def test_train_max_steps_is_not_incremental(self):
with ops.Graph().as_default() as g, self.test_session(g):
with ops.control_dependencies(self._build_inference_graph()):
train_op = state_ops.assign_add(variables_lib.get_global_step(), 1)
learn.graph_actions.train(
g,
output_dir=self._output_dir,
train_op=train_op,
loss_op=constant_op.constant(2.0),
max_steps=10)
step = checkpoint_utils.load_variable(
self._output_dir, variables_lib.get_global_step().name)
self.assertEqual(10, step)
with ops.Graph().as_default() as g, self.test_session(g):
with ops.control_dependencies(self._build_inference_graph()):
train_op = state_ops.assign_add(variables_lib.get_global_step(), 1)
learn.graph_actions.train(
g,
output_dir=self._output_dir,
train_op=train_op,
loss_op=constant_op.constant(2.0),
max_steps=15)
step = checkpoint_utils.load_variable(
self._output_dir, variables_lib.get_global_step().name)
self.assertEqual(15, step)
def test_train_skip_train_if_max_step_already_saved(self):
with ops.Graph().as_default() as g, self.test_session(g):
with ops.control_dependencies(self._build_inference_graph()):
train_op = state_ops.assign_add(variables_lib.get_global_step(), 1)
learn.graph_actions._monitored_train( # pylint: disable=protected-access
g,
output_dir=self._output_dir,
train_op=train_op,
loss_op=constant_op.constant(2.0),
max_steps=10)
step = checkpoint_utils.load_variable(
self._output_dir, variables_lib.get_global_step().name)
self.assertEqual(10, step)
with ops.Graph().as_default() as g, self.test_session(g):
with ops.control_dependencies(self._build_inference_graph()):
train_op = state_ops.assign_add(variables_lib.get_global_step(), 1)
learn.graph_actions._monitored_train( # pylint: disable=protected-access
g,
output_dir=self._output_dir,
train_op=train_op,
loss_op=constant_op.constant(2.0),
max_steps=10)
step = checkpoint_utils.load_variable(
self._output_dir, variables_lib.get_global_step().name)
self.assertEqual(10, step)
def test_save_steps_saves_periodically(self):
with self.graph.as_default():
monitor = learn.monitors.CheckpointSaver(
self.model_dir, save_steps=2, scaffold=self.scaffold)
monitor.begin()
self.scaffold.finalize()
with session_lib.Session() as sess:
sess.run(self.scaffold.init_op)
self._run(monitor, 1, self.train_op, sess)
self._run(monitor, 2, self.train_op, sess)
# Not saved
self.assertEqual(1,
checkpoint_utils.load_variable(self.model_dir,
self.global_step.name))
self._run(monitor, 3, self.train_op, sess)
# saved
self.assertEqual(3,
checkpoint_utils.load_variable(self.model_dir,
self.global_step.name))
self._run(monitor, 4, self.train_op, sess)
# Not saved
self.assertEqual(3,
checkpoint_utils.load_variable(self.model_dir,
self.global_step.name))
self._run(monitor, 5, self.train_op, sess)
# saved
self.assertEqual(5,
checkpoint_utils.load_variable(self.model_dir,
self.global_step.name))
def test_save_steps_saves_periodically(self):
with self.graph.as_default():
hook = basic_session_run_hooks.CheckpointSaverHook(
self.model_dir, save_steps=2, scaffold=self.scaffold)
hook.begin()
self.scaffold.finalize()
with session_lib.Session() as sess:
sess.run(self.scaffold.init_op)
mon_sess = monitored_session._HookedSession(sess, [hook])
mon_sess.run(self.train_op)
mon_sess.run(self.train_op)
# Not saved
self.assertEqual(1,
checkpoint_utils.load_variable(self.model_dir,
self.global_step.name))
mon_sess.run(self.train_op)
# saved
self.assertEqual(3,
checkpoint_utils.load_variable(self.model_dir,
self.global_step.name))
mon_sess.run(self.train_op)
# Not saved
self.assertEqual(3,
checkpoint_utils.load_variable(self.model_dir,
self.global_step.name))
mon_sess.run(self.train_op)
# saved
self.assertEqual(5,
checkpoint_utils.load_variable(self.model_dir,
self.global_step.name))
def testGetTensor(self):
checkpoint_dir = self.get_temp_dir()
with self.test_session() as session:
v1, v2, v3, v4 = _create_checkpoints(session, checkpoint_dir)
self.assertAllEqual(
checkpoint_utils.load_variable(checkpoint_dir, "var1"), v1)
self.assertAllEqual(
checkpoint_utils.load_variable(checkpoint_dir, "var2"), v2)
self.assertAllEqual(
checkpoint_utils.load_variable(checkpoint_dir, "var3"), v3)
self.assertAllEqual(
checkpoint_utils.load_variable(checkpoint_dir, "useful_scope/var4"), v4)
def testNoTensor(self):
checkpoint_dir = self.get_temp_dir()
with self.test_session() as session:
_, _, _, _ = _create_checkpoints(session, checkpoint_dir)
with self.assertRaises(errors_impl.OpError):
self.assertAllEqual(
checkpoint_utils.load_variable(checkpoint_dir, "var5"), [])
def print_tensors_in_checkpoint_file(file_name, tensor_name):
"""Prints tensors in a checkpoint file.
If no `tensor_name` is provided, prints the tensor names and shapes
in the checkpoint file.
If `tensor_name` is provided, prints the content of the tensor.
Args:
file_name: Name of the checkpoint file.
tensor_name: Name of the tensor in the checkpoint file to print.
"""
try:
if not tensor_name:
variables = checkpoint_utils.list_variables(file_name)
for name, shape in variables:
print("%s\t%s" % (name, str(shape)))
else:
print("tensor_name: ", tensor_name)
print(checkpoint_utils.load_variable(file_name, tensor_name))
except Exception as e: # pylint: disable=broad-except
print(str(e))
if "corrupted compressed block contents" in str(e):
print("It's likely that your checkpoint file has been compressed "
"with SNAPPY.")
def test_save_secs_saves_periodically(self, mock_time):
# Let's have a realistic start time
current_time = 1484695987.209386
with self.graph.as_default():
mock_time.return_value = current_time
hook = basic_session_run_hooks.CheckpointSaverHook(
self.model_dir, save_secs=2, scaffold=self.scaffold)
hook.begin()
self.scaffold.finalize()
with session_lib.Session() as sess:
sess.run(self.scaffold.init_op)
mon_sess = monitored_session._HookedSession(sess, [hook])
mock_time.return_value = current_time
mon_sess.run(self.train_op) # Saved.
mock_time.return_value = current_time + 0.5
mon_sess.run(self.train_op) # Not saved.
self.assertEqual(1,
checkpoint_utils.load_variable(self.model_dir,
self.global_step.name))
# Simulate 2.5 seconds of sleep.
mock_time.return_value = current_time + 2.5
mon_sess.run(self.train_op) # Saved.
mock_time.return_value = current_time + 2.6
mon_sess.run(self.train_op) # Not saved.
mock_time.return_value = current_time + 2.7
mon_sess.run(self.train_op) # Not saved.
self.assertEqual(3,
checkpoint_utils.load_variable(self.model_dir,
self.global_step.name))
# Simulate 7.5 more seconds of sleep (10 seconds from start.
mock_time.return_value = current_time + 10
mon_sess.run(self.train_op) # Saved.
self.assertEqual(6,
checkpoint_utils.load_variable(self.model_dir,
self.global_step.name))
def test_saves_when_saver_and_scaffold_both_missing(self):
with self.graph.as_default():
hook = basic_session_run_hooks.CheckpointSaverHook(
self.model_dir, save_steps=1)
hook.begin()
self.scaffold.finalize()
with session_lib.Session() as sess:
sess.run(self.scaffold.init_op)
mon_sess = monitored_session._HookedSession(sess, [hook])
mon_sess.run(self.train_op)
self.assertEqual(1,
checkpoint_utils.load_variable(self.model_dir,
self.global_step.name))
def test_save_steps_saves_in_first_step(self):
with self.graph.as_default():
monitor = learn.monitors.CheckpointSaver(self.model_dir,
save_steps=2,
scaffold=self.scaffold)
monitor.begin()
self.scaffold.finalize()
with session_lib.Session() as sess:
sess.run(self.scaffold.init_op)
self._run(monitor, 1, self.train_op, sess)
self.assertEqual(
1,
checkpoint_utils.load_variable(self.model_dir,
self.global_step.name))
def test_save_steps_saves_in_first_step(self):
with self.graph.as_default():
hook = basic_session_run_hooks.CheckpointSaverHook(
self.model_dir, save_steps=2, scaffold=self.scaffold)
hook.begin()
self.scaffold.finalize()
with session_lib.Session() as sess:
sess.run(self.scaffold.init_op)
mon_sess = monitored_session._HookedSession(sess, [hook])
mon_sess.run(self.train_op)
self.assertEqual(
1,
checkpoint_utils.load_variable(self.model_dir,
self.global_step.name))
def test_save_saves_at_end(self):
with self.graph.as_default():
monitor = learn.monitors.CheckpointSaver(
self.model_dir, save_secs=2, scaffold=self.scaffold)
monitor.begin()
self.scaffold.finalize()
with session_lib.Session() as sess:
sess.run(self.scaffold.init_op)
self._run(monitor, 1, self.train_op, sess)
self._run(monitor, 2, self.train_op, sess)
monitor.end(sess)
self.assertEqual(2,
checkpoint_utils.load_variable(self.model_dir,
self.global_step.name))
def disabled_test_save_secs_saves_periodically(self):
with self.graph.as_default():
monitor = learn.monitors.CheckpointSaver(self.model_dir,
save_secs=2,
scaffold=self.scaffold)
monitor.begin()
self.scaffold.finalize()
with session_lib.Session() as sess:
sess.run(self.scaffold.init_op)
self._run(monitor, 1, self.train_op, sess)
self._run(monitor, 2, self.train_op, sess)
# Not saved
self.assertEqual(
1,
checkpoint_utils.load_variable(self.model_dir,
self.global_step.name))
time.sleep(2.5)
self._run(monitor, 3, self.train_op, sess)
# saved
self.assertEqual(
3,
checkpoint_utils.load_variable(self.model_dir,
self.global_step.name))
self._run(monitor, 4, self.train_op, sess)
self._run(monitor, 5, self.train_op, sess)
# Not saved
self.assertEqual(
3,
checkpoint_utils.load_variable(self.model_dir,
self.global_step.name))
time.sleep(2.5)
self._run(monitor, 6, self.train_op, sess)
# saved
self.assertEqual(
6,
checkpoint_utils.load_variable(self.model_dir,
self.global_step.name))
def load_trained_classifier(self, sess, run_id, target_category,
iteration):
tvars = tf.trainable_variables(scope=self.classifierName)
file_path = pathlib.Path(__file__).parent.absolute()
model_folder = os.path.join(file_path, "..", "models", target_category)
checkpoint_folder = os.path.join(
model_folder, "lstm{0}_iteration{1}".format(run_id, iteration))
model_path = os.path.join(
checkpoint_folder,
"lstm{0}_iteration{1}.ckpt".format(run_id, iteration))
saved_vars = checkpoint_utils.list_variables(checkpoint_dir=model_path)
for var in tvars:
# assert len([_var for _var in saved_vars if _var.name == var.name]) == 1
# if "Adam" in var.name:
# continue
var_name = var.name[len(self.classifierName) + 1:]
source_array = checkpoint_utils.load_variable(
checkpoint_dir=model_path, name=var_name)
tf.assign(var, source_array).eval(session=sess)
def load_fp32_weights_into_fp16_vars(checkpoint_path: Path) -> List:
"""Load fp32 weights from checkpoint path into fp16 variables.
Assumes that caller has executed `tf.run(tf.global_variables_initializer())`
Args:
checkpoint_path: Checkpoint path
Returns:
Collection of ops to use to restore the weights in the graph.
"""
checkpoint_variables = [var_name for var_name, _ in list_variables(checkpoint_path)]
for graph_var in tf.global_variables():
if graph_var.op.name in checkpoint_variables:
var = load_variable(checkpoint_path, graph_var.op.name)
weights = tf.cast(var, tf.float16) if var.dtype == np.float32 else var
tf.add_to_collection('restore_ops', graph_var.assign(weights))
return tf.get_collection('restore_ops')
def init_graph():
restore_vars = []
layer_num = 0
regex = re.compile('layer_\d+')
amaxTotalNum = 0
for name, shape in var_list:
var = checkpoint_utils.load_variable(tf.flags.FLAGS.init_checkpoint, name)
if "intermediate/dense/kernel" in name and amaxTotalNum == 0:
amaxTotalNum = ACTIVATION_AMAX_NUM + 9*shape[0]
print(amaxTotalNum, shape[0])
recon_dtype = var.dtype
restore_vars.append(tf.get_variable(name, shape=shape, dtype=var.dtype))
tmp = regex.findall(name)
if len(tmp) < 1:
continue
num_tmp = int(tmp[0].replace("layer_", ""))
if layer_num < num_tmp:
layer_num = num_tmp
layer_num = layer_num + 1
#add new var for amax
for i in range(layer_num):
tf.get_variable("bert/encoder/layer_{}/amaxList".format(i), shape=[amaxTotalNum], dtype=tf.float32)
return layer_num, amaxTotalNum, restore_vars
def train(self):
if tf.gfile.Exists(self.cluster_conf):
cluster_spec, run_config, num_workers = self.gen_run_config()
server = tf.train.Server(cluster_spec,
job_name=self.job_name,
task_index=self.task_index)
else:
run_config = tf.contrib.learn.RunConfig(
save_checkpoints_secs=self.save_checkpoints_secs)
num_workers = 1
server = None
print ('-' * 40)
print ('run_config =', run_config)
if self.job_name == 'ps':
print ("ps start")
server.join()
elif self.job_name == 'worker':
print ("worker start")
train_steps = self.model_parameters.get("train_steps")
self.model_fn_param['run_config'] = run_config
builder = WideNDeepModelBuilder(model_desc=self.model_desc_obj)
model = builder.build_estimator(self.model_fn_param)
hooks = []
try:
global_step = checkpoint_utils.load_variable(
self.checkpoint_path, tf.GraphKeys.GLOBAL_STEP)
except:
global_step = 0
print ('global_step =', global_step)
if self.init_dir and global_step < 100:
print ('InitEmbeddings from %s' % self.init_dir)
hooks.append(InitEmbeddingsHook(self.init_dir))
model.train(input_fn=self.input_fn_train,
steps=train_steps,
hooks=hooks)
import tensorflow as tf
from tensorflow.python.ops import data_flow_ops
from tensorflow.python.framework import dtypes
from tensorflow.python.user_ops import user_ops
from tensorflow.python.training import queue_runner
from tensorflow.contrib.framework.python.framework import checkpoint_utils
#print checkpoint_utils.list_variables("hdfs://tensorflow-on-yarn-test/user/yonggang.myg/wnd_plugin_test/kafka_reader_test/checkpoint")
checkpoint_utils.list_variables("checkpoint")
print checkpoint_utils.load_variable("checkpoint",
"queue_runner_checkpoint_var")
# tvars = tf.trainable_variables()
# tvars_vals = sess.run(tvars)
#
# for var, val in zip(tvars, tvars_vals):
# print(var.name, val)
def testNoCheckpoints(self):
checkpoint_dir = self.get_temp_dir() + "/no_checkpoints"
with self.assertRaises(errors_impl.OpError):
self.assertAllEqual(
checkpoint_utils.load_variable(checkpoint_dir, "var1"), [])
def covariances(self):
"""Returns the covariances."""
return checkpoint_utils.load_variable(
self.model_dir, gmm_ops.GmmAlgorithm.CLUSTERS_COVS_VARIABLE)
def checkpoint_quantization(in_checkpoint_file, out_checkpoint_file,
per_channel_quantization):
var_list = checkpoint_utils.list_variables(tf.flags.FLAGS.init_checkpoint)
def init_graph():
restore_vars = []
layer_num = 0
regex = re.compile('layer_\d+')
amaxTotalNum = 0
for name, shape in var_list:
var = checkpoint_utils.load_variable(
tf.flags.FLAGS.init_checkpoint, name)
if "intermediate/dense/kernel" in name and amaxTotalNum == 0:
amaxTotalNum = ACTIVATION_AMAX_NUM + 9 * shape[
0] + INT8O_GEMM_NUM + TRT_FUSED_MHA_AMAX_NUM
print(amaxTotalNum, shape[0])
recon_dtype = var.dtype
restore_vars.append(
tf.get_variable(name, shape=shape, dtype=var.dtype))
tmp = regex.findall(name)
if len(tmp) < 1:
continue
num_tmp = int(tmp[0].replace("layer_", ""))
if layer_num < num_tmp:
layer_num = num_tmp
layer_num = layer_num + 1
#add new var for amax
for i in range(layer_num):
tf.get_variable("bert/encoder/layer_{}/amaxList".format(i),
shape=[amaxTotalNum],
dtype=tf.float32)
return layer_num, amaxTotalNum, restore_vars
layer_num, amaxTotalNum, restore_vars = init_graph()
restorer = tf.train.Saver(restore_vars)
saver = tf.train.Saver()
config = tf.ConfigProto()
config.gpu_options.allow_growth = True
with tf.Session(config=config) as sess:
restorer.restore(sess, in_checkpoint_file)
kernel_name_list = [
"attention/self/query", "attention/self/key",
"attention/self/value", "attention/output/dense",
"intermediate/dense", "output/dense"
]
#input_scale, 0
amax_name_list = [
"attention/self/query/input_quantizer",
#Q_aftergemm_scale, 1
"attention/self/query/aftergemm_quantizer",
#Qbias_scale, 2
"attention/self/matmul_q_input_quantizer",
#K_aftergemm_scale, 3
"attention/self/key/aftergemm_quantizer",
#Kbias_scale, 4
"attention/self/matmul_k_input_quantizer",
#V_aftergemm_scale, 5
"attention/self/value/aftergemm_quantizer",
#Vbias_scale, 6
"attention/self/matmul_v_input_quantizer",
#bmm1_scale, 7
"attention/self/softmax_input_quantizer",
#Softmax_scale, 8
"attention/self/matmul_a_input_quantizer",
#bmm2_scale, 9
"attention/output/dense/input_quantizer",
#Proj_aftergemm_scale, 10
"attention/output/dense/aftergemm_quantizer",
#ProjBiasNorm_scale, 11
"intermediate/dense/input_quantizer",
#FC1_aftergemm_scale, 12
"intermediate/dense/aftergemm_quantizer",
#F1Bias_scale, 13
"output/dense/input_quantizer",
#FC2_aftergemm_scale, 14
"output/dense/aftergemm_quantizer",
#F2Bias_scale, 15
"special_F2Bias_scale",
]
int8O_gemm_weight_amax_list = [0 for i in range(INT8O_GEMM_NUM)]
#Q_aftergemm
int8O_gemm_weight_list = [
"attention/self/query",
#K_aftergemm
"attention/self/key",
#V_aftergemm
"attention/self/value",
#bmm1_aftergemm
"attention/self/matmul_k_input_quantizer",
#bmm2_aftergemm
"attention/self/matmul_v_input_quantizer",
#Proj_aftergemm
"attention/output/dense",
#FC1_aftergemm
"intermediate/dense",
#FC2_aftergemm
"output/dense"
]
int8O_gemm_input_amax_list = [0 for i in range(INT8O_GEMM_NUM)]
#Q_aftergemm
int8O_gemm_input_list = [
"attention/self/query/input_quantizer",
#K_aftergemm
"attention/self/key/input_quantizer",
#V_aftergemm
"attention/self/value/input_quantizer",
#bmm1_aftergemm
"attention/self/matmul_q_input_quantizer",
#bmm2_aftergemm
"attention/self/matmul_a_input_quantizer",
#Proj_aftergemm
"attention/output/dense/input_quantizer",
#FC1_aftergemm
"intermediate/dense/input_quantizer",
#FC2_aftergemm
"output/dense/input_quantizer"
]
int8O_gemm_output_amax_list = [0 for i in range(INT8O_GEMM_NUM)]
#Q_aftergemm
int8O_gemm_output_list = [
"attention/self/query/aftergemm_quantizer",
#K_aftergemm
"attention/self/key/aftergemm_quantizer",
#V_aftergemm
"attention/self/value/aftergemm_quantizer",
#bmm1_aftergemm
"attention/self/softmax_input_quantizer",
#bmm2_aftergemm
"attention/output/dense/input_quantizer",
#Proj_aftergemm
"attention/output/dense/aftergemm_quantizer",
#FC1_aftergemm
"intermediate/dense/aftergemm_quantizer",
#FC2_aftergemm
"output/dense/aftergemm_quantizer"
]
factor = 1000000.0
for i in range(layer_num):
amaxList = np.zeros([amaxTotalNum])
amax_id = 0
for amax_name in amax_name_list:
if amax_name == "special_F2Bias_scale":
if i != layer_num - 1:
name = "bert/encoder/layer_{}/{}/quant_max:0".format(
i + 1, amax_name_list[0])
quant_max = checkpoint_utils.load_variable(
tf.flags.FLAGS.init_checkpoint, name)
name = "bert/encoder/layer_{}/{}/quant_min:0".format(
i + 1, amax_name_list[0])
quant_min = checkpoint_utils.load_variable(
tf.flags.FLAGS.init_checkpoint, name)
if abs(quant_max) > abs(quant_min):
amax = abs(
quant_max) #int(abs(quant_max)*factor)/factor
else:
amax = abs(
quant_min) #int(abs(quant_min)*factor)/factor
else:
#not used, placeholder
amax = 1.0
amaxList[amax_id] = amax
amax_id += 1
amaxList[amax_id] = amax / 127.0
amax_id += 1
amaxList[amax_id] = amax / 127.0 / 127.0
amax_id += 1
amaxList[amax_id] = 127.0 / amax
amax_id += 1
continue
name = "bert/encoder/layer_{}/{}/quant_max:0".format(
i, amax_name)
quant_max = checkpoint_utils.load_variable(
tf.flags.FLAGS.init_checkpoint, name)
name = "bert/encoder/layer_{}/{}/quant_min:0".format(
i, amax_name)
quant_min = checkpoint_utils.load_variable(
tf.flags.FLAGS.init_checkpoint, name)
if abs(quant_max) > abs(quant_min):
amax = abs(quant_max) #int(abs(quant_max)*factor)/factor
else:
amax = abs(quant_min) #int(abs(quant_min)*factor)/factor
if amax_name in int8O_gemm_input_list:
int8O_gemm_input_amax_list[int8O_gemm_input_list.index(
amax_name)] = amax
if amax_name == "attention/self/query/input_quantizer":
int8O_gemm_input_amax_list[int8O_gemm_input_list.index(
"attention/self/key/input_quantizer")] = amax
int8O_gemm_input_amax_list[int8O_gemm_input_list.index(
"attention/self/value/input_quantizer")] = amax
if amax_name in int8O_gemm_output_list:
int8O_gemm_output_amax_list[int8O_gemm_output_list.index(
amax_name)] = amax
if amax_name in int8O_gemm_weight_list:
int8O_gemm_weight_amax_list[int8O_gemm_weight_list.index(
amax_name)] = amax
amaxList[amax_id] = amax
amax_id += 1
amaxList[amax_id] = amax / 127.0
amax_id += 1
amaxList[amax_id] = amax / 127.0 / 127.0
amax_id += 1
amaxList[amax_id] = 127.0 / amax
amax_id += 1
print("done process layer_{} activation amax".format(i))
#kernel amax starts from ACTIVATION_AMAX_NUM
amax_id = ACTIVATION_AMAX_NUM
for kernel_id, kernel_name in enumerate(kernel_name_list):
kernel = tf.get_default_graph().get_tensor_by_name(
"bert/encoder/layer_{}/{}/kernel:0".format(i, kernel_name))
name = "bert/encoder/layer_{}/{}/kernel_quantizer/quant_max:0".format(
i, kernel_name)
quant_max2 = tf.convert_to_tensor(
checkpoint_utils.load_variable(
tf.flags.FLAGS.init_checkpoint, name))
name = "bert/encoder/layer_{}/{}/kernel_quantizer/quant_min:0".format(
i, kernel_name)
quant_min2 = tf.convert_to_tensor(
checkpoint_utils.load_variable(
tf.flags.FLAGS.init_checkpoint, name))
kernel_processed, quant_max_processed = transformer_op_module.weight_quantize(
kernel,
quant_max2,
quant_min2,
per_channel_quantization=per_channel_quantization)
kernel_processed_, quant_max_processed_ = sess.run(
[kernel_processed, quant_max_processed])
sess.run(tf.assign(kernel, kernel_processed_))
if kernel_name in int8O_gemm_weight_list:
int8O_gemm_weight_amax_list[int8O_gemm_weight_list.index(
kernel_name)] = quant_max_processed_[0]
for e in quant_max_processed_:
amaxList[amax_id] = e
amax_id += 1
#for int8O gemm deQuant
for j in range(INT8O_GEMM_NUM):
amaxList[amax_id] = (int8O_gemm_input_amax_list[j] *
int8O_gemm_weight_amax_list[j]) / (
127.0 *
int8O_gemm_output_amax_list[j])
amax_id += 1
#for trt fused MHA amax
#### QKV_addBias_amax
amaxList[amax_id] = np.maximum(
np.maximum(amaxList[8], amaxList[16]), amaxList[24])
amax_id += 1
#### softmax amax
amaxList[amax_id] = amaxList[32]
amax_id += 1
#### bmm2 amax
amaxList[amax_id] = amaxList[36]
amax_id += 1
amaxL = tf.get_default_graph().get_tensor_by_name(
"bert/encoder/layer_{}/amaxList:0".format(i))
sess.run(tf.assign(amaxL, amaxList))
print("done process layer_{} kernel weight".format(i))
saver.save(sess, out_checkpoint_file)
def load_variable(checkpoint_dir, name):
"""See `tf.contrib.framework.load_variable`."""
return checkpoint_utils.load_variable(checkpoint_dir, name)
def clusters(self):
"""Returns cluster centers."""
clusters = checkpoint_utils.load_variable(
self.model_dir, gmm_ops.GmmAlgorithm.CLUSTERS_VARIABLE)
return np.squeeze(clusters, 1)
def init_graph():
for name, shape in var_list:
var = checkpoint_utils.load_variable(
tf.flags.FLAGS.init_checkpoint, name)
recon_dtype = tf.float16 if var.dtype == np.float32 else var.dtype
tf.get_variable(name, shape=shape, dtype=recon_dtype)
def testNoCheckpoints(self):
checkpoint_dir = self.get_temp_dir() + "/no_checkpoints"
with self.assertRaises(errors_impl.OpError):
self.assertAllEqual(
checkpoint_utils.load_variable(checkpoint_dir, "var1"), [])
def weights(self):
"""Returns the cluster weights."""
return checkpoint_utils.load_variable(
self.model_dir, gmm_ops.GmmAlgorithm.CLUSTERS_WEIGHT)
def _get_checkpointed_vars(checkpoints, name, fn=lambda x: x):
return {
k: fn(checkpoint_utils.load_variable(checkpoint_dir=k, name=name))
for k in checkpoints
}
def covariances(self):
"""Returns the covariances."""
return checkpoint_utils.load_variable(
self.model_dir, gmm_ops.GmmAlgorithm.CLUSTERS_COVS_VARIABLE)
def clusters(self):
"""Returns cluster centers."""
clusters = checkpoint_utils.load_variable(
self.model_dir, gmm_ops.GmmAlgorithm.CLUSTERS_VARIABLE)
return np.squeeze(clusters, 1)
def weights(self):
"""Returns the cluster weights."""
return checkpoint_utils.load_variable(
self.model_dir, gmm_ops.GmmAlgorithm.CLUSTERS_WEIGHT)
def load_variable(checkpoint_dir, name): """See `tf.contrib.framework.load_variable`.""" return checkpoint_utils.load_variable(checkpoint_dir, name)
