def test_fusedbatchnorm_nchw():
#Test 1: tf_model TF-native
with tf.compat.v1.Session(config=config) as sess_tf:
ngraph_bridge.disable()
tf_out, in_0 = tf_model()
feed_dict = {in_0: k_np}
tf_outval = sess_tf.run(tf_out, feed_dict=feed_dict)
#Test 2: model2 with ngraph, NNP backend
with tf.compat.v1.Session(config=config) as sess_ng:
ngraph_bridge.enable()
ngraph_bridge.update_config(config)
os.environ['NGRAPH_TF_DISABLE_DEASSIGN_CLUSTERS'] = '1'
ng_out, in_0 = ng_model()
feed_dict = {in_0: k_np}
ng_outval = sess_ng.run(ng_out, feed_dict=feed_dict)
# transpose TF output from NHWC to NCHW for comparison with ngraph output
result1_bool = np.allclose(
np.transpose(tf_outval[0], (0, 3, 1, 2)),
ng_outval[0],
rtol=0,
atol=1e-02)
# these TF outputs do not need to be transposed since they have only 1 dimension
result2_bool = np.allclose(tf_outval[1], ng_outval[1], rtol=0, atol=1e-02)
result3_bool = np.allclose(tf_outval[2], ng_outval[2], rtol=0, atol=1e-02)
assert (result1_bool and result2_bool and result3_bool)
def on_ngraph_change(self, change):
if change['type'] == 'change' and change['name'] == 'value':
i = self.ngraph_backends.index(change['new'])
if self.ngraph_backends[i] == 'DISABLED':
self.use_ngraph = False
ngraph_bridge.disable()
else:
self.use_ngraph = True
ngraph_bridge.enable()
ngraph_bridge.set_backend(self.ngraph_backends[i])
def set_os_env(select_device):
if select_device == 'CPU':
# run on TF only
ngraph_bridge.disable()
else:
if not ngraph_bridge.is_enabled():
ngraph_bridge.enable()
assert select_device[:
7] == "NGRAPH_", "Expecting device name to start with NGRAPH_"
back_end = select_device.split("NGRAPH_")
os.environ['NGRAPH_TF_BACKEND'] = back_end[1]
def without_ngraph(self, l, config=tf.ConfigProto()):
ngraph_tf_disable_deassign_clusters = os.environ.pop(
'NGRAPH_TF_DISABLE_DEASSIGN_CLUSTERS', None)
ngraph_bridge.disable()
with tf.Session(config=config) as sess:
retval = l(sess)
if ngraph_tf_disable_deassign_clusters is not None:
os.environ['NGRAPH_TF_DISABLE_DEASSIGN_CLUSTERS'] = \
ngraph_tf_disable_deassign_clusters
return retval
def test_conv2dbackpropinput_nhwc(padding):
np_filter = np.random.rand(*filter_size_hwio).astype('f')
n_np_out = np.random.rand(*out_backprop_in_sizes[padding]).astype('f')
with tf.compat.v1.Session(config=config) as sess_tf:
ngraph_bridge.disable()
tf_out, filter_size, out_backprop = tf_model(padding)
feed_dict = {filter_size: np_filter, out_backprop: n_np_out}
tf_outval = sess_tf.run(tf_out, feed_dict=feed_dict)
#Test 2: model2 with ngraph, NNP backend
with tf.compat.v1.Session(config=config) as sess_ng:
ngraph_bridge.enable()
ngraph_bridge.update_config(config)
os.environ['NGRAPH_TF_DISABLE_DEASSIGN_CLUSTERS'] = '1'
ng_out, filter_size, out_backprop = ng_model(padding)
feed_dict = {filter_size: np_filter, out_backprop: n_np_out}
ng_outval = sess_ng.run(ng_out, feed_dict=feed_dict)
assert np.allclose(tf_outval, ng_outval, rtol=0, atol=1e-02)
def test_mnist_training(self, optimizer):
class mnist_training_flags:
def __init__(self, data_dir, model_dir, training_iterations,
training_batch_size, validation_batch_size,
make_deterministic, training_optimizer):
self.data_dir = data_dir
self.model_dir = model_dir
self.train_loop_count = training_iterations
self.batch_size = training_batch_size
self.test_image_count = validation_batch_size
self.make_deterministic = make_deterministic
self.optimizer = optimizer
data_dir = '/tmp/' + getpass.getuser() + 'tensorflow/mnist/input_data'
train_loop_count = 50
batch_size = 50
test_image_count = None
make_deterministic = True
model_dir = './mnist_trained/'
FLAGS = mnist_training_flags(data_dir, model_dir, train_loop_count,
batch_size, test_image_count,
make_deterministic, optimizer)
# Run on nGraph
ng_loss_values, ng_test_accuracy = train_mnist_cnn(FLAGS)
ng_values = ng_loss_values + [ng_test_accuracy]
# Reset the Graph
tf.compat.v1.reset_default_graph()
# disable ngraph-tf
ngraph_bridge.disable()
tf_loss_values, tf_test_accuracy = train_mnist_cnn(FLAGS)
tf_values = tf_loss_values + [tf_test_accuracy]
# compare values
assert np.allclose(
ng_values, tf_values,
atol=1e-3), "Loss or Accuracy values don't match"
def test_conv2d():
#Test 1: tf_model TF-native
with tf.compat.v1.Session(config=config) as sess_tf:
ngraph_bridge.disable()
tf_out, input_data = tf_model()
feed_dict = {input_data: t_np}
tf_outval = sess_tf.run(tf_out, feed_dict=feed_dict)
#Test 2: model2 with ngraph, NNP backend
with tf.compat.v1.Session(config=config) as sess_ng:
ngraph_bridge.enable()
ngraph_bridge.update_config(config)
os.environ['NGRAPH_TF_DISABLE_DEASSIGN_CLUSTERS'] = '1'
ng_out, input_data = ng_model()
feed_dict = {input_data: n_np}
ng_outval = sess_ng.run(ng_out, feed_dict=feed_dict)
assert np.allclose(np.transpose(tf_outval, (0, 3, 1, 2)),
ng_outval,
rtol=0,
atol=1e-02)
def test_maxpoolbackprop_nhwc(padding):
g_np = grad_nhwc[padding]
o_np = output_nhwc[padding]
#Test 1: tf_model TF-native
with tf.Session(config=config) as sess_tf:
ngraph_bridge.disable()
tf_out, orig_in, orig_out, grad = tf_model(padding)
feed_dict = {orig_in: i_np, orig_out: o_np, grad: g_np}
tf_outval = sess_tf.run(tf_out, feed_dict=feed_dict)
#Test 2: model2 with ngraph, NNP backend
with tf.Session(config=config) as sess_ng:
ngraph_bridge.enable()
ngraph_bridge.update_config(config)
os.environ['NGRAPH_TF_DISABLE_DEASSIGN_CLUSTERS'] = '1'
ng_out, orig_in, orig_out, grad = ng_model(padding)
feed_dict = {orig_in: i_np, orig_out: o_np, grad: g_np}
ng_outval = sess_ng.run(ng_out, feed_dict=feed_dict)
assert (np.allclose(tf_outval, ng_outval, rtol=0, atol=1e-02))
def test_fusedbatchnorm_nhwc():
#Test 1: tf_model TF-native
with tf.Session(config=config) as sess_tf:
ngraph_bridge.disable()
tf_out, in_0 = tf_model()
feed_dict = {in_0: k_np}
tf_outval = sess_tf.run(tf_out, feed_dict=feed_dict)
#Test 2: model2 with ngraph, NNP backend
with tf.Session(config=config) as sess_ng:
ngraph_bridge.enable()
ngraph_bridge.update_config(config)
os.environ['NGRAPH_TF_DISABLE_DEASSIGN_CLUSTERS'] = '1'
ng_out, in_0 = ng_model()
feed_dict = {in_0: k_np}
ng_outval = sess_ng.run(ng_out, feed_dict=feed_dict)
result1_bool = np.allclose(tf_outval[0], ng_outval[0], rtol=0, atol=1e-02)
result2_bool = np.allclose(tf_outval[1], ng_outval[1], rtol=0, atol=1e-02)
result3_bool = np.allclose(tf_outval[2], ng_outval[2], rtol=0, atol=1e-02)
assert (result1_bool and result2_bool and result3_bool)
def setup_ngraph_bridge(self, backend):
# Enviornment variables
os.environ['PLAIDML_USE_STRIPE'] = '1'
if self.workers < 1:
os.environ['OMP_NUM_THREADS'] = 1
else:
# Use default
if os.getenv('OMP_NUM_THREADS') is not None:
del os.environ['OMP_NUM_THREADS']
import ngraph_bridge
if backend == 'DISABLED' or backend == 'TF':
ngraph_bridge.disable()
elif backend == 'CPU':
ngraph_bridge.set_backend('CPU')
ngraph_bridge.enable()
elif backend == 'PLAIDML':
ngraph_bridge.set_backend('PLAIDML')
ngraph_bridge.enable()
else:
print("ERROR: Unsupported backend " + backend + " selected.")
def test_conv2dbackpropfilter_nchw(padding):
n_np_inp = np.random.rand(*input_sizes_nchw).astype('f')
n_np_out = np.random.rand(*out_backprop_in_sizes[padding]).astype('f')
#Reshape to NHWC for TF
t_np_inp = np.transpose(n_np_inp, (0, 2, 3, 1))
t_np_out = np.transpose(n_np_out, (0, 2, 3, 1))
with tf.compat.v1.Session(config=config) as sess_tf:
ngraph_bridge.disable()
tf_out, input_data, out_backprop = tf_model(padding)
feed_dict = {input_data: t_np_inp, out_backprop: t_np_out}
tf_outval = sess_tf.run(tf_out, feed_dict=feed_dict)
#Test 2: model2 with ngraph, NNP backend
with tf.compat.v1.Session(config=config) as sess_ng:
ngraph_bridge.enable()
ngraph_bridge.update_config(config)
os.environ['NGRAPH_TF_DISABLE_DEASSIGN_CLUSTERS'] = '1'
ng_out, input_data, out_backprop = ng_model(padding)
feed_dict = {input_data: n_np_inp, out_backprop: n_np_out}
ng_outval = sess_ng.run(ng_out, feed_dict=feed_dict)
assert np.allclose(tf_outval, ng_outval, rtol=0, atol=1e-02)
def test_disable(self):
ngraph_bridge.disable()
assert ngraph_bridge.is_enabled() == 0
def run(self):
"""run benchmark with optimized graph"""
print("Run inference with dummy data")
config = tf.compat.v1.ConfigProto()
config.intra_op_parallelism_threads = self.args.num_intra_threads
config.inter_op_parallelism_threads = self.args.num_inter_threads
config.use_per_session_threads = True
data_graph = tf.Graph()
with data_graph.as_default():
input_shape = [
self.args.batch_size, RESNET_IMAGE_SIZE, RESNET_IMAGE_SIZE, 3
]
images = tf.random.uniform(
input_shape,
0.0,
255.0,
dtype=tf.float32,
seed=42,
name='synthetic_images')
infer_graph = tf.Graph()
with infer_graph.as_default():
graph_def = tf.compat.v1.GraphDef()
with tf.io.gfile.GFile(self.args.input_graph, 'rb') as input_file:
input_graph_content = input_file.read()
graph_def.ParseFromString(input_graph_content)
print(
"Optimizing graph %s for inference..." % self.args.input_graph)
output_graph = optimize_for_inference(
graph_def, [INPUTS], [OUTPUTS], dtypes.float32.as_datatype_enum,
False)
tf.import_graph_def(output_graph, name='')
input_tensor = infer_graph.get_tensor_by_name('input_tensor:0')
output_tensor = infer_graph.get_tensor_by_name('softmax_tensor:0')
# Run without nGraph first
print("Run inference (without nGraph)")
ngraph_bridge.disable()
data_sess = tf.compat.v1.Session(graph=data_graph, config=config)
infer_sess = tf.compat.v1.Session(graph=infer_graph, config=config)
iteration = 0
num_processed_images = 0
num_remaining_images = self.args.num_images
tf_time = 0.0
tf_labels = np.array([], dtype=np.int32)
while num_remaining_images >= self.args.batch_size:
np_images = data_sess.run(images)
if iteration > self.args.warmup_iters:
num_processed_images += self.args.batch_size
num_remaining_images -= self.args.batch_size
tf_start_time = time.time()
predictions = infer_sess.run(output_tensor,
{input_tensor: np_images})
tf_elapsed_time = time.time() - tf_start_time
if iteration > self.args.warmup_iters:
tf_time += tf_elapsed_time
tf_labels = np.append(tf_labels, np.argmax(
predictions, axis=-1))
iteration += 1
print("Total execution time (TF): ", tf_time)
# Run with nGraph now
print("Run inference (with nGraph)")
ngraph_bridge.enable()
data_sess = tf.compat.v1.Session(graph=data_graph, config=config)
infer_sess = tf.compat.v1.Session(graph=infer_graph, config=config)
iteration = 0
num_processed_images = 0
num_remaining_images = self.args.num_images
ngtf_time = 0.0
ngtf_labels = np.array([], dtype=np.int32)
while num_remaining_images >= self.args.batch_size:
np_images = data_sess.run(images)
if iteration > self.args.warmup_iters:
num_processed_images += self.args.batch_size
num_remaining_images -= self.args.batch_size
ngtf_start_time = time.time()
predictions = infer_sess.run(output_tensor,
{input_tensor: np_images})
ngtf_elapsed_time = time.time() - ngtf_start_time
if iteration > self.args.warmup_iters:
ngtf_time += ngtf_elapsed_time
ngtf_labels = np.append(ngtf_labels,
np.argmax(predictions, axis=-1))
iteration += 1
print("Total execution time (NGTF): ", ngtf_time)
print("Processed %d images. Batch size = %d" % (num_processed_images,
self.args.batch_size))
print("Avg throughput (TF): %0.4f img/s" %
(num_processed_images / tf_time))
print("Avg throughput (NGTF): %0.4f img/s" %
(num_processed_images / ngtf_time))
assert ((tf_labels == ngtf_labels).all())