def main(_):
if len(sys.argv) < 3:
print(
'Usage: ucdoc_saved_model.py [--model_version=y] --data_dir=xxx --ckpt_dir=xxx --saved_dir=xxx'
)
sys.exit(-1)
if FLAGS.training_iteration <= 0:
print('Please specify a positive value for training iteration.')
sys.exit(-1)
if FLAGS.model_version <= 0:
print('Please specify a positive value for version number.')
sys.exit(-1)
# create deploy model first
with tf.variable_scope('input') as inp_scope:
with tf.device("/cpu:0"):
#inp = VarFeeder.read_vars("data/vars")
inp = VarFeeder.read_vars(FLAGS.data_dir)
pipe = InputPipe(inp,
ucdoc_features(inp),
inp.hits.shape[0],
mode=ModelMode.PREDICT,
batch_size=FLAGS.batch_size,
n_epoch=1,
verbose=False,
train_completeness_threshold=0.01,
predict_window=FLAGS.predict_window,
predict_completeness_threshold=0.0,
train_window=FLAGS.train_window,
back_offset=FLAGS.predict_window + 1)
asgd_decay = 0.99 if FLAGS.asgd else None
if FLAGS.n_models == 1:
model = Model(pipe,
build_from_set(FLAGS.hparam_set),
is_train=False,
seed=1,
asgd_decay=asgd_decay)
else:
models = []
for i in range(FLAGS.n_models):
prefix = f"m_{i}"
with tf.variable_scope(prefix) as scope:
models.append(
Model(pipe,
build_from_set(FLAGS.hparam_set),
is_train=False,
seed=1,
asgd_decay=asgd_decay,
graph_prefix=prefix))
model = models[FLAGS.target_model]
# load checkpoint model from training
#ckpt_path = FLAGS.ckpt_dir
print('loading checkpoint model...')
ckpt_file = tf.train.latest_checkpoint(FLAGS.ckpt_dir)
#graph = tf.Graph()
graph = model.predictions.graph
saver = tf.train.Saver(name='deploy_saver', var_list=None)
with tf.Session(config=tf.ConfigProto(gpu_options=tf.GPUOptions(
allow_growth=True))) as sess:
pipe.load_vars(sess)
pipe.init_iterator(sess)
saver.restore(sess, ckpt_file)
print('Done loading checkpoint model')
export_path_base = FLAGS.saved_dir
export_path = os.path.join(
tf.compat.as_bytes(export_path_base),
tf.compat.as_bytes(str(FLAGS.model_version)))
print('Exporting trained model to', export_path)
if os.path.isdir(export_path):
shutil.rmtree(export_path)
builder = tf.saved_model.builder.SavedModelBuilder(export_path)
true_x = tf.saved_model.utils.build_tensor_info(model.inp.true_x)
time_x = tf.saved_model.utils.build_tensor_info(model.inp.time_x)
norm_x = tf.saved_model.utils.build_tensor_info(model.inp.norm_x)
lagged_x = tf.saved_model.utils.build_tensor_info(model.inp.lagged_x)
true_y = tf.saved_model.utils.build_tensor_info(model.inp.true_y)
time_y = tf.saved_model.utils.build_tensor_info(model.inp.time_y)
norm_y = tf.saved_model.utils.build_tensor_info(model.inp.norm_y)
norm_mean = tf.saved_model.utils.build_tensor_info(model.inp.norm_mean)
norm_std = tf.saved_model.utils.build_tensor_info(model.inp.norm_std)
pg_features = tf.saved_model.utils.build_tensor_info(
model.inp.ucdoc_features)
page_ix = tf.saved_model.utils.build_tensor_info(model.inp.page_ix)
pred = tf.saved_model.utils.build_tensor_info(model.predictions)
labeling_signature = (
tf.saved_model.signature_def_utils.build_signature_def(
inputs={
"truex": true_x,
"timex": time_x,
"normx": norm_x,
"laggedx": lagged_x,
"truey": true_y,
"timey": time_y,
"normy": norm_y,
"normmean": norm_mean,
"normstd": norm_std,
"page_features": pg_features,
"pageix": page_ix,
},
outputs={"pred": pred},
method_name="tensorflow/serving/predict"))
legacy_init_op = tf.group(tf.tables_initializer(),
name='legacy_init_op')
builder.add_meta_graph_and_variables(
sess, [tf.saved_model.tag_constants.SERVING],
signature_def_map={
tf.saved_model.signature_constants.DEFAULT_SERVING_SIGNATURE_DEF_KEY:
labeling_signature
},
main_op=tf.tables_initializer(),
strip_default_attrs=True)
builder.save()
print("Build Done")
def predict(checkpoints, hparams, return_x=False, verbose=False, predict_window=10, back_offset=10, n_models=1,
target_model=0, asgd=False, seed=1, batch_size=1024):
with tf.variable_scope('input') as inp_scope:
with tf.device("/cpu:0"):
inp = VarFeeder.read_vars("data/vars")
pipe = InputPipe(inp, ucdoc_features(inp), inp.n_pages, mode=ModelMode.PREDICT, batch_size=batch_size,
n_epoch=1, verbose=verbose,
train_completeness_threshold=0.01,
predict_window=predict_window,
predict_completeness_threshold=0.0, train_window=hparams.train_window,
back_offset=back_offset)
asgd_decay = 0.99 if asgd else None
if n_models == 1:
model = Model(pipe, hparams, is_train=False, seed=seed, asgd_decay=asgd_decay)
else:
models = []
for i in range(n_models):
prefix = f"m_{i}"
with tf.variable_scope(prefix) as scope:
models.append(Model(pipe, hparams, is_train=False, seed=seed, asgd_decay=asgd_decay, graph_prefix=prefix))
model = models[target_model]
if asgd:
var_list = model.ema.variables_to_restore()
prefix = f"m_{target_model}"
for var in list(var_list.keys()):
if var.endswith('ExponentialMovingAverage') and not var.startswith(prefix):
del var_list[var]
else:
var_list = None
saver = tf.train.Saver(name='eval_saver', var_list=var_list)
x_buffer = []
predictions = None
with tf.Session(config=tf.ConfigProto(gpu_options=tf.GPUOptions(allow_growth=True))) as sess:
pipe.load_vars(sess)
for checkpoint in checkpoints:
pred_buffer = []
pipe.init_iterator(sess)
saver.restore(sess, checkpoint)
cnt = 0
e = []
while True:
try:
if return_x:
#truex, timex, normx, laggedx, truey, timey, normy, normmean, normstd, pgfeatures, pageix
pgfeatures, y ,pred, x, pname, truex , timex, timey= sess.run([model.inp.ucdoc_features, model.inp.true_y, model.predictions, model.inp.true_x, model.inp.page_ix, model.inp.true_x, model.inp.time_x, model.inp.time_y ])
# print(pname, '\n', "true_y:", '\n', np.round(np.expm1(y)), '\n', "prediction:", '\n',
# np.round(np.expm1(pred)), np.average(np.divide(np.abs(np.subtract(np.round(np.expm1(pred)),np.round(np.expm1(y)))),np.round(np.expm1(y)))))
error = np.average(np.divide(np.abs(np.subtract(np.round(np.expm1(pred)),np.round(np.expm1(y)))),np.round(np.expm1(y))))
e.append(error)
# if pname == b'magazinelock,1,3G,g_f,2,pt,1004,icc,2,11':
# print(pgfeatures, y ,pred, x, pname, truex , timex, timey)
# raise NotImplementedError()
# ,np.average(np.divide(np.abs(np.subtract(np.round(np.expm1(y)),np.round(np.expm1(pred)))),np.round(np.expm1(y)) )),
# "time_y:",timey, "time_x:" , timex)
# print(np.average(np.divide(np.abs(np.subtract(np.round(np.expm1(y)),np.round(np.expm1(pred)))),np.round(np.expm1(y)) )))
else:
pred, pname = sess.run([model.predictions, model.inp.page_ix])
utf_names = [str(name, 'utf-8') for name in pname]
pred_df = pd.DataFrame(index=utf_names, data=np.expm1(pred))
pred_buffer.append(pred_df)
if return_x:
# noinspection PyUnboundLocalVariable
x_values = pd.DataFrame(index=utf_names, data=np.round(np.expm1(x)).astype(np.int64))
x_buffer.append(x_values)
newline = cnt % 80 == 0
if cnt > 0:
log.info('.') #, end='\n' if newline else '' , flush=True
if newline:
log.info(cnt) #, end='\n'
cnt += 1
except tf.errors.OutOfRangeError:
log.info('🎉')
break
print(np.average(e))
cp_predictions = pd.concat(pred_buffer)
if predictions is None:
predictions = cp_predictions
else:
predictions += cp_predictions
predictions /= len(checkpoints)
offset = pd.Timedelta(back_offset, 'D')
start_prediction = inp.data_end + pd.Timedelta('1D') - offset
end_prediction = start_prediction + pd.Timedelta(predict_window - 1, 'D')
predictions.columns = pd.date_range(start_prediction, end_prediction)
if return_x:
x = pd.concat(x_buffer)
start_data = inp.data_end - pd.Timedelta(hparams.train_window - 1, 'D') - back_offset
end_data = inp.data_end - back_offset
x.columns = pd.date_range(start_data, end_data)
return predictions, x
else:
return predictions
def main(_):
if not FLAGS.server:
print('please specify server host:port')
return
channel = grpc.insecure_channel(FLAGS.server)
stub = prediction_service_pb2_grpc.PredictionServiceStub(channel)
request = predict_pb2.PredictRequest()
request.model_spec.name = "ucdoc"
request.model_spec.signature_name = "serving_default"
with tf.variable_scope('input') as inp_scope:
with tf.device("/cpu:0"):
inp = VarFeeder.read_vars("data/vars")
pipe = InputPipe(inp,
ucdoc_features(inp),
inp.n_pages,
mode=ModelMode.PREDICT,
batch_size=FLAGS.batch_size,
n_epoch=1,
verbose=FLAGS.verbose,
train_completeness_threshold=0.01,
predict_window=FLAGS.predict_window,
predict_completeness_threshold=0.0,
train_window=FLAGS.train_window,
back_offset=FLAGS.predict_window + 1)
with tf.Session(config=tf.ConfigProto(gpu_options=tf.GPUOptions(
allow_growth=True))) as sess:
pipe.load_vars(sess)
pipe.init_iterator(sess)
while True:
try:
truex, timex, normx, laggedx, truey, timey, normy, normmean, normstd, pgfeatures, pageix = \
sess.run([pipe.true_x, pipe.time_x, pipe.norm_x, pipe.lagged_x, pipe.true_y, pipe.time_y,
pipe.norm_y, pipe.norm_mean, pipe.norm_std, pipe.ucdoc_features, pipe.page_ix])
request.inputs["truex"].CopyFrom(tf.make_tensor_proto(truex))
request.inputs["timex"].CopyFrom(tf.make_tensor_proto(timex))
request.inputs["normx"].CopyFrom(tf.make_tensor_proto(normx))
request.inputs["laggedx"].CopyFrom(
tf.make_tensor_proto(laggedx))
request.inputs["truey"].CopyFrom(tf.make_tensor_proto(truey))
request.inputs["timey"].CopyFrom(tf.make_tensor_proto(timey))
request.inputs["normy"].CopyFrom(tf.make_tensor_proto(normy))
request.inputs["normmean"].CopyFrom(
tf.make_tensor_proto(normmean))
request.inputs["normstd"].CopyFrom(
tf.make_tensor_proto(normstd))
request.inputs["page_features"].CopyFrom(
tf.make_tensor_proto(pgfeatures))
request.inputs["pageix"].CopyFrom(tf.make_tensor_proto(pageix))
response = stub.Predict(request, 10)
tensor_proto = response.outputs['pred']
if not 'pred_result' in locals():
pred_result = tf.contrib.util.make_ndarray(tensor_proto)
else:
pred_result = np.concatenate([
pred_result,
tf.contrib.util.make_ndarray(tensor_proto)
])
except tf.errors.OutOfRangeError:
print('done with prediction')
break
pred_result = np.expm1(pred_result) + 0.5
pred_result = pred_result.astype(int)
if not os.path.exists(FLAGS.result_dir):
os.mkdir(FLAGS.result_dir)
result_file = os.path.join(FLAGS.result_dir, "predict.pkl")
pickle.dump(pred_result, open(result_file, "wb"))
print('finished prediction')
def train(name, hparams, multi_gpu=False, n_models=1, train_completeness_threshold=0.01,
seed=None, logdir='data/logs', max_epoch=100, patience=2, train_sampling=1.0,
eval_sampling=1.0, eval_memsize=5, gpu=0, gpu_allow_growth=False, save_best_model=False,
forward_split=False, write_summaries=False, verbose=False, asgd_decay=None, tqdm=True,
side_split=True, max_steps=None, save_from_step=None, do_eval=True, predict_window=63):
eval_k = int(round(2621 * eval_memsize / n_models))
eval_batch_size = int(
eval_k / (hparams.rnn_depth * hparams.encoder_rnn_layers)) # 128 -> 1024, 256->512, 512->256
eval_pct = 0.2
batch_size = hparams.batch_size
train_window = hparams.train_window
tf.reset_default_graph()
if seed:
tf.set_random_seed(seed)
with tf.device("/cpu:0"):
inp = VarFeeder.read_vars("data/vars")
if side_split:
splitter = Splitter(ucdoc_features(inp),inp.page_map, 3, train_sampling=train_sampling,
test_sampling=eval_sampling, seed=seed)
else:
splitter = FakeSplitter(ucdoc_features(inp), 3, seed=seed, test_sampling=eval_sampling)
real_train_pages = splitter.splits[0].train_size
real_eval_pages = splitter.splits[0].test_size
items_per_eval = real_eval_pages * eval_pct
eval_batches = int(np.ceil(items_per_eval / eval_batch_size))
steps_per_epoch = real_train_pages // batch_size
eval_every_step = int(round(steps_per_epoch * eval_pct))
# eval_every_step = int(round(items_per_eval * train_sampling / batch_size))
global_step = tf.train.get_or_create_global_step()
inc_step = tf.assign_add(global_step, 1)
all_models: List[ModelTrainerV2] = []
def create_model(scope, index, prefix, seed):
with tf.variable_scope('input') as inp_scope:
with tf.device("/cpu:0"):
split = splitter.splits[index]
pipe = InputPipe(inp, features=split.train_set, n_pages=split.train_size,
mode=ModelMode.TRAIN, batch_size=batch_size, n_epoch=None, verbose=verbose,
train_completeness_threshold=train_completeness_threshold,
predict_completeness_threshold=train_completeness_threshold, train_window=train_window,
predict_window=predict_window,
rand_seed=seed, train_skip_first=hparams.train_skip_first,
back_offset=predict_window if forward_split else 0)
inp_scope.reuse_variables()
if side_split:
side_eval_pipe = InputPipe(inp, features=split.test_set, n_pages=split.test_size,
mode=ModelMode.EVAL, batch_size=eval_batch_size, n_epoch=None,
verbose=verbose, predict_window=predict_window,
train_completeness_threshold=0.01, predict_completeness_threshold=0,
train_window=train_window, rand_seed=seed, runs_in_burst=eval_batches,
back_offset=predict_window * (2 if forward_split else 1))
else:
side_eval_pipe = None
if forward_split:
forward_eval_pipe = InputPipe(inp, features=split.test_set, n_pages=split.test_size,
mode=ModelMode.EVAL, batch_size=eval_batch_size, n_epoch=None,
verbose=verbose, predict_window=predict_window,
train_completeness_threshold=0.01, predict_completeness_threshold=0,
train_window=train_window, rand_seed=seed, runs_in_burst=eval_batches,
back_offset=predict_window)
else:
forward_eval_pipe = None
avg_sgd = asgd_decay is not None
#asgd_decay = 0.99 if avg_sgd else None
train_model = Model(pipe, hparams, is_train=True, graph_prefix=prefix, asgd_decay=asgd_decay, seed=seed)
scope.reuse_variables()
eval_stages = []
if side_split:
side_eval_model = Model(side_eval_pipe, hparams, is_train=False,
#loss_mask=np.concatenate([np.zeros(50, dtype=np.float32), np.ones(10, dtype=np.float32)]),
seed=seed)
eval_stages.append((Stage.EVAL_SIDE, side_eval_model))
if avg_sgd:
eval_stages.append((Stage.EVAL_SIDE_EMA, side_eval_model))
if forward_split:
forward_eval_model = Model(forward_eval_pipe, hparams, is_train=False, seed=seed)
eval_stages.append((Stage.EVAL_FRWD, forward_eval_model))
if avg_sgd:
eval_stages.append((Stage.EVAL_FRWD_EMA, forward_eval_model))
if write_summaries:
summ_path = f"{logdir}/{name}_{index}"
if os.path.exists(summ_path):
shutil.rmtree(summ_path)
summ_writer = tf.summary.FileWriter(summ_path) # , graph=tf.get_default_graph()
else:
summ_writer = None
if do_eval and forward_split:
stop_metric = lambda metrics: metrics[Stage.EVAL_FRWD]['SMAPE'].avg_epoch
else:
stop_metric = None
return ModelTrainerV2(train_model, eval_stages, index, patience=patience,
stop_metric=stop_metric,
summary_writer=summ_writer)
if n_models == 1:
with tf.device(f"/gpu:{gpu}"):
scope = tf.get_variable_scope()
all_models = [create_model(scope, 0, None, seed=seed)]
else:
for i in range(n_models):
device = f"/gpu:{i}" if multi_gpu else f"/gpu:{gpu}"
with tf.device(device):
prefix = f"m_{i}"
with tf.variable_scope(prefix) as scope:
all_models.append(create_model(scope, i, prefix=prefix, seed=seed + i))
trainer = MultiModelTrainer(all_models, inc_step)
if save_best_model or save_from_step:
saver_path = f'data/cpt/{name}'
if os.path.exists(saver_path):
shutil.rmtree(saver_path)
os.makedirs(saver_path)
saver = tf.train.Saver(max_to_keep=10, name='train_saver')
else:
saver = None
avg_sgd = asgd_decay is not None
if avg_sgd:
from itertools import chain
def ema_vars(model):
ema = model.train_model.ema
return {ema.average_name(v):v for v in model.train_model.ema._averages}
ema_names = dict(chain(*[ema_vars(model).items() for model in all_models]))
#ema_names = all_models[0].train_model.ema.variables_to_restore()
ema_loader = tf.train.Saver(var_list=ema_names, max_to_keep=1, name='ema_loader')
ema_saver = tf.train.Saver(max_to_keep=1, name='ema_saver')
else:
ema_loader = None
init = tf.global_variables_initializer()
if forward_split and do_eval:
eval_smape = trainer.metric(Stage.EVAL_FRWD, 'SMAPE')
eval_mae = trainer.metric(Stage.EVAL_FRWD, 'MAE')
else:
eval_smape = DummyMetric()
eval_mae = DummyMetric()
if side_split and do_eval:
eval_mae_side = trainer.metric(Stage.EVAL_SIDE, 'MAE')
eval_smape_side = trainer.metric(Stage.EVAL_SIDE, 'SMAPE')
else:
eval_mae_side = DummyMetric()
eval_smape_side = DummyMetric()
train_smape = trainer.metric(Stage.TRAIN, 'SMAPE')
train_mae = trainer.metric(Stage.TRAIN, 'MAE')
grad_norm = trainer.metric(Stage.TRAIN, 'GrNorm')
eval_stages = []
ema_eval_stages = []
if forward_split and do_eval:
eval_stages.append(Stage.EVAL_FRWD)
ema_eval_stages.append(Stage.EVAL_FRWD_EMA)
if side_split and do_eval:
eval_stages.append(Stage.EVAL_SIDE)
ema_eval_stages.append(Stage.EVAL_SIDE_EMA)
# gpu_options=tf.GPUOptions(allow_growth=False),
with tf.Session(config=tf.ConfigProto(allow_soft_placement=True,
gpu_options=tf.GPUOptions(allow_growth=gpu_allow_growth))) as sess:
sess.run(init)
# pipe.load_vars(sess)
inp.restore(sess)
for model in all_models:
model.init(sess)
# if beholder:
# visualizer = Beholder(session=sess, logdir=summ_path)
step = 0
prev_top = np.inf
best_smape = np.inf
# Contains best value (first item) and subsequent values
best_epoch_smape = []
for epoch in range(max_epoch):
# n_steps = pusher.n_pages // batch_size
if tqdm:
#tqr = trange(steps_per_epoch, desc="%2d" % (epoch + 1), leave=False)
tqr = trange(steps_per_epoch, desc="%2d" % (epoch + 1), leave=False,
file=logging.root.handlers[0].stream)
else:
tqr = range(steps_per_epoch)
for _ in tqr:
try:
step = trainer.train_step(sess, epoch)
pred, time_y, true_y, true_x, time_x, page_ix , norm_mean ,norm_std, lagged_ix =sess.run([trainer.trainers[0].train_model.predictions, trainer.trainers[0].train_model.inp.time_y,trainer.trainers[0].train_model.inp.true_y,
trainer.trainers[0].train_model.inp.true_x,trainer.trainers[0].train_model.inp.time_x, trainer.trainers[0].train_model.inp.page_ix , trainer.trainers[0].train_model.inp.norm_mean, trainer.trainers[0].train_model.inp.norm_std
,trainer.trainers[0].train_model.inp.lagged_x])
#sess.run(trainer.trainers[0].train_model.inp.inp.hits)
#inp = all_models[0].train_model.inp.inp,
pred_exp = np.round(np.expm1(pred))
true_exp = np.expm1(true_y )
error_exp= np.mean(np.abs(true_exp-pred_exp) /(true_exp))
error= np.mean(np.abs(true_y -pred )/(true_y))
# page_ix = sess.run([trainer.trainers[0].train_model.inp.page_ix])[0][0]
# true_x = sess.run([trainer.trainers[0].train_model.inp.true_x])[0][0]
last_error = error_exp
epsilon = 0.1 # Smoothing factor, helps SMAPE to be well-behaved near zero
true_o = np.expm1(true_y)
pred_o = np.expm1(pred)
summ = np.maximum(np.abs(true_o) + epsilon, 0.5 + epsilon)
smape = np.mean(np.abs(pred_o - true_o) / summ)
except tf.errors.OutOfRangeError:
break
# if beholder:
# if step % 5 == 0:
# noinspection PyUnboundLocalVariable
# visualizer.update()
if step % eval_every_step == 0:
if eval_stages:
trainer.eval_step(sess, epoch, step, eval_batches, stages=eval_stages)
if save_best_model and epoch > 0 and eval_smape.last < best_smape:
best_smape = eval_smape.last
saver.save(sess, f'data/cpt/{name}/cpt', global_step=step)
if save_from_step and step >= save_from_step:
saver.save(sess, f'data/cpt/{name}/cpt', global_step=step)
if avg_sgd and ema_eval_stages:
ema_saver.save(sess, 'data/cpt_tmp/ema', write_meta_graph=False)
# restore ema-backed vars
ema_loader.restore(sess, 'data/cpt_tmp/ema')
trainer.eval_step(sess, epoch, step, eval_batches, stages=ema_eval_stages)
# restore normal vars
ema_saver.restore(sess, 'data/cpt_tmp/ema')
MAE = "%.3f/%.3f/%.3f" % (eval_mae.last, eval_mae_side.last, train_mae.last)
improvement = '↑' if eval_smape.improved else ' '
SMAPE = "%s%.3f/%.3f/%.3f" % (improvement, eval_smape.last, eval_smape_side.last, train_smape.last)
if tqdm:
tqr.set_postfix(gr=grad_norm.last, MAE=MAE, SMAPE=SMAPE)
if not trainer.has_active() or (max_steps and step > max_steps):
break
if tqdm:
tqr.close()
trainer.end_epoch()
if not best_epoch_smape or eval_smape.avg_epoch < best_epoch_smape[0]:
best_epoch_smape = [eval_smape.avg_epoch]
else:
best_epoch_smape.append(eval_smape.avg_epoch)
current_top = eval_smape.top
if prev_top > current_top:
prev_top = current_top
has_best_indicator = '↑'
else:
has_best_indicator = ' '
status = "%2d: Best top SMAPE=%.3f%s (%s)" % (
epoch + 1, current_top, has_best_indicator,
",".join(["%.3f" % m.top for m in eval_smape.metrics]))
if trainer.has_active():
status += ", frwd/side best MAE=%.3f/%.3f, SMAPE=%.3f/%.3f; avg MAE=%.3f/%.3f, SMAPE=%.3f/%.3f, %d am ,Error=%3f " % \
(eval_mae.best_epoch, eval_mae_side.best_epoch, eval_smape.best_epoch, eval_smape_side.best_epoch,
eval_mae.avg_epoch, eval_mae_side.avg_epoch, eval_smape.avg_epoch, eval_smape_side.avg_epoch,
trainer.has_active(), last_error)
log.info(status)
else:
log.info(status)
log.info("Early stopping!")
break
if max_steps and step > max_steps:
log.info("Max steps calculated")
break
sys.stderr.flush()
# noinspection PyUnboundLocalVariable
return np.mean(best_epoch_smape, dtype=np.float64)
def main(_):
if len(sys.argv) < 3:
print(
'Usage: saved_model.py [--model_version=y] --data_dir=xxx --ckpt_dir=xxx --saved_dir=xxx'
)
sys.exit(-1)
if FLAGS.training_iteration <= 0:
print('Please specify a positive value for training iteration.')
sys.exit(-1)
if FLAGS.model_version <= 0:
print('Please specify a positive value for version number.')
sys.exit(-1)
with open(FLAGS.config_file, 'r') as ymlfile:
cfg = yaml.load(ymlfile)
holiday_list = cfg['pipeline']['normalization']['holidays']
if FLAGS.back_offset < FLAGS.predict_window:
extend_inp(FLAGS.data_dir, FLAGS.predict_window, holiday_list)
# create deploy model first
back_offset_ = FLAGS.back_offset
with tf.variable_scope('input') as inp_scope:
with tf.device("/cpu:0"):
if FLAGS.back_offset < FLAGS.predict_window:
inp = VarFeeder.read_vars(
os.path.join(FLAGS.data_dir, 'predict_future'))
back_offset_ += FLAGS.predict_window
else:
inp = VarFeeder.read_vars(FLAGS.data_dir)
pipe = InputPipe(inp,
ucdoc_features(inp),
inp.hits.shape[0],
mode=ModelMode.PREDICT,
batch_size=FLAGS.batch_size,
n_epoch=1,
verbose=False,
train_completeness_threshold=0.01,
predict_window=FLAGS.predict_window,
predict_completeness_threshold=0.0,
train_window=FLAGS.train_window,
back_offset=back_offset_)
asgd_decay = 0.99 if FLAGS.asgd else None
if FLAGS.n_models == 1:
model = Model(pipe,
build_from_set(FLAGS.hparam_set),
is_train=False,
seed=1,
asgd_decay=asgd_decay)
else:
models = []
for i in range(FLAGS.n_models):
prefix = f"m_{i}"
with tf.variable_scope(prefix) as scope:
models.append(
Model(pipe,
build_from_set(FLAGS.hparam_set),
is_train=False,
seed=1,
asgd_decay=asgd_decay,
graph_prefix=prefix))
model = models[FLAGS.target_model]
if FLAGS.asgd:
var_list = model.ema.variables_to_restore()
if FLAGS.n_models > 1:
prefix = f"m_{target_model}"
for var in list(var_list.keys()):
if var.endswith('ExponentialMovingAverage'
) and not var.startswith(prefix):
del var_list[var]
else:
var_list = None
# load checkpoint model from training
#ckpt_path = FLAGS.ckpt_dir
print('loading checkpoint model...')
ckpt_file = tf.train.latest_checkpoint(FLAGS.ckpt_dir)
#graph = tf.Graph()
graph = model.predictions.graph
init = tf.global_variables_initializer()
saver = tf.train.Saver(name='deploy_saver', var_list=var_list)
with tf.Session(config=tf.ConfigProto(gpu_options=tf.GPUOptions(
allow_growth=True))) as sess:
sess.run(init)
pipe.load_vars(sess)
pipe.init_iterator(sess)
saver.restore(sess, ckpt_file)
print('Done loading checkpoint model')
export_path_base = FLAGS.saved_dir
export_path = os.path.join(
tf.compat.as_bytes(export_path_base),
tf.compat.as_bytes(str(FLAGS.model_version)))
print('Exporting trained model to', export_path)
if os.path.isdir(export_path):
shutil.rmtree(export_path)
builder = tf.saved_model.builder.SavedModelBuilder(export_path)
true_x = tf.saved_model.utils.build_tensor_info(
model.inp.true_x) # pipe.true_x
time_x = tf.saved_model.utils.build_tensor_info(
model.inp.time_x) # pipe.time_x
norm_x = tf.saved_model.utils.build_tensor_info(
model.inp.norm_x) # pipe.norm_x
lagged_x = tf.saved_model.utils.build_tensor_info(
model.inp.lagged_x) # pipe.lagged_x
true_y = tf.saved_model.utils.build_tensor_info(
model.inp.true_y) # pipe.true_y
time_y = tf.saved_model.utils.build_tensor_info(
model.inp.time_y) # pipe.time_y
norm_y = tf.saved_model.utils.build_tensor_info(
model.inp.norm_y) # pipe.norm_y
norm_mean = tf.saved_model.utils.build_tensor_info(
model.inp.norm_mean) # pipe.norm_mean
norm_std = tf.saved_model.utils.build_tensor_info(
model.inp.norm_std) # pipe.norm_std
pg_features = tf.saved_model.utils.build_tensor_info(
model.inp.ucdoc_features) # pipe.ucdoc_features
page_ix = tf.saved_model.utils.build_tensor_info(
model.inp.page_ix) # pipe.page_ix
#pred = tf.saved_model.utils.build_tensor_info(graph.get_operation_by_name('m_0/add').outputs[0])
pred = tf.saved_model.utils.build_tensor_info(model.predictions)
labeling_signature = (
tf.saved_model.signature_def_utils.build_signature_def(
inputs={
"truex": true_x,
"timex": time_x,
"normx": norm_x,
"laggedx": lagged_x,
"truey": true_y,
"timey": time_y,
"normy": norm_y,
"normmean": norm_mean,
"normstd": norm_std,
"page_features": pg_features,
"pageix": page_ix,
},
outputs={"predictions": pred},
method_name="tensorflow/serving/predict"))
legacy_init_op = tf.group(tf.tables_initializer(),
name='legacy_init_op')
builder.add_meta_graph_and_variables(
sess, [tf.saved_model.tag_constants.SERVING],
signature_def_map={
tf.saved_model.signature_constants.DEFAULT_SERVING_SIGNATURE_DEF_KEY:
labeling_signature
},
main_op=tf.tables_initializer(),
strip_default_attrs=True)
builder.save()
print("Build Done")