ctx_ids, word_ids)
progress.update(args.batch_size)
epoch_step += 1
global_step += 1
# if not early stop, evaluate last state of the model
if not args.early_stop or patience < 3:
evaluation(model_runner, progress, epoch, epoch_step)
ppl_file.close()
if args.save_model:
model_runner.save_model(model_name=model_path,
step=global_step,
write_state=False)
model_runner.close_session()
progress.close()
tf.reset_default_graph()
except Exception as e:
traceback.print_exc()
os.remove(ppl_file.name)
os.remove(param_file.name)
raise e
def run(**kwargs):
arg_dict.from_dict(kwargs)
args = arg_dict.to_namespace()
# ======================================================================================
# Load Corpus & Vocab
# ======================================================================================
corpus = PTBReader(path=args.corpus, mark_eos=args.mark_eos)
corpus_stats = h5py.File(os.path.join(args.corpus, "ptb_stats.hdf5"), mode='r')
vocab = marisa_trie.Trie(corpus_stats["vocabulary"])
to_ngrams_batch = partial(to_ngrams,
vocab=vocab,
ngram_size=args.ngram_size,
batch_size=args.batch_size,
epochs=1,
shuffle=False,
shuffle_buffer_size=args.shuffle_buffer_size,
enum_epoch=False)
training_len = sum(1 for _ in to_ngrams_batch(corpus.training_set, batch_size=1))
validation_len = None
test_len = None
if args.eval_progress:
validation_len = sum(1 for _ in to_ngrams_batch(corpus.validation_set, batch_size=1))
test_len = sum(1 for _ in to_ngrams_batch(corpus.test_set, batch_size=1))
# ======================================================================================
# Load Params, Prepare results assets
# ======================================================================================
# Experiment parameter summary
res_param_filename = os.path.join(args.out_dir, "params_{id}_{run}.csv".format(id=args.id, run=args.run))
with open(res_param_filename, "w") as param_file:
writer = csv.DictWriter(f=param_file, fieldnames=arg_dict.keys())
writer.writeheader()
writer.writerow(arg_dict)
param_file.flush()
# make dir for model checkpoints
if args.save_model:
model_ckpt_dir = os.path.join(args.out_dir, "model_{id}_{run}".format(id=args.id, run=args.run))
os.makedirs(model_ckpt_dir, exist_ok=True)
model_path = os.path.join(model_ckpt_dir, "nnlm_{id}_{run}.ckpt".format(id=args.id, run=args.run))
# start perplexity file
ppl_header = ["id", "run", "epoch", "step", "lr", "dataset", "perplexity"]
ppl_filename = os.path.join(args.out_dir, "perplexity_{id}_{run}.csv".format(id=args.id, run=args.run))
ppl_file = open(ppl_filename, "w")
ppl_writer = csv.DictWriter(f=ppl_file, fieldnames=ppl_header)
ppl_writer.writeheader()
# ======================================================================================
# MODEL
# ======================================================================================
# Configure weight initializers based on activation functions
if args.h_act == "relu":
h_act = tx.relu
h_init = tx.he_normal_init()
elif args.h_act == "tanh":
h_act = tx.tanh
h_init = tx.glorot_uniform()
elif args.h_act == "elu":
h_act = tx.elu
h_init = tx.he_normal_init()
elif args.h_act == "selu":
h_act = tf.nn.selu
h_init = tx.glorot_uniform()
# Configure embedding and logit weight initializers
if args.embed_init == "normal":
embed_init = tx.random_normal(mean=0.,
stddev=args.embed_init_val)
elif args.embed_init == "uniform":
embed_init = tx.random_uniform(minval=-args.embed_init_val,
maxval=args.embed_init_val)
if args.logit_init == "normal":
logit_init = tx.random_normal(mean=0.,
stddev=args.logit_init_val)
elif args.logit_init == "uniform":
logit_init = tx.random_uniform(minval=-args.logit_init_val,
maxval=args.logit_init_val)
f_init = None
if args.use_f_predict:
if args.f_init == "normal":
f_init = tx.random_normal(mean=0., stddev=args.f_init_val)
elif args.f_init == "uniform":
f_init = tx.random_uniform(minval=-args.f_init_val, maxval=args.f_init_val)
inputs = tx.Input(args.ngram_size - 1, dtype=tf.int64, name="ctx_inputs")
labels = tx.Input(1, dtype=tf.int64, name="ctx_inputs")
model = NNLM(inputs=inputs,
label_inputs=labels,
vocab_size=len(vocab),
embed_dim=args.embed_dim,
embed_init=embed_init,
embed_share=args.embed_share,
logit_init=logit_init,
h_dim=args.h_dim,
num_h=args.num_h,
h_activation=h_act,
h_init=h_init,
use_dropout=args.dropout,
drop_probability=args.drop_probability,
embed_dropout=args.embed_dropout,
l2_loss=args.l2_loss,
l2_weight=args.l2_loss_coef,
use_f_predict=args.use_f_predict,
f_init=f_init,
logit_bias=args.logit_bias,
use_nce=False)
# Input params can be changed during training by setting their value
# lr_param = tx.InputParam(init_value=args.lr)
lr_param = tensorx.train.EvalStepDecayParam(value=args.lr,
improvement_threshold=args.eval_threshold,
less_is_better=True,
decay_rate=args.lr_decay_rate,
decay_threshold=args.lr_decay_threshold)
if args.optimizer == "sgd":
optimizer = tf.train.GradientDescentOptimizer(learning_rate=lr_param.tensor)
elif args.optimizer == "adam":
optimizer = tf.train.AdamOptimizer(learning_rate=lr_param.tensor,
beta1=args.optimizer_beta1,
beta2=args.optimizer_beta2,
epsilon=args.optimizer_epsilon)
elif args.optimizer == "ams":
optimizer = tx.AMSGrad(learning_rate=lr_param.tensor,
beta1=args.optimizer_beta1,
beta2=args.optimizer_beta2,
epsilon=args.optimizer_epsilon)
def clip_grad_global(grads):
grads, _ = tf.clip_by_global_norm(grads, 12)
return grads
def clip_grad_local(grad):
return tf.clip_by_norm(grad, args.clip_value)
if args.clip_grads:
if args.clip_local:
clip_fn = clip_grad_local
else:
clip_fn = clip_grad_global
if args.clip_grads:
model.config_optimizer(optimizer, optimizer_params=lr_param,
gradient_op=clip_fn,
global_gradient_op=not args.clip_local)
else:
model.config_optimizer(optimizer, optimizer_params=lr_param)
# ======================================================================================
# EVALUATION
# ======================================================================================
def eval_model(model, dataset_it, len_dataset=None, display_progress=False):
if display_progress:
pb = tqdm(total=len_dataset, ncols=60, position=1)
batches_processed = 0
sum_loss = 0
for batch in dataset_it:
batch = np.array(batch, dtype=np.int64)
ctx = batch[:, :-1]
target = batch[:, -1:]
mean_loss = model.eval({inputs: ctx, labels: target})
sum_loss += mean_loss
if display_progress:
pb.update(args.batch_size)
batches_processed += 1
if display_progress:
pb.close()
return np.exp(sum_loss / batches_processed)
def evaluation(model: tx.Model, progress_bar, cur_epoch, step, display_progress=False):
ppl_validation = eval_model(model,
to_ngrams_batch(corpus.validation_set),
validation_len,
display_progress)
res_row = {"id": args.id, "run": args.run, "epoch": cur_epoch, "step": step, "lr": lr_param.value,
"dataset": "validation",
"perplexity": ppl_validation}
ppl_writer.writerow(res_row)
if args.eval_test:
# pb.write("[Eval Test Set]")
ppl_test = eval_model(model, to_ngrams(corpus.test_set), test_len, display_progress)
res_row = {"id": args.id, "run": args.run, "epoch": cur_epoch, "step": step, "lr": lr_param.value,
"dataset": "test",
"perplexity": ppl_test}
ppl_writer.writerow(res_row)
ppl_file.flush()
if args.eval_test:
progress_bar.set_postfix({"test PPL ": ppl_test})
# pb.write("valid. ppl = {}".format(ppl_validation))
return ppl_validation
# ======================================================================================
# TRAINING LOOP
# ======================================================================================
# print("Starting TensorFlow Session")
# preparing evaluation steps
# I use ceil because I make sure we have padded batches at the end
epoch_step = 0
global_step = 0
current_epoch = 0
patience = 0
cfg = tf.ConfigProto()
cfg.gpu_options.allow_growth = True
sess = tf.Session(config=cfg)
model.set_session(sess)
model.init_vars()
progress = tqdm(total=training_len * args.epochs, position=args.pid + 1, disable=not args.display_progress)
training_data = to_ngrams_batch(corpus.training_set,
epochs=args.epochs,
shuffle=args.shuffle,
enum_epoch=True)
evaluations = []
try:
for i, ngram_batch in training_data:
epoch = i + 1
# Start New Epoch
if epoch != current_epoch:
current_epoch = epoch
epoch_step = 0
if args.display_progress:
progress.set_postfix({"epoch": current_epoch})
# ================================================
# EVALUATION
# ================================================
if epoch_step == 0:
current_eval = evaluation(model, progress, epoch, global_step,
display_progress=args.eval_progress)
evaluations.append(current_eval)
lr_param.update(current_eval)
# print(lr_param.eval_history)
# print("improvement ", lr_param.eval_improvement())
if global_step > 0:
if args.early_stop and epoch > 1:
if lr_param.eval_improvement() < lr_param.improvement_threshold:
if patience >= 3:
break
patience += 1
else:
patience = 0
# ================================================
# TRAIN MODEL
# ================================================
ngram_batch = np.array(ngram_batch, dtype=np.int64)
ctx_ids = ngram_batch[:, :-1]
word_ids = ngram_batch[:, -1:]
model.train({inputs: ctx_ids, labels: word_ids})
progress.update(args.batch_size)
epoch_step += 1
global_step += 1
# if not early stop, evaluate last state of the model
if not args.early_stop or patience < 3:
current_eval = evaluation(model, progress, epoch, epoch_step)
evaluations.append(current_eval)
ppl_file.close()
if args.save_model:
model.save_model(model_name=model_path, step=global_step, write_state=False)
model.close_session()
progress.close()
tf.reset_default_graph()
# return the best validation evaluation
return min(evaluations)
except Exception as e:
traceback.print_exc()
os.remove(ppl_file.name)
os.remove(param_file.name)
raise e
if args.use_f_predict:
if args.f_init == "normal":
f_init = tx.random_normal(mean=0., stddev=args.f_init_val)
elif args.f_init == "uniform":
f_init = tx.random_uniform(minval=-args.f_init_val,
maxval=args.f_init_val)
model = NNLM(ctx_size=args.ngram_size - 1,
vocab_size=len(vocab),
embed_dim=args.embed_dim,
embed_init=embed_init,
embed_share=args.embed_share,
logit_init=logit_init,
h_dim=args.h_dim,
num_h=args.num_h,
h_activation=h_act,
h_init=h_init,
use_dropout=args.dropout,
drop_probability=args.keep_prob,
embed_dropout=args.embed_dropout,
l2_loss=args.l2_loss,
l2_weight=args.l2_loss_coef,
use_f_predict=args.use_f_predict,
f_init=f_init)
model_runner = tx.ModelRunner(model)
# we use an InputParam because we might want to change it during training
lr_param = tx.InputParam(value=args.lr)
if args.optimizer == "sgd":
optimizer = tf.train.GradientDescentOptimizer(
def test_nnlm(self):
vocab_size = 4
ctx_size = 22
batch_size = 2
embed_dim = 512
h_dim = 128
inputs = tx.Input(ctx_size, dtype=tf.int64, name="ctx_inputs")
labels = tx.Input(1, dtype=tf.int64, name="ctx_inputs")
model = NNLM(inputs=inputs,
label_inputs=labels,
vocab_size=vocab_size,
embed_dim=embed_dim,
embed_share=True,
use_f_predict=True,
h_dim=h_dim,
use_dropout=False,
drop_probability=0.9,
embed_dropout=False,
use_nce=True,
nce_samples=2)
model.set_session(runtime_stats=True)
model.set_log_dir("/tmp/")
model.log_graph()
options = tf.RunOptions(trace_level=tf.RunOptions.FULL_TRACE)
# options = None
model.set_session(runtime_stats=True, run_options=options)
model.config_optimizer(
tf.train.AdamOptimizer(learning_rate=0.005),
gradient_op=lambda grad: tf.clip_by_norm(grad, 4.0))
input_data = np.random.randint(0, vocab_size, [batch_size, ctx_size])
label_data = np.random.randint(0, vocab_size, [batch_size, 1])
with self.cached_session():
for _ in tqdm(range(1)):
eval1 = model.eval({inputs: input_data, labels: label_data})
eval2 = model.eval({inputs: input_data, labels: label_data})
result = model.train({
inputs: input_data,
labels: label_data
},
write_summaries=True)
# print(list(map(str,model.train_graph.output_layers)))
self.assertArrayEqual(eval1, eval2)
self.assertArrayNotEqual(result, eval2)