def evaluate_single():
# PreProcess
if not FLAGS.image:
raise ValueError('Please supply a image')
if tf.gfile.Exists(FLAGS.image) == False:
raise ValueError('Image not found.')
if tf.gfile.Exists(FLAGS.ckpt) == False:
raise ValueError('Please supply a checkpoint')
if FLAGS.num_of_classes == None:
raise ValueError('Please supply num_of_classes.')
if FLAGS.event_dir == None:
FLAGS.event_dir = os.path.join(FLAGS.ckpt, 'event')
#raise ValueError('Please supply a event_dir')
if tf.gfile.Exists(FLAGS.event_dir):
tf.gfile.DeleteRecursively(FLAGS.event_dir)
tf.gfile.MakeDirs(FLAGS.event_dir)
with tf.Graph().as_default() as g:
nn = CNN(FLAGS.network, FLAGS.num_of_classes, 1, FLAGS.image_size,
FLAGS.image_crop_size, True, FLAGS.grayscale, True,
FLAGS.use_fp16)
with tf.device("/cpu:0"):
file_data = tf.read_file(FLAGS.image)
image = tf.image.decode_jpeg(file_data, channels=3)
image = nn.input(image, True, False)
dtype = tf.float16 if FLAGS.use_fp16 else tf.float32
keep_prob = tf.constant(1.0, dtype=dtype)
batch_size = tf.constant(1)
logits = nn.inference(image, keep_prob, batch_size)
# Calculate predictions.
logits = tf.cast(logits, tf.float32)
softmax = tf.nn.softmax(logits)
# Restore the moving average version of the learned variables for eval.
variable_averages = tf.train.ExponentialMovingAverage(
CNN.MOVING_AVERAGE_DECAY)
variables_to_restore = variable_averages.variables_to_restore()
saver = tf.train.Saver(variables_to_restore)
# Build the summary operation based on the TF collection of Summaries.
info = [['Image', str(FLAGS.image)],
['Image size', str(FLAGS.image_size)],
['Image Crop size',
str(FLAGS.image_crop_size)],
['Grayscale', str(FLAGS.grayscale)],
['Use Float16', str(FLAGS.use_fp16)]]
info_summary = tf.summary.text('Info',
tf.convert_to_tensor(info),
collections=[])
summary_op = tf.summary.merge_all()
summary_writer = tf.summary.FileWriter(FLAGS.event_dir, g)
with tf.Session() as sess:
summary_writer.add_summary(sess.run(info_summary))
ckpt = tf.train.get_checkpoint_state(FLAGS.ckpt)
if ckpt and ckpt.model_checkpoint_path:
saver.restore(sess, ckpt.model_checkpoint_path)
global_step = ckpt.model_checkpoint_path.split('/')[-1].split(
'-')[-1]
else:
print('No checkpoint file found')
return
result = sess.run(softmax)
result = result[0]
print('- Result')
for i, v in enumerate(result):
print('[%d]: %f' % (i, v))
summary = tf.Summary()
summary.ParseFromString(sess.run(summary_op))
summary_writer.add_summary(summary, global_step)
def export():
output_dir = os.path.join(FLAGS.ckpt, "export")
if tf.gfile.Exists(output_dir) == True:
tf.gfile.DeleteRecursively(output_dir)
if tf.gfile.Exists(output_dir) == False:
tf.gfile.MakeDirs(output_dir)
### EDIT TRAINING GRAPH ###
with tf.Graph().as_default() as g:
nn = CNN(FLAGS.network, FLAGS.num_of_classes, 1, FLAGS.image_size,
FLAGS.image_crop_size, FLAGS.log_input, FLAGS.grayscale,
FLAGS.log_feature, FLAGS.use_fp16)
with tf.device("/cpu:0"):
image = tf.placeholder(tf.float32,
shape=(None, None, 3),
name="image")
image = nn.input(image, True, False)
dtype = tf.float16 if FLAGS.use_fp16 else tf.float32
keep_prob = tf.constant(1.0, dtype=dtype)
batch_size = tf.constant(1)
logits = nn.inference(image, keep_prob, batch_size)
# Calculate predictions.
logits = tf.cast(logits, tf.float32)
softmax = tf.nn.softmax(logits)
# Restore the moving average version of the learned variables for eval.
variable_averages = tf.train.ExponentialMovingAverage(
CNN.MOVING_AVERAGE_DECAY)
variables_to_restore = variable_averages.variables_to_restore()
saver = tf.train.Saver(variables_to_restore)
# Build the summary operation based on the TF collection of Summaries.
"""
summary_op = tf.summary.merge_all()
summary_writer = tf.summary.FileWriter(output_dir, g)
"""
with tf.Session() as sess:
ckpt = tf.train.get_checkpoint_state(FLAGS.ckpt)
if ckpt and ckpt.model_checkpoint_path:
saver.restore(sess, ckpt.model_checkpoint_path)
global_step = ckpt.model_checkpoint_path.split('/')[-1].split(
'-')[-1]
else:
print('No checkpoint file found')
return
tf.train.Saver().save(
sess,
os.path.join(output_dir, 'model.ckpt'),
global_step=tf.convert_to_tensor(global_step))
tf.train.write_graph(sess.graph.as_graph_def(),
output_dir,
'graph.pbtxt',
as_text=True)
"""
summary = tf.Summary()
summary.ParseFromString(sess.run(summary_op))
summary_writer.add_summary(summary, global_step)
"""
### EXPORT MODEL ###
graph_path = os.path.join(output_dir, 'graph.pbtxt')
if tf.gfile.Exists(graph_path) == False:
raise ValueError('Graph not found({})'.format(graph_path))
ckpt = tf.train.get_checkpoint_state(output_dir)
ckpt_path = ckpt.model_checkpoint_path
if ckpt == False or ckpt_path == False:
raise ValueError('Check point not found.')
output_path = os.path.join(output_dir, 'frozen.pb')
optimized_output_path = os.path.join(output_dir, 'optimized.pb')
freeze_graph.freeze_graph(input_graph=graph_path,
input_saver="",
input_binary=False,
input_checkpoint=ckpt_path,
output_node_names="softmax_linear/softmax",
restore_op_name="save/restore_all",
filename_tensor_name="save/Const:0",
output_graph=output_path,
clear_devices=True,
initializer_nodes="")
input_graph_def = tf.GraphDef()
with tf.gfile.Open(output_path, "r") as f:
data = f.read()
input_graph_def.ParseFromString(data)
output_graph_def = optimize_for_inference_lib.optimize_for_inference(
input_graph_def, ['image'], ["softmax_linear/softmax"],
tf.float32.as_datatype_enum)
f = tf.gfile.FastGFile(optimized_output_path, "w")
f.write(output_graph_def.SerializeToString())
output_size = os.path.getsize(output_path)
optimized_output_size = os.path.getsize(optimized_output_path)
print('Model Exported successfuly.')
print('- Frozen Model: {} ({})'.format(output_path,
_humansize(output_size)))
print('- Optimized Model: {} ({})'.format(
optimized_output_path, _humansize(optimized_output_size)))
def export_serving():
output_dir = os.path.join(FLAGS.ckpt, "export_serving")
if tf.gfile.Exists(output_dir) == True:
tf.gfile.DeleteRecursively(output_dir)
if tf.gfile.Exists(output_dir) == False:
tf.gfile.MakeDirs(output_dir)
### EDIT TRAINING GRAPH ###
with tf.Graph().as_default() as g:
nn = CNN(FLAGS.network, FLAGS.num_of_classes, 1, FLAGS.image_size,
FLAGS.image_crop_size, FLAGS.log_input, FLAGS.grayscale,
FLAGS.log_feature, FLAGS.use_fp16)
with tf.device("/cpu:0"):
image = tf.placeholder(tf.float32,
shape=(None, None, 3),
name="image")
image = nn.input(image, True, False)
dtype = tf.float16 if FLAGS.use_fp16 else tf.float32
keep_prob = tf.constant(1.0, dtype=dtype)
batch_size = tf.constant(1)
logits = nn.inference(image, keep_prob, batch_size)
# Calculate predictions.
logits = tf.cast(logits, tf.float32)
softmax = tf.nn.softmax(logits)
# Restore the moving average version of the learned variables for eval.
variable_averages = tf.train.ExponentialMovingAverage(
CNN.MOVING_AVERAGE_DECAY)
variables_to_restore = variable_averages.variables_to_restore()
saver = tf.train.Saver(variables_to_restore)
# Build the summary operation based on the TF collection of Summaries.
"""
summary_op = tf.summary.merge_all()
summary_writer = tf.summary.FileWriter(output_dir, g)
"""
with tf.Session() as sess:
ckpt = tf.train.get_checkpoint_state(FLAGS.ckpt)
if ckpt and ckpt.model_checkpoint_path:
saver.restore(sess, ckpt.model_checkpoint_path)
global_step = ckpt.model_checkpoint_path.split('/')[-1].split(
'-')[-1]
else:
print('No checkpoint file found')
return
tf.train.Saver().save(
sess,
os.path.join(output_dir, 'model.ckpt'),
global_step=tf.convert_to_tensor(global_step))
tf.train.write_graph(sess.graph.as_graph_def(),
output_dir,
'graph.pbtxt',
as_text=True)
"""
summary = tf.Summary()
summary.ParseFromString(sess.run(summary_op))
summary_writer.add_summary(summary, global_step)
"""
### EXPORT MODEL ###
export_path_base = output_dir
export_path = os.path.join(
tf.compat.as_bytes(export_path_base),
tf.compat.as_bytes(str(FLAGS.model_version)))
print('Exporting model to '.format(export_path))
builder = tf.saved_model.builder.SavedModelBuilder(export_path)
# Build the signature_def_map.
classification_inputs = tf.saved_model.utils.build_tensor_info(
image)
classification_outputs_classes = tf.saved_model.utils.build_tensor_info(
softmax)
classification_outputs_scores = tf.saved_model.utils.build_tensor_info(
softmax)
classification_signature = (
tf.saved_model.signature_def_utils.build_signature_def(
inputs={
tf.saved_model.signature_constants.CLASSIFY_INPUTS:
classification_inputs
},
outputs={
tf.saved_model.signature_constants.CLASSIFY_OUTPUT_CLASSES:
classification_outputs_classes,
tf.saved_model.signature_constants.CLASSIFY_OUTPUT_SCORES:
classification_outputs_scores
},
method_name=tf.saved_model.signature_constants.
CLASSIFY_METHOD_NAME))
tensor_info_x = tf.saved_model.utils.build_tensor_info(image)
tensor_info_y = tf.saved_model.utils.build_tensor_info(softmax)
prediction_signature = (
tf.saved_model.signature_def_utils.build_signature_def(
inputs={'image': tensor_info_x},
outputs={'score': tensor_info_y},
method_name=tf.saved_model.signature_constants.
PREDICT_METHOD_NAME))
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={
'predict_image':
prediction_signature,
tf.saved_model.signature_constants.DEFAULT_SERVING_SIGNATURE_DEF_KEY:
classification_signature,
},
legacy_init_op=legacy_init_op)
builder.save()
print('Done exporting!')
