def _v1_single_metagraph_saved_model(self, use_resource):
export_graph = ops.Graph()
with export_graph.as_default():
start = array_ops.placeholder(
shape=[None], dtype=dtypes.float32, name="start")
if use_resource:
distractor = variables.RefVariable(-1., name="distractor")
v = resource_variable_ops.ResourceVariable(3., name="v")
else:
# "distractor" gets saved in the checkpoint and so used in the restore
# function, but not in the pruned function for the signature. This tests
# node naming: it needs to be consistent (and ideally always the same as
# the node in the original GraphDef) for the resource manager to find
# the right variable.
distractor = variables.RefVariable(-1., name="distractor")
v = variables.RefVariable(3., name="v")
local_variable = variables.VariableV1(
1.,
collections=[ops.GraphKeys.LOCAL_VARIABLES],
trainable=False,
use_resource=True)
output = array_ops.identity(start * v * local_variable, name="output")
with session_lib.Session() as session:
session.run([v.initializer, distractor.initializer,
local_variable.initializer])
path = os.path.join(self.get_temp_dir(), "saved_model", str(ops.uid()))
simple_save.simple_save(
session,
path,
inputs={"start": start},
outputs={"output": output},
legacy_init_op=local_variable.initializer)
return path
def _v1_nested_while_saved_model(self):
export_graph = ops.Graph()
with export_graph.as_default():
def _inner_while(loop_iterations):
_, output = control_flow_ops.while_loop(
lambda index, accum: index <= loop_iterations,
lambda index, accum: (index + 1, accum + index),
[constant_op.constant(0), constant_op.constant(0)])
return output
loop_iterations = array_ops.placeholder(
name="loop_iterations", shape=[], dtype=dtypes.int32)
_, output = control_flow_ops.while_loop(
lambda index, accum: index <= loop_iterations,
lambda index, accum: (index + 1, accum + _inner_while(index)),
[constant_op.constant(0), constant_op.constant(0)])
with session_lib.Session() as session:
path = os.path.join(self.get_temp_dir(), "saved_model", str(ops.uid()))
simple_save.simple_save(
session,
path,
inputs={"loop_iterations": loop_iterations},
outputs={"output": output})
return path
def _model_with_defun(self):
"""Generate a graph with a Defun and serialize in V1 format."""
export_graph = ops.Graph()
with export_graph.as_default():
@framework_function.Defun(dtypes.int64)
def z(x):
return x + 1
@framework_function.Defun(dtypes.int64)
def g(x):
return z(x) + 1
@framework_function.Defun(dtypes.int64)
def f(x):
return g(x) + 1
in_placeholder = array_ops.placeholder(dtype=dtypes.int64,
shape=[1])
out = f(in_placeholder)
with session_lib.Session() as session:
path = os.path.join(self.get_temp_dir(), "saved_model",
str(ops.uid()))
simple_save.simple_save(session,
path,
inputs={"start": in_placeholder},
outputs={"output": out})
return path
def _v1_single_metagraph_saved_model(self, use_resource):
export_graph = ops.Graph()
with export_graph.as_default():
start = array_ops.placeholder(
shape=None, dtype=dtypes.float32, name="start")
if use_resource:
distractor = variables.RefVariable(-1., name="distractor")
v = resource_variable_ops.ResourceVariable(3., name="v")
else:
# "distractor" gets saved in the checkpoint and so used in the restore
# function, but not in the pruned function for the signature. This tests
# node naming: it needs to be consistent (and ideally always the same as
# the node in the original GraphDef) for the resource manager to find
# the right variable.
distractor = variables.RefVariable(-1., name="distractor")
v = variables.RefVariable(3., name="v")
local_variable = variables.VariableV1(
1.,
collections=[ops.GraphKeys.LOCAL_VARIABLES],
trainable=False,
use_resource=True)
output = array_ops.identity(start * v * local_variable, name="output")
with session_lib.Session() as session:
session.run([v.initializer, distractor.initializer,
local_variable.initializer])
path = os.path.join(self.get_temp_dir(), "saved_model", str(ops.uid()))
simple_save.simple_save(
session,
path,
inputs={"start": start},
outputs={"output": output},
legacy_init_op=local_variable.initializer)
return path
def test_resave_signature(self):
# Tests that signatures saved using TF1 can be resaved with TF2.
# See b/211666001 for context.
export_graph = ops.Graph()
with export_graph.as_default():
a = array_ops.placeholder(shape=[None, 1],
dtype=dtypes.float32,
name="input_2")
b = array_ops.placeholder(shape=[None, 2],
dtype=dtypes.float32,
name="input_1")
c = array_ops.identity(a)
with session_lib.Session() as session:
path = os.path.join(self.get_temp_dir(), "saved_model",
str(ops.uid()))
simple_save.simple_save(session,
path,
inputs={
"a": a,
"b": b
},
outputs={"c": c})
imported = load.load(path)
path2 = os.path.join(self.get_temp_dir(), "saved_model",
str(ops.uid()))
save.save(imported, path2, imported.signatures)
imported2 = load.load(path2)
self.assertEqual(
imported2.signatures["serving_default"](
a=constant_op.constant([5.]),
b=constant_op.constant([1., 3.]))["c"].numpy(), 5.)
def test_load_and_restore_partitioned_variables(self):
export_graph = ops.Graph()
with export_graph.as_default():
partitioned_var = variable_scope.get_variable(
"a",
shape=[6],
initializer=init_ops.constant_initializer(13),
partitioner=partitioned_variables.fixed_size_partitioner(2),
use_resource=True)
x = array_ops.placeholder(shape=[], dtype=dtypes.float32)
y = x * partitioned_var
with session_lib.Session() as session:
session.run(variables.global_variables_initializer())
path = os.path.join(self.get_temp_dir(), "saved_model",
str(ops.uid()))
simple_save.simple_save(session,
path,
inputs={"x": x},
outputs={"y": y})
# Create a name-based checkpoint with different values.
session.run(partitioned_var.assign([[5, 4, 3], [2, 1, 0]]))
ckpt_path = os.path.join(self.get_temp_dir(), "restore_ckpt")
saver.Saver().save(session, ckpt_path)
imported = load.load(path)
self.assertAllClose(self.evaluate(imported.variables),
[[13, 13, 13], [13, 13, 13]])
self.evaluate(imported.restore(ckpt_path))
self.assertAllClose(self.evaluate(imported.variables),
[[5, 4, 3], [2, 1, 0]])
self.assertAllClose(
self.evaluate(imported.signatures["serving_default"](
constant_op.constant(2.))), {"y": [10, 8, 6, 4, 2, 0]})
def _v1_asset_saved_model(self):
export_graph = ops.Graph()
vocab_path = os.path.join(self.get_temp_dir(), "vocab.txt")
with open(vocab_path, "w") as f:
f.write("alpha\nbeta\ngamma\n")
with export_graph.as_default():
initializer = lookup_ops.TextFileInitializer(
vocab_path,
key_dtype=dtypes.string,
key_index=lookup_ops.TextFileIndex.WHOLE_LINE,
value_dtype=dtypes.int64,
value_index=lookup_ops.TextFileIndex.LINE_NUMBER)
table = lookup_ops.HashTable(initializer, default_value=-1)
start = array_ops.placeholder(shape=None,
dtype=dtypes.string,
name="in")
output = table.lookup(start, name="out")
with session_lib.Session() as session:
session.run([table.initializer])
path = os.path.join(self.get_temp_dir(), "saved_model",
str(ops.uid()))
simple_save.simple_save(session,
path,
inputs={"start": start},
outputs={"output": output},
legacy_init_op=table.initializer)
file_io.delete_file(vocab_path)
return path
def _v1_nested_while_saved_model(self):
export_graph = ops.Graph()
with export_graph.as_default():
def _inner_while(loop_iterations):
_, output = control_flow_ops.while_loop(
lambda index, accum: index <= loop_iterations,
lambda index, accum: (index + 1, accum + index),
[constant_op.constant(0),
constant_op.constant(0)])
return output
loop_iterations = array_ops.placeholder(name="loop_iterations",
shape=[],
dtype=dtypes.int32)
_, output = control_flow_ops.while_loop(
lambda index, accum: index <= loop_iterations,
lambda index, accum: (index + 1, accum + _inner_while(index)),
[constant_op.constant(0),
constant_op.constant(0)])
with session_lib.Session() as session:
path = os.path.join(self.get_temp_dir(), "saved_model",
str(ops.uid()))
simple_save.simple_save(
session,
path,
inputs={"loop_iterations": loop_iterations},
outputs={"output": output})
return path
def _v1_asset_saved_model(self):
export_graph = ops.Graph()
vocab_path = os.path.join(self.get_temp_dir(), "vocab.txt")
with open(vocab_path, "w") as f:
f.write("alpha\nbeta\ngamma\n")
with export_graph.as_default():
initializer = lookup_ops.TextFileInitializer(
vocab_path,
key_dtype=dtypes.string,
key_index=lookup_ops.TextFileIndex.WHOLE_LINE,
value_dtype=dtypes.int64,
value_index=lookup_ops.TextFileIndex.LINE_NUMBER)
table = lookup_ops.HashTable(
initializer, default_value=-1)
start = array_ops.placeholder(
shape=None, dtype=dtypes.string, name="in")
output = table.lookup(start, name="out")
with session_lib.Session() as session:
session.run([table.initializer])
path = os.path.join(self.get_temp_dir(), "saved_model", str(ops.uid()))
simple_save.simple_save(
session,
path,
inputs={"start": start},
outputs={"output": output},
legacy_init_op=table.initializer)
file_io.delete_file(vocab_path)
return path
def _v1_single_metagraph_saved_model(self, use_resource):
export_graph = ops.Graph()
with export_graph.as_default():
start = array_ops.placeholder(shape=[None],
dtype=dtypes.float32,
name="start")
if use_resource:
v = resource_variable_ops.ResourceVariable(3.)
else:
v = variables.RefVariable(3.)
local_variable = variables.VariableV1(
1.,
collections=[ops.GraphKeys.LOCAL_VARIABLES],
trainable=False,
use_resource=True)
output = array_ops.identity(start * v * local_variable,
name="output")
with session_lib.Session() as session:
session.run([v.initializer, local_variable.initializer])
path = os.path.join(self.get_temp_dir(), "saved_model",
str(ops.uid()))
simple_save.simple_save(
session,
path,
inputs={"start": start},
outputs={"output": output},
legacy_init_op=local_variable.initializer)
return path
def _model_with_ragged_input(self):
"""Generate a graph with a RaggedTensor input and serialize in V1 format."""
export_graph = ops.Graph()
with export_graph.as_default():
x = ragged_factory_ops.placeholder(dtypes.float32, 1, [])
y = x * 2
with session_lib.Session() as sess:
path = os.path.join(self.get_temp_dir(), "saved_model", str(ops.uid()))
simple_save.simple_save(sess, path, inputs={"x": x}, outputs={"y": y})
return path
def save(self, directory="export_dir"):
if os.path.isdir(directory):
print("DIRECTORY {} IS NOT EMPTY. REMOVING IN 5 SEC...".format(
directory))
time.sleep(5)
shutil.rmtree(directory)
simple_save(self.session,
directory,
inputs={"inputs_validation": self.inputs_validation},
outputs={"logits_validation": self.logits_validation})
def testSimpleSave(self):
"""Test simple_save that uses the default parameters."""
export_dir = os.path.join(test.get_temp_dir(), "test_simple_save")
# Force the test to run in graph mode.
# This tests a deprecated v1 API that both requires a session and uses
# functionality that does not work with eager tensors (such as
# build_tensor_info as called by predict_signature_def).
with ops.Graph().as_default():
# Initialize input and output variables and save a prediction graph using
# the default parameters.
with self.session(graph=ops.Graph()) as sess:
var_x = self._init_and_validate_variable("var_x", 1)
var_y = self._init_and_validate_variable("var_y", 2)
inputs = {"x": var_x}
outputs = {"y": var_y}
simple_save.simple_save(sess, export_dir, inputs, outputs)
# Restore the graph with a valid tag and check the global variables and
# signature def map.
with self.session(graph=ops.Graph()) as sess:
graph = loader.load(sess, [tag_constants.SERVING], export_dir)
collection_vars = ops.get_collection(
ops.GraphKeys.GLOBAL_VARIABLES)
# Check value and metadata of the saved variables.
self.assertEqual(len(collection_vars), 2)
self.assertEqual(1, collection_vars[0].eval())
self.assertEqual(2, collection_vars[1].eval())
self._check_variable_info(collection_vars[0], var_x)
self._check_variable_info(collection_vars[1], var_y)
# Check that the appropriate signature_def_map is created with the
# default key and method name, and the specified inputs and outputs.
signature_def_map = graph.signature_def
self.assertEqual(1, len(signature_def_map))
self.assertEqual(
signature_constants.DEFAULT_SERVING_SIGNATURE_DEF_KEY,
list(signature_def_map.keys())[0])
signature_def = signature_def_map[
signature_constants.DEFAULT_SERVING_SIGNATURE_DEF_KEY]
self.assertEqual(signature_constants.PREDICT_METHOD_NAME,
signature_def.method_name)
self.assertEqual(1, len(signature_def.inputs))
self._check_tensor_info(signature_def.inputs["x"], var_x)
self.assertEqual(1, len(signature_def.outputs))
self._check_tensor_info(signature_def.outputs["y"], var_y)
def _model_with_sparse_output(self):
"""Generate a graph with a SparseTensor output and serialize in V1 format"""
export_graph = ops.Graph()
with export_graph.as_default():
in_placeholder = array_ops.placeholder(dtype=dtypes.int64, shape=[1])
out_sparse_tensor = sparse_tensor.SparseTensor(
indices=[[0]], values=in_placeholder, dense_shape=[1]) * 2
with session_lib.Session() as session:
path = os.path.join(self.get_temp_dir(), "saved_model", str(ops.uid()))
simple_save.simple_save(
session,
path,
inputs={"start": in_placeholder},
outputs={"output": out_sparse_tensor})
return path
def _v1_cond_saved_model(self):
export_graph = ops.Graph()
with export_graph.as_default():
branch_selector = array_ops.placeholder(
name="branch_selector", shape=[], dtype=dtypes.bool)
output = control_flow_ops.cond(
branch_selector,
lambda: array_ops.ones([]),
lambda: array_ops.zeros([]))
with session_lib.Session() as session:
path = os.path.join(self.get_temp_dir(), "saved_model", str(ops.uid()))
simple_save.simple_save(
session,
path,
inputs={"branch_selector": branch_selector},
outputs={"output": output})
return path
def _v1_cond_saved_model(self):
export_graph = ops.Graph()
with export_graph.as_default():
branch_selector = array_ops.placeholder(
name="branch_selector", shape=[], dtype=dtypes.bool)
output = control_flow_ops.cond(
branch_selector,
lambda: array_ops.ones([]),
lambda: array_ops.zeros([]))
with session_lib.Session() as session:
path = os.path.join(self.get_temp_dir(), "saved_model", str(ops.uid()))
simple_save.simple_save(
session,
path,
inputs={"branch_selector": branch_selector},
outputs={"output": output})
return path
def main():
# create instance of config
config = Configuration()
config.dir_model = sys.argv[1]
target_dir = sys.argv[2]
# build model
model = CodePoSModel(config)
model.build()
model.restore_session(config.dir_model)
# fully export model
simple_save(model.sess,
target_dir,
inputs=model.get_input_dict(),
outputs=model.get_output_dict())
with open(os.path.join(target_dir, 'configuration.json'), 'w') as _:
json.dumps(config.__dict__)
def testSimpleSave(self):
"""Test simple_save that uses the default parameters."""
export_dir = os.path.join(test.get_temp_dir(),
"test_simple_save")
# Initialize input and output variables and save a prediction graph using
# the default parameters.
with self.session(graph=ops.Graph()) as sess:
var_x = self._init_and_validate_variable(sess, "var_x", 1)
var_y = self._init_and_validate_variable(sess, "var_y", 2)
inputs = {"x": var_x}
outputs = {"y": var_y}
simple_save.simple_save(sess, export_dir, inputs, outputs)
# Restore the graph with a valid tag and check the global variables and
# signature def map.
with self.session(graph=ops.Graph()) as sess:
graph = loader.load(sess, [tag_constants.SERVING], export_dir)
collection_vars = ops.get_collection(ops.GraphKeys.GLOBAL_VARIABLES)
# Check value and metadata of the saved variables.
self.assertEqual(len(collection_vars), 2)
self.assertEqual(1, collection_vars[0].eval())
self.assertEqual(2, collection_vars[1].eval())
self._check_variable_info(collection_vars[0], var_x)
self._check_variable_info(collection_vars[1], var_y)
# Check that the appropriate signature_def_map is created with the
# default key and method name, and the specified inputs and outputs.
signature_def_map = graph.signature_def
self.assertEqual(1, len(signature_def_map))
self.assertEqual(signature_constants.DEFAULT_SERVING_SIGNATURE_DEF_KEY,
list(signature_def_map.keys())[0])
signature_def = signature_def_map[
signature_constants.DEFAULT_SERVING_SIGNATURE_DEF_KEY]
self.assertEqual(signature_constants.PREDICT_METHOD_NAME,
signature_def.method_name)
self.assertEqual(1, len(signature_def.inputs))
self._check_tensor_info(signature_def.inputs["x"], var_x)
self.assertEqual(1, len(signature_def.outputs))
self._check_tensor_info(signature_def.outputs["y"], var_y)
def main(_):
# Build the inference graph.
g = tf.Graph()
with g.as_default():
model = ShowAndTellModel(ModelConfig(), mode='inference')
restore_fn = model.build_graph_from_config(FLAGS.checkpoint_path)
# Create the vocabulary.
vocab = Vocabulary(FLAGS.vocab_file)
filenames = []
for file_pattern in FLAGS.input_files.split(","):
filenames.extend(tf.gfile.Glob(file_pattern))
tf.logging.info("Running caption generation on %d files matching %s",
len(filenames), FLAGS.input_files)
with tf.Session(graph=g) as sess:
# Load the model from checkpoint.
restore_fn(sess)
# Prepare the caption generator. Here we are implicitly using the default
# beam search parameters. See caption_generator.py for a description of the
# available beam search parameters.
# generator = CaptionGenerator(model, vocab)
for filename in filenames:
with tf.gfile.GFile(filename, "rb") as f:
image = f.read()
sess.run(model.sentence.initializer)
beamed = sess.run(fetches=["loopx/Exit_1:0"],
feed_dict={"image_bytes:0": image})
#print(beamed)
sentence = [vocab.id_to_word(w) for w in beamed[0]]
sentence = " ".join(sentence)
print(sentence[3:-4])
#continue
simple_save.simple_save(sess, "./saved_model/0010",
{"image": model.image_feed},
{"caption": model.output})
break
def test_task(endpoint: str, bucket: str, model_file: str, examples_file: str) -> str:
"""Connects to served model and tests example MNIST images."""
from minio import Minio
from pathlib import Path
from retrying import retry
from tensorflow.python.keras.backend import get_session
from tensorflow.python.keras.saving import load_model
from tensorflow.python.saved_model.simple_save import simple_save
import numpy as np
import requests
mclient = Minio(
endpoint,
access_key=Path('/secrets/accesskey').read_text(),
secret_key=Path('/secrets/secretkey').read_text(),
secure=False,
)
print('Downloading model')
mclient.fget_object(bucket, model_file, '/models/model.h5')
mclient.fget_object(bucket, examples_file, '/models/examples.npz')
print('Downloaded model, converting it to serving format')
with get_session() as sess:
model = load_model('/models/model.h5')
simple_save(
sess,
'/output/mnist/1/',
inputs={'input_image': model.input},
outputs={t.name: t for t in model.outputs},
)
model_url = 'http://localhost:9001/v1/models/mnist'
@retry(stop_max_delay=30 * 1000)
def wait_for_model():
requests.get(f'{model_url}/versions/1').raise_for_status()
wait_for_model()
response = requests.get(f'{model_url}/metadata')
response.raise_for_status()
assert response.json() == {
'model_spec': {'name': 'mnist', 'signature_name': '', 'version': '1'},
'metadata': {
'signature_def': {
'signature_def': {
'serving_default': {
'inputs': {
'input_image': {
'dtype': 'DT_FLOAT',
'tensor_shape': {
'dim': [
{'size': '-1', 'name': ''},
{'size': '28', 'name': ''},
{'size': '28', 'name': ''},
{'size': '1', 'name': ''},
],
'unknown_rank': False,
},
'name': 'conv2d_input:0',
}
},
'outputs': {
'dense_1/Softmax:0': {
'dtype': 'DT_FLOAT',
'tensor_shape': {
'dim': [{'size': '-1', 'name': ''}, {'size': '10', 'name': ''}],
'unknown_rank': False,
},
'name': 'dense_1/Softmax:0',
}
},
'method_name': 'tensorflow/serving/predict',
}
}
}
},
}
examples = np.load('/models/examples.npz')
assert examples['X'].shape == (10, 28, 28, 1)
assert examples['y'].shape == (10, 10)
response = requests.post(
f'{model_url}:predict', json={'instances': examples['X'].tolist()}
)
response.raise_for_status()
predicted = np.argmax(response.json()['predictions'], axis=1).tolist()
actual = np.argmax(examples['y'], axis=1).tolist()
accuracy = sum(1 for (p, a) in zip(predicted, actual) if p == a) / len(predicted)
if accuracy >= 0.8:
print(f'Got accuracy of {accuracy:0.2f} in mnist model')
else:
raise Exception(f'Low accuracy in mnist model: {accuracy}')
def main(save_dir, distant_dir, walltime):
"""
Atari games have two kinds of inputs,
ram: size (128,)
human: size (screen_height,screen_width,3)
Since the preprocessing stacks several frames (RL_config.num_recent_obs) together,
the input size should be modified accordingly.
"""
game = "SpaceInvaders-v0"
print("Playing game {}".format(game.split("-")[0]))
env = gym.make(game)
checkpoint = os.path.join(save_dir, game + "-dqn.ckpt")
time0 = time.time()
num_actions = env.action_space.n
print("Number of actions is {}".format(num_actions))
rl_conf = RL_config()
rl_model = Prioritized_replay(env, rl_conf)
rl_model.initialize_replay()
obs_shape = rl_model.obs_shape
print("Shape of the input:{}".format(obs_shape))
if len(obs_shape) == 3:
obs_h, obs_w, obs_c = rl_model.obs_shape
obs_ph = tf.placeholder(tf.float32,
shape=(None, obs_h, obs_w, obs_c),
name="obs_ph")
obs_dim = 3
elif obs_shape < 3:
if len(obs_shape) == 1:
obs_h = obs_shape[0]
obs_c = 1
else:
obs_h, obs_c = obs_shape
obs_ph = tf.placeholder(tf.float32,
shape=(None, obs_h, obs_c),
name="obs_ph")
obs_dim = 2
else:
print("obs_shape inconsistent with the model: {}".format(obs_shape))
sys.exit()
indexed_action_ph = tf.placeholder(tf.int32,
shape=(None, 2),
name="indexed_action_ph")
y_ph = tf.placeholder(tf.float32, shape=(None, 1), name="y_ph")
#ph for weights of importance sampling (IS)
is_weight_ph = tf.placeholder(tf.float32, shape=None, name="is_weight_ph")
#3 cnn layers
if obs_dim == 3:
w1 = tf.get_variable("cnn_w1",[8,8,obs_c,32],dtype=tf.float32,\
initializer=tf.contrib.layers.xavier_initializer())
w2 = tf.get_variable("cnn_w2",[4,4,32,64],dtype=tf.float32,\
initializer=tf.contrib.layers.xavier_initializer())
w3 = tf.get_variable("cnn_w3",[3,3,64,64],dtype=tf.float32,\
initializer=tf.contrib.layers.xavier_initializer())
else:
w1 = tf.get_variable("cnn_w1",[8,obs_c,32],dtype=tf.float32,\
initializer=tf.contrib.layers.xavier_initializer())
w2 = tf.get_variable("cnn_w2",[4,32,64],dtype=tf.float32,\
initializer=tf.contrib.layers.xavier_initializer())
w3 = tf.get_variable("cnn_w3",[3,64,64],dtype=tf.float32,\
initializer=tf.contrib.layers.xavier_initializer())
tf.add_to_collection(tf.GraphKeys.REGULARIZATION_LOSSES,
1 / 2 * tf.norm(w1))
if obs_dim == 3:
z1 = tf.nn.conv2d(obs_ph, w1, strides=[1, 4, 4, 1], padding="VALID")
else:
z1 = tf.nn.conv1d(obs_ph, w1, strides=[1, 4, 1], padding="VALID")
b1 = tf.get_variable("cnn_b1",z1.get_shape().as_list()[1:],dtype=tf.float32,\
initializer=tf.zeros_initializer())
a1 = tf.nn.relu(z1 + b1)
tf.add_to_collection(tf.GraphKeys.REGULARIZATION_LOSSES,
1 / 2 * tf.norm(w2))
if obs_dim == 3:
z2 = tf.nn.conv2d(a1, w2, strides=[1, 2, 2, 1], padding="VALID")
else:
z2 = tf.nn.conv1d(a1, w2, strides=[1, 2, 1], padding="VALID")
b2 = tf.get_variable("cnn_b2",z2.get_shape().as_list()[1:],dtype=tf.float32,\
initializer=tf.zeros_initializer())
a2 = tf.nn.relu(z2 + b2)
tf.add_to_collection(tf.GraphKeys.REGULARIZATION_LOSSES,
1 / 2 * tf.norm(w3))
if obs_dim == 3:
z3 = tf.nn.conv2d(a2, w3, strides=[1, 1, 1, 1], padding="VALID")
else:
z3 = tf.nn.conv1d(a2, w3, strides=[1, 1, 1], padding="VALID")
b3 = tf.get_variable("cnn_b3",z3.get_shape().as_list()[1:],dtype=tf.float32,\
initializer=tf.zeros_initializer())
a3 = tf.nn.relu(z3 + b3)
#fully connected relu
a3_flat = tf.contrib.layers.flatten(a3)
w4 = tf.get_variable("fc_w4",[a3_flat.get_shape().as_list()[1],512],dtype=tf.float32,\
initializer=tf.contrib.layers.xavier_initializer())
tf.add_to_collection(tf.GraphKeys.REGULARIZATION_LOSSES,
1 / 2 * tf.norm(w4))
b4 = tf.get_variable("fc_b4",[1,512],dtype=tf.float32,\
initializer=tf.zeros_initializer())
a4 = tf.nn.relu(tf.matmul(a3_flat, w4) + b4)
#linear output layer
wo = tf.get_variable("wo",[512,num_actions],dtype=tf.float32,\
initializer=tf.contrib.layers.xavier_initializer())
tf.add_to_collection(tf.GraphKeys.REGULARIZATION_LOSSES,
1 / 2 * tf.norm(wo))
bo = tf.get_variable("bo",[1,num_actions],dtype=tf.float32,\
initializer=tf.zeros_initializer())
#q is an array with the size of num_actions
q = tf.add(tf.matmul(a4, wo), bo, name="q")
print("Shape of q: {}".format(q.get_shape().as_list()))
#huber loss
preds = tf.reshape(tf.gather_nd(q, indexed_action_ph), [-1, 1],
name="preds")
loss = tf.losses.huber_loss(y_ph, preds, weights=is_weight_ph)
reg_loss = loss + tf.reduce_sum(
tf.get_collection(tf.GraphKeys.REGULARIZATION_LOSSES))
#Adam Optimizer
train_op = tf.train.AdamOptimizer(
learning_rate=rl_conf.learning_rate).minimize(reg_loss)
variable_set = [w1, b1, w2, b2, w3, b3, w4, b4, wo, bo]
with tf.Session() as sess:
saver = tf.train.Saver()
if tf.train.checkpoint_exists(checkpoint):
saver.restore(sess, checkpoint)
print("Restore model from checkpoint.")
else:
sess.run(tf.global_variables_initializer())
print("Files saved to {}".format(save_dir))
#Save model state
print("Save model to save_dir:")
inputs_dict = {
"obs_ph": obs_ph,
"indexed_action_ph": indexed_action_ph,
"y_ph": y_ph,
"is_weight_ph": is_weight_ph
}
outputs_dict = {"preds": preds}
model_dir = os.path.join(save_dir, "model")
if os.path.isdir(model_dir):
model_dir = os.path.join(save_dir, "model-0")
os.makedirs(model_dir)
simple_save(sess, model_dir, inputs_dict, outputs_dict)
q_file = os.path.join(save_dir, "q_values.txt")
reward_file = os.path.join(save_dir, "rewards.txt")
#compute priorites for initial replay
batch_idx = 0
priorities = np.zeros((len(rl_model.replay), 1))
variable_values_q1 = sess.run(variable_set)
var_feed_dict_list_q1 = [
(key, val) for key, val in zip(variable_set, variable_values_q1)
]
while batch_idx < len(rl_model.replay) / rl_conf.batch:
if (batch_idx + 1) * rl_conf.batch > len(rl_model.replay):
samples = rl_model.replay[batch_idx * rl_conf.batch:]
else:
samples = rl_model.replay[batch_idx *
rl_conf.batch:(batch_idx + 1) *
rl_conf.batch]
sample_size = len(samples)
obs_samples = np.array([sample[0] for sample in samples])
obs_next_samples = np.array([sample[3] for sample in samples])
rew_samples = np.array([sample[2] for sample in samples]).reshape(
(sample_size, 1))
done_samples = np.array([sample[4] for sample in samples]).reshape(
(sample_size, 1))
action_samples = np.array([sample[1]
for sample in samples]).astype(int)
q_targets = sess.run(q, feed_dict={obs_ph: obs_next_samples})
q_for_priorities = sess.run(q, feed_dict={obs_ph: obs_samples})
q_targets_selected_actions = np.amax(q_targets,
axis=-1,
keepdims=True)
labels = np.where(
done_samples, rew_samples,
rew_samples + rl_conf.gamma * q_targets_selected_actions)
if (batch_idx + 1) * rl_conf.batch > len(rl_model.replay):
q_priorities_selected_actions = q_for_priorities[
np.arange(len(action_samples)), action_samples].reshape(
(sample_size, 1))
priorities[batch_idx * rl_conf.batch:] = np.absolute(
labels - q_priorities_selected_actions)**(
rl_conf.alpha_prioritized_replay)
else:
q_priorities_selected_actions = q_for_priorities[
np.arange(len(action_samples)), action_samples].reshape(
(sample_size, 1))
priorities[batch_idx * rl_conf.batch:(batch_idx + 1) *
rl_conf.batch] = np.absolute(
labels - q_priorities_selected_actions)**(
rl_conf.alpha_prioritized_replay)
batch_idx += 1
rl_model.initialize_priority_list(priorities.flatten())
update_q1_steps = 0
for episode in np.arange(1, rl_conf.max_episodes + 1):
#Make sure the running time does not exceed walltime
start = time.time()
if start - time0 > float(walltime) * 0.9:
saver.save(sess, checkpoint)
print("Running time limit achieves.")
sys.exit()
#metrics
total_rew = 0.0
average_q = 0.0
s0 = preprocess_frame(env.reset(), rl_conf.shape_of_frame)
zero_pad = np.zeros(s0.shape)
if rl_conf.num_recent_obs == 1:
sequence = []
else:
sequence = [zero_pad] * (rl_conf.num_recent_obs - 1)
sequence.append(s0)
#epsilon schedule
epsilon = max(
1.0 - float(episode) / rl_conf.final_exploration_episodes, 0.1)
#beta_prioritized_replay schedule
beta_prioritized_replay = min(
rl_conf.beta_prioritized_replay +
float(episode) / rl_conf.final_exploration_episodes, 1.0)
for step in np.arange(1, rl_conf.max_steps):
#whether to update variables of q1
if update_q1_steps % rl_conf.steps_for_updating_q1 == 0:
variable_values_q1 = sess.run(variable_set)
var_feed_dict_list_q1 = [
(key, val)
for key, val in zip(variable_set, variable_values_q1)
]
#whether to recompute priority list
if len(
rl_model.replay
) >= rl_conf.replay_capacity and update_q1_steps % rl_model.frequency_to_update_priorities_globally == 0:
rl_model.compute_priority_list(rl_model.priority_list)
if len(
rl_model.replay
) >= rl_conf.replay_capacity and rl_model.group_rbs == []:
rl_model.compute_priority_list(rl_model.priority_list)
update_q1_steps += 1
#double dqn
samples, sample_p_idxes = rl_model.sample_from_replay()
obs_samples = np.array([sample[0] for sample in samples])
obs_next_samples = np.array([sample[3] for sample in samples])
rew_samples = np.array([sample[2]
for sample in samples]).reshape(
(rl_conf.batch, 1))
done_samples = np.array([sample[4]
for sample in samples]).reshape(
(rl_conf.batch, 1))
action_samples = np.array([sample[1]
for sample in samples]).astype(int)
indexed_action_samples = np.array(
[[idx, sample[1]]
for idx, sample in enumerate(samples)]).astype(int)
q_for_selecting_actions = sess.run(
q, feed_dict={obs_ph: obs_next_samples})
selected_actions = np.argmax(q_for_selecting_actions, axis=-1)
q_for_priorities = sess.run(q, feed_dict={obs_ph: obs_samples})
q_targets = sess.run(
q,
feed_dict=dict([(obs_ph, obs_next_samples)] +
var_feed_dict_list_q1))
q_targets_selected_actions = q_targets[
np.arange(len(selected_actions)),
selected_actions].reshape((rl_conf.batch, 1))
labels = np.where(
done_samples, rew_samples,
rew_samples + rl_conf.gamma * q_targets_selected_actions)
#update priorities for replay
if sample_p_idxes is not None:
priorities = np.absolute(labels - q_for_priorities[
np.arange(len(action_samples)),
action_samples].reshape((rl_conf.batch, 1)))**(
rl_conf.alpha_prioritized_replay)
p_indexed_priorities = np.concatenate(
(sample_p_idxes.reshape(
(rl_conf.batch, 1)), priorities),
axis=1)
rl_model.update_priority_after_sampling(
p_indexed_priorities)
weights = (len(rl_model.replay) *
priorities)**(-beta_prioritized_replay)
weights = weights / np.amax(weights)
else:
weights = 1.0
sess.run(train_op,
feed_dict={
obs_ph: obs_samples,
y_ph: labels,
indexed_action_ph: indexed_action_samples,
is_weight_ph: weights
})
#update replay
if step % rl_conf.action_repeat == 1:
#calculate q1 and get optimal action
if step == 1:
obs_input = preprocess_obs(
sequence[:rl_conf.num_recent_obs])
q_values_for_new_sample = sess.run(
q, feed_dict={obs_ph: np.array([obs_input])})
max_action_for_new_sample = np.argmax(
q_values_for_new_sample, axis=-1)
average_q += q_values_for_new_sample[0][
max_action_for_new_sample]
#epsilon greedy algo
dice = np.random.uniform()
if dice < epsilon:
action = np.random.randint(num_actions)
else:
action = max_action_for_new_sample
obs, rew, done, _ = env.step(action)
total_rew += rew
sequence.append(preprocess_frame(obs, rl_conf.shape_of_frame))
last_obs_input = obs_input
obs_input = preprocess_obs(sequence[step:step +
rl_conf.num_recent_obs])
#calculate priority for new sample
if done:
priority = np.absolute(
rew - q_values_for_new_sample[0][action])**(
rl_conf.alpha_prioritized_replay)
else:
q_target_for_new_sample = sess.run(
q,
feed_dict=dict([(obs_ph, np.array([obs_input]))] +
var_feed_dict_list_q1))
priority = np.absolute(rew + rl_conf.gamma*q_target_for_new_sample[0][max_action_for_new_sample] - q_values_for_new_sample[0][action]\
)**(rl_conf.alpha_prioritized_replay)
if np.isscalar(priority):
priority = np.array([priority])
else:
priority = priority.flatten()
rl_model.update_replay(
np.array([(last_obs_input, action, rew, obs_input, done)]),
priority)
if done:
with open(q_file, "a+") as out:
out.write(str(average_q * 4 / step) + "\n")
with open(reward_file, "a+") as out:
out.write(str(total_rew) + "\n")
break
if episode % rl_conf.episodes_to_save == 0:
saver.save(sess, checkpoint)
save_file_to_distant_dir = """cp {} {}""".format(
os.path.join(save_dir, "*"), distant_dir)
subprocess.call(save_file_to_distant_dir, shell=True)
print("Time taken to complete this episode: {}s.".format(
time.time() - start))
#Evaluate the performance of the agent with total rewards
if episode % rl_conf.episodes_to_validate == 0:
total_rew_eval = []
epsilon = 0.05
first_actions = {}
for _ in np.arange(rl_conf.evaluation_trials):
s0 = preprocess_frame(env.reset(), rl_conf.shape_of_frame)
zero_pad = np.zeros(s0.shape)
if rl_conf.num_recent_obs == 1:
sequence = []
else:
sequence = [zero_pad] * (rl_conf.num_recent_obs - 1)
sequence.append(s0)
rew_eval = 0
for step in np.arange(1, rl_conf.max_steps):
if step % rl_conf.action_repeat == 1:
#calculate q1 and get optimal action
if step == 1:
obs_input = preprocess_obs(
sequence[:rl_conf.num_recent_obs])
q_values_for_evaluation = sess.run(
q, feed_dict={obs_ph: np.array([obs_input])})
max_action_for_evaluation = np.argmax(
q_values_for_evaluation, axis=-1)
#epsilon greedy algo
dice = np.random.uniform()
if dice < epsilon:
action = np.random.randint(num_actions)
else:
action = max_action_for_evaluation
obs, rew, done, _ = env.step(action)
rew_eval += rew
sequence.append(
preprocess_frame(obs, rl_conf.shape_of_frame))
last_obs_input = obs_input
obs_input = preprocess_obs(
sequence[step:step + rl_conf.num_recent_obs])
if done:
total_rew_eval.append(rew_eval)
break
print("Evaluating the agent at episode-{}:".format(episode))
total_rew_eval = np.array(total_rew_eval)
print("Total rewards of trials have max-{}, average-{}, std-{}.".format(np.amax(total_rew_eval),\
np.mean(total_rew_eval),np.std(total_rew_eval)))
decoder = tf.matmul(noise_input,
weights['decoder_h1']) + biases['decoder_b1']
decoder = tf.nn.tanh(decoder)
decoder = tf.matmul(decoder,
weights['decoder_out']) + biases['decoder_out']
decoder = tf.nn.sigmoid(decoder)
# Building a manifold of generated digits
n = 20
x_axis = np.linspace(-3, 3, n)
y_axis = np.linspace(-3, 3, n)
canvas = np.empty((28 * n, 28 * n))
for i, yi in enumerate(x_axis):
for j, xi in enumerate(y_axis):
z_mu = np.array([[xi, yi]] * batch_size)
x_mean = sess.run(decoder, feed_dict={noise_input: z_mu})
canvas[(n - i - 1) * 28:(n - i) * 28, j * 28:(j + 1) * 28] = \
x_mean[0].reshape(28, 28)
plt.figure(figsize=(8, 10))
Xi, Yi = np.meshgrid(x_axis, y_axis)
plt.imshow(canvas, origin="upper", cmap="gray")
plt.show()
print("Saving the model")
simple_save(sess,
export_dir='./saved_variational_autoencoder',
inputs={"inp": input_image},
outputs={"out": decoder_export})
def save_model(sess,path_to_model,X,Y):
try:
shutil.rmtree(path_to_model)
except:
pass
simple_save(sess, path_to_model, inputs={"myInput": X}, outputs={"myOutput": Y})
def main(mode: str, user_feedback=None, detections=None):
"""
Trains a model or uses an existent model to make predictions.
Args:
mode: Defines whether a model should be trained
or an existent model should be used for making predictions.
user_feedback: Used in live-train-mode to train the model based on user input.
detections: Used in detection-mode to give a list of detections
for which predictions should be generated.
"""
gpu_options = tf.GPUOptions(per_process_gpu_memory_fraction=0.1)
accept_prob_model_dir = os.path.join(model_dir, 'accept_prob_predictor')
if not os.path.exists(accept_prob_model_dir):
os.mkdir(accept_prob_model_dir)
existent_checkpoints = os.listdir(accept_prob_model_dir)
existent_checkpoints.sort(key=int)
x = tf.placeholder(tf.float32, [None, 1005])
y = prob_model(x)
_y = tf.placeholder(tf.float32, [None, 1])
loss = tf.reduce_mean(tf.nn.sigmoid_cross_entropy_with_logits(logits=y, labels=_y))
saver = tf.train.Saver()
while True:
actual_checkpoint_dir = ''
if len(existent_checkpoints) == 0:
new_checkpoint_dir = os.path.join(accept_prob_model_dir, '1')
break
actual_checkpoint = existent_checkpoints[len(existent_checkpoints) - 1]
actual_checkpoint_dir = os.path.join(accept_prob_model_dir, actual_checkpoint)
if len(os.listdir(actual_checkpoint_dir)) > 0:
new_checkpoint_dir = os.path.join(accept_prob_model_dir,
str(int(actual_checkpoint) + 1))
break
existent_checkpoints.remove(actual_checkpoint)
shutil.rmtree(actual_checkpoint_dir)
if (FLAGS and FLAGS.mode == 'train') or mode == 'train':
# initial train mode with fixed training data
if not user_feedback:
iterations = FLAGS.iterations
learning_rate = FLAGS.learning_rate
base = FLAGS.path_to_training_data
path_to_train_data = '{}_{}'.format(base, 'train.tfrecords')
train_features, train_labels = parse_tf_records(path_to_train_data)
# live train mode with user feedback
else:
iterations = 1
learning_rate = 0.001
batch_size = 64
train_features, train_labels = tf.train.shuffle_batch([user_feedback['x'],
user_feedback['y_']],
batch_size=batch_size,
capacity=50000,
min_after_dequeue=0,
allow_smaller_final_batch=True)
test_feat = []
test_lbl = []
example = tf.train.Example()
for record in tf.python_io.tf_record_iterator(path_to_test_data):
example.ParseFromString(record)
f = example.features.feature
test_feat.append(np.asarray(f['test/feature'].float_list.value))
test_lbl.append(f['test/label'].float_list.value[0])
test_feat = np.reshape(test_feat, (-1, 1005))
test_lbl = np.reshape(test_lbl, (-1, 1))
init_op = tf.group(tf.global_variables_initializer(),
tf.local_variables_initializer())
optimizer = tf.train.GradientDescentOptimizer(learning_rate=learning_rate)
train_op = optimizer.minimize(loss=loss, global_step=tf.train.get_global_step())
best_acc = 0.0
best_acc_ann = 0.0
best_acc_ver = 0.0
early_stopping_counter = 0
with tf.Session(config=tf.ConfigProto(gpu_options=gpu_options)) as sess:
sess.run(init_op)
if len(existent_checkpoints) > 0:
saver.restore(sess, os.path.join(actual_checkpoint_dir, 'prob_predictor.ckpt'))
coord = tf.train.Coordinator()
threads = tf.train.start_queue_runners(coord=coord)
for batch_index in range(iterations):
try:
feat, lbl = sess.run([train_features, train_labels])
except OutOfRangeError:
print('No more Training Data available')
break
train_op.run(feed_dict={x: np.reshape(feat, (-1, 1005)),
_y: np.reshape(lbl, (-1, 1))})
if batch_index % 100 == 0 or user_feedback:
y_test = y.eval(feed_dict={x: test_feat, _y: test_lbl})
acc, acc_ann, acc_ver = evaluate_prediction_record(y_test, test_lbl)
print('step {},\t Annotation acc.:{}\tVerification acc.:{}'.format(batch_index,
acc_ann,
acc_ver))
if user_feedback:
if os.path.exists(new_checkpoint_dir):
shutil.rmtree(new_checkpoint_dir)
simple_save(sess,
new_checkpoint_dir,
inputs={'inputs': x},
outputs={'outputs': y})
saver.save(sess, os.path.join(new_checkpoint_dir, 'prob_predictor.ckpt'))
if acc_ann + acc_ver > best_acc_ann + best_acc_ver:
if os.path.exists(new_checkpoint_dir):
shutil.rmtree(new_checkpoint_dir)
best_acc = acc
best_acc_ann = acc_ann
best_acc_ver = acc_ver
early_stopping_counter = 0
simple_save(sess,
new_checkpoint_dir,
inputs={'inputs': x},
outputs={'outputs': y})
saver.save(sess, os.path.join(new_checkpoint_dir, 'prob_predictor.ckpt'))
# elif early_stopping_counter == 50:
# print('Stopped early at batch {}/{}'.format(batch_index, iterations))
# break
else:
early_stopping_counter += 1
print('Accuracy:\t{}'.format(best_acc))
print('Accuracy Ann:\t{}'.format(best_acc_ann))
print('Accuracy Ver:\t{}'.format(best_acc_ver))
# Stop the threads
coord.request_stop()
# Wait for threads to stop
coord.join(threads)
sess.close()
elif mode == 'predict' and len(existent_checkpoints) > 0:
feature_data = []
for key in detections:
for i, _ in enumerate(detections[key]):
feature_data.append(compute_feature_vector(detections[key][i]))
with tf.Session(config=tf.ConfigProto(gpu_options=gpu_options)) as sess:
saver.restore(sess, os.path.join(actual_checkpoint_dir, 'prob_predictor.ckpt'))
result = y.eval(feed_dict={
x: np.reshape(feature_data, (-1, 1005))})
prediction = tf.round(result).eval()
for i, val in enumerate(prediction):
print('{}: {}'.format(i, val))
return prediction
def main():
logging.getLogger().setLevel(logging.DEBUG)
# Reference variable setup
sess = tf.Session()
K.set_session(sess)
model_version = lib.get_batch_name()
K.set_learning_phase(0)
# Data setup
logging.info('Loading data')
dataframe = pandas.read_csv("iris_with_header.csv")
numerical_cols = ['sepal_length', 'sepal_width']
input_data = list()
for col in numerical_cols:
input_data.append(dataframe[col].values)
logging.info('OHE-ing response variable')
encoder = LabelEncoder()
encoder.fit(dataframe.values[:, 4])
encoded_Y = encoder.transform(dataframe.values[:, 4])
one_hot_labels = np_utils.to_categorical(encoded_Y)
# Model setup
logging.info('Creating model')
input_layers = list()
for col in numerical_cols:
logging.info('Creating input for {}'.format(col))
if len(dataframe[col].shape) > 1:
shape = dataframe[col].shape[1]
else:
shape = 1
logging.info('Inferring variable {} has shape: {}'.format(col, shape))
input_layers.append(Input(shape=(shape, ),
name='{}_input'.format(col)))
layers = Concatenate()(input_layers)
layers = Dense(32)(layers)
layers = Dense(3)(layers)
model = Model(input_layers, layers)
model.compile(loss=losses.categorical_crossentropy, optimizer='adam')
model.fit(input_data, one_hot_labels)
tf_inputs = list()
tf_examples = list()
# Register input placholders
for col in numerical_cols:
logging.info('Creating tf placeholder for col: {}'.format(col))
if len(dataframe[col].shape) > 1:
shape = dataframe[col].shape[1]
else:
shape = 1
logging.info('Inferring variable {} has shape: {}'.format(col, shape))
serialized_tf_example = tf.placeholder(tf.string, name=col)
tf_examples.append(serialized_tf_example)
# TODO Better type lookup based on numpy types
feature_configs = {
col: tf.FixedLenFeature(shape=shape, dtype=tf.float32),
}
tf_example = tf.parse_example(serialized_tf_example, feature_configs)
tf_inputs.append(tf.identity(tf_example[col], name=col))
# Generate output tensor by feeding inputs into model
y = model(tf_inputs)
# Generate classification signature definition
labels = encoder.classes_
values, indices = tf.nn.top_k(y, len(labels))
OHE_index_to_string_lookup_table = tf.contrib.lookup.index_to_string_table_from_tensor(
tf.constant(labels))
prediction_classes = OHE_index_to_string_lookup_table.lookup(
tf.to_int64(indices))
# Save model
output_path = './' + model_version
logging.info('Saving model to {}'.format(output_path))
simple_save_inputs = dict(zip(numerical_cols, tf_inputs))
logging.info('Inputs to simple save: {}'.format(simple_save_inputs))
simple_save(sess, output_path, inputs=simple_save_inputs, outputs={'y': y})
# Can now get signature w/ https://www.tensorflow.org/guide/saved_model#cli_to_inspect_and_execute_savedmodel
# TODO Create tensorflow-serving signature
# TODO validate signatures
# TODO Serialize tensorflow-serving elements
# TODO Save the graph
# TODO Return path to serialized graph
pass
# Reshape data to get 28 seq of 28 elements
batch_x = batch_x.reshape((batch_size, timesteps, num_input))
# Run optimization op (backprop)
sess.run(train_op, feed_dict={X: batch_x, Y: batch_y})
if step % display_step == 0 or step == 1:
# Calculate batch loss and accuracy
loss, acc = sess.run([loss_op, accuracy],
feed_dict={
X: batch_x,
Y: batch_y
})
print("Step " + str(step) + ", Minibatch Loss= " + \
"{:.4f}".format(loss) + ", Training Accuracy= " + \
"{:.3f}".format(acc))
print("Optimization Finished!")
# Calculate accuracy for 128 mnist test images
test_len = 128
test_data = mnist.test.images[:test_len].reshape(
(-1, timesteps, num_input))
test_label = mnist.test.labels[:test_len]
print("Testing Accuracy:", \
sess.run(accuracy, feed_dict={X: test_data, Y: test_label}))
print("Saving the model")
simple_save(sess,
export_dir='./saved_recurrent_network',
inputs={"images": X},
outputs={"out": prediction})
def train(Xl, yl, Xt, yt):
""" Main train
:param Xl training examples tensor, shape [num_examples, num_inputs, num_timesteps]
:param yl one-hot encoded training labels tensor, shape [num_examples, num_classes]
:param Xt test examples tensor, shape [num_examples, num_inputs, num_timesteps]
:param yt one-hot encoded test labels tensor, shape [num_examples, num_classes]
"""
batch_size = 64
num_timesteps = Xl.shape[
2] # number of rows (each row in the image is considered as a timestep)
num_inputs = Xl.shape[1] # length of each row
num_hidden = 128
num_classes = yl.shape[1]
training_steps = 500 # TODO temporary to make faster, 10000 to actually train
learning_rate = 0.001
num_examples = Xl.shape[0]
with tf.Session(graph=tf.Graph()) as sess:
X, lstm_out = lstm_and_dense_layer(num_timesteps,
num_inputs,
num_classes=num_classes,
num_hidden=num_hidden)
y = tf.placeholder(tf.float32, shape=(None, num_classes))
loss_op = tf.reduce_mean(
tf.nn.softmax_cross_entropy_with_logits_v2(logits=lstm_out,
labels=y))
optimizer = tf.train.GradientDescentOptimizer(
learning_rate=learning_rate)
train_op = optimizer.minimize(loss_op)
prediction = tf.nn.softmax(lstm_out, name='prediction')
correct_pred = tf.equal(tf.argmax(prediction, 1), tf.argmax(y, 1))
accuracy_op = tf.reduce_mean(tf.cast(correct_pred, tf.float32))
sess.run(variables.global_variables_initializer())
epoch = -1
batch_start = 0
batch_end = batch_size
def next_batch():
nonlocal epoch, batch_start, batch_end, Xl, yl
if batch_end > num_examples or epoch == -1:
epoch += 1
batch_start = 0
batch_end = batch_size
perm0 = np.arange(num_examples)
np.random.shuffle(perm0)
Xl = Xl[perm0]
yl = yl[perm0]
Xi_ = Xl[batch_start:batch_end, :, :]
yi_ = yl[batch_start:batch_end, :]
batch_start = batch_end
batch_end = batch_start + batch_size
return {X: Xi_, y: yi_}
for step in range(training_steps + 1):
batch = next_batch()
sess.run(train_op, feed_dict=batch)
if step % 100 == 0:
loss_, acc_ = sess.run((loss_op, accuracy_op), feed_dict=batch)
print("epoch", epoch, "step", step, "loss",
"{:.4f}".format(loss_), "acc", "{:.2f}".format(acc_))
print("Optimization Finished!")
print("Testing Accuracy:",
sess.run(accuracy_op, feed_dict={
X: Xt,
y: yt
}))
save_model = False
if save_model:
print("Saving the model")
simple_save(sess,
export_dir='./lstm',
inputs={"images": X},
outputs={"out": prediction})
# Reshape data to get 28 seq of 28 elements
batch_x = batch_x.reshape((batch_size, timesteps, num_input))
# Run optimization op (backprop)
sess.run(train_op, feed_dict={X: batch_x, Y: batch_y})
if step % display_step == 0 or step == 1:
# Calculate batch loss and accuracy
loss, acc = sess.run([loss_op, accuracy],
feed_dict={
X: batch_x,
Y: batch_y
})
print("Step " + str(step) + ", Minibatch Loss= " + \
"{:.4f}".format(loss) + ", Training Accuracy= " + \
"{:.3f}".format(acc))
print("Optimization Finished!")
# Calculate accuracy for 128 mnist test images
test_len = 128
test_data = mnist.test.images[:test_len].reshape(
(-1, timesteps, num_input))
test_label = mnist.test.labels[:test_len]
print("Testing Accuracy:", \
sess.run(accuracy, feed_dict={X: test_data, Y: test_label}))
print("Saving the model")
simple_save(sess,
export_dir='./saved_bidirectional_rnn',
inputs={"inp": X},
outputs={"out": prediction})
def amazon_attribute_train(generator: rmc_att_topic.generator,
discriminator: rmc_att_topic.discriminator,
oracle_loader: RealDataAmazonLoader, config, args):
batch_size = config['batch_size']
num_sentences = config['num_sentences']
vocab_size = config['vocabulary_size']
seq_len = config['seq_len']
dataset = config['dataset']
npre_epochs = config['npre_epochs']
n_topic_pre_epochs = config['n_topic_pre_epochs']
nadv_steps = config['nadv_steps']
temper = config['temperature']
adapt = config['adapt']
# changed to match resources path
data_dir = resources_path(config['data_dir'], "Amazon_Attribute")
log_dir = resources_path(config['log_dir'])
sample_dir = resources_path(config['sample_dir'])
# filename
oracle_file = os.path.join(sample_dir, 'oracle_{}.txt'.format(dataset))
gen_file = os.path.join(sample_dir, 'generator.txt')
gen_text_file = os.path.join(sample_dir, 'generator_text.txt')
gen_text_file_print = os.path.join(sample_dir, 'gen_text_file_print.txt')
json_file = os.path.join(sample_dir, 'json_file.txt')
json_file_validation = os.path.join(sample_dir, 'json_file_validation.txt')
csv_file = os.path.join(log_dir, 'experiment-log-rmcgan.csv')
data_file = os.path.join(data_dir, '{}.txt'.format(dataset))
test_file = os.path.join(data_dir, 'test.csv')
# create necessary directories
if not os.path.exists(data_dir):
os.makedirs(data_dir)
if not os.path.exists(sample_dir):
os.makedirs(sample_dir)
if not os.path.exists(log_dir):
os.makedirs(log_dir)
# placeholder definitions
x_real = tf.placeholder(tf.int32, [batch_size, seq_len],
name="x_real") # tokens of oracle sequences
x_user = tf.placeholder(tf.int32, [batch_size], name="x_user")
x_product = tf.placeholder(tf.int32, [batch_size], name="x_product")
x_rating = tf.placeholder(tf.int32, [batch_size], name="x_rating")
temperature = tf.Variable(1., trainable=False, name='temperature')
x_real_onehot = tf.one_hot(x_real,
vocab_size) # batch_size x seq_len x vocab_size
assert x_real_onehot.get_shape().as_list() == [
batch_size, seq_len, vocab_size
]
# generator and discriminator outputs
generator_obj = generator(x_real=x_real,
temperature=temperature,
x_user=x_user,
x_product=x_product,
x_rating=x_rating)
discriminator_real = discriminator(
x_onehot=x_real_onehot) # , with_out=False)
discriminator_fake = discriminator(
x_onehot=generator_obj.gen_x_onehot_adv) # , with_out=False)
# GAN / Divergence type
log_pg, g_loss, d_loss = get_losses(generator_obj, discriminator_real,
discriminator_fake, config)
# Global step
global_step = tf.Variable(0, trainable=False)
global_step_op = global_step.assign_add(1)
# Train ops
g_pretrain_op, g_train_op, d_train_op, d_topic_pretrain_op = get_train_ops(
config, generator_obj.pretrain_loss, g_loss, d_loss, None, log_pg,
temperature, global_step)
# Record wall clock time
time_diff = tf.placeholder(tf.float32)
Wall_clock_time = tf.Variable(0., trainable=False)
update_Wall_op = Wall_clock_time.assign_add(time_diff)
# Temperature placeholder
temp_var = tf.placeholder(tf.float32)
update_temperature_op = temperature.assign(temp_var)
# Loss summaries
loss_summaries = [
tf.summary.scalar('adv_loss/discriminator/total', d_loss),
tf.summary.scalar('adv_loss/generator/total_g_loss', g_loss),
tf.summary.scalar('adv_loss/log_pg', log_pg),
tf.summary.scalar('adv_loss/Wall_clock_time', Wall_clock_time),
tf.summary.scalar('adv_loss/temperature', temperature),
]
loss_summary_op = tf.summary.merge(loss_summaries)
# Metric Summaries
config['bleu_amazon'] = True
config['bleu_amazon_validation'] = True
metrics_pl, metric_summary_op = get_metric_summary_op(config)
# Summaries
gen_pretrain_loss_summary = CustomSummary(name='pretrain_loss',
scope='generator')
gen_sentences_summary = CustomSummary(name='generated_sentences',
scope='generator',
summary_type=tf.summary.text,
item_type=tf.string)
topic_discr_pretrain_summary = CustomSummary(name='pretrain_loss',
scope='topic_discriminator')
topic_discr_accuracy_summary = CustomSummary(name='pretrain_accuracy',
scope='topic_discriminator')
run_information = CustomSummary(name='run_information',
scope='info',
summary_type=tf.summary.text,
item_type=tf.string)
custom_summaries = [
gen_pretrain_loss_summary, gen_sentences_summary,
topic_discr_pretrain_summary, topic_discr_accuracy_summary,
run_information
]
# To save the trained model
saver = tf.train.Saver()
# ------------- initial the graph --------------
with init_sess() as sess:
variables_dict = get_parameters_division()
log = open(csv_file, 'w')
sum_writer = tf.summary.FileWriter(os.path.join(log_dir, 'summary'),
sess.graph)
for custom_summary in custom_summaries:
custom_summary.set_file_writer(sum_writer, sess)
run_information.write_summary(str(args), 0)
print("Information stored in the summary!")
metrics = get_metrics(config, oracle_loader, sess, json_file,
json_file_validation, generator_obj)
gc.collect()
# Check if there is a pretrained generator saved
model_dir = "PretrainGenerator"
model_path = resources_path(
os.path.join("checkpoint_folder", model_dir))
try:
new_saver = tf.train.import_meta_graph(
os.path.join(model_path, "model.ckpt.meta"))
new_saver.restore(sess, os.path.join(model_path, "model.ckpt"))
print("Used saved model for generator pretrain")
except OSError:
print('Start pre-training...')
# pre-training
# Pre-train the generator using MLE for one epoch
progress = tqdm(range(npre_epochs))
for epoch in progress:
g_pretrain_loss_np = generator_obj.pretrain_epoch(
oracle_loader, sess, g_pretrain_op=g_pretrain_op)
gen_pretrain_loss_summary.write_summary(
g_pretrain_loss_np, epoch)
# Test
ntest_pre = 40
if np.mod(epoch, ntest_pre) == 0:
generator_obj.generated_num = 200
json_object = generator_obj.generate_samples(
sess, oracle_loader, dataset="train")
write_json(json_file, json_object)
json_object = generator_obj.generate_samples(
sess, oracle_loader, dataset="validation")
write_json(json_file_validation, json_object)
# take sentences from saved files
sent = generator.get_sentences(json_object)
sent = take_sentences_attribute(json_object)
gen_sentences_summary.write_summary(sent, epoch)
# write summaries
scores = [metric.get_score() for metric in metrics]
metrics_summary_str = sess.run(metric_summary_op,
feed_dict=dict(
zip(metrics_pl,
scores)))
sum_writer.add_summary(metrics_summary_str, epoch)
msg = 'pre_gen_epoch:' + str(
epoch) + ', g_pre_loss: %.4f' % g_pretrain_loss_np
metric_names = [metric.get_name() for metric in metrics]
for (name, score) in zip(metric_names, scores):
msg += ', ' + name + ': %.4f' % score
tqdm.write(msg)
log.write(msg)
log.write('\n')
gc.collect()
print('Start adversarial training...')
progress = tqdm(range(nadv_steps))
for adv_epoch in progress:
gc.collect()
niter = sess.run(global_step)
t0 = time.time()
# Adversarial training
for _ in range(config['gsteps']):
user, product, rating, sentence = oracle_loader.random_batch(
dataset="train")
feed_dict = {
generator_obj.x_user: user,
generator_obj.x_product: product,
generator_obj.x_rating: rating
}
sess.run(g_train_op, feed_dict=feed_dict)
for _ in range(config['dsteps']):
user, product, rating, sentence = oracle_loader.random_batch(
dataset="train")
n = np.zeros((batch_size, seq_len))
for ind, el in enumerate(sentence):
n[ind] = el
feed_dict = {
generator_obj.x_user: user,
generator_obj.x_product: product,
generator_obj.x_rating: rating,
x_real: n
}
sess.run(d_train_op, feed_dict=feed_dict)
t1 = time.time()
sess.run(update_Wall_op, feed_dict={time_diff: t1 - t0})
# temperature
temp_var_np = get_fixed_temperature(temper, niter, nadv_steps,
adapt)
sess.run(update_temperature_op, feed_dict={temp_var: temp_var_np})
user, product, rating, sentence = oracle_loader.random_batch(
dataset="train")
n = np.zeros((batch_size, seq_len))
for ind, el in enumerate(sentence):
n[ind] = el
feed_dict = {
generator_obj.x_user: user,
generator_obj.x_product: product,
generator_obj.x_rating: rating,
x_real: n
}
g_loss_np, d_loss_np, loss_summary_str = sess.run(
[g_loss, d_loss, loss_summary_op], feed_dict=feed_dict)
sum_writer.add_summary(loss_summary_str, niter)
sess.run(global_step_op)
progress.set_description('g_loss: %4.4f, d_loss: %4.4f' %
(g_loss_np, d_loss_np))
# Test
if np.mod(adv_epoch, 300) == 0 or adv_epoch == nadv_steps - 1:
generator_obj.generated_num = generator_obj.batch_size * 10
json_object = generator_obj.generate_samples(sess,
oracle_loader,
dataset="train")
write_json(json_file, json_object)
json_object = generator_obj.generate_samples(
sess, oracle_loader, dataset="validation")
write_json(json_file_validation, json_object)
# take sentences from saved files
sent = take_sentences_attribute(json_object)
gen_sentences_summary.write_summary(sent,
adv_epoch + npre_epochs)
# write summaries
scores = [metric.get_score() for metric in metrics]
metrics_summary_str = sess.run(metric_summary_op,
feed_dict=dict(
zip(metrics_pl, scores)))
sum_writer.add_summary(metrics_summary_str,
adv_epoch + npre_epochs)
# tqdm.write("in {} seconds".format(time.time() - t))
msg = 'pre_gen_epoch:' + str(
adv_epoch) + ', g_pre_loss: %.4f' % g_pretrain_loss_np
metric_names = [metric.get_name() for metric in metrics]
for (name, score) in zip(metric_names, scores):
msg += ', ' + name + ': %.4f' % score
tqdm.write(msg)
log.write(msg)
log.write('\n')
gc.collect()
sum_writer.close()
save_model = True
if save_model:
model_dir = datetime.datetime.now().strftime("%Y%m%d%H%M%S")
model_path = os.path.join(resources_path("trained_models"),
model_dir)
simple_save(sess,
model_path,
inputs={
"x_user": x_user,
"x_rating": x_rating,
"x_product": x_product
},
outputs={"gen_x": generator_obj.gen_x})
# save_path = saver.save(sess, os.path.join(model_path, "model.ckpt"))
print("Model saved in path: %s" % model_path)
def customer_reviews_train(generator: ReviewGenerator,
discriminator_positive: ReviewDiscriminator,
discriminator_negative: ReviewDiscriminator,
oracle_loader: RealDataCustomerReviewsLoader,
config, args):
batch_size = config['batch_size']
num_sentences = config['num_sentences']
vocab_size = config['vocabulary_size']
seq_len = config['seq_len']
dataset = config['dataset']
npre_epochs = config['npre_epochs']
nadv_steps = config['nadv_steps']
temper = config['temperature']
adapt = config['adapt']
# changed to match resources path
data_dir = resources_path(config['data_dir'], "Amazon_Attribute")
log_dir = resources_path(config['log_dir'])
sample_dir = resources_path(config['sample_dir'])
# filename
json_file = os.path.join(sample_dir, 'json_file.txt')
csv_file = os.path.join(log_dir, 'experiment-log-rmcgan.csv')
# create necessary directories
if not os.path.exists(data_dir):
os.makedirs(data_dir)
if not os.path.exists(sample_dir):
os.makedirs(sample_dir)
if not os.path.exists(log_dir):
os.makedirs(log_dir)
# placeholder definitions
x_real = tf.placeholder(tf.int32, [batch_size, seq_len], name="x_real")
x_pos = tf.placeholder(tf.int32, [batch_size, seq_len], name="x_pos")
x_neg = tf.placeholder(tf.int32, [batch_size, seq_len], name="x_neg")
x_sentiment = tf.placeholder(tf.int32, [batch_size], name="x_sentiment")
temperature = tf.Variable(1., trainable=False, name='temperature')
x_real_pos_onehot = tf.one_hot(
x_pos, vocab_size) # batch_size x seq_len x vocab_size
x_real_neg_onehot = tf.one_hot(
x_neg, vocab_size) # batch_size x seq_len x vocab_size
assert x_real_pos_onehot.get_shape().as_list() == [
batch_size, seq_len, vocab_size
]
# generator and discriminator outputs
generator_obj = generator(x_real=x_real,
temperature=temperature,
x_sentiment=x_sentiment)
# discriminator for positive sentences
discriminator_positive_real_pos = discriminator_positive(
x_onehot=x_real_pos_onehot)
discriminator_positive_real_neg = discriminator_positive(
x_onehot=x_real_neg_onehot)
discriminator_positive_fake = discriminator_positive(
x_onehot=generator_obj.gen_x_onehot_adv)
# discriminator for negative sentences
discriminator_negative_real_pos = discriminator_negative(
x_onehot=x_real_pos_onehot)
discriminator_negative_real_neg = discriminator_negative(
x_onehot=x_real_neg_onehot)
discriminator_negative_fake = discriminator_negative(
x_onehot=generator_obj.gen_x_onehot_adv)
# GAN / Divergence type
log_pg, g_loss, d_loss = get_losses(
generator_obj, discriminator_positive_real_pos,
discriminator_positive_real_neg, discriminator_positive_fake,
discriminator_negative_real_pos, discriminator_negative_real_neg,
discriminator_negative_fake)
# Global step
global_step = tf.Variable(0, trainable=False)
global_step_op = global_step.assign_add(1)
# Train ops
g_pretrain_op, g_train_op, d_train_op, d_topic_pretrain_op = get_train_ops(
config, generator_obj.pretrain_loss, g_loss, d_loss, None, log_pg,
temperature, global_step)
# Record wall clock time
time_diff = tf.placeholder(tf.float32)
Wall_clock_time = tf.Variable(0., trainable=False)
update_Wall_op = Wall_clock_time.assign_add(time_diff)
# Temperature placeholder
temp_var = tf.placeholder(tf.float32)
update_temperature_op = temperature.assign(temp_var)
# Loss summaries
loss_summaries = [
tf.summary.scalar('adv_loss/discriminator/total', d_loss),
tf.summary.scalar('adv_loss/generator/total_g_loss', g_loss),
tf.summary.scalar('adv_loss/log_pg', log_pg),
tf.summary.scalar('adv_loss/Wall_clock_time', Wall_clock_time),
tf.summary.scalar('adv_loss/temperature', temperature),
]
loss_summary_op = tf.summary.merge(loss_summaries)
# Metric Summaries
metrics_pl, metric_summary_op = get_metric_summary_op(config)
# Summaries
gen_pretrain_loss_summary = CustomSummary(name='pretrain_loss',
scope='generator')
gen_sentences_summary = CustomSummary(name='generated_sentences',
scope='generator',
summary_type=tf.summary.text,
item_type=tf.string)
run_information = CustomSummary(name='run_information',
scope='info',
summary_type=tf.summary.text,
item_type=tf.string)
custom_summaries = [
gen_pretrain_loss_summary, gen_sentences_summary, run_information
]
# To save the trained model
# ------------- initial the graph --------------
with init_sess() as sess:
# count parameters
log = open(csv_file, 'w')
summary_dir = os.path.join(log_dir, 'summary', str(time.time()))
if not os.path.exists(summary_dir):
os.makedirs(summary_dir)
sum_writer = tf.summary.FileWriter(summary_dir, sess.graph)
for custom_summary in custom_summaries:
custom_summary.set_file_writer(sum_writer, sess)
run_information.write_summary(str(args), 0)
print("Information stored in the summary!")
def get_metrics():
# set up evaluation metric
metrics = []
if config['nll_gen']:
nll_gen = NllReview(oracle_loader,
generator_obj,
sess,
name='nll_gen_review')
metrics.append(nll_gen)
if config['KL']:
KL_div = KL_divergence(oracle_loader,
json_file,
name='KL_divergence')
metrics.append(KL_div)
if config['jaccard_similarity']:
Jaccard_Sim = JaccardSimilarity(oracle_loader,
json_file,
name='jaccard_similarity')
metrics.append(Jaccard_Sim)
if config['jaccard_diversity']:
Jaccard_Sim = JaccardDiversity(oracle_loader,
json_file,
name='jaccard_diversity')
metrics.append(Jaccard_Sim)
return metrics
metrics = get_metrics()
generator_obj.generated_num = 200 #num_sentences
gc.collect()
# Check if there is a pretrained generator saved
model_dir = "PretrainGenerator"
model_path = resources_path(
os.path.join("checkpoint_folder", model_dir))
try:
new_saver = tf.train.import_meta_graph(
os.path.join(model_path, "model.ckpt.meta"))
new_saver.restore(sess, os.path.join(model_path, "model.ckpt"))
print("Used saved model for generator pretrain")
except OSError:
print('Start pre-training...')
# pre-training
# Pre-train the generator using MLE for one epoch
progress = tqdm(range(npre_epochs))
for epoch in progress:
oracle_loader.reset_pointer()
g_pretrain_loss_np = generator_obj.pretrain_epoch(
oracle_loader, sess, g_pretrain_op=g_pretrain_op)
gen_pretrain_loss_summary.write_summary(
g_pretrain_loss_np, epoch)
msg = 'pre_gen_epoch:' + str(
epoch) + ', g_pre_loss: %.4f' % g_pretrain_loss_np
progress.set_description(msg)
# Test
ntest_pre = 30
if np.mod(epoch, ntest_pre) == 0 or epoch == npre_epochs - 1:
json_object = generator_obj.generate_json(
oracle_loader, sess)
write_json(json_file, json_object)
# take sentences from saved files
sent = take_sentences_json(json_object)
gen_sentences_summary.write_summary(sent, epoch)
# write summaries
scores = [metric.get_score() for metric in metrics]
metrics_summary_str = sess.run(metric_summary_op,
feed_dict=dict(
zip(metrics_pl,
scores)))
sum_writer.add_summary(metrics_summary_str, epoch)
msg = 'pre_gen_epoch:' + str(
epoch) + ', g_pre_loss: %.4f' % g_pretrain_loss_np
metric_names = [metric.get_name() for metric in metrics]
for (name, score) in zip(metric_names, scores):
msg += ', ' + name + ': %.4f' % score
tqdm.write(msg)
log.write(msg)
log.write('\n')
gc.collect()
gc.collect()
print('Start adversarial training...')
progress = tqdm(range(nadv_steps))
for adv_epoch in progress:
gc.collect()
niter = sess.run(global_step)
t0 = time.time()
# Adversarial training
for _ in range(config['gsteps']):
sentiment, sentence = oracle_loader.random_batch()
n = np.zeros((generator_obj.batch_size, generator_obj.seq_len))
for ind, el in enumerate(sentence):
n[ind] = el
sess.run(g_pretrain_op,
feed_dict={
generator_obj.x_real: n,
generator_obj.x_sentiment: sentiment
})
for _ in range(config['dsteps']):
sentiment, sentence, pos, neg = oracle_loader.get_positive_negative_batch(
)
n1 = np.zeros(
(generator_obj.batch_size, generator_obj.seq_len))
n2 = np.zeros(
(generator_obj.batch_size, generator_obj.seq_len))
n3 = np.zeros(
(generator_obj.batch_size, generator_obj.seq_len))
for ind, (s, p, n) in enumerate(zip(sentence, pos, neg)):
n1[ind] = s
n2[ind] = p[0]
n3[ind] = n[0]
feed_dict = {
x_real: n1,
x_pos: n2,
x_neg: n3,
x_sentiment: sentiment
}
sess.run(d_train_op, feed_dict=feed_dict)
t1 = time.time()
sess.run(update_Wall_op, feed_dict={time_diff: t1 - t0})
# temperature
temp_var_np = get_fixed_temperature(temper, niter, nadv_steps,
adapt)
sess.run(update_temperature_op, feed_dict={temp_var: temp_var_np})
sentiment, sentence, pos, neg = oracle_loader.get_positive_negative_batch(
)
n1 = np.zeros((generator_obj.batch_size, generator_obj.seq_len))
n2 = np.zeros((generator_obj.batch_size, generator_obj.seq_len))
n3 = np.zeros((generator_obj.batch_size, generator_obj.seq_len))
for ind, (s, p, n) in enumerate(zip(sentence, pos, neg)):
n1[ind] = s
n2[ind] = p[0]
n3[ind] = n[0]
feed_dict = {
x_real: n1,
x_pos: n2,
x_neg: n3,
x_sentiment: sentiment
}
g_loss_np, d_loss_np, loss_summary_str = sess.run(
[g_loss, d_loss, loss_summary_op], feed_dict=feed_dict)
sum_writer.add_summary(loss_summary_str, niter)
sess.run(global_step_op)
progress.set_description('g_loss: %4.4f, d_loss: %4.4f' %
(g_loss_np, d_loss_np))
# Test
# print("N_iter: {}, test every {} epochs".format(niter, config['ntest']))
if np.mod(adv_epoch, 100) == 0 or adv_epoch == nadv_steps - 1:
json_object = generator_obj.generate_json(oracle_loader, sess)
write_json(json_file, json_object)
# take sentences from saved files
sent = take_sentences_json(json_object)
gen_sentences_summary.write_summary(
sent, niter + config['npre_epochs'])
# write summaries
scores = [metric.get_score() for metric in metrics]
metrics_summary_str = sess.run(metric_summary_op,
feed_dict=dict(
zip(metrics_pl, scores)))
sum_writer.add_summary(metrics_summary_str,
niter + config['npre_epochs'])
msg = 'adv_step: ' + str(niter)
metric_names = [metric.get_name() for metric in metrics]
for (name, score) in zip(metric_names, scores):
msg += ', ' + name + ': %.4f' % score
tqdm.write(msg)
log.write(msg)
log.write('\n')
gc.collect()
sum_writer.close()
save_model = False
if save_model:
model_dir = datetime.datetime.now().strftime("%Y%m%d%H%M%S")
model_path = os.path.join(resources_path("trained_models"),
model_dir)
simple_save(sess,
model_path,
inputs={"x_topic": x_topic},
outputs={"gen_x": x_fake})
# save_path = saver.save(sess, os.path.join(model_path, "model.ckpt"))
print("Model saved in path: %s" % model_path)
def real_topic_train(generator_obj, discriminator_obj, topic_discriminator_obj,
oracle_loader: RealDataTopicLoader, config, args):
batch_size = config['batch_size']
num_sentences = config['num_sentences']
vocab_size = config['vocab_size']
seq_len = config['seq_len']
dataset = config['dataset']
npre_epochs = config['npre_epochs']
n_topic_pre_epochs = config['n_topic_pre_epochs']
nadv_steps = config['nadv_steps']
temper = config['temperature']
adapt = config['adapt']
# changed to match resources path
data_dir = resources_path(config['data_dir'])
log_dir = resources_path(config['log_dir'])
sample_dir = resources_path(config['sample_dir'])
# filename
oracle_file = os.path.join(sample_dir, 'oracle_{}.txt'.format(dataset))
gen_file = os.path.join(sample_dir, 'generator.txt')
gen_text_file = os.path.join(sample_dir, 'generator_text.txt')
json_file = os.path.join(sample_dir, 'json_file.txt')
csv_file = os.path.join(log_dir, 'experiment-log-rmcgan.csv')
data_file = os.path.join(data_dir, '{}.txt'.format(dataset))
if dataset == 'image_coco':
test_file = os.path.join(data_dir, 'testdata/test_coco.txt')
elif dataset == 'emnlp_news':
test_file = os.path.join(data_dir, 'testdata/test_emnlp.txt')
else:
raise NotImplementedError('Unknown dataset!')
# create necessary directories
if not os.path.exists(data_dir):
os.makedirs(data_dir)
if not os.path.exists(sample_dir):
os.makedirs(sample_dir)
if not os.path.exists(log_dir):
os.makedirs(log_dir)
# placeholder definitions
x_real = tf.placeholder(tf.int32, [batch_size, seq_len],
name="x_real") # tokens of oracle sequences
x_topic = tf.placeholder(tf.float32,
[batch_size, oracle_loader.vocab_size + 1],
name="x_topic") # todo stessa cosa del +1
x_topic_random = tf.placeholder(tf.float32,
[batch_size, oracle_loader.vocab_size + 1],
name="x_topic_random")
temperature = tf.Variable(1., trainable=False, name='temperature')
x_real_onehot = tf.one_hot(x_real,
vocab_size) # batch_size x seq_len x vocab_size
assert x_real_onehot.get_shape().as_list() == [
batch_size, seq_len, vocab_size
]
# generator and discriminator outputs
generator = generator_obj(x_real=x_real,
temperature=temperature,
x_topic=x_topic)
d_real = discriminator_obj(x_onehot=x_real_onehot)
d_fake = discriminator_obj(x_onehot=generator.gen_x_onehot_adv)
if not args.no_topic:
d_topic_real_pos = topic_discriminator_obj(x_onehot=x_real_onehot,
x_topic=x_topic)
d_topic_real_neg = topic_discriminator_obj(x_onehot=x_real_onehot,
x_topic=x_topic_random)
d_topic_fake = topic_discriminator_obj(
x_onehot=generator.gen_x_onehot_adv, x_topic=x_topic)
else:
d_topic_real_pos = None
d_topic_real_neg = None
d_topic_fake = None
# GAN / Divergence type
losses = get_losses(generator, d_real, d_fake, d_topic_real_pos,
d_topic_real_neg, d_topic_fake, config)
if not args.no_topic:
d_topic_loss = losses['d_topic_loss_real_pos'] + losses[
'd_topic_loss_real_neg'] # only from real data for pretrain
d_topic_accuracy = get_accuracy(d_topic_real_pos.logits,
d_topic_real_neg.logits)
else:
d_topic_loss = None
d_topic_accuracy = None
# Global step
global_step = tf.Variable(0, trainable=False)
global_step_op = global_step.assign_add(1)
# Train ops
g_pretrain_op, g_train_op, d_train_op, d_topic_pretrain_op = get_train_ops(
config, generator.pretrain_loss, losses['g_loss'], losses['d_loss'],
d_topic_loss, losses['log_pg'], temperature, global_step)
generator.g_pretrain_op = g_pretrain_op
# Record wall clock time
time_diff = tf.placeholder(tf.float32)
Wall_clock_time = tf.Variable(0., trainable=False)
update_Wall_op = Wall_clock_time.assign_add(time_diff)
# Temperature placeholder
temp_var = tf.placeholder(tf.float32)
update_temperature_op = temperature.assign(temp_var)
# Loss summaries
loss_summaries = [
tf.summary.scalar('adv_loss/discriminator/classic/d_loss_real',
losses['d_loss_real']),
tf.summary.scalar('adv_loss/discriminator/classic/d_loss_fake',
losses['d_loss_fake']),
tf.summary.scalar('adv_loss/discriminator/total', losses['d_loss']),
tf.summary.scalar('adv_loss/generator/g_sentence_loss',
losses['g_sentence_loss']),
tf.summary.scalar('adv_loss/generator/total_g_loss', losses['g_loss']),
tf.summary.scalar('adv_loss/log_pg', losses['log_pg']),
tf.summary.scalar('adv_loss/Wall_clock_time', Wall_clock_time),
tf.summary.scalar('adv_loss/temperature', temperature),
]
if not args.no_topic:
loss_summaries += [
tf.summary.scalar(
'adv_loss/discriminator/topic_discriminator/d_topic_loss_real_pos',
losses['d_topic_loss_real_pos']),
tf.summary.scalar(
'adv_loss/discriminator/topic_discriminator/d_topic_loss_fake',
losses['d_topic_loss_fake']),
tf.summary.scalar('adv_loss/generator/g_topic_loss',
losses['g_topic_loss'])
]
loss_summary_op = tf.summary.merge(loss_summaries)
# Metric Summaries
metrics_pl, metric_summary_op = get_metric_summary_op(config)
# Summaries
gen_pretrain_loss_summary = CustomSummary(name='pretrain_loss',
scope='generator')
gen_sentences_summary = CustomSummary(name='generated_sentences',
scope='generator',
summary_type=tf.summary.text,
item_type=tf.string)
topic_discr_pretrain_summary = CustomSummary(name='pretrain_loss',
scope='topic_discriminator')
topic_discr_accuracy_summary = CustomSummary(name='pretrain_accuracy',
scope='topic_discriminator')
run_information = CustomSummary(name='run_information',
scope='info',
summary_type=tf.summary.text,
item_type=tf.string)
custom_summaries = [
gen_pretrain_loss_summary, gen_sentences_summary,
topic_discr_pretrain_summary, topic_discr_accuracy_summary,
run_information
]
# To save the trained model
saver = tf.compat.v1.train.Saver()
# ------------- initial the graph --------------
with init_sess() as sess:
variables_dict = get_parameters_division()
print("Total paramter number: {}".format(
np.sum([
np.prod(v.get_shape().as_list())
for v in tf.trainable_variables()
])))
log = open(csv_file, 'w')
now = datetime.datetime.now()
additional_text = now.strftime(
"%Y-%m-%d_%H-%M") + "_" + args.summary_name
summary_dir = os.path.join(log_dir, 'summary', additional_text)
if not os.path.exists(summary_dir):
os.makedirs(summary_dir)
sum_writer = tf.compat.v1.summary.FileWriter(os.path.join(summary_dir),
sess.graph)
for custom_summary in custom_summaries:
custom_summary.set_file_writer(sum_writer, sess)
run_information.write_summary(str(args), 0)
print("Information stored in the summary!")
# generate oracle data and create batches
oracle_loader.create_batches(oracle_file)
metrics = get_metrics(config, oracle_loader, test_file, gen_text_file,
generator.pretrain_loss, x_real, x_topic, sess,
json_file)
gc.collect()
# Check if there is a pretrained generator saved
model_dir = "PretrainGenerator"
model_path = resources_path(
os.path.join("checkpoint_folder", model_dir))
try:
new_saver = tf.train.import_meta_graph(
os.path.join(model_path, "model.ckpt.meta"))
new_saver.restore(sess, os.path.join(model_path, "model.ckpt"))
print("Used saved model for generator pretrain")
except OSError:
print('Start pre-training...')
progress = tqdm(range(npre_epochs), dynamic_ncols=True)
for epoch in progress:
# pre-training
g_pretrain_loss_np = generator.pretrain_epoch(
sess, oracle_loader)
gen_pretrain_loss_summary.write_summary(
g_pretrain_loss_np, epoch)
# Test
ntest_pre = 30
if np.mod(epoch, ntest_pre) == 0:
json_object = generator.generate_samples_topic(
sess, oracle_loader, num_sentences)
write_json(json_file, json_object)
# take sentences from saved files
sent = generator.get_sentences(json_object)
gen_sentences_summary.write_summary(sent, epoch)
# write summaries
t = time.time()
scores = [metric.get_score() for metric in metrics]
metrics_summary_str = sess.run(metric_summary_op,
feed_dict=dict(
zip(metrics_pl,
scores)))
sum_writer.add_summary(metrics_summary_str, epoch)
msg = 'pre_gen_epoch:' + str(
epoch) + ', g_pre_loss: %.4f' % g_pretrain_loss_np
metric_names = [metric.get_name() for metric in metrics]
for (name, score) in zip(metric_names, scores):
score = score * 1e5 if 'Earth' in name else score
msg += ', ' + name + ': %.4f' % score
progress.set_description(
msg + " in {:.2f} sec".format(time.time() - t))
log.write(msg)
log.write('\n')
gc.collect()
if not args.no_topic:
gc.collect()
print('Start Topic Discriminator pre-training...')
progress = tqdm(range(n_topic_pre_epochs))
for epoch in progress:
# pre-training and write loss
# Pre-train the generator using MLE for one epoch
supervised_g_losses = []
supervised_accuracy = []
oracle_loader.reset_pointer()
for it in range(oracle_loader.num_batch):
text_batch, topic_batch = oracle_loader.next_batch(
only_text=False)
_, topic_loss, accuracy = sess.run(
[d_topic_pretrain_op, d_topic_loss, d_topic_accuracy],
feed_dict={
x_real: text_batch,
x_topic: topic_batch,
x_topic_random: oracle_loader.random_topic()
})
supervised_g_losses.append(topic_loss)
supervised_accuracy.append(accuracy)
d_topic_pretrain_loss = np.mean(supervised_g_losses)
accuracy_mean = np.mean(supervised_accuracy)
topic_discr_pretrain_summary.write_summary(
d_topic_pretrain_loss, epoch)
topic_discr_accuracy_summary.write_summary(
accuracy_mean, epoch)
progress.set_description(
'topic_loss: %4.4f, accuracy: %4.4f' %
(d_topic_pretrain_loss, accuracy_mean))
print('Start adversarial training...')
progress = tqdm(range(nadv_steps))
for adv_epoch in progress:
gc.collect()
niter = sess.run(global_step)
t0 = time.time()
# Adversarial training
for _ in range(config['gsteps']):
text_batch, topic_batch = oracle_loader.random_batch(
only_text=False)
sess.run(g_train_op,
feed_dict={
x_real: text_batch,
x_topic: topic_batch
})
for _ in range(config['dsteps']):
# normal + topic discriminator together
text_batch, topic_batch = oracle_loader.random_batch(
only_text=False)
sess.run(d_train_op,
feed_dict={
x_real: text_batch,
x_topic: topic_batch,
x_topic_random: oracle_loader.random_topic()
})
t1 = time.time()
sess.run(update_Wall_op, feed_dict={time_diff: t1 - t0})
# temperature
temp_var_np = get_fixed_temperature(temper, niter, nadv_steps,
adapt)
sess.run(update_temperature_op, feed_dict={temp_var: temp_var_np})
text_batch, topic_batch = oracle_loader.random_batch(
only_text=False)
feed = {
x_real: text_batch,
x_topic: topic_batch,
x_topic_random: oracle_loader.random_topic()
}
g_loss_np, d_loss_np, loss_summary_str = sess.run(
[losses['g_loss'], losses['d_loss'], loss_summary_op],
feed_dict=feed)
sum_writer.add_summary(loss_summary_str, niter)
sess.run(global_step_op)
progress.set_description('g_loss: %4.4f, d_loss: %4.4f' %
(g_loss_np, d_loss_np))
# Test
if np.mod(adv_epoch, 400) == 0 or adv_epoch == nadv_steps - 1:
json_object = generator.generate_samples_topic(
sess, oracle_loader, num_sentences)
write_json(json_file, json_object)
# take sentences from saved files
sent = generator.get_sentences(json_object)
gen_sentences_summary.write_summary(sent, adv_epoch)
# write summaries
scores = [metric.get_score() for metric in metrics]
metrics_summary_str = sess.run(metric_summary_op,
feed_dict=dict(
zip(metrics_pl, scores)))
sum_writer.add_summary(metrics_summary_str,
niter + config['npre_epochs'])
msg = 'adv_step: ' + str(niter)
metric_names = [metric.get_name() for metric in metrics]
for (name, score) in zip(metric_names, scores):
msg += ', ' + name + ': %.4f' % score
tqdm.write(msg)
log.write(msg)
log.write('\n')
sum_writer.close()
save_model = False
if save_model:
model_dir = datetime.datetime.now().strftime("%Y%m%d%H%M%S")
model_path = os.path.join(resources_path("trained_models"),
model_dir)
simple_save(sess,
model_path,
inputs={"x_topic": x_topic},
outputs={"gen_x": generator.gen_x})
# save_path = saver.save(sess, os.path.join(model_path, "model.ckpt"))
print("Model saved in path: %s" % model_path)
def real_topic_train_NoDiscr(generator, oracle_loader: RealDataTopicLoader,
config, args):
batch_size = config['batch_size']
num_sentences = config['num_sentences']
vocab_size = config['vocab_size']
seq_len = config['seq_len']
dataset = config['dataset']
npre_epochs = config['npre_epochs']
nadv_steps = config['nadv_steps']
temper = config['temperature']
adapt = config['adapt']
# changed to match resources path
data_dir = resources_path(config['data_dir'])
log_dir = resources_path(config['log_dir'])
sample_dir = resources_path(config['sample_dir'])
# filename
oracle_file = os.path.join(sample_dir, 'oracle_{}.txt'.format(dataset))
gen_file = os.path.join(sample_dir, 'generator.txt')
gen_text_file = os.path.join(sample_dir, 'generator_text.txt')
gen_text_file_print = os.path.join(sample_dir, 'gen_text_file_print.txt')
json_file = os.path.join(sample_dir,
'json_file{}.txt'.format(args.json_file))
csv_file = os.path.join(log_dir, 'experiment-log-rmcgan.csv')
data_file = os.path.join(data_dir, '{}.txt'.format(dataset))
if dataset == 'image_coco':
test_file = os.path.join(data_dir, 'testdata/test_coco.txt')
elif dataset == 'emnlp_news':
test_file = os.path.join(data_dir, 'testdata/test_emnlp.txt')
else:
raise NotImplementedError('Unknown dataset!')
# create necessary directories
if not os.path.exists(data_dir):
os.makedirs(data_dir)
if not os.path.exists(sample_dir):
os.makedirs(sample_dir)
if not os.path.exists(log_dir):
os.makedirs(log_dir)
# placeholder definitions
x_real = placeholder(tf.int32, [batch_size, seq_len],
name="x_real") # tokens of oracle sequences
x_topic = placeholder(tf.float32,
[batch_size, oracle_loader.vocab_size + 1],
name="x_topic") # todo stessa cosa del +1
x_topic_random = placeholder(tf.float32,
[batch_size, oracle_loader.vocab_size + 1],
name="x_topic_random")
temperature = tf.Variable(1., trainable=False, name='temperature')
x_real_onehot = tf.one_hot(x_real,
vocab_size) # batch_size x seq_len x vocab_size
assert x_real_onehot.get_shape().as_list() == [
batch_size, seq_len, vocab_size
]
# generator and discriminator outputs
x_fake_onehot_appr, x_fake, g_pretrain_loss, gen_o, \
lambda_values_returned, gen_x_no_lambda = generator(x_real=x_real,
temperature=temperature,
x_topic=x_topic)
# A function to calculate the gradients and get training operations
def get_train_ops(config, g_pretrain_loss):
gpre_lr = config['gpre_lr']
g_vars = tf.get_collection(tf.GraphKeys.TRAINABLE_VARIABLES,
scope='generator')
grad_clip = 5.0 # keep the same with the previous setting
# generator pre-training
pretrain_opt = tf.train.AdamOptimizer(gpre_lr,
beta1=0.9,
beta2=0.999,
name="gen_pretrain_adam")
pretrain_grad, _ = tf.clip_by_global_norm(
tf.gradients(g_pretrain_loss, g_vars, name="gradients_g_pretrain"),
grad_clip,
name="g_pretrain_clipping") # gradient clipping
g_pretrain_op = pretrain_opt.apply_gradients(zip(
pretrain_grad, g_vars))
return g_pretrain_op
# Train ops
g_pretrain_op = get_train_ops(config, g_pretrain_loss)
# Metric Summaries
metrics_pl, metric_summary_op = get_metric_summary_op(config)
# Summaries
gen_pretrain_loss_summary = CustomSummary(name='pretrain_loss',
scope='generator')
gen_sentences_summary = CustomSummary(name='generated_sentences',
scope='generator',
summary_type=tf.summary.text,
item_type=tf.string)
run_information = CustomSummary(name='run_information',
scope='info',
summary_type=tf.summary.text,
item_type=tf.string)
custom_summaries = [
gen_pretrain_loss_summary, gen_sentences_summary, run_information
]
# ------------- initial the graph --------------
with init_sess() as sess:
variables_dict = get_parameters_division()
log = open(csv_file, 'w')
sum_writer = tf.summary.FileWriter(os.path.join(log_dir, 'summary'),
sess.graph)
for custom_summary in custom_summaries:
custom_summary.set_file_writer(sum_writer, sess)
run_information.write_summary(str(args), 0)
print("Information stored in the summary!")
oracle_loader.create_batches(oracle_file)
metrics = get_metrics(config, oracle_loader, test_file, gen_text_file,
g_pretrain_loss, x_real, x_topic, sess,
json_file)
gc.collect()
# Check if there is a pretrained generator saved
model_dir = "PretrainGenerator"
model_path = resources_path(
os.path.join("checkpoint_folder", model_dir))
try:
new_saver = tf.train.import_meta_graph(
os.path.join(model_path, "model.ckpt.meta"))
new_saver.restore(sess, os.path.join(model_path, "model.ckpt"))
print("Used saved model for generator pretrain")
except OSError:
print('Start pre-training...')
progress = tqdm(range(npre_epochs))
for epoch in progress:
# pre-training
g_pretrain_loss_np = pre_train_epoch(sess, g_pretrain_op,
g_pretrain_loss, x_real,
oracle_loader, x_topic)
gen_pretrain_loss_summary.write_summary(g_pretrain_loss_np, epoch)
progress.set_description(
"Pretrain_loss: {}".format(g_pretrain_loss_np))
# Test
ntest_pre = 40
if np.mod(epoch, ntest_pre) == 0:
json_object = generate_sentences(sess,
x_fake,
batch_size,
num_sentences,
oracle_loader=oracle_loader,
x_topic=x_topic)
write_json(json_file, json_object)
with open(gen_text_file, 'w') as outfile:
i = 0
for sent in json_object['sentences']:
if i < 200:
outfile.write(sent['generated_sentence'] + "\n")
else:
break
# take sentences from saved files
sent = take_sentences_json(json_object,
first_elem='generated_sentence',
second_elem='real_starting')
gen_sentences_summary.write_summary(sent, epoch)
# write summaries
scores = [metric.get_score() for metric in metrics]
metrics_summary_str = sess.run(metric_summary_op,
feed_dict=dict(
zip(metrics_pl, scores)))
sum_writer.add_summary(metrics_summary_str, epoch)
msg = 'pre_gen_epoch:' + str(
epoch) + ', g_pre_loss: %.4f' % g_pretrain_loss_np
metric_names = [metric.get_name() for metric in metrics]
for (name, score) in zip(metric_names, scores):
msg += ', ' + name + ': %.4f' % score
tqdm.write(msg)
log.write(msg)
log.write('\n')
gc.collect()
gc.collect()
sum_writer.close()
save_model = True
if save_model:
model_dir = datetime.datetime.now().strftime("%Y%m%d%H%M%S")
model_path = os.path.join(resources_path("trained_models"),
model_dir)
simple_save(sess,
model_path,
inputs={"x_topic": x_topic},
outputs={"gen_x": x_fake})
# save_path = saver.save(sess, os.path.join(model_path, "model.ckpt"))
print("Model saved in path: %s" % model_path)
def test_task(
model_file: InputBinaryFile(str),
examples_file: InputBinaryFile(str),
confusion_matrix: OutputTextFile(str),
results: OutputTextFile(str),
):
"""Connects to served model and tests example MNIST images."""
import time
import json
import numpy as np
import requests
from tensorflow.python.keras.backend import get_session
from tensorflow.python.keras.saving import load_model
from tensorflow.python.saved_model.simple_save import simple_save
with get_session() as sess:
model = load_model(model_file)
simple_save(
sess,
'/output/mnist/1/',
inputs={'input_image': model.input},
outputs={t.name: t
for t in model.outputs},
)
model_url = 'http://localhost:9001/v1/models/mnist'
for _ in range(60):
try:
requests.get(f'{model_url}/versions/1').raise_for_status()
break
except requests.RequestException:
time.sleep(5)
else:
raise Exception("Waited too long for sidecar to come up!")
response = requests.get(f'{model_url}/metadata')
response.raise_for_status()
assert response.json() == {
'model_spec': {
'name': 'mnist',
'signature_name': '',
'version': '1'
},
'metadata': {
'signature_def': {
'signature_def': {
'serving_default': {
'inputs': {
'input_image': {
'dtype': 'DT_FLOAT',
'tensor_shape': {
'dim': [
{
'size': '-1',
'name': ''
},
{
'size': '28',
'name': ''
},
{
'size': '28',
'name': ''
},
{
'size': '1',
'name': ''
},
],
'unknown_rank':
False,
},
'name': 'conv2d_input:0',
}
},
'outputs': {
'dense_1/Softmax:0': {
'dtype': 'DT_FLOAT',
'tensor_shape': {
'dim': [{
'size': '-1',
'name': ''
}, {
'size': '10',
'name': ''
}],
'unknown_rank':
False,
},
'name': 'dense_1/Softmax:0',
}
},
'method_name': 'tensorflow/serving/predict',
}
}
}
},
}
examples = np.load(examples_file)
assert examples['val_x'].shape == (100, 28, 28, 1)
assert examples['val_y'].shape == (100, 10)
response = requests.post(f'{model_url}:predict',
json={'instances': examples['val_x'].tolist()})
response.raise_for_status()
predicted = np.argmax(response.json()['predictions'], axis=1).tolist()
actual = np.argmax(examples['val_y'], axis=1).tolist()
zipped = list(zip(predicted, actual))
accuracy = sum(1 for (p, a) in zipped if p == a) / len(predicted)
print(f"Accuracy: {accuracy:0.2f}")
y: batch_y,
seqlen: batch_seqlen
})
if step % display_step == 0 or step == 1:
# Calculate batch accuracy & loss
acc, loss = sess.run([accuracy, cost],
feed_dict={
x: batch_x,
y: batch_y,
seqlen: batch_seqlen
})
print("Step " + str(step*batch_size) + ", Minibatch Loss= " + \
"{:.6f}".format(loss) + ", Training Accuracy= " + \
"{:.5f}".format(acc))
print("Optimization Finished!")
# Calculate accuracy
test_data = testset.data
test_label = testset.labels
test_seqlen = testset.seqlen
print("Testing Accuracy:", \
sess.run(accuracy, feed_dict={x: test_data, y: test_label,
seqlen: test_seqlen}))
print("Saving the model")
simple_save(sess,
export_dir='./saved_dynamic_rnn',
inputs={"inp": x},
outputs={"out": prediction})
[train_op, cost, weighted_action_loss],
feed_dict={
food: x_train[i:i + step],
individual_values: individual_values_train[i:i + step],
reward: reward_train[i:i + step],
actions_performed: actions_train[i:i + step],
actions_target: actions_target_train,
next_pred_reward: next_y_pred_v,
keep_prob: 0.99
})
print("cost_train: " + str(cost_train) + " reward_loss_v: " +
str(reward_loss_v) + " weighted_action_loss_v: " +
str(weighted_action_loss_v))
shutil.rmtree('model', ignore_errors=True)
simple_save(sess,
"model",
inputs={"input": food},
outputs={
"action_pred": action_pred,
"reward_pred": reward_pred
})
save_path = saver.save(sess, "model_tmp/model.ckpt")
r.flushall()
#action_pred_v = sess.run(
# [action_pred],
# feed_dict={food: x_train, individual_values: individual_values_train, keep_prob: 1.0})
#print(action_pred_v)