def main(args):
tf.disable_eager_execution()
if os.path.exists(args.output_path):
shutil.rmtree(args.output_path)
os.makedirs(args.output_path)
model_dic, sess = load_model(args)
dirs = os.listdir(args.input_path)
for folder in dirs:
path = os.path.join(args.input_path, folder)
if os.path.isdir(path):
output_path = os.path.join(args.output_path, folder)
if not os.path.exists(output_path):
os.makedirs(output_path)
print('========== input %s ============' % (folder))
positive_img_names = tf.gfile.Glob(os.path.join(
path, '*_0001.jpg')) # positive image
print('Positive: %s' % (os.path.basename(positive_img_names[0])))
#get positive image
img_positive = io.imread(positive_img_names[0]) / 255.
img_positive = rescale(img_positive,
112. / 600.,
order=5,
multichannel=True)
img_positive = prep(img_positive) #[1,3,112,112]
fdict = {
model_dic['image_input']: img_positive,
model_dic['keep_prob']: 1.0,
model_dic['is_train']: False
}
pos_embedding_np = sess.run(model_dic['embedding'],
feed_dict=fdict) # [2,512]
name_list = []
img_list = []
img_names = tf.gfile.Glob(os.path.join(
path, '*.jpg')) #tf.gfile.Glob得到的是一个list
for file in img_names:
if file != positive_img_names[0]:
print('%s' % (os.path.basename(file)))
img = prep(
rescale(io.imread(file) / 255.,
112. / 600.,
order=5,
multichannel=True)) #[1,3,112,112]
name_list.append(os.path.basename(file)[:-4])
img_list.append(img)
feed_img = np.squeeze(np.array(img_list), axis=1) #[?,3,112,112]
fdict1 = {
model_dic['image_input']: feed_img,
model_dic['keep_prob']: 1.0,
model_dic['is_train']: False,
model_dic['positive_embedding']: pos_embedding_np[0]
}
last_conv_np, grad_conv_np, grad_input_np, loss_np = sess.run(
[
model_dic['stage4_unit3_conv2'], model_dic['grad_conv'],
model_dic['grad_input'], model_dic['loss']
],
feed_dict=fdict1)
#add neg and pos embedings
csv_file_name = os.path.join(output_path, 'distance.csv')
write_csv(name_list, loss_np * (-1), csv_file_name)
#cam, norm_grad_input_0_1:[112,112]
norm_grad_input = cacu_norm_grad(grad_input_np)
cam, norm_grads_conv = calcu_cam(last_conv_np, grad_conv_np)
cam_dir = os.path.join(output_path, 'cam')
grad_dir = os.path.join(output_path, 'grad')
# Filter with threshold
find_adv_image(args, norm_grad_input, grad_dir, feed_img,
name_list, sess, pos_embedding_np[0], model_dic)
limitations under the License.
"""
import logging as log
import os
import re
from distutils.version import LooseVersion
from mo.utils.error import Error, FrameworkError
from mo.utils.utils import refer_to_faq_msg
from mo.utils.versions_checker import get_environment_setup
try:
import tensorflow.compat.v1 as tf_v1
# disable eager execution of TensorFlow 2 environment immediately
tf_v1.disable_eager_execution()
import tensorflow as tf
from tensorflow.python.framework.convert_to_constants import convert_variables_to_constants_v2
except ImportError:
import tensorflow as tf_v1
from google.protobuf import text_format
from mo.graph.graph import fill_graph_with_nodes, Graph
from mo.utils.summarize_graph import summarize_graph
def freeze_checkpoints(graph_def: tf_v1.GraphDef, checkpoint_dir: str,
output_node_names: list):
"""
Loads all the variables in a graph and stores them in a separate dictionary. Freezes output nodes in the graph
:param graph_def: GraphDef object holding the network.
def main(): # 训练程序的主函数
if not os.path.exists(args.snapshot_dir): # 如果保存模型参数的文件夹不存在则创建
os.makedirs(args.snapshot_dir)
if not os.path.exists(args.out_dir): # 如果保存训练中可视化输出的文件夹不存在则创建
os.makedirs(args.out_dir)
train_picture_list = glob.glob(os.path.join(args.train_picture_path,
"*")) # 得到训练输入图像路径名称列表
tf.set_random_seed(args.random_seed) # 初始一下随机数
tf.disable_eager_execution() #tf2.0版本和1.x版本的区别
train_picture = tf.placeholder(
tf.float32,
shape=[1, args.image_size, args.image_size, 3],
name='train_picture') # 输入的训练图像
train_label = tf.placeholder(
tf.float32,
shape=[1, args.image_size, args.image_size, 3],
name='train_label') # 输入的与训练图像匹配的标签
gen_label = generator(image=train_picture,
gf_dim=64,
reuse=False,
name='generator') # 得到生成器的输出
dis_real = discriminator(image=train_picture,
targets=train_label,
df_dim=64,
reuse=False,
name="discriminator") # 判别器返回的对真实标签的判别结果
dis_fake = discriminator(image=train_picture,
targets=gen_label,
df_dim=64,
reuse=True,
name="discriminator") # 判别器返回的对生成(虚假的)标签判别结果
#原损失函数
gen_loss_GAN = tf.reduce_mean(
-tf.log(1 - dis_fake + EPS)) # 计算生成器损失中的GAN_loss部分
gen_loss_L1 = tf.reduce_mean(l1_loss(gen_label,
train_label)) # 计算生成器损失中的L1_loss部分
#测试用损失函数
#gen_loss_L2 = tf.reduce_mean(tf.square(tf.abs(gen_label - train_label + EPS))) # 计算生成器损失中的L2_loss部分
#gen_loss_ssim = tf.reduce_mean(1 - tf.image.ssim(gen_label, train_label, max_val=255)) # 计算生成器损失中的ssim_loss部分
# 原生成器损失函数
gen_loss = gen_loss_GAN * args.lamda_gan_weight + gen_loss_L1 * args.lamda_l1_weight # 计算生成器的loss
# 测试用生成器损失函数
#gen_loss = gen_loss_GAN * args.lamda_gan_weight + gen_loss_L2 * args.lamda_l1_weight
##最小二乘损失
#gen_loss = tf.reduce_mean(tf.square(dis_fake+ EPS)) * args.lamda_gan_weight + gen_loss_l1 * args.lamda_l1_weight
# 原判别器损失函数
dis_loss = tf.reduce_mean(
-(tf.log(1 - dis_real + EPS) + tf.log(dis_fake + EPS))) # 计算判别器的loss
# 测试用判别器损失函数
##最小二乘损失
# dis_loss = tf.reduce_mean(tf.square(dis_fake+ EPS))
gen_loss_sum = tf.summary.scalar("gen_loss", gen_loss) # 记录生成器loss的日志
dis_loss_sum = tf.summary.scalar("dis_loss", dis_loss) # 记录判别器loss的日志
summary_writer = tf.summary.FileWriter(
args.log_dir, graph=tf.get_default_graph()) # 日志记录器
g_vars = [v for v in tf.trainable_variables()
if 'generator' in v.name] # 所有生成器的可训练参数
d_vars = [
v for v in tf.trainable_variables() if 'discriminator' in v.name
] # 所有判别器的可训练参数
d_optim = tf.train.AdamOptimizer(args.base_lr, beta1=args.beta1) # 判别器训练器
g_optim = tf.train.AdamOptimizer(args.base_lr, beta1=args.beta1) # 生成器训练器
d_grads_and_vars = d_optim.compute_gradients(dis_loss,
var_list=d_vars) # 计算判别器参数梯度
d_train = d_optim.apply_gradients(d_grads_and_vars) # 更新判别器参数
g_grads_and_vars = g_optim.compute_gradients(gen_loss,
var_list=g_vars) # 计算生成器参数梯度
g_train = g_optim.apply_gradients(g_grads_and_vars) # 更新生成器参数
train_op = tf.group(d_train, g_train) # train_op表示了参数更新操作
config = tf.ConfigProto()
config.gpu_options.allow_growth = True # 设定显存不超量使用
sess = tf.Session(config=config) # 新建会话层
#init = tf.compat.v1.global_variables_initializer() # 参数初始化器
init = tf.global_variables_initializer() # 参数初始化器
sess.run(init) # 初始化所有可训练参数
# saver = tf.compat.v1.train.Saver(var_list=tf.compat.v1.global_variables(), max_to_keep=50) # 模型保存器
saver = tf.train.Saver(var_list=tf.global_variables(),
max_to_keep=50) # 模型保存器
counter = 0 # counter记录训练步数
for epoch in range(args.epoch): # 训练epoch数
shuffle(train_picture_list) # 每训练一个epoch，就打乱一下输入的顺序
for step in range(len(train_picture_list)): # 每个训练epoch中的训练step数
counter += 1
picture_name, _ = os.path.splitext(
os.path.basename(
train_picture_list[step])) # 获取不包含路径和格式的输入图片名称
# 读取一张训练图片，一张训练标签，以及相应的高和宽
picture_resize, label_resize, picture_height, picture_width = ImageReader(
file_name=picture_name,
picture_path=args.train_picture_path,
label_path=args.train_label_path,
picture_format=args.train_picture_format,
label_format=args.train_label_format,
size=args.image_size)
batch_picture = np.expand_dims(np.array(picture_resize).astype(
np.float32),
axis=0) # 填充维度
batch_label = np.expand_dims(np.array(label_resize).astype(
np.float32),
axis=0) # 填充维度
feed_dict = {
train_picture: batch_picture,
train_label: batch_label
} # 构造feed_dict
gen_loss_value, dis_loss_value, _ = sess.run(
[gen_loss, dis_loss, train_op],
feed_dict=feed_dict) # 得到每个step中的生成器和判别器loss
if counter % args.save_pred_every == 0: # 每过save_pred_every次保存模型
save(saver, sess, args.snapshot_dir, counter)
if counter % args.summary_pred_every == 0: # 每过summary_pred_every次保存训练日志
gen_loss_sum_value, discriminator_sum_value = sess.run(
[gen_loss_sum, dis_loss_sum], feed_dict=feed_dict)
summary_writer.add_summary(gen_loss_sum_value, counter)
summary_writer.add_summary(discriminator_sum_value, counter)
if counter % args.write_pred_every == 0: # 每过write_pred_every次写一下训练的可视化结果
gen_label_value = sess.run(gen_label,
feed_dict=feed_dict) # run出生成器的输出
write_image = get_write_picture(picture_resize,
gen_label_value, label_resize,
picture_height,
picture_width) # 得到训练的可视化结果
write_image_name = args.out_dir + "out" + str(
counter) + ".png" # 待保存的训练可视化结果路径与名称
cv2.imwrite(write_image_name, write_image) # 保存训练的可视化结果
print(
'epoch {:d} step {:d} \t gen_loss = {:.3f}, dis_loss = {:.3f}'.
format(epoch, step, gen_loss_value, dis_loss_value))
def __init__(self, board_width, board_height, num_ships, model_file=None):
self.board_width = board_width
self.board_height = board_height
self.num_ships = num_ships
self.num_input_dimension = self.num_ships + 1
self.board_size = self.board_width * self.board_height
self.hidden_units = self.board_size
self.output_units = self.board_size
self.type = tf.float32
tf.reset_default_graph()
tf.disable_eager_execution()
# Define the tensorflow neural network
# 1. Input:
self.input_dimensions = tf.placeholder(
self.type,
shape=[1, self.num_input_dimension * board_height * board_width])
self.input_dimensions_reshaped = tf.reshape(
self.input_dimensions,
[-1, board_height, board_width, self.num_input_dimension])
# 2. Networks Layers
self.layer1 = tf.layers.conv2d(inputs=self.input_dimensions_reshaped,
filters=32,
kernel_size=[3, 3],
padding="same",
activation=tf.nn.relu)
self.layer2 = tf.layers.conv2d(inputs=self.layer1,
filters=64,
kernel_size=[3, 3],
padding="same",
activation=tf.nn.relu)
self.layer3 = tf.layers.conv2d(inputs=self.layer2,
filters=self.num_input_dimension,
kernel_size=[1, 1],
padding="same",
activation=tf.nn.relu)
self.layer3_reshaped = tf.reshape(
self.layer3,
[-1, board_height * board_width * self.num_input_dimension])
# Alternative network structure, simple but not effective compared to CNN
# self.layer1 = tf.layers.dense(inputs=self.input_dimensions, units=self.num_input_dimension * self.board_size, activation=tf.nn.relu)
# self.layer2 = tf.layers.dense(inputs=self.layer1, units=self.num_input_dimension * self.board_size, activation=tf.nn.relu)
self.logits = tf.layers.dense(inputs=self.layer3_reshaped,
units=self.board_size)
self.probabilities = tf.nn.softmax(self.logits)
# Training step
self.labels = tf.placeholder(tf.int64)
self.learning_rate = tf.placeholder(self.type, shape=[])
self.cross_entropy = tf.nn.sparse_softmax_cross_entropy_with_logits(
logits=self.logits, labels=self.labels)
# Define loss
self.loss = tf.reduce_sum(
tf.multiply(self.learning_rate, self.cross_entropy))
self.optimizer = tf.train.AdamOptimizer(learning_rate=0.0005).minimize(
self.loss)
# Start TF session
self.session = tf.Session()
init = tf.initialize_all_variables()
self.session.run(init)
# For saving and restoring
self.saver = tf.train.Saver()
if model_file is not None:
print('load model', model_file)
self.restoreModel(model_file)
def get_network_graph(num_features,
num_classes,
num_layers,
activation_function,
learning_rate=0.1):
n = Network()
num_hidden_neurons = 20 # Neurons per hidden layer
tf.disable_eager_execution()
tf.reset_default_graph()
# x and y of Network n are placeholders.
# n.x : placeholder. no matter the number of row, it works.
n.x = tf.placeholder(tf.float32, shape=[None, num_features], name="images")
n.y_ = tf.placeholder(tf.int32, shape=[None], name="labels")
current_layer = n.x
# Add hidden layers.
# First hidden layer
# variables are in the scope of Layer1:
with tf.variable_scope("Layer1"):
# not to forget that the dimension of the weight matrix and bias vector
# is transposed.
w = tf.get_variable("weights",
shape=[num_features, num_hidden_neurons])
b = tf.get_variable("offsets", shape=[1, num_hidden_neurons])
# w(weights) shape of (num_features, num_hidden_neurons) added.
# b(biases) shape of (1, num_hidden_neurons) biases added.
n.weights.append(w)
n.biases.append(b)
# check the dimension of tf.matmul and activation function.
current_layer = activation_function(tf.matmul(current_layer, w) + b)
# Other hidden layers
for i in range(2, num_layers + 1):
with tf.variable_scope("Layer" + str(i)):
w = tf.get_variable("weights",
shape=[num_hidden_neurons, num_hidden_neurons])
b = tf.get_variable("offsets", shape=[1, num_hidden_neurons])
n.weights.append(w)
n.biases.append(b)
current_layer = activation_function(
tf.matmul(current_layer, w) + b)
# Output layer
with tf.variable_scope("LayerOutput"):
w = tf.get_variable("weights", shape=[num_hidden_neurons, num_classes])
b = tf.get_variable("offsets", shape=[1, num_classes])
n.weights.append(w)
n.biases.append(b)
y_unscaled = tf.matmul(current_layer, w) + b
correct_labels = n.y_
predicted_labels = tf.argmax(y_unscaled, axis=1, output_type=tf.int32)
# change the true/falses to float32.
correct_prediction = tf.cast(tf.equal(correct_labels, predicted_labels),
tf.float32)
n.accuracy = tf.reduce_sum(correct_prediction) / tf.cast(
tf.shape(correct_prediction)[0], tf.float32)
# loss, train_step, variables initilizer
n.loss = tf.reduce_mean(
tf.nn.sparse_softmax_cross_entropy_with_logits(logits=y_unscaled,
labels=n.y_))
n.train_step = tf.train.GradientDescentOptimizer(learning_rate).minimize(
n.loss)
n.init_op = tf.global_variables_initializer()
# return the network.
return n
def load_tf_graph_def(graph_file_name: str = "",
is_binary: bool = True,
checkpoint: str = "",
model_dir: str = "",
saved_model_tags: list = [],
meta_graph_file: str = "",
user_output_node_names_list: list = []):
# As a provisional solution, use a native TF methods to load a model protobuf
graph_def = tf_v1.GraphDef()
if isinstance(graph_file_name, str) and (re.match(r'.*\.(ckpt|meta)$',
graph_file_name)):
print(
'[ WARNING ] The value for the --input_model command line parameter ends with ".ckpt" or ".meta" '
'extension.\n'
'It means that the model is not frozen.\n'
'To load non frozen model to Model Optimizer run:'
'\n\n1. For "*.ckpt" file:'
'\n- if inference graph is in binary format'
'\npython3 mo_tf.py --input_model "path/to/inference_graph.pb" --input_checkpoint "path/to/*.ckpt"'
'\n- if inference graph is in text format'
'\npython3 mo_tf.py --input_model "path/to/inference_graph.pbtxt" --input_model_is_text '
'--input_checkpoint "path/to/*.ckpt"'
'\n\n2. For "*.meta" file:'
'\npython3 mo_tf.py --input_meta_graph "path/to/*.meta"')
variables_values = {}
try:
if graph_file_name and not meta_graph_file and not checkpoint:
# frozen graph
return read_file_to_graph_def(
graph_def, graph_file_name,
is_binary), variables_values, 'tf', None
if graph_file_name and not meta_graph_file and checkpoint:
# inference graph and checkpoint
graph_def = read_file_to_graph_def(graph_def, graph_file_name,
is_binary)
outputs = get_output_node_names_list(graph_def,
user_output_node_names_list)
if os.path.isfile(checkpoint):
graph_def = freeze_checkpoint(graph_def=graph_def,
checkpoint=checkpoint,
output_node_names=outputs)
elif os.path.isdir(checkpoint):
graph_def, variables_values = freeze_checkpoints(
graph_def=graph_def,
checkpoint_dir=checkpoint,
output_node_names=outputs)
# we are sure that checkpoint is existing file or directory due to cli_parser configuration
return graph_def, variables_values, 'tf', None
if not graph_file_name and meta_graph_file:
meta_graph_file = deducing_metagraph_path(meta_graph_file)
input_meta_graph_def = read_file_to_graph_def(
tf_v1.MetaGraphDef(), meta_graph_file, is_binary)
# Since version 2.2 TF can fail with internal error while loading graph from .meta file.
# It happens because some operation may has an _output_shapes attribute inconsistent with the GraphDef
# calculated value. To avoid this problem we must delete `_output_shapes` attributes from operations
for node in input_meta_graph_def.graph_def.node:
if '_output_shapes' in node.attr:
del node.attr['_output_shapes']
# pylint: disable=no-member
with tf_v1.Session() as sess:
restorer = tf_v1.train.import_meta_graph(input_meta_graph_def)
restorer.restore(sess, re.sub(r'\.meta$', '', meta_graph_file))
outputs = get_output_node_names_list(
input_meta_graph_def.graph_def,
user_output_node_names_list)
graph_def = tf_v1.graph_util.convert_variables_to_constants(
sess, input_meta_graph_def.graph_def, outputs)
return graph_def, variables_values, 'tf', None
if model_dir:
# saved model directory
try:
env_setup = get_environment_setup("tf")
# enable eager execution temporarily while TensorFlow 2 model is being loaded
tf_v1.enable_eager_execution()
try:
# Code to extract Keras model.
# tf.keras.models.load_model function throws TypeError,KeyError or IndexError
# for TF 1.x SavedModel format in case TF 1.x installed
imported = tf.keras.models.load_model(model_dir,
compile=False)
except:
imported = tf.saved_model.load(model_dir, saved_model_tags) # pylint: disable=E1120
# to get a signature by key throws KeyError for TF 1.x SavedModel format in case TF 2.x installed
concrete_func = imported.signatures[
tf.saved_model.DEFAULT_SERVING_SIGNATURE_DEF_KEY]
# the aggressive inlining parameter needs to freeze a table of embeddings for Keras Embedding operation
# and a model with Embedding operation cannot properly converted to IR without this function parameter
if "tensorflow" in env_setup and env_setup[
"tensorflow"] >= LooseVersion("2.2.0"):
frozen_func = convert_variables_to_constants_v2(
concrete_func,
lower_control_flow=False,
aggressive_inlining=True) # pylint: disable=E1123
else:
frozen_func = convert_variables_to_constants_v2(
concrete_func, lower_control_flow=False) # pylint: disable=E1123
graph_def = frozen_func.graph.as_graph_def(add_shapes=True)
# disable eager execution since next steps are executed with a graph in non-eager mode
tf_v1.disable_eager_execution()
input_names = []
if hasattr(imported, 'inputs'):
# Extract tensor names order from Keras model
input_names = [tensor.name for tensor in imported.inputs]
# After model freezing output tensor names are changing and recieve "Func/PartitionedCall" prefix,
# so output_names from saved_model cannot be used. Here tensor names from frozen graph are used,
# as TF adds indexed Identity nodes during freezing to each output, so this indexing is used for
# order alignment.
output_names = [tensor.name for tensor in frozen_func.outputs]
inputs_outputs_order = (input_names, output_names)
return graph_def, variables_values, 'tf2', inputs_outputs_order
except:
# disable eager execution since TensorFlow 1 model is handled
tf_v1.disable_eager_execution()
# code to extract GraphDef for TF 1.0 SavedModel format
tags = saved_model_tags if saved_model_tags is not None else [
tf_v1.saved_model.tag_constants.SERVING
]
with tf_v1.Session() as sess:
meta_graph_def = tf_v1.saved_model.loader.load(
sess, tags, model_dir)
outputs = get_output_node_names_list(
meta_graph_def.graph_def, user_output_node_names_list)
graph_def = tf_v1.graph_util.convert_variables_to_constants(
sess, meta_graph_def.graph_def, outputs)
return graph_def, variables_values, 'tf', None
except Exception as e:
raise FrameworkError('Cannot load input model: {}', e) from e
raise Error("Unknown configuration of input model parameters")
# -*- coding: utf-8 -*-
from tensorflow.compat.v1 import placeholder, Session, disable_eager_execution
from tensorflow import float32
disable_eager_execution()
a = placeholder(float32)
b = placeholder(float32)
c = placeholder(float32)
x = (a + b) / c
sess = Session()
result = sess.run([x], {a: 2, b: 3, c: 4})
print(result) # [1.25]
if FLAGS.mode == 'eval':
for ckpt in tf.train.checkpoints_iterator(
run_config.model_dir, min_interval_secs=15):
try:
result = perform_evaluation(
estimator=estimator,
input_fn=data_lib.build_input_fn(builder, False),
eval_steps=eval_steps,
model=model,
num_classes=num_classes,
checkpoint_path=ckpt)
except tf.errors.NotFoundError:
continue
if result['global_step'] >= train_steps:
return
else:
estimator.train(
data_lib.build_input_fn(builder, True), max_steps=train_steps)
if FLAGS.mode == 'train_then_eval':
perform_evaluation(
estimator=estimator,
input_fn=data_lib.build_input_fn(builder, False),
eval_steps=eval_steps,
model=model,
num_classes=num_classes)
if __name__ == '__main__':
tf.disable_eager_execution() # Disable eager mode when running with TF2.
app.run(main)
def main(_):
if FLAGS.strategy == 'horovod':
import horovod.tensorflow as hvd # pylint: disable=g-import-not-at-top
logging.info('Use horovod with multi gpus')
hvd.init()
os.environ['CUDA_VISIBLE_DEVICES'] = str(hvd.local_rank())
import tensorflow.compat.v1 as tf # pylint: disable=g-import-not-at-top
tf.enable_v2_tensorshape()
tf.disable_eager_execution()
if FLAGS.strategy == 'tpu':
tpu_cluster_resolver = tf.distribute.cluster_resolver.TPUClusterResolver(
FLAGS.tpu, zone=FLAGS.tpu_zone, project=FLAGS.gcp_project)
tpu_grpc_url = tpu_cluster_resolver.get_master()
tf.Session.reset(tpu_grpc_url)
else:
tpu_cluster_resolver = None
# Check data path
if FLAGS.mode in (
'train', 'train_and_eval') and FLAGS.training_file_pattern is None:
raise RuntimeError(
'You must specify --training_file_pattern for training.')
if FLAGS.mode in ('eval', 'train_and_eval'):
if FLAGS.validation_file_pattern is None:
raise RuntimeError('You must specify --validation_file_pattern '
'for evaluation.')
# Parse and override hparams
config = hparams_config.get_detection_config(FLAGS.model_name)
config.override(FLAGS.hparams)
if FLAGS.num_epochs: # NOTE: remove this flag after updating all docs.
config.num_epochs = FLAGS.num_epochs
# Parse image size in case it is in string format.
config.image_size = utils.parse_image_size(config.image_size)
# The following is for spatial partitioning. `features` has one tensor while
# `labels` had 4 + (`max_level` - `min_level` + 1) * 2 tensors. The input
# partition is performed on `features` and all partitionable tensors of
# `labels`, see the partition logic below.
# In the TPUEstimator context, the meaning of `shard` and `replica` is the
# same; follwing the API, here has mixed use of both.
if FLAGS.use_spatial_partition:
# Checks input_partition_dims agrees with num_cores_per_replica.
if FLAGS.num_cores_per_replica != np.prod(FLAGS.input_partition_dims):
raise RuntimeError(
'--num_cores_per_replica must be a product of array'
'elements in --input_partition_dims.')
labels_partition_dims = {
'mean_num_positives': None,
'source_ids': None,
'groundtruth_data': None,
'image_scales': None,
}
# The Input Partition Logic: We partition only the partition-able tensors.
# Spatial partition requires that the to-be-partitioned tensors must have a
# dimension that is a multiple of `partition_dims`. Depending on the
# `partition_dims` and the `image_size` and the `max_level` in config, some
# high-level anchor labels (i.e., `cls_targets` and `box_targets`) cannot
# be partitioned. For example, when `partition_dims` is [1, 4, 2, 1], image
# size is 1536, `max_level` is 9, `cls_targets_8` has a shape of
# [batch_size, 6, 6, 9], which cannot be partitioned (6 % 4 != 0). In this
# case, the level-8 and level-9 target tensors are not partition-able, and
# the highest partition-able level is 7.
feat_sizes = utils.get_feat_sizes(config.get('image_size'),
config.get('max_level'))
for level in range(config.get('min_level'),
config.get('max_level') + 1):
def _can_partition(spatial_dim):
partitionable_index = np.where(
spatial_dim % np.array(FLAGS.input_partition_dims) == 0)
return len(partitionable_index[0]) == len(
FLAGS.input_partition_dims)
spatial_dim = feat_sizes[level]
if _can_partition(spatial_dim['height']) and _can_partition(
spatial_dim['width']):
labels_partition_dims['box_targets_%d' %
level] = FLAGS.input_partition_dims
labels_partition_dims['cls_targets_%d' %
level] = FLAGS.input_partition_dims
else:
labels_partition_dims['box_targets_%d' % level] = None
labels_partition_dims['cls_targets_%d' % level] = None
num_cores_per_replica = FLAGS.num_cores_per_replica
input_partition_dims = [
FLAGS.input_partition_dims, labels_partition_dims
]
num_shards = FLAGS.num_cores // num_cores_per_replica
else:
num_cores_per_replica = None
input_partition_dims = None
num_shards = FLAGS.num_cores
params = dict(config.as_dict(),
model_name=FLAGS.model_name,
iterations_per_loop=FLAGS.iterations_per_loop,
model_dir=FLAGS.model_dir,
num_shards=num_shards,
num_examples_per_epoch=FLAGS.num_examples_per_epoch,
strategy=FLAGS.strategy,
backbone_ckpt=FLAGS.backbone_ckpt,
ckpt=FLAGS.ckpt,
val_json_file=FLAGS.val_json_file,
testdev_dir=FLAGS.testdev_dir,
mode=FLAGS.mode)
config_proto = tf.ConfigProto(allow_soft_placement=True,
log_device_placement=False)
if FLAGS.use_xla and FLAGS.strategy != 'tpu':
config_proto.graph_options.optimizer_options.global_jit_level = (
tf.OptimizerOptions.ON_1)
config_proto.gpu_options.allow_growth = True
tpu_config = tf.estimator.tpu.TPUConfig(
FLAGS.iterations_per_loop,
num_shards=num_shards,
num_cores_per_replica=num_cores_per_replica,
input_partition_dims=input_partition_dims,
per_host_input_for_training=tf.estimator.tpu.InputPipelineConfig.
PER_HOST_V2)
if FLAGS.strategy == 'horovod':
model_dir = FLAGS.model_dir if hvd.rank() == 0 else None
else:
model_dir = FLAGS.model_dir
run_config = tf.estimator.tpu.RunConfig(
cluster=tpu_cluster_resolver,
evaluation_master=FLAGS.eval_master,
model_dir=model_dir,
log_step_count_steps=FLAGS.iterations_per_loop,
session_config=config_proto,
tpu_config=tpu_config,
tf_random_seed=FLAGS.tf_random_seed,
)
model_fn_instance = det_model_fn.get_model_fn(FLAGS.model_name)
max_instances_per_image = config.max_instances_per_image
use_tpu = (FLAGS.strategy == 'tpu')
# TPU Estimator
logging.info(params)
if FLAGS.mode == 'train':
train_estimator = tf.estimator.tpu.TPUEstimator(
model_fn=model_fn_instance,
use_tpu=use_tpu,
train_batch_size=FLAGS.train_batch_size,
config=run_config,
params=params)
train_estimator.train(
input_fn=dataloader.InputReader(
FLAGS.training_file_pattern,
is_training=True,
use_fake_data=FLAGS.use_fake_data,
max_instances_per_image=max_instances_per_image),
max_steps=int((config.num_epochs * FLAGS.num_examples_per_epoch) /
FLAGS.train_batch_size))
if FLAGS.eval_after_training:
# Run evaluation after training finishes.
eval_params = dict(
params,
strategy=FLAGS.strategy,
input_rand_hflip=False,
is_training_bn=False,
mixed_precision=None,
)
eval_estimator = tf.estimator.tpu.TPUEstimator(
model_fn=model_fn_instance,
use_tpu=use_tpu,
train_batch_size=FLAGS.train_batch_size,
eval_batch_size=FLAGS.eval_batch_size,
config=run_config,
params=eval_params)
eval_results = eval_estimator.evaluate(
input_fn=dataloader.InputReader(
FLAGS.validation_file_pattern,
is_training=False,
max_instances_per_image=max_instances_per_image),
steps=FLAGS.eval_samples // FLAGS.eval_batch_size,
name=FLAGS.eval_name)
logging.info('Eval results: %s', eval_results)
ckpt = tf.train.latest_checkpoint(FLAGS.model_dir)
utils.archive_ckpt(eval_results, eval_results['AP'], ckpt)
elif FLAGS.mode == 'eval':
# Override the default options: disable randomization in the input pipeline
# and don't run on the TPU.
eval_params = dict(
params,
strategy=FLAGS.strategy,
input_rand_hflip=False,
is_training_bn=False,
mixed_precision=None,
)
eval_estimator = tf.estimator.tpu.TPUEstimator(
model_fn=model_fn_instance,
use_tpu=use_tpu,
train_batch_size=FLAGS.train_batch_size,
eval_batch_size=FLAGS.eval_batch_size,
config=run_config,
params=eval_params)
def terminate_eval():
logging.info('Terminating eval after %d seconds of no checkpoints',
FLAGS.eval_timeout)
return True
# Run evaluation when there's a new checkpoint
for ckpt in tf.train.checkpoints_iterator(
FLAGS.model_dir,
min_interval_secs=FLAGS.min_eval_interval,
timeout=FLAGS.eval_timeout,
timeout_fn=terminate_eval):
logging.info('Starting to evaluate.')
try:
eval_results = eval_estimator.evaluate(
input_fn=dataloader.InputReader(
FLAGS.validation_file_pattern,
is_training=False,
max_instances_per_image=max_instances_per_image),
steps=FLAGS.eval_samples // FLAGS.eval_batch_size,
name=FLAGS.eval_name)
logging.info('Eval results: %s', eval_results)
# Terminate eval job when final checkpoint is reached.
try:
current_step = int(os.path.basename(ckpt).split('-')[1])
except IndexError:
logging.info('%s has no global step info: stop!', ckpt)
break
utils.archive_ckpt(eval_results, eval_results['AP'], ckpt)
total_step = int(
(config.num_epochs * FLAGS.num_examples_per_epoch) /
FLAGS.train_batch_size)
if current_step >= total_step:
logging.info('Evaluation finished after training step %d',
current_step)
break
except tf.errors.NotFoundError:
# Since the coordinator is on a different job than the TPU worker,
# sometimes the TPU worker does not finish initializing until long after
# the CPU job tells it to start evaluating. In this case, the checkpoint
# file could have been deleted already.
logging.info(
'Checkpoint %s no longer exists, skipping checkpoint',
ckpt)
elif FLAGS.mode == 'train_and_eval':
for cycle in range(config.num_epochs):
logging.info('Starting training cycle, epoch: %d.', cycle)
train_estimator = tf.estimator.tpu.TPUEstimator(
model_fn=model_fn_instance,
use_tpu=use_tpu,
train_batch_size=FLAGS.train_batch_size,
config=run_config,
params=params)
train_estimator.train(input_fn=dataloader.InputReader(
FLAGS.training_file_pattern,
is_training=True,
use_fake_data=FLAGS.use_fake_data,
max_instances_per_image=max_instances_per_image),
steps=int(FLAGS.num_examples_per_epoch /
FLAGS.train_batch_size))
logging.info('Starting evaluation cycle, epoch: %d.', cycle)
# Run evaluation after every epoch.
eval_params = dict(
params,
strategy=FLAGS.strategy,
input_rand_hflip=False,
is_training_bn=False,
)
eval_estimator = tf.estimator.tpu.TPUEstimator(
model_fn=model_fn_instance,
use_tpu=use_tpu,
train_batch_size=FLAGS.train_batch_size,
eval_batch_size=FLAGS.eval_batch_size,
config=run_config,
params=eval_params)
eval_results = eval_estimator.evaluate(
input_fn=dataloader.InputReader(
FLAGS.validation_file_pattern,
is_training=False,
max_instances_per_image=max_instances_per_image),
steps=FLAGS.eval_samples // FLAGS.eval_batch_size,
name=FLAGS.eval_name)
logging.info('Evaluation results: %s', eval_results)
ckpt = tf.train.latest_checkpoint(FLAGS.model_dir)
utils.archive_ckpt(eval_results, eval_results['AP'], ckpt)
else:
logging.info('Mode not found.')
def main(argv=None):
# if tf.gfile.Exists(FLAGS.save_dir):
# tf.gfile.DeleteRecursively(FLAGS.save_dir)
# tf.gfile.MakeDirs(FLAGS.save_dir)
# if tf.gfile.Exists(FLAGS.gen_frm_dir):
# tf.gfile.DeleteRecursively(FLAGS.gen_frm_dir)
# tf.gfile.MakeDirs(FLAGS.gen_frm_dir)
# if os.path.exists(FLAGS.save_dir):
# shutil.rmtree(FLAGS.save_dir)
# os.mkdir(FLAGS.save_dir)
# if os.path.exists(FLAGS.gen_frm_dir):
# shutil.rmtree(FLAGS.gen_frm_dir)
# os.mkdir(FLAGS.gen_frm_dir)
tf.disable_eager_execution() #toegevoegd anders error
# load data
train_input_handle, test_input_handle = datasets_factory.data_provider(
FLAGS.dataset_name, FLAGS.train_data_paths, FLAGS.valid_data_paths,
FLAGS.batch_size, FLAGS.img_width)
#exit(0) #deze exit gebruikt voor het maken van de auto.py dataset-maker
print("Initializing models")
model = Model()
lr = FLAGS.lr
delta = 0.00002
base = 0.99998
eta = 1
for itr in xrange(1, FLAGS.max_iterations + 1):
if train_input_handle.no_batch_left():
train_input_handle.begin(do_shuffle=True)
ims = train_input_handle.get_batch()
ims = preprocess.reshape_patch(ims, FLAGS.patch_size)
if itr < 50000:
eta -= delta
else:
eta = 0.0
random_flip = np.random.random_sample(
(FLAGS.batch_size, FLAGS.seq_length - FLAGS.input_length - 1))
true_token = (random_flip < eta)
#true_token = (random_flip < pow(base,itr))
ones = np.ones((int(FLAGS.img_height / FLAGS.patch_size),
int(FLAGS.img_width / FLAGS.patch_size),
FLAGS.patch_size**2 * FLAGS.img_channel))
zeros = np.zeros((int(FLAGS.img_height / FLAGS.patch_size),
int(FLAGS.img_width / FLAGS.patch_size),
FLAGS.patch_size**2 * FLAGS.img_channel))
mask_true = []
for i in xrange(FLAGS.batch_size):
for j in xrange(FLAGS.seq_length - FLAGS.input_length - 1):
if true_token[i, j]:
mask_true.append(ones)
else:
mask_true.append(zeros)
mask_true = np.array(mask_true)
mask_true = np.reshape(
mask_true,
(FLAGS.batch_size, FLAGS.seq_length - FLAGS.input_length - 1,
int(FLAGS.img_height / FLAGS.patch_size),
int(FLAGS.img_width / FLAGS.patch_size),
FLAGS.patch_size**2 * FLAGS.img_channel))
print("Learning rate: " + str(lr))
cost = model.train(ims, lr, mask_true)
if itr == 32000 or itr == 52000:
lr /= 10
if FLAGS.reverse_input:
ims_rev = ims[:, ::-1]
cost += model.train(ims_rev, lr, mask_true)
cost = cost / 2
if itr % FLAGS.display_interval == 0:
print('itr: ' + str(itr))
print('training loss: ' + str(cost))
if itr % FLAGS.test_interval == 0:
print('test...')
test_input_handle.begin(do_shuffle=False)
res_path = os.path.join(FLAGS.gen_frm_dir, str(itr))
os.mkdir(res_path)
avg_mse = 0
batch_id = 0
img_mse, ssim, psnr, fmae, sharp = [], [], [], [], []
for i in xrange(FLAGS.seq_length - FLAGS.input_length):
img_mse.append(0)
ssim.append(0)
psnr.append(0)
fmae.append(0)
sharp.append(0)
mask_true = np.zeros(
(FLAGS.batch_size, FLAGS.seq_length - FLAGS.input_length - 1,
int(FLAGS.img_height / FLAGS.patch_size),
int(FLAGS.img_width / FLAGS.patch_size),
FLAGS.patch_size**2 * FLAGS.img_channel))
while (test_input_handle.no_batch_left() == False):
batch_id = batch_id + 1
test_ims = test_input_handle.get_batch()
test_dat = preprocess.reshape_patch(test_ims, FLAGS.patch_size)
img_gen = model.test(test_dat, mask_true)
# concat outputs of different gpus along batch
img_gen = np.concatenate(img_gen)
img_gen = preprocess.reshape_patch_back(
img_gen, FLAGS.patch_size)
# MSE per frame
for i in xrange(FLAGS.seq_length - FLAGS.input_length):
x = test_ims[:, i + FLAGS.input_length, :, :, 0]
gx = img_gen[:, i, :, :, 0]
fmae[i] += metrics.batch_mae_frame_float(gx, x)
gx = np.maximum(gx, 0)
gx = np.minimum(gx, 1)
mse = np.square(x - gx).sum()
img_mse[i] += mse
avg_mse += mse
real_frm = np.uint8(x * 255)
pred_frm = np.uint8(gx * 255)
psnr[i] += metrics.batch_psnr(pred_frm, real_frm)
for b in xrange(FLAGS.batch_size):
sharp[i] += np.max(
cv2.convertScaleAbs(cv2.Laplacian(pred_frm[b], 3)))
score, _ = compare_ssim(pred_frm[b],
real_frm[b],
full=True)
ssim[i] += score
# save prediction examples
if batch_id <= 10:
path = os.path.join(res_path, str(batch_id))
os.mkdir(path)
for i in xrange(FLAGS.seq_length):
name = 'gt' + str(i + 1) + '.png'
file_name = os.path.join(path, name)
img_gt = np.uint8(test_ims[0, i, :, :, :] * 255)
cv2.imwrite(file_name, img_gt)
for i in xrange(FLAGS.seq_length - FLAGS.input_length):
name = 'pd' + str(i + 1 + FLAGS.input_length) + '.png'
file_name = os.path.join(path, name)
img_pd = img_gen[0, i, :, :, :]
img_pd = np.maximum(img_pd, 0)
img_pd = np.minimum(img_pd, 1)
img_pd = np.uint8(img_pd * 255)
cv2.imwrite(file_name, img_pd)
test_input_handle.next()
avg_mse = avg_mse / (batch_id * FLAGS.batch_size)
print('mse per seq: ' + str(avg_mse))
for i in xrange(FLAGS.seq_length - FLAGS.input_length):
print(img_mse[i] / (batch_id * FLAGS.batch_size))
psnr = np.asarray(psnr, dtype=np.float32) / batch_id
fmae = np.asarray(fmae, dtype=np.float32) / batch_id
ssim = np.asarray(ssim,
dtype=np.float32) / (FLAGS.batch_size * batch_id)
sharp = np.asarray(
sharp, dtype=np.float32) / (FLAGS.batch_size * batch_id)
print('psnr per frame: ' + str(np.mean(psnr)))
for i in xrange(FLAGS.seq_length - FLAGS.input_length):
print(psnr[i])
print('fmae per frame: ' + str(np.mean(fmae)))
for i in xrange(FLAGS.seq_length - FLAGS.input_length):
print(fmae[i])
print('ssim per frame: ' + str(np.mean(ssim)))
for i in xrange(FLAGS.seq_length - FLAGS.input_length):
print(ssim[i])
print('sharpness per frame: ' + str(np.mean(sharp)))
for i in xrange(FLAGS.seq_length - FLAGS.input_length):
print(sharp[i])
if itr % FLAGS.snapshot_interval == 0:
model.save(itr)
train_input_handle.next()
import tensorflow.compat.v1 as tensorflow
tensorflow.disable_eager_execution()
config = tensorflow.ConfigProto()
config.gpu_options.allow_growth = True
import random
import numpy
import pickle
from Dataset import Dataset
from Embeddings.TransE import TransE
from Embeddings.ComplEx import ComplEx
from Embeddings.TuckER import TuckER
import argparse
parser = argparse.ArgumentParser()
parser.add_argument('-d', '--dataset', type=str, default="NELL995")
parser.add_argument('-t',
'--task',
type=str,
default="concept_agentbelongstoorganization")
parser.add_argument('-x', '--embedding-method', type=str, default="TransE")
parser.add_argument('-e', '--embedding-size', type=int, default=100)
parser.add_argument('-m', '--margin', type=float, default=1.0)
parser.add_argument('-r', '--learning-rate', type=float, default=1e-3)
parser.add_argument('-b', '--batch-size', type=int, default=1024)
parser.add_argument('-g', '--sampling-type', type=str, default='bernoulli')
"""
# pylint: disable=missing-docstring
from __future__ import absolute_import
from __future__ import division
from __future__ import print_function
import os
import re
import sys
import tarfile
from six.moves import urllib
#修改以适应TF2.0
import tensorflow.compat.v1 as tf
tf.disable_eager_execution(
) #tensorflow2.0运行1.0的代码会报错误：“AttributeError: module 'tensorflow' has no attribute 'placeholder'”
import cifar10_input
FLAGS = tf.app.flags.FLAGS
# Basic model parameters.
tf.app.flags.DEFINE_integer('batch_size', 128,
"""Number of images to process in a batch.""")
tf.app.flags.DEFINE_string('data_dir', 'cifar10_data',
"""Path to the CIFAR-10 data directory.""")
tf.app.flags.DEFINE_boolean('use_fp16', False,
"""Train the model using fp16.""")
# Global constants describing the CIFAR-10 data set.
IMAGE_SIZE = cifar10_input.IMAGE_SIZE
def main(argv=None):
tf.disable_eager_execution() #toegevoegd anders error
# load data
_, test_input_handle = datasets_factory.data_provider(
FLAGS.dataset_name, FLAGS.train_data_paths, FLAGS.test_data_paths,
FLAGS.batch_size, FLAGS.img_width)
print("Initializing models")
model = Model()
lr = FLAGS.lr
print('test...')
test_input_handle.begin(do_shuffle=False)
res_path = os.path.join(FLAGS.gen_frm_dir, 'test')
os.mkdir(res_path)
avg_mse = 0
batch_id = 0
img_mse, ssim, psnr, fmae, sharp = [], [], [], [], []
for i in xrange(FLAGS.seq_length - FLAGS.input_length):
img_mse.append(0)
ssim.append(0)
psnr.append(0)
fmae.append(0)
sharp.append(0)
mask_true = np.zeros(
(FLAGS.batch_size, FLAGS.seq_length - FLAGS.input_length - 1,
int(FLAGS.img_height / FLAGS.patch_size),
int(FLAGS.img_width / FLAGS.patch_size),
FLAGS.patch_size**2 * FLAGS.img_channel))
while (test_input_handle.no_batch_left() == False):
batch_id = batch_id + 1
test_ims = test_input_handle.get_batch()
test_dat = preprocess.reshape_patch(test_ims, FLAGS.patch_size)
img_gen = model.test(test_dat, mask_true)
# concat outputs of different gpus along batch
img_gen = np.concatenate(img_gen)
img_gen = preprocess.reshape_patch_back(img_gen, FLAGS.patch_size)
# MSE per frame
for i in xrange(FLAGS.seq_length - FLAGS.input_length):
x = test_ims[:, i + FLAGS.input_length, :, :, 0]
gx = img_gen[:, i, :, :, 0]
fmae[i] += metrics.batch_mae_frame_float(gx, x)
gx = np.maximum(gx, 0)
gx = np.minimum(gx, 1)
mse = np.square(x - gx).sum()
img_mse[i] += mse
avg_mse += mse
real_frm = np.uint8(x * 255)
pred_frm = np.uint8(gx * 255)
psnr[i] += metrics.batch_psnr(pred_frm, real_frm)
for b in xrange(FLAGS.batch_size):
sharp[i] += np.max(
cv2.convertScaleAbs(cv2.Laplacian(pred_frm[b], 3)))
score, _ = compare_ssim(pred_frm[b], real_frm[b], full=True)
ssim[i] += score
# save prediction examples
if batch_id <= 10:
path = os.path.join(res_path, str(batch_id))
os.mkdir(path)
for i in xrange(FLAGS.seq_length):
name = 'gt' + str(i + 1) + '.png'
file_name = os.path.join(path, name)
img_gt = np.uint8(test_ims[0, i, :, :, :] * 255)
cv2.imwrite(file_name, img_gt)
for i in xrange(FLAGS.seq_length - FLAGS.input_length):
name = 'pd' + str(i + 1 + FLAGS.input_length) + '.png'
file_name = os.path.join(path, name)
img_pd = img_gen[0, i, :, :, :]
img_pd = np.maximum(img_pd, 0)
img_pd = np.minimum(img_pd, 1)
img_pd = np.uint8(img_pd * 255)
cv2.imwrite(file_name, img_pd)
test_input_handle.next()
avg_mse = avg_mse / (batch_id * FLAGS.batch_size)
print('mse per seq: ' + str(avg_mse))
for i in xrange(FLAGS.seq_length - FLAGS.input_length):
print(img_mse[i] / (batch_id * FLAGS.batch_size))
psnr = np.asarray(psnr, dtype=np.float32) / batch_id
fmae = np.asarray(fmae, dtype=np.float32) / batch_id
ssim = np.asarray(ssim, dtype=np.float32) / (FLAGS.batch_size * batch_id)
sharp = np.asarray(sharp, dtype=np.float32) / (FLAGS.batch_size * batch_id)
print('psnr per frame: ' + str(np.mean(psnr)))
for i in xrange(FLAGS.seq_length - FLAGS.input_length):
print(psnr[i])
print('fmae per frame: ' + str(np.mean(fmae)))
for i in xrange(FLAGS.seq_length - FLAGS.input_length):
print(fmae[i])
print('ssim per frame: ' + str(np.mean(ssim)))
for i in xrange(FLAGS.seq_length - FLAGS.input_length):
print(ssim[i])
print('sharpness per frame: ' + str(np.mean(sharp)))
for i in xrange(FLAGS.seq_length - FLAGS.input_length):
print(sharp[i])
def __init__(self, log_dir):
"""Create a summary writer logging to log_dir."""
tf.disable_eager_execution()
self.writer = tf.summary.FileWriter(log_dir)
def main(argv):
del argv # unused.
tf.reset_default_graph()
tf.disable_eager_execution()
# set random seed.
tf.set_random_seed(FLAGS.tf_rand_seed)
np.random.seed(FLAGS.np_rand_seed)
dataset_name = FLAGS.dataset_name
n_labels = num_labels[dataset_name]
num_train = num_train_dict[dataset_name]
num_test = num_test_dict[dataset_name]
matching = FLAGS.matching
# load training and test data.
print('--- Loading {} ---'.format(dataset_name))
# print('loading dataset')
x_train, y_train, x_test, y_test = get_data(dataset_name)
# define the model.
model_name = FLAGS.model_name
if model_name == 'cnn':
print('--- Building CNN model ---')
model = models.CnnModel(n_labels)
elif model_name == 'vgg':
print('--- Building VGG model ---')
model = models.VggModel(n_labels)
else:
raise NotImplementedError
# num of batches to eval test data.
eval_batch_size = FLAGS.eval_batch_size
num_eval_batch = int(num_test / eval_batch_size)
# num of batches to eval train data.
num_eval_train_batch = int(num_train / eval_batch_size)
batch_size = FLAGS.batch_size
num_steps_per_epoch = int(num_train / batch_size)
training_epochs = FLAGS.training_epochs
# define training operations.
global_step = tf.train.get_or_create_global_step()
trainable_vars = tf.trainable_variables()
grads = tf.gradients(model.total_loss, trainable_vars)
learning_rate = FLAGS.learning_rate
train_step = tf.train.GradientDescentOptimizer(learning_rate).minimize(
model.total_loss, global_step=global_step)
# define matching operations
if matching != 'none':
l_rate_match = get_matching_lr(matching, learning_rate, batch_size)
trainable_vars_ph_0 = []
trainable_vars_ph_1 = []
for var in trainable_vars:
trainable_vars_ph_0.append(
tf.placeholder(tf.float32, shape=var.shape))
trainable_vars_ph_1.append(
tf.placeholder(tf.float32, shape=var.shape))
matching_step = []
for var, ph_0, ph_1 in zip(trainable_vars, trainable_vars_ph_0,
trainable_vars_ph_1):
matching_step.append(
tf.assign_add(var, l_rate_match * (ph_0 - ph_1)))
# checkpoint saver
saver = tf.train.Saver(max_to_keep=1)
checkpoint_dir = FLAGS.checkpoint_dir
# start training.
with tf.Session() as sess:
sess.run(tf.global_variables_initializer())
# First, if checkpoint_dir does not exist, create a folder.
if not tf.gfile.IsDirectory(checkpoint_dir):
# checkpoint directory does not exist; create directory.
tf.gfile.MakeDirs(checkpoint_dir)
epoch = 0
train_acc_record = []
train_loss_record = []
test_acc_record = []
test_loss_record = []
while epoch < training_epochs:
# compute train/test acc/loss and log, add to records.
# compute test acc and loss on whole dataset.
test_acc = 0.0
test_loss = 0.0
for b_id in range(num_eval_batch):
data_dict_eval = {
model.x_input:
x_test[b_id * eval_batch_size:(b_id + 1) * eval_batch_size,
...],
model.y_input:
y_test[b_id * eval_batch_size:(b_id + 1) * eval_batch_size]
}
test_acc_batch, test_loss_batch = sess.run(
[model.accuracy, model.total_loss],
feed_dict=data_dict_eval)
test_acc += test_acc_batch
test_loss += test_loss_batch
test_acc /= float(num_eval_batch)
test_loss /= float(num_eval_batch)
test_acc_record.append(test_acc)
test_loss_record.append(test_loss)
# compute train acc, loss, and full gradient on whole dataset.
train_acc = 0.0
train_loss = 0.0
for b_id in range(num_eval_train_batch):
data_dict_eval_train = {
model.x_input:
x_train[b_id * eval_batch_size:(b_id + 1) *
eval_batch_size, ...],
model.y_input:
y_train[b_id * eval_batch_size:(b_id + 1) *
eval_batch_size]
}
train_acc_batch, train_loss_batch = sess.run(
[model.accuracy, model.total_loss],
feed_dict=data_dict_eval_train)
train_acc += train_acc_batch
train_loss += train_loss_batch
train_acc /= float(num_eval_train_batch)
train_loss /= float(num_eval_train_batch)
train_acc_record.append(train_acc)
train_loss_record.append(train_loss)
print('--- epoch {}, train acc {:.4f}, test acc {:.4f} ---'.format(
epoch, train_acc, test_acc))
if (epoch + 1) % FLAGS.num_checkpoint_epochs == 0:
# write a checkpoint.
print('--- Saving checkpoint ---')
# save checkpoints.
saver.save(sess,
os.path.join(checkpoint_dir, 'model.ckpt'),
global_step=global_step)
# save training and test records.
print('--- Saving records ---')
# save train/test records.
np.save(checkpoint_dir + '/train_acc.npy', train_acc_record)
np.save(checkpoint_dir + '/train_loss.npy', train_loss_record)
np.save(checkpoint_dir + '/test_acc.npy', test_acc_record)
np.save(checkpoint_dir + '/test_loss.npy', test_loss_record)
# Actual training step
for _ in range(num_steps_per_epoch):
# sample a batch without replacement
selected_index = np.random.choice(num_train,
size=batch_size,
replace=False)
x_batch = x_train[selected_index, ...]
x_aug = data_augment(x_batch)
y_batch = y_train[selected_index]
data_dict = {model.x_input: x_aug, model.y_input: y_batch}
sess.run(train_step, feed_dict=data_dict)
# If using matching schemes, do these after the gradient descent step.
if matching == 'lr_matching':
selected_index_0 = np.random.choice(num_train,
size=batch_size,
replace=True)
selected_index_1 = np.random.choice(num_train,
size=batch_size,
replace=True)
elif matching == 'batch_matching':
selected_index_0 = np.random.choice(
num_train,
size=FLAGS.matching_batch_size,
replace=True)
selected_index_1 = np.random.choice(
num_train,
size=FLAGS.matching_batch_size,
replace=True)
if matching != 'none':
x_batch_0 = x_train[selected_index_0, ...]
x_batch_1 = x_train[selected_index_1, ...]
x_aug_0 = data_augment(x_batch_0)
x_aug_1 = data_augment(x_batch_1)
y_batch_0 = y_train[selected_index_0]
y_batch_1 = y_train[selected_index_1]
data_dict_0 = {
model.x_input: x_aug_0,
model.y_input: y_batch_0
}
data_dict_1 = {
model.x_input: x_aug_1,
model.y_input: y_batch_1
}
# compute gradients and difference
grads_val_0 = sess.run(grads, feed_dict=data_dict_0)
grads_val_1 = sess.run(grads, feed_dict=data_dict_1)
# add to variables
grads_dict = {}
for ph_0, ph_1, array_0, array_1 in zip(
trainable_vars_ph_0, trainable_vars_ph_1,
grads_val_0, grads_val_1):
grads_dict[ph_0] = array_0
grads_dict[ph_1] = array_1
sess.run(matching_step, feed_dict=grads_dict)
# This epoch finished, update the epoch counter.
epoch += 1
def create_client_batches(clients: List[ReptileClient],
batch_size: int) -> List[List[ReptileClient]]:
if batch_size == -1:
client_batches = [clients]
else:
client_batches = [
clients[i:i + batch_size]
for i in range(0, len(clients), batch_size)
]
return client_batches
num_clients_train = 10000
num_clients_test = 1000
num_classes_per_client = 5
num_shots_per_class = 5
eval_iters = 10
reptile_args = ReptileTrainingArgs(model=OmniglotModel,
inner_optimizer=optim.Adam,
inner_learning_rate=0.001,
num_inner_steps=5,
num_inner_steps_eval=50,
log_every_n_steps=3,
inner_batch_size=10,
meta_batch_size=5,
meta_learning_rate_initial=1,
meta_learning_rate_final=0,
num_meta_steps=3000)
experiment_logger = create_tensorboard_logger(
context.name, "dataloading_ours;models_ours")
# Load and prepare Omniglot data
data_dir = REPO_ROOT / 'data' / 'omniglot'
#######
tf.disable_eager_execution()
#######
omniglot_train_clients, omniglot_test_clients = load_omniglot_datasets(
str(data_dir.absolute()),
num_clients_train=num_clients_train,
num_clients_test=num_clients_test,
num_classes_per_client=num_classes_per_client,
num_shots_per_class=num_shots_per_class,
inner_batch_size=reptile_args.inner_batch_size,
random_seed=RANDOM_SEED)
# Prepare ModelArgs for task training
inner_optimizer_args = OptimizerArgs(
optimizer_class=reptile_args.inner_optimizer,
lr=reptile_args.inner_learning_rate,
betas=(0, 0.999))
inner_model_args = ModelArgs(reptile_args.model,
inner_optimizer_args,
num_classes=num_classes_per_client)
dummy_optimizer_args = OptimizerArgs(optimizer_class=optim.SGD)
meta_model_args = ModelArgs(reptile_args.model,
dummy_optimizer_args,
num_classes=num_classes_per_client)
"""
# Set up clients
# Since we are doing meta-learning, we need separate sets of training and
# test clients
train_clients = initialize_reptile_clients(context, train_datasets)
test_clients = initialize_reptile_clients(context, test_datasets)
# Set up server
server = ReptileServer(
participant_name='initial_server',
model_args=context.meta_model_args,
context=context,
initial_model_state=initial_model_state
)"""
torch_model = OmniglotModel(num_classes=num_classes_per_client)
torch_optimizer = inner_optimizer_args.optimizer_class(
torch_model.parameters(), **inner_optimizer_args.optimizer_kwargs)
reptile = Reptile(model=torch_model,
optimizer=torch_optimizer,
inner_iterations=reptile_args.num_inner_steps,
inner_iterations_eval=reptile_args.num_inner_steps_eval)
for i in range(reptile_args.num_meta_steps):
frac_done = i / reptile_args.num_meta_steps
cur_meta_step_size = frac_done * reptile_args.meta_learning_rate_final + (
1 - frac_done) * reptile_args.meta_learning_rate_initial
meta_batch = {
k: omniglot_train_clients.train_data_local_dict[k]
for k in cyclerange(
i * reptile_args.meta_batch_size %
len(omniglot_train_clients.train_data_local_dict), (i + 1) *
reptile_args.meta_batch_size %
len(omniglot_train_clients.train_data_local_dict),
len(omniglot_train_clients.train_data_local_dict))
}
reptile.train_step(meta_batch=meta_batch,
meta_step_size=cur_meta_step_size)
if i % eval_iters == 0:
accuracies = []
k = RANDOM.randrange(
len(omniglot_train_clients.train_data_local_dict))
train_train = omniglot_train_clients.train_data_local_dict[k]
train_test = omniglot_train_clients.test_data_local_dict[k]
k = RANDOM.randrange(
len(omniglot_test_clients.train_data_local_dict))
test_train = omniglot_test_clients.train_data_local_dict[k]
test_test = omniglot_test_clients.test_data_local_dict[k]
for train_dl, test_dl in [(train_train, train_test),
(test_train, test_test)]:
correct = reptile.evaluate(train_dl, test_dl)
accuracies.append(correct / num_classes_per_client)
print('batch %d: train=%f test=%f' %
(i, accuracies[0], accuracies[1]))
# Write to TensorBoard
experiment_logger.experiment.add_scalar(
'train-test/acc/{}/mean'.format('global_model'),
accuracies[0],
global_step=i)
experiment_logger.experiment.add_scalar(
'test-test/acc/{}/mean'.format('global_model'),
accuracies[1],
global_step=i)
"""
#shenjingtest3.py
import tensorflow.compat.v1 as tf
tf.disable_eager_execution() #启用动态图机制
#构建两个常数
matrix1 = tf.constant([[3, 3]])
matrix2 = tf.constant([[2], [2]])
product = tf.matmul(matrix1, matrix2) #矩阵相乘
##part1
# sess = tf.Session()
# result = sess.run(product)
# print(result)
# sess.close()#关闭会话
#part2
with tf.Session() as sess: #进入Session会话
result = sess.run(product)
print(result)
import resnet
import data as data_lib
import model as model_lib
import model_util as model_util
import tensorflow.compat.v1 as tf
import tensorflow_datasets as tfds
import tensorflow_hub as hub
#boostx :
'''
the tf2.0 have a issue with mode.save https://github.com/tensorflow/tensorflow/issues/26814
use tf.keras.models.save() instead of model.save()
'''
'''
tf.disable_eager_execution()
with tf.Session() as sess:
saver = tf.train.import_meta_graph('./tmp/simclr_test_ft/model.ckpt-196.meta')
saver.restore(sess, tf.train.latest_checkpoint('./tmp/simclr_test_ft'))
graph = tf.get_default_graph()
print(graph.get_operations())
'''
FLAGS = flags.FLAGS
flags.DEFINE_float('learning_rate', 0.3,
'Initial learning rate per batch size of 256.')
def testPCgradBasic(self, denylist, allowlist, pcgrad_var_idx):
tf.disable_eager_execution()
for dtype in [tf.dtypes.float32, tf.dtypes.float64]:
with self.session(graph=tf.Graph()):
var0_np = np.array([1.0, 2.0], dtype=dtype.as_numpy_dtype)
var1_np = np.array([3.0, 4.0], dtype=dtype.as_numpy_dtype)
const0_np = np.array([1., 0.], dtype=dtype.as_numpy_dtype)
const1_np = np.array([-1., -1.], dtype=dtype.as_numpy_dtype)
const2_np = np.array([-1., 1.], dtype=dtype.as_numpy_dtype)
var0 = tf.Variable(var0_np, dtype=dtype, name='first_var/var0')
var1 = tf.Variable(var1_np,
dtype=dtype,
name='second_var/var1')
const0 = tf.constant(const0_np)
const1 = tf.constant(const1_np)
const2 = tf.constant(const2_np)
loss0 = tf.tensordot(var0, const0, 1) + tf.tensordot(
var1, const2, 1)
loss1 = tf.tensordot(var0, const1, 1) + tf.tensordot(
var1, const0, 1)
learning_rate = lambda: 0.001
opt = tf.train.GradientDescentOptimizer(learning_rate)
losses = loss0 + loss1
opt_grads = opt.compute_gradients(losses,
var_list=[var0, var1])
pcgrad_opt = pcgrad.PCGrad(
tf.train.GradientDescentOptimizer(learning_rate),
denylist=denylist,
allowlist=allowlist)
pcgrad_col_opt = pcgrad.PCGrad(
tf.train.GradientDescentOptimizer(learning_rate),
use_collection_losses=True,
denylist=denylist,
allowlist=allowlist)
losses = [loss0, loss1]
pcgrad_grads = pcgrad_opt.compute_gradients(
losses, var_list=[var0, var1])
tf.add_to_collection(pcgrad.PCGRAD_LOSSES_COLLECTION, loss0)
tf.add_to_collection(pcgrad.PCGRAD_LOSSES_COLLECTION, loss1)
pcgrad_grads_collection = pcgrad_col_opt.compute_gradients(
None, var_list=[var0, var1])
with tf.Graph().as_default():
# Shouldn't return non-slot variables from other graphs.
self.assertEmpty(opt.variables())
self.evaluate(tf.global_variables_initializer())
grad_vec, pcgrad_vec, pcgrad_col_vec = self.evaluate(
[opt_grads, pcgrad_grads, pcgrad_grads_collection])
# Make sure that both methods take grads of the same vars.
self.assertAllCloseAccordingToType(pcgrad_vec, pcgrad_col_vec)
results = [{
'var': var0,
'pcgrad_vec': [0.5, -1.5],
'result': [0.9995, 2.0015]
}, {
'var': var1,
'pcgrad_vec': [0.5, 1.5],
'result': [2.9995, 3.9985]
}]
grad_var_idx = {0, 1}.difference(pcgrad_var_idx)
self.assertAllCloseAccordingToType(grad_vec[0][0], [0.0, -1.0],
atol=1e-5)
self.assertAllCloseAccordingToType(grad_vec[1][0], [0.0, 1.0],
atol=1e-5)
pcgrad_vec_idx = 0
for var_idx in pcgrad_var_idx:
self.assertAllCloseAccordingToType(
pcgrad_vec[pcgrad_vec_idx][0],
results[var_idx]['pcgrad_vec'],
atol=1e-5)
pcgrad_vec_idx += 1
for var_idx in grad_var_idx:
self.assertAllCloseAccordingToType(
pcgrad_vec[pcgrad_vec_idx][0],
grad_vec[var_idx][0],
atol=1e-5)
pcgrad_vec_idx += 1
self.evaluate(opt.apply_gradients(pcgrad_grads))
self.assertAllCloseAccordingToType(
self.evaluate(
[results[idx]['var'] for idx in pcgrad_var_idx]),
[results[idx]['result'] for idx in pcgrad_var_idx])
def __init__(
self,
estimator: "SPEECH_RECOGNIZER_TYPE",
masker: "PsychoacousticMasker",
eps: float = 2000.0,
learning_rate_1: float = 100.0,
max_iter_1: int = 1000,
alpha: float = 0.05,
learning_rate_2: float = 1.0,
max_iter_2: int = 4000,
loss_theta_min: float = 0.05,
decrease_factor_eps: float = 0.8,
num_iter_decrease_eps: int = 10,
increase_factor_alpha: float = 1.2,
num_iter_increase_alpha: int = 20,
decrease_factor_alpha: float = 0.8,
num_iter_decrease_alpha: int = 50,
batch_size: int = 1,
) -> None:
"""
Create an instance of the :class:`.ImperceptibleASR`.
The default parameters assume that audio input is in `int16` range. If using normalized audio input, parameters
`eps` and `learning_rate_{1,2}` need to be scaled with a factor `2^-15`
:param estimator: A trained speech recognition estimator.
:param masker: A Psychoacoustic masker.
:param eps: Initial max norm bound for adversarial perturbation.
:param learning_rate_1: Learning rate for stage 1 of attack.
:param max_iter_1: Number of iterations for stage 1 of attack.
:param alpha: Initial alpha value for balancing stage 2 loss.
:param learning_rate_2: Learning rate for stage 2 of attack.
:param max_iter_2: Number of iterations for stage 2 of attack.
:param loss_theta_min: If imperceptible loss reaches minimum, stop early. Works best with `batch_size=1`.
:param decrease_factor_eps: Decrease factor for epsilon (Paper default: 0.8).
:param num_iter_decrease_eps: Iterations after which to decrease epsilon if attack succeeds (Paper default: 10).
:param increase_factor_alpha: Increase factor for alpha (Paper default: 1.2).
:param num_iter_increase_alpha: Iterations after which to increase alpha if attack succeeds (Paper default: 20).
:param decrease_factor_alpha: Decrease factor for alpha (Paper default: 0.8).
:param num_iter_decrease_alpha: Iterations after which to decrease alpha if attack fails (Paper default: 50).
:param batch_size: Batch size.
"""
# Super initialization
super().__init__(estimator=estimator)
self.masker = masker
self.eps = eps
self.learning_rate_1 = learning_rate_1
self.max_iter_1 = max_iter_1
self.alpha = alpha
self.learning_rate_2 = learning_rate_2
self.max_iter_2 = max_iter_2
self._targeted = True
self.batch_size = batch_size
self.loss_theta_min = loss_theta_min
self.decrease_factor_eps = decrease_factor_eps
self.num_iter_decrease_eps = num_iter_decrease_eps
self.increase_factor_alpha = increase_factor_alpha
self.num_iter_increase_alpha = num_iter_increase_alpha
self.decrease_factor_alpha = decrease_factor_alpha
self.num_iter_decrease_alpha = num_iter_decrease_alpha
self._check_params()
# init some aliases
self._window_size = masker.window_size
self._hop_size = masker.hop_size
self._sample_rate = masker.sample_rate
self._framework: Optional[str] = None
if isinstance(self.estimator, TensorFlowV2Estimator):
import tensorflow.compat.v1 as tf1
# set framework attribute
self._framework = "tensorflow"
# disable eager execution and use tensorflow.compat.v1 API, e.g. Lingvo uses TF2v1 AP
tf1.disable_eager_execution()
# TensorFlow placeholders
self._delta = tf1.placeholder(tf1.float32,
shape=[None, None],
name="art_delta")
self._power_spectral_density_maximum_tf = tf1.placeholder(
tf1.float32, shape=[None], name="art_psd_max")
self._masking_threshold_tf = tf1.placeholder(
tf1.float32,
shape=[None, None, None],
name="art_masking_threshold")
# TensorFlow loss gradient ops
self._loss_gradient_masking_threshold_op_tf = self._loss_gradient_masking_threshold_tf(
self._delta, self._power_spectral_density_maximum_tf,
self._masking_threshold_tf)
elif isinstance(self.estimator, PyTorchEstimator):
# set framework attribute
self._framework = "pytorch"
def test_sum(self):
tf.disable_eager_execution()
nodes = tf.constant([1, 3])
nodes2 = tf.constant([1, 3])
fused = efficientdet_arch.fuse_features([nodes, nodes2], 'sum')
self.assertAllCloseAccordingToType(fused, [2, 6])
def main(_):
tf.disable_eager_execution()
if not FLAGS.dataset_dir:
raise ValueError('You must supply the dataset directory with --dataset_dir')
tf.logging.set_verbosity(tf.logging.INFO)
with tf.Graph().as_default():
######################
# Select the dataset #
######################
dataset = dataset_factory.get_dataset(
FLAGS.dataset_name, FLAGS.dataset_split_name, FLAGS.dataset_dir)
#########################
# Configure the network #
#########################
inception_params = network_params.InceptionV3FCNParams(
receptive_field_size=FLAGS.receptive_field_size,
prelogit_dropout_keep_prob=0.8,
depth_multiplier=0.1,
min_depth=16,
inception_fcn_stride=0,
)
conv_params = network_params.ConvScopeParams(
dropout=False,
dropout_keep_prob=0.8,
batch_norm=True,
batch_norm_decay=0.99,
l2_weight_decay=4e-05,
)
network_fn = inception_v3_fcn.get_inception_v3_fcn_network_fn(
inception_params,
conv_params,
num_classes=dataset.num_classes,
is_training=True,
)
#####################################
# Select the preprocessing function #
#####################################
image_preprocessing_fn = preprocessing_factory.get_preprocessing(
'inception_v3', is_training=True)
##############################################################
# Create a dataset provider that loads data from the dataset #
##############################################################
provider = slim.dataset_data_provider.DatasetDataProvider(
dataset,
num_readers=DATASET_READERS,
common_queue_capacity=20 * FLAGS.batch_size,
common_queue_min=10 * FLAGS.batch_size)
[image, label] = provider.get(['image', 'label'])
train_image_size = FLAGS.receptive_field_size
image = image_preprocessing_fn(image, train_image_size, train_image_size)
images, labels = tf.train.batch([image, label],
batch_size=FLAGS.batch_size,
num_threads=PREPROCESSING_THREADS,
capacity=5 * FLAGS.batch_size)
labels = slim.one_hot_encoding(labels, dataset.num_classes)
####################
# Define the model #
####################
logits, _ = network_fn(images)
slim.losses.softmax_cross_entropy(logits, labels)
total_loss = slim.losses.get_total_loss()
tf.summary.scalar('losses/Total_Loss', total_loss)
optimizer = tf.train.RMSPropOptimizer(0.01)
train_op = slim.learning.create_train_op(
total_loss,
optimizer,
variables_to_train=_get_variables_to_train())
###########################
# Kicks off the training. #
###########################
slim.learning.train(
train_op,
logdir=FLAGS.train_dir,
init_fn=_get_init_fn(),
number_of_steps=FLAGS.max_number_of_steps,
log_every_n_steps=FLAGS.log_every_n_steps,
save_summaries_secs=FLAGS.save_summaries_secs,
save_interval_secs=FLAGS.save_interval_secs,
session_config=tf.ConfigProto(allow_soft_placement=True))
def validate_yolo_model_pb(model_path, image_file, anchors, class_names,
model_image_size, elim_grid_sense, v5_decode,
loop_count):
# check tf version to be compatible with TF 2.x
global tf
if tf.__version__.startswith('2'):
import tensorflow.compat.v1 as tf
tf.disable_eager_execution()
# NOTE: TF 1.x frozen pb graph need to specify input/output tensor name
# so we hardcode the input/output tensor names here to get them from model
if len(anchors) == 6:
output_tensor_names = [
'graph/predict_conv_1/BiasAdd:0', 'graph/predict_conv_2/BiasAdd:0'
]
elif len(anchors) == 9:
output_tensor_names = [
'graph/predict_conv_1/BiasAdd:0', 'graph/predict_conv_2/BiasAdd:0',
'graph/predict_conv_3/BiasAdd:0'
]
elif len(anchors) == 5:
# YOLOv2 use 5 anchors and have only 1 prediction
output_tensor_names = ['graph/predict_conv/BiasAdd:0']
else:
raise ValueError('invalid anchor number')
# assume only 1 input tensor for image
input_tensor_name = 'graph/image_input:0'
#load frozen pb graph
def load_pb_graph(model_path):
# We parse the graph_def file
with tf.gfile.GFile(model_path, "rb") as f:
graph_def = tf.GraphDef()
graph_def.ParseFromString(f.read())
# We load the graph_def in the default graph
with tf.Graph().as_default() as graph:
tf.import_graph_def(graph_def,
input_map=None,
return_elements=None,
name="graph",
op_dict=None,
producer_op_list=None)
return graph
graph = load_pb_graph(model_path)
# We can list operations, op.values() gives you a list of tensors it produces
# op.name gives you the name. These op also include input & output node
# print output like:
# prefix/Placeholder/inputs_placeholder
# ...
# prefix/Accuracy/predictions
#
# NOTE: prefix/Placeholder/inputs_placeholder is only op's name.
# tensor name should be like prefix/Placeholder/inputs_placeholder:0
#for op in graph.get_operations():
#print(op.name, op.values())
image_input = graph.get_tensor_by_name(input_tensor_name)
output_tensors = [
graph.get_tensor_by_name(output_tensor_name)
for output_tensor_name in output_tensor_names
]
batch, height, width, channel = image_input.shape
model_image_size = (int(height), int(width))
img = Image.open(image_file)
image = np.array(img, dtype='uint8')
image_data = preprocess_image(img, model_image_size)
#origin image shape, in (height, width) format
image_shape = tuple(reversed(img.size))
# predict once first to bypass the model building time
with tf.Session(graph=graph) as sess:
prediction = sess.run(output_tensors,
feed_dict={image_input: image_data})
start = time.time()
for i in range(loop_count):
with tf.Session(graph=graph) as sess:
prediction = sess.run(output_tensors,
feed_dict={image_input: image_data})
end = time.time()
print("Average Inference time: {:.8f}ms".format(
(end - start) * 1000 / loop_count))
prediction.sort(key=lambda x: len(x[0]))
handle_prediction(prediction, image_file, image, image_shape, anchors,
class_names, model_image_size, elim_grid_sense,
v5_decode)
def load_tf_weights_in_bert_generation(
model, tf_hub_path, model_class, is_encoder_named_decoder=False, is_encoder=False
):
try:
import numpy as np
import tensorflow.compat.v1 as tf
import tensorflow_hub as hub
import tensorflow_text # noqa: F401
tf.disable_eager_execution()
except ImportError:
logger.error(
"Loading a TensorFlow model in PyTorch, requires TensorFlow to be installed. Please see "
"https://www.tensorflow.org/install/ for installation instructions."
)
raise
tf_model = hub.Module(tf_hub_path)
init = tf.global_variables_initializer()
with tf.Session() as sess:
init.run()
all_variables = tf_model.variable_map
keep_track_variables = all_variables.copy()
for key in list(all_variables.keys()):
if "global" in key:
logger.info(f"Skipping {key}...")
continue
if not is_encoder:
model_pointer = getattr(model, model_class)
else:
model_pointer = model
is_embedding = False
logger.info(f"Trying to match {key}...")
# remove start_string = "module/bert/"
sub_layers = key.split("/")[2:]
if is_encoder_named_decoder and sub_layers[0] == "encoder":
logger.info(f"Skipping encoder layer {key} for decoder")
continue
if is_encoder and sub_layers[0] == "decoder":
logger.info(f"Skipping decoder layer {key} for encoder")
continue
for i, sub_layer in enumerate(sub_layers):
if sub_layer == "embeddings":
is_embedding = True
elif sub_layer == "LayerNorm":
is_embedding = False
if "layer" in sub_layer:
model_pointer = model_pointer.layer[int(sub_layer.split("_")[-1])]
elif sub_layer in ["kernel", "gamma"]:
model_pointer = model_pointer.weight
elif sub_layer == "beta":
model_pointer = model_pointer.bias
elif sub_layer == "encdec":
model_pointer = model_pointer.crossattention.self
elif sub_layer == "encdec_output":
model_pointer = model_pointer.crossattention.output
elif is_encoder_named_decoder and sub_layer == "decoder":
model_pointer = model_pointer.encoder
else:
if sub_layer == "attention" and "encdec" in sub_layers[i + 1]:
continue
try:
model_pointer = getattr(model_pointer, sub_layer)
except AttributeError:
logger.info(f"Skipping to initialize {key} at {sub_layer}...")
raise AttributeError
array = np.asarray(sess.run(all_variables[key]))
if not is_embedding:
logger.info(f"Transposing numpy weight of shape {array.shape} for {key}")
array = np.transpose(array)
else:
model_pointer = model_pointer.weight
if model_pointer.shape != array.shape:
raise ValueError(f"Pointer shape {model_pointer.shape} and array shape {array.shape} mismatched")
logger.info(f"Initialize PyTorch weight {key}")
model_pointer.data = torch.from_numpy(array.astype(np.float32))
keep_track_variables.pop(key, None)
logger.info(f"Weights not copied to PyTorch model: {', '.join(keep_track_variables.keys())}")
return model
def setUpModule():
tf.disable_eager_execution()
# encoding: utf-8
#import tensorflow as tf
import tensorflow.compat.v1 as tf
tf.disable_eager_execution() # 忽略占位符
import numpy as np
#from auxiliaryTools.ExtractData import Dataset
#from auxiliaryTools.GetTest import get_test_list
from ExtractData import Dataset
from GetTest import get_test_list
from time import time
import math, os
def get_train_instance(train):
user_input, item_input, rates = [], [], []
for (u, i) in train.keys():
# positive instance
user_input.append(u)
item_input.append(i)
rates.append(train[u, i])
return user_input, item_input, rates
def get_train_instance_batch_change(count, batch_size, user_input, item_input,
ratings, user_reviews, item_reviews):
users_batch, items_batch, user_input_batch, item_input_batch, labels_batch = [], [], [], [], []
for idx in range(batch_size):
index = (count * batch_size + idx) % len(user_input)
users_batch.append(user_input[index])
def load_tf_graph_def(graph_file_name: str = "",
is_binary: bool = True,
checkpoint: str = "",
model_dir: str = "",
saved_model_tags: list = [],
meta_graph_file: str = "",
user_output_node_names_list: list = []):
# As a provisional solution, use a native TF methods to load a model protobuf
graph_def = tf_v1.GraphDef()
if isinstance(graph_file_name, str) and (re.match('.*\.(ckpt|meta)$',
graph_file_name)):
print(
'[ WARNING ] The value for the --input_model command line parameter ends with ".ckpt" or ".meta" '
'extension.\n'
'It means that the model is not frozen.\n'
'To load non frozen model to Model Optimizer run:'
'\n\n1. For "*.ckpt" file:'
'\n- if inference graph is in binary format'
'\npython3 mo_tf.py --input_model "path/to/inference_graph.pb" --input_checkpoint "path/to/*.ckpt"'
'\n- if inference graph is in text format'
'\npython3 mo_tf.py --input_model "path/to/inference_graph.pbtxt" --input_model_is_text '
'--input_checkpoint "path/to/*.ckpt"'
'\n\n2. For "*.meta" file:'
'\npython3 mo_tf.py --input_meta_graph "path/to/*.meta"')
variables_values = {}
try:
if graph_file_name and not meta_graph_file and not checkpoint:
# frozen graph
return read_file_to_graph_def(graph_def, graph_file_name,
is_binary), variables_values
if graph_file_name and not meta_graph_file and checkpoint:
# inference graph and checkpoint
graph_def = read_file_to_graph_def(graph_def, graph_file_name,
is_binary)
outputs = get_output_node_names_list(graph_def,
user_output_node_names_list)
if os.path.isfile(checkpoint):
graph_def = freeze_checkpoint(graph_def=graph_def,
checkpoint=checkpoint,
output_node_names=outputs)
elif os.path.isdir(checkpoint):
graph_def, variables_values = freeze_checkpoints(
graph_def=graph_def,
checkpoint_dir=checkpoint,
output_node_names=outputs)
# we are sure that checkpoint is existing file or directory due to cli_parser configuration
return graph_def, variables_values
if not graph_file_name and meta_graph_file:
meta_graph_file = deducing_metagraph_path(meta_graph_file)
input_meta_graph_def = read_file_to_graph_def(
tf_v1.MetaGraphDef(), meta_graph_file, is_binary)
# pylint: disable=no-member
with tf_v1.Session() as sess:
restorer = tf_v1.train.import_meta_graph(input_meta_graph_def)
restorer.restore(sess, re.sub('\.meta$', '', meta_graph_file))
outputs = get_output_node_names_list(
input_meta_graph_def.graph_def,
user_output_node_names_list)
graph_def = tf_v1.graph_util.convert_variables_to_constants(
sess, input_meta_graph_def.graph_def, outputs)
return graph_def, variables_values
if model_dir:
# saved model directory
try:
env_setup = get_environment_setup()
# enable eager execution temporarily while TensorFlow 2 model is being loaded
tf_v1.enable_eager_execution()
# code to extract GraphDef for TF 2.0 SavedModel format
# tf.saved_model.load function throws TypeError for TF 1.x SavedModel format in case TF 1.x installed
imported = tf.saved_model.load(model_dir, saved_model_tags) # pylint: disable=E1120
# to get a signature by key throws KeyError for TF 1.x SavedModel format in case TF 2.x installed
concrete_func = imported.signatures[
tf.saved_model.DEFAULT_SERVING_SIGNATURE_DEF_KEY]
# the aggressive inlining parameter needs to freeze a table of embeddings for Keras Embedding operation
# and a model with Embedding operation cannot properly converted to IR without this function parameter
if "tensorflow" in env_setup and env_setup[
"tensorflow"] >= LooseVersion("2.2.0"):
frozen_func = convert_variables_to_constants_v2(
concrete_func,
lower_control_flow=False,
aggressive_inlining=True) # pylint: disable=E1123
else:
frozen_func = convert_variables_to_constants_v2(
concrete_func, lower_control_flow=False) # pylint: disable=E1123
graph_def = frozen_func.graph.as_graph_def(add_shapes=True)
# disable eager execution since next steps are executed with a graph in non-eager mode
tf_v1.disable_eager_execution()
return graph_def, variables_values
except (TypeError, KeyError):
# disable eager execution since TensorFlow 1 model is handled
tf_v1.disable_eager_execution()
# code to extract GraphDef for TF 1.0 SavedModel format
tags = saved_model_tags if saved_model_tags is not None else [
tf_v1.saved_model.tag_constants.SERVING
]
with tf_v1.Session() as sess:
meta_graph_def = tf_v1.saved_model.loader.load(
sess, tags, model_dir)
outputs = get_output_node_names_list(
meta_graph_def.graph_def, user_output_node_names_list)
graph_def = tf_v1.graph_util.convert_variables_to_constants(
sess, meta_graph_def.graph_def, outputs)
return graph_def, variables_values
except Exception as e:
raise FrameworkError('SavedModel format load failure: {}',
e) from e
except Exception as e:
raise FrameworkError('Cannot load input model: {}', e) from e
raise Error("Unknown configuration of input model parameters")
def main(_):
if FLAGS.strategy == 'tpu':
tf.disable_eager_execution()
tpu_cluster_resolver = tf.distribute.cluster_resolver.TPUClusterResolver(
FLAGS.tpu, zone=FLAGS.tpu_zone, project=FLAGS.gcp_project)
tpu_grpc_url = tpu_cluster_resolver.get_master()
tf.Session.reset(tpu_grpc_url)
else:
tpu_cluster_resolver = None
# Check data path
if FLAGS.mode in ('train', 'train_and_eval'):
if FLAGS.training_file_pattern is None:
raise RuntimeError('Must specify --training_file_pattern for train.')
if FLAGS.mode in ('eval', 'train_and_eval'):
if FLAGS.validation_file_pattern is None:
raise RuntimeError('Must specify --validation_file_pattern for eval.')
# Parse and override hparams
config = hparams_config.get_detection_config(FLAGS.model_name)
config.override(FLAGS.hparams)
if FLAGS.num_epochs: # NOTE: remove this flag after updating all docs.
config.num_epochs = FLAGS.num_epochs
# Parse image size in case it is in string format.
config.image_size = utils.parse_image_size(config.image_size)
# The following is for spatial partitioning. `features` has one tensor while
# `labels` had 4 + (`max_level` - `min_level` + 1) * 2 tensors. The input
# partition is performed on `features` and all partitionable tensors of
# `labels`, see the partition logic below.
# In the TPUEstimator context, the meaning of `shard` and `replica` is the
# same; follwing the API, here has mixed use of both.
if FLAGS.use_spatial_partition:
# Checks input_partition_dims agrees with num_cores_per_replica.
if FLAGS.num_cores_per_replica != np.prod(FLAGS.input_partition_dims):
raise RuntimeError('--num_cores_per_replica must be a product of array'
'elements in --input_partition_dims.')
labels_partition_dims = {
'mean_num_positives': None,
'source_ids': None,
'groundtruth_data': None,
'image_scales': None,
'image_masks': None,
}
# The Input Partition Logic: We partition only the partition-able tensors.
feat_sizes = utils.get_feat_sizes(
config.get('image_size'), config.get('max_level'))
for level in range(config.get('min_level'), config.get('max_level') + 1):
def _can_partition(spatial_dim):
partitionable_index = np.where(
spatial_dim % np.array(FLAGS.input_partition_dims) == 0)
return len(partitionable_index[0]) == len(FLAGS.input_partition_dims)
spatial_dim = feat_sizes[level]
if _can_partition(spatial_dim['height']) and _can_partition(
spatial_dim['width']):
labels_partition_dims['box_targets_%d' %
level] = FLAGS.input_partition_dims
labels_partition_dims['cls_targets_%d' %
level] = FLAGS.input_partition_dims
else:
labels_partition_dims['box_targets_%d' % level] = None
labels_partition_dims['cls_targets_%d' % level] = None
num_cores_per_replica = FLAGS.num_cores_per_replica
input_partition_dims = [FLAGS.input_partition_dims, labels_partition_dims]
num_shards = FLAGS.num_cores // num_cores_per_replica
else:
num_cores_per_replica = None
input_partition_dims = None
num_shards = FLAGS.num_cores
params = dict(
config.as_dict(),
model_name=FLAGS.model_name,
iterations_per_loop=FLAGS.iterations_per_loop,
model_dir=FLAGS.model_dir,
num_shards=num_shards,
num_examples_per_epoch=FLAGS.num_examples_per_epoch,
strategy=FLAGS.strategy,
backbone_ckpt=FLAGS.backbone_ckpt,
ckpt=FLAGS.ckpt,
val_json_file=FLAGS.val_json_file,
testdev_dir=FLAGS.testdev_dir,
profile=FLAGS.profile,
mode=FLAGS.mode)
config_proto = tf.ConfigProto(
allow_soft_placement=True, log_device_placement=False)
if FLAGS.strategy != 'tpu':
if FLAGS.use_xla:
config_proto.graph_options.optimizer_options.global_jit_level = (
tf.OptimizerOptions.ON_1)
config_proto.gpu_options.allow_growth = True
model_dir = FLAGS.model_dir
model_fn_instance = det_model_fn.get_model_fn(FLAGS.model_name)
max_instances_per_image = config.max_instances_per_image
if FLAGS.eval_samples:
eval_steps = int((FLAGS.eval_samples + FLAGS.eval_batch_size - 1) //
FLAGS.eval_batch_size)
else:
eval_steps = None
total_examples = int(config.num_epochs * FLAGS.num_examples_per_epoch)
train_steps = total_examples // FLAGS.train_batch_size
logging.info(params)
if not tf.io.gfile.exists(model_dir):
tf.io.gfile.makedirs(model_dir)
config_file = os.path.join(model_dir, 'config.yaml')
if not tf.io.gfile.exists(config_file):
tf.io.gfile.GFile(config_file, 'w').write(str(config))
train_input_fn = dataloader.InputReader(
FLAGS.training_file_pattern,
is_training=True,
use_fake_data=FLAGS.use_fake_data,
max_instances_per_image=max_instances_per_image)
eval_input_fn = dataloader.InputReader(
FLAGS.validation_file_pattern,
is_training=False,
use_fake_data=FLAGS.use_fake_data,
max_instances_per_image=max_instances_per_image)
if FLAGS.strategy == 'tpu':
tpu_config = tf.estimator.tpu.TPUConfig(
FLAGS.iterations_per_loop if FLAGS.strategy == 'tpu' else 1,
num_cores_per_replica=num_cores_per_replica,
input_partition_dims=input_partition_dims,
per_host_input_for_training=tf.estimator.tpu.InputPipelineConfig
.PER_HOST_V2)
run_config = tf.estimator.tpu.RunConfig(
cluster=tpu_cluster_resolver,
model_dir=model_dir,
log_step_count_steps=FLAGS.iterations_per_loop,
session_config=config_proto,
tpu_config=tpu_config,
save_checkpoints_steps=FLAGS.save_checkpoints_steps,
tf_random_seed=FLAGS.tf_random_seed,
)
# TPUEstimator can do both train and eval.
train_est = tf.estimator.tpu.TPUEstimator(
model_fn=model_fn_instance,
train_batch_size=FLAGS.train_batch_size,
eval_batch_size=FLAGS.eval_batch_size,
config=run_config,
params=params)
eval_est = train_est
else:
strategy = None
if FLAGS.strategy == 'gpus':
strategy = tf.distribute.MirroredStrategy()
run_config = tf.estimator.RunConfig(
model_dir=model_dir,
train_distribute=strategy,
log_step_count_steps=FLAGS.iterations_per_loop,
session_config=config_proto,
save_checkpoints_steps=FLAGS.save_checkpoints_steps,
tf_random_seed=FLAGS.tf_random_seed,
)
def get_estimator(global_batch_size):
params['num_shards'] = getattr(strategy, 'num_replicas_in_sync', 1)
params['batch_size'] = global_batch_size // params['num_shards']
return tf.estimator.Estimator(
model_fn=model_fn_instance, config=run_config, params=params)
# train and eval need different estimator due to different batch size.
train_est = get_estimator(FLAGS.train_batch_size)
eval_est = get_estimator(FLAGS.eval_batch_size)
# start train/eval flow.
if FLAGS.mode == 'train':
train_est.train(input_fn=train_input_fn, max_steps=train_steps)
if FLAGS.eval_after_training:
eval_est.evaluate(input_fn=eval_input_fn, steps=eval_steps)
elif FLAGS.mode == 'eval':
# Run evaluation when there's a new checkpoint
for ckpt in tf.train.checkpoints_iterator(
FLAGS.model_dir,
min_interval_secs=FLAGS.min_eval_interval,
timeout=FLAGS.eval_timeout):
logging.info('Starting to evaluate.')
try:
eval_results = eval_est.evaluate(eval_input_fn, steps=eval_steps)
# Terminate eval job when final checkpoint is reached.
try:
current_step = int(os.path.basename(ckpt).split('-')[1])
except IndexError:
logging.info('%s has no global step info: stop!', ckpt)
break
utils.archive_ckpt(eval_results, eval_results['AP'], ckpt)
if current_step >= train_steps:
logging.info('Eval finished step %d/%d', current_step, train_steps)
break
except tf.errors.NotFoundError:
# Checkpoint might be not already deleted by the time eval finished.
# We simply skip ssuch case.
logging.info('Checkpoint %s no longer exists, skipping.', ckpt)
elif FLAGS.mode == 'train_and_eval':
ckpt = tf.train.latest_checkpoint(FLAGS.model_dir)
try:
step = int(os.path.basename(ckpt).split('-')[1])
current_epoch = (
step * FLAGS.train_batch_size // FLAGS.num_examples_per_epoch)
logging.info('found ckpt at step %d (epoch %d)', step, current_epoch)
except (IndexError, TypeError):
logging.info('Folder %s has no ckpt with valid step.', FLAGS.model_dir)
current_epoch = 0
def run_train_and_eval(e):
print('\n =====> Starting training, epoch: %d.' % e)
train_est.train(
input_fn=train_input_fn,
max_steps=e * FLAGS.num_examples_per_epoch // FLAGS.train_batch_size)
print('\n =====> Starting evaluation, epoch: %d.' % e)
eval_results = eval_est.evaluate(input_fn=eval_input_fn, steps=eval_steps)
ckpt = tf.train.latest_checkpoint(FLAGS.model_dir)
utils.archive_ckpt(eval_results, eval_results['AP'], ckpt)
epochs_per_cycle = 1 # higher number has less graph construction overhead.
for e in range(current_epoch + 1, config.num_epochs + 1, epochs_per_cycle):
if FLAGS.run_epoch_in_child_process:
p = multiprocessing.Process(target=run_train_and_eval, args=(e,))
p.start()
p.join()
if p.exitcode != 0:
return p.exitcode
else:
tf.compat.v1.reset_default_graph()
run_train_and_eval(e)
else:
logging.info('Invalid mode: %s', FLAGS.mode)
tensorrt=FLAGS.tensorrt,
use_xla=FLAGS.use_xla,
ckpt_path=FLAGS.ckpt_path,
export_ckpt=FLAGS.export_ckpt,
saved_model_dir=FLAGS.saved_model_dir,
tflite_path=FLAGS.tflite_path,
batch_size=FLAGS.batch_size,
hparams=FLAGS.hparams,
score_thresh=FLAGS.min_score_thresh,
max_output_size=FLAGS.max_boxes_to_draw,
nms_method=FLAGS.nms_method)
inspector.run_model(FLAGS.runmode,
input_image=FLAGS.input_image,
output_image_dir=FLAGS.output_image_dir,
input_video=FLAGS.input_video,
output_video=FLAGS.output_video,
line_thickness=FLAGS.line_thickness,
max_boxes_to_draw=FLAGS.max_boxes_to_draw,
min_score_thresh=FLAGS.min_score_thresh,
nms_method=FLAGS.nms_method,
bm_runs=FLAGS.bm_runs,
threads=FLAGS.threads,
trace_filename=FLAGS.trace_filename)
if __name__ == '__main__':
logging.set_verbosity(logging.WARNING)
tf.enable_v2_tensorshape()
tf.disable_eager_execution()
app.run(main)
def main(_):
logging.info("Seed: %d", FLAGS.seed)
np.random.seed(FLAGS.seed)
torch.random.manual_seed(FLAGS.seed)
if FLAGS.save_freq % FLAGS.small_eval_freq != 0:
raise ValueError(
("Save frequency ({}) must be a multiple of evaluation frequency ({})."
" Allows choosing checkpoints based on their evaluation scores.")
.format(FLAGS.save_freq, FLAGS.small_eval_freq))
if FLAGS.full_eval_freq % FLAGS.small_eval_freq != 0:
raise ValueError(
("Full evaluation frequency ({}) must be a multiple of small"
" evaluation frequency ({}) so that their values can be compared.")
.format(FLAGS.full_eval_freq, FLAGS.small_eval_freq))
exp_dir = os.path.join(FLAGS.experiment_base_dir, FLAGS.experiment_name)
common_utils.create_experiment_directory(exp_dir, FLAGS.force_overwrite)
tensorboard_dir = os.path.join(exp_dir, "tensorboard")
tf.disable_eager_execution()
tb_writer = tf.summary.FileWriter(tensorboard_dir)
predictions_dir = os.path.join(exp_dir, "predictions")
os.makedirs(predictions_dir, exist_ok=True)
checkpoints_dir = os.path.join(exp_dir, "checkpoints")
os.makedirs(checkpoints_dir, exist_ok=True)
evict_trace_dir = os.path.join(exp_dir, "evictions")
os.makedirs(evict_trace_dir, exist_ok=True)
model_config = cfg.Config.from_files_and_bindings(
FLAGS.model_configs, FLAGS.model_bindings)
logging.info("Model config: %s", model_config)
with open(os.path.join(exp_dir, "model_config.json"), "w") as f:
model_config.to_file(f)
cache_config = cfg.Config.from_files_and_bindings(
FLAGS.cache_configs, FLAGS.cache_bindings)
logging.info("Cache config: %s", cache_config)
with open(os.path.join(exp_dir, "cache_config.json"), "w") as f:
cache_config.to_file(f)
dagger_schedule_config = cfg.Config.from_files_and_bindings(
FLAGS.dagger_schedule_configs, FLAGS.dagger_schedule_bindings)
logging.info("DAgger config: %s", dagger_schedule_config)
with open(os.path.join(exp_dir, "dagger_config.json"), "w") as f:
dagger_schedule_config.to_file(f)
dagger_schedule = schedule_from_config(dagger_schedule_config)
# Process everything on GPU if available
device = torch.device("cpu")
if torch.cuda.is_available():
torch.set_default_tensor_type(torch.cuda.FloatTensor)
device = torch.device("cuda:0")
logging.info("Device: %s", device)
policy_model = model.EvictionPolicyModel.from_config(model_config).to(device)
optimizer = optim.Adam(policy_model.parameters(), lr=model_config.get("lr"))
step = 0
get_step = lambda: step
oracle_valid_data, hit_rates = next(measure_cache_hit_rate(
FLAGS.valid_memtrace, cache_config, policy_model,
schedules.ConstantSchedule(0), get_step,
os.path.join(evict_trace_dir, "oracle_valid.txt")))
log_hit_rates(tb_writer, "cache_hit_rate/oracle_valid", hit_rates, step)
with tqdm.tqdm(total=FLAGS.total_steps) as pbar:
while True: # loop for waiting until steps == FLAGS.total_steps
# Optimization: Instead of passing through the whole memory trace for
# training and only using update_freq many of them, we lazily gather k *
# update_freq batches and still train on a subsample of update_freq.
# The value of k=collection_multiplier trades off between:
# - The set of k * update_freq examples are all consecutive in the
# memory trace. As k gets small, the set of these examples becomes less
# i.i.d., as they are temporally correlated. The examples cannot be
# random access within the memory trace, since at time t, we require the
# previous cache accesses to compute the cache state at time t.
# - As k gets large, training becomes slower, as we must perform k times
# as much collecting work than training work.
max_examples = (dagger_schedule_config.get("update_freq") *
FLAGS.collection_multiplier * FLAGS.batch_size)
train_data_generator = measure_cache_hit_rate(
FLAGS.train_memtrace, cache_config, policy_model, dagger_schedule,
get_step, os.path.join(evict_trace_dir, "mixture-train-{}.txt"),
max_examples=max_examples)
for train_data, hit_rates in train_data_generator:
log_hit_rates(
tb_writer, "cache_hit_rate/train_mixture_policy", hit_rates, step)
utils.log_scalar(
tb_writer, "cache_hit_rate/mixture_parameter",
dagger_schedule.value(step), step)
for batch_num, batch in enumerate(utils.as_batches(
[train_data], FLAGS.batch_size,
model_config.get("sequence_length"))):
def evaluate_helper(eval_size, suffix):
"""Evaluates the model on train / valid data on and off-policy.
Args:
eval_size (int): the number of examples to evaluate on.
suffix (str): appended to all logging and tensorboard paths.
"""
evaluate(policy_model, oracle_valid_data[-eval_size:], step,
"off_policy_valid" + suffix, tb_writer, predictions_dir)
# train_data is defined in the loop, but evaluate_helper is only
# called in the same loop iteration.
# pylint: disable=cell-var-from-loop
evaluate(policy_model, train_data[-eval_size:],
step, "train" + suffix, tb_writer, predictions_dir)
# pylint: enable=cell-var-from-loop
# Log the cache hit rates on portions of train / valid
_, hit_rates = next(measure_cache_hit_rate(
FLAGS.train_memtrace, cache_config, policy_model,
schedules.ConstantSchedule(1), get_step,
os.path.join(
evict_trace_dir, "train{}-{}.txt".format(suffix, step)),
max_examples=eval_size, use_oracle_scores=False))
log_hit_rates(
tb_writer, "cache_hit_rate/train" + suffix, hit_rates, step)
# Use oracle scores, since eviction trace in log_evaluate_stats will
# log with on-policy scores.
on_policy_valid_data, hit_rates = next(measure_cache_hit_rate(
FLAGS.valid_memtrace, cache_config, policy_model,
schedules.ConstantSchedule(1), get_step,
os.path.join(
evict_trace_dir, "valid{}-{}.txt".format(suffix, step)),
max_examples=eval_size))
log_hit_rates(
tb_writer, "cache_hit_rate/valid" + suffix, hit_rates, step)
evaluate(policy_model, on_policy_valid_data[-eval_size:], step,
"on_policy_valid" + suffix, tb_writer, predictions_dir)
if step % FLAGS.small_eval_freq == 0:
evaluate_helper(FLAGS.small_eval_size, "")
if step % FLAGS.full_eval_freq == 0:
evaluate_helper(len(oracle_valid_data), "_full")
if step % FLAGS.save_freq == 0 and step != 0:
save_path = os.path.join(checkpoints_dir, "{}.ckpt".format(step))
with open(save_path, "wb") as save_file:
checkpoint_buffer = io.BytesIO()
torch.save(policy_model.state_dict(), checkpoint_buffer)
logging.info("Saving model checkpoint to: %s", save_path)
save_file.write(checkpoint_buffer.getvalue())
optimizer.zero_grad()
losses = policy_model.loss(
batch, model_config.get("sequence_length") // 2)
total_loss = sum(losses.values())
total_loss.backward()
optimizer.step()
pbar.update(1)
step += 1
if step % FLAGS.tb_freq == 0:
utils.log_scalar(tb_writer, "loss/total", total_loss, step)
for loss_name, loss_value in losses.items():
utils.log_scalar(
tb_writer, "loss/{}".format(loss_name), loss_value, step)
if step == FLAGS.total_steps:
return
# Break out of inner-loop to get next set of k * update_freq batches
if batch_num == dagger_schedule_config.get("update_freq"):
break