def export_train_graph(modelClass, optimizerClass, height, width, channels,
classes):
graph = tf.Graph()
with graph.as_default():
# 1. instantiate the model
model = modelClass(width, height, channels, classes)
# 4. export train graph
filename = "%s_%dx%dx%d_%d.meta" % (model.name.lower(), height, width,
channels, classes)
if os.path.exists(filename):
print("file %s exists. skipping" % filename)
return
# 2. instantiate the optimizer
optimizer = optimizerClass()
# 3. instantiate the train wrapper
trainStrategy = train.ImageClassificationTrainStrategy(
graph, model, optimizer)
saver = tf.train.Saver()
init = tf.global_variables_initializer()
tf.add_to_collection("init", init.name)
tf.add_to_collection("train", trainStrategy.optimize)
tf.add_to_collection("logits", trainStrategy.logits)
tf.add_to_collection("summaries", trainStrategy.summary_op)
tf.add_to_collection("predictions", model.predictions)
meta = json.dumps({
"inputs": {
"batch_image_input": trainStrategy.inputs.name,
"categorical_labels": trainStrategy.labels.name
},
"outputs": {
"categorical_logits": model.logits.name
},
"metrics": {
"accuracy": trainStrategy.accuracy.name,
"total_loss": trainStrategy.loss.name
},
"parameters": {
"global_step": trainStrategy.global_step.name,
"learning_rate": trainStrategy._optimizer.learning_rate.name,
"momentum": trainStrategy._optimizer.momentum.name
},
})
tf.add_to_collection("meta", meta)
tf.train.export_meta_graph(filename=filename,
saver_def=saver.as_saver_def())
print("model exported to ", filename)
def generate_train_graph(model_class, optimizer_class,
width, height, channels, classes, add_summaries=False):
graph = tf.Graph()
with graph.as_default():
tf.placeholder(tf.bool, name="global_is_training")
# 1. instantiate the model
model = model_class(width, height, channels, classes)
batch_size = tf.placeholder(tf.float32, [], name="batch_size")
# 2. instantiate the optimizer
optimizer = optimizer_class()
# 3. instantiate the train wrapper
train_strategy = train.ImageClassificationTrainStrategy(
graph, model, optimizer, batch_size, add_summaries=add_summaries)
return train_strategy
def generate_train_graph(modelClass, optimizerClass, width, height, channels,
classes):
graph = tf.Graph()
with graph.as_default():
# 1. instantiate the model
model = modelClass(width, height, channels, classes)
# 2. instantiate the optimizer
optimizer = optimizerClass()
# 3. instantiate the train wrapper
trainStrategy = train.ImageClassificationTrainStrategy(
graph, model, optimizer)
init = tf.global_variables_initializer()
tf.add_to_collection("init", init.name)
return trainStrategy
def export_train_graph(model_class, optimizer_class,
height, width, channels, classes, output=None):
graph = tf.Graph()
with graph.as_default():
global_is_training = tf.placeholder(tf.bool, name="global_is_training")
batch_size = tf.placeholder(tf.float32, [], name="batch_size")
# 1. instantiate the model
model = model_class(width, height, channels, classes)
# 2. export train graph
filename = "%s_%dx%dx%d_%d.meta" % (model.name.lower(),
width,
height,
channels,
classes)
if output is not None:
if not output.endswith('/'):
output = output + '/'
filename = output + filename
if os.path.exists(filename):
print("file %s exists. skipping" % filename)
return
# 3. instantiate the optimizer
optimizer = optimizer_class()
# 4. instantiate the train wrapper
train_strategy = train.ImageClassificationTrainStrategy(
graph,
model,
optimizer,
batch_size,
add_summaries=True,
)
saver = tf.train.Saver()
init = tf.global_variables_initializer()
tf.add_to_collection(ops.GraphKeys.INIT_OP, init.name)
tf.add_to_collection(ops.GraphKeys.TRAIN_OP, train_strategy.optimize)
tf.add_to_collection("logits", train_strategy.logits)
tf.add_to_collection(ops.GraphKeys.SUMMARY_OP, train_strategy.summary_op)
tf.add_to_collection("predictions", model.predictions)
tf.add_to_collection(ops.GraphKeys.LOSSES, model.predictions)
basic_params = {
"inputs": {"batch_image_input": train_strategy.inputs.name,
"categorical_labels": train_strategy.labels.name},
"outputs": {"categorical_logits": model.logits.name,
# This can be removed when TF Java API implements get_operations
"layers": ','.join([m.name for m in graph.get_operations()])},
"metrics": {"accuracy": train_strategy.accuracy.name,
"total_loss": train_strategy.loss.name},
"parameters": {
"global_step": train_strategy.global_step.name,
"learning_rate": train_strategy._optimizer.learning_rate.name,
"momentum": train_strategy._optimizer.momentum.name,
"batch_size": train_strategy._batch_size.name,
"global_is_training": global_is_training.name},
}
if hasattr(model, 'hidden_dropout'):
basic_params['parameters']['hidden_dropout'] = model._hidden_dropout.name
if hasattr(model, 'input_dropout'):
basic_params['parameters']['input_dropout'] = model._input_dropout.name
if hasattr(model, 'activations'):
basic_params['parameters']['activations'] = model._activations.name
meta = json.dumps(basic_params)
tf.add_to_collection("meta", meta)
tf.train.export_meta_graph(filename=filename,
saver_def=saver.as_saver_def())
print("model exported to ", filename)