label_path = download_testdata(label_map_url, label_map, module='data')
######################################################################
# Import model
# ------------
# Creates tensorflow graph definition from protobuf file.
with tf_compat_v1.gfile.GFile(model_path, 'rb') as f:
graph_def = tf_compat_v1.GraphDef()
graph_def.ParseFromString(f.read())
graph = tf.import_graph_def(graph_def, name='')
# Call the utility to import the graph definition into default graph.
graph_def = tf_testing.ProcessGraphDefParam(graph_def)
# Add shapes to the graph.
with tf_compat_v1.Session() as sess:
graph_def = tf_testing.AddShapesToGraphDef(sess, 'softmax')
######################################################################
# Decode image
# ------------
# .. note::
#
# tensorflow frontend import doesn't support preprocessing ops like JpegDecode.
# JpegDecode is bypassed (just return source node).
# Hence we supply decoded frame to TVM instead.
#
from PIL import Image
image = Image.open(img_path).resize((299, 299))
from PIL import Image
try:
tf_compat_v1 = tf.compat.v1
except ImportError:
tf_compat_v1 = tf
# Load Model
MODEL_PATH = "./model/yolov3_visdrone.pb"
with tf_compat_v1.gfile.GFile(MODEL_PATH, "rb") as f:
graph_def = tf_compat_v1.GraphDef()
graph_def.ParseFromString(f.read())
graph = tf.import_graph_def(graph_def, name="")
# Call the utility to import the graph definition into default graph.
graph_def = tf_testing.ProcessGraphDefParam(graph_def)
with tf_compat_v1.Session() as sess:
graph_def = tf_testing.AddShapesToGraphDef(sess,
"pred_multi_scale/concat")
IMAGE_PATH = "./images/road.jpg"
image = Image.open(IMAGE_PATH, 'r').resize((416, 416))
x = np.array(image)
x = x[np.newaxis, ...]
shape_dict = {"input/input_data": x.shape}
mod, params = relay.frontend.from_tensorflow(graph_def,
layout=None,
shape=shape_dict)
print("Tensorflow protobuf imported to relay frontend.")
target = 'llvm'
target_host = 'llvm'
with tvm.transform.PassContext(opt_level=1):
lib = relay.build(mod,
"",
model_path,
as_text=False)
"""
Create the graph in TVM and compile it
"""
with tf.io.gfile.GFile(model_path, "rb") as f:
graph_def = tf.compat.v1.GraphDef()
graph_def.ParseFromString(f.read())
graph = tf.import_graph_def(graph_def, name="")
graph_def = tf_testing.ProcessGraphDefParam(graph_def)
# Add shapes to the graph.
with tf.Session() as sess:
graph_def = tf_testing.AddShapesToGraphDef(sess, "output")
# Import TF graph definition to Relay frontend
shape_dict = {"X": input_shape}
mod, parameters = relay.frontend.from_tensorflow(graph_def,
layout=layout,
shape=shape_dict)
# Compile the graph
with relay.build_config(opt_level=3):
graph, lib, params = relay.build_module.build(mod,
target=target,
params=parameters)
""" Execute and evaluate the graph """
ctx = tvm.cpu()
data_tvm = tvm.nd.array((np.random.uniform(size=input_shape)).astype(dtype))
x = CustomLayer(10, name='MyLayer')(x)
w = tf.Variable(tf.zeros([10, 8]))
b = tf.Variable(tf.zeros([8]))
x = tf.nn.sigmoid(tf.matmul(x, w) + b, name='custom_layer')
outputs = tf.layers.dense(x, 5, activation='softmax', name='Outputs')
losses = tf.keras.losses.categorical_crossentropy(ph_ytrue, outputs)
# model = keras.models.Model(inputs, outputs)
# model.compile('sgd', loss=tf.keras.losses.CategoricalCrossentropy())
# model.summary()
sess = tf.Session()
sess.run(tf.global_variables_initializer())
graph_def = tf_testing.AddShapesToGraphDef(sess, 'Outputs/Softmax')
mod, params = relay.frontend.from_tensorflow(graph_def, layout=None)
with relay.build_config(opt_level=3):
graph, lib, params = relay.build(mod, target='llvm', params=params)
ctx = tvm.cpu()
e = relay.create_executor(mod=mod, ctx=ctx)
m = tvm.contrib.graph_runtime.create(graph, lib, ctx)
x = tvm.nd.array(np.ones([1, 10]).astype(np.float32))
m.set_input(**params)
m.run(**{'Input_ph': x})
out = m.get_output(0)
print(out)
# model_path = "/home/lesliefang/tvm/test_script/mobilenet_v1/mobilenet_v1_1.0_224_frozen-with-shapes.pb"
# model_path = "/root/.tvm_test_data/tf/InceptionV1/classify_image_graph_def-with_shapes.pb"
with tf_compat_v1.gfile.GFile(model_path, 'rb') as f:
graph_def = tf_compat_v1.GraphDef()
graph_def.ParseFromString(f.read())
graph = tf_compat_v1.import_graph_def(graph_def, name='')
# Call the utility to import the graph definition into default graph.
graph_def = tf_testing.ProcessGraphDefParam(graph_def)
# Add shapes to the graph.
with tf_compat_v1.Session() as sess:
graph_def = tf_testing.AddShapesToGraphDef(sess, OUTPUTS)
#graph_def = tf_testing.AddShapesToGraphDef(sess, 'MobilenetV1/Predictions/Reshape_1')
#graph_def = tf_testing.AddShapesToGraphDef(sess, 'softmax')
shape_dict = {INPUTS: (1, 784)}
#shape_dict = None
mod, params = relay.frontend.from_tensorflow(graph_def,
layout=layout,
shape=shape_dict)
print("Tensorflow protobuf imported to relay frontend.")
with tvm.transform.PassContext(opt_level=3):
graph, lib, params = relay.build(mod,
target=target,
def export_module(opts):
# Target settings
layout = "NCHW"
# Download required files
from tvm.contrib.download import download_testdata
model_path = download_testdata(model_url, model_file_name, module=['tf', 'keyword_spotting'])
label_path = download_testdata(label_url, label_name, module=['data'])
# Import model
with tf_compat_v1.gfile.GFile(model_path, 'rb') as f:
graph_def = tf_compat_v1.GraphDef()
graph_def.ParseFromString(f.read())
graph = tf.import_graph_def(graph_def, name='')
graph_def = tf_testing.ProcessGraphDefParam(graph_def)
with tf_compat_v1.Session() as sess:
graph_def = tf_testing.AddShapesToGraphDef(sess, 'labels_softmax')
build_dir = opts.out_dir
if not os.path.exists(build_dir):
os.makedirs(build_dir)
##save original TF graph
if DEBUG_LOG:
with open(os.path.join(build_dir, f'{model_name}_graph_original.log'), 'w') as orig_file:
orig_file.write(str(graph_def))
##remove pre-processing nodes and fix begining
nodes = []
##add first op
input_dim0 = 1
input_dim1 = 49
input_dim2 = 10
new_input = graph_def.node.add()
new_input.op = 'Placeholder'
new_input.name = 'Mfcc'
new_input.attr["dtype"].CopyFrom(attr_value_pb2.AttrValue(
type=dtypes.float32.as_datatype_enum))
nodes.append(new_input)
removed_count = 0
for ii, node in enumerate(graph_def.node, start=0):
if node.op == 'DecodeWav' \
or node.op == 'AudioSpectrogram' \
or node.op == 'Mfcc' \
or node.op == 'Placeholder' \
or node.op == 'wav_data':
removed_count += 1
pass
else:
nodes.append(node)
print(f'NUM of layers removed: {removed_count}')
new_graph = tf_compat_v1.GraphDef()
new_graph.node.extend(nodes)
##log new graph
if DEBUG_LOG:
with open(os.path.join(build_dir, f'{model_name}_graph_new.log'), 'w') as new_graph_log:
new_graph_log.write(str(new_graph))
##get mod and params with new graph
shape_dict = {'Mfcc': (1, 49, 10)}
mod, params = relay.frontend.from_tensorflow(new_graph,
layout=layout,
shape=shape_dict)
if DEBUG_LOG:
with open(os.path.join(build_dir, f'{model_name}_mod.log'), 'w') as mod_file:
mod_file.write(str(mod))
with open(os.path.join(build_dir, f'{model_name}_param.log'), 'w') as param_log:
param_log.write(str(params))
#quantization
if opts.quantize:
if not opts.global_scale:
raise RuntimeError('Global Scale is not valid!')
global_scale = float(opts.global_scale)
print('INFO: Quantizing...')
print(f'INFO: Global Scale: {global_scale}')
with relay.quantize.qconfig(calibrate_mode='global_scale',
global_scale=global_scale,
skip_conv_layers=[0]):
mod = relay.quantize.quantize(mod, params)
if DEBUG_LOG:
with open(os.path.join(build_dir, f'{model_name}_mod_quantized.log'), 'w') as mod_log:
mod_log.write(str(mod))
#save module
if opts.quantize:
file_path = f'{build_dir}/module_gs_{global_scale}.pickle'
with open(file_path, 'wb') as h1:
pickle.dump(mod, h1, protocol=pickle.HIGHEST_PROTOCOL)
print(f'INFO: {file_path} saved!')
with open(f'{build_dir}/module_gs_{global_scale}.txt', 'w') as f:
f.write(mod.astext())
else:
file_path = f'{build_dir}/module.pickle'
with open(file_path, 'wb') as h1:
pickle.dump(mod, h1, protocol=pickle.HIGHEST_PROTOCOL)
print(f'INFO: {file_path} saved!')
param_path = f'{build_dir}/params.bin'
with open(param_path, 'wb') as f_params:
f_params.write(relay.save_param_dict(params))
print(f'INFO: {param_path} saved!')
with open(f'{build_dir}/module.txt', 'w') as f:
f.write(mod.astext())
return mod, params