def call_zac(txt_path, model_path):
# Read in the csv with the file names you would want to predict on
filepath = txt_path
with open(filepath) as fp:
all_filenames = fp.readlines()
file_names = [x.strip() for x in all_filenames]
# file_names = np.loadtxt(txt_path, dtype=np.str)
# file_names = pd.read_csv(
# '/media/data/Track_2/New_Labels_For_Track_2.csv',
# dtype=object,
# keep_default_na=False,
# na_values=[]).as_matrix()
# We trained on the first 4 subjects, so we predict on the rest
# file_names = file_names[-N_VALIDATION_SUBJECTS:]
config = tf.ConfigProto()
config.gpu_options.allow_growth = True
nn = tf.estimator.Estimator(
model_fn=model_fn,
model_dir=model_path,
params={"learning_rate": 1e-3},
config=tf.estimator.RunConfig(session_config=config))
reader_params = {
'n_examples': 1,
'example_size': RESIZE_SIZE,
'extract_examples': True
}
reader_example_shapes = {
'features': {
'x': reader_params['example_size'] + [1]
},
'labels': {
'y': []
}
}
reader = Reader(read_fn, {
'features': {
'x': tf.float32
},
'labels': {
'y': tf.int32
}
})
input_fn, qinit_hook = reader.get_inputs(
file_references=file_names,
mode=tf.estimator.ModeKeys.PREDICT,
example_shapes=reader_example_shapes,
batch_size=1,
shuffle_cache_size=1,
params=reader_params)
features_to_save = []
labels_to_save = []
for i in tqdm(nn.predict(input_fn, hooks=[qinit_hook])):
features_to_save.append(i['features'])
labels_to_save.append(i['y_'])
return np.array(features_to_save), np.array(labels_to_save)
def train(args):
np.random.seed(42)
tf.set_random_seed(42)
print('Setting up...')
# Parse csv files for file names
all_filenames = pd.read_csv(args.train_csv,
dtype=object,
keep_default_na=False,
na_values=[]).as_matrix()
train_filenames = all_filenames[1:10]
val_filenames = all_filenames[10:12]
# Set up a data reader to handle the file i/o.
reader_params = {
'n_examples': 16,
'example_size': [1, 64, 64],
'extract_examples': True
}
reader_example_shapes = {
'features': {
'x': reader_params['example_size'] + [
NUM_CHANNELS,
]
},
'labels': {
'y': reader_params['example_size']
}
}
reader = Reader(read_fn, {
'features': {
'x': tf.float32
},
'labels': {
'y': tf.int32
}
})
# Get input functions and queue initialisation hooks for training and
# validation data
train_input_fn, train_qinit_hook = reader.get_inputs(
file_references=train_filenames,
mode=tf.estimator.ModeKeys.TRAIN,
example_shapes=reader_example_shapes,
batch_size=BATCH_SIZE,
shuffle_cache_size=SHUFFLE_CACHE_SIZE,
params=reader_params)
val_input_fn, val_qinit_hook = reader.get_inputs(
file_references=val_filenames,
mode=tf.estimator.ModeKeys.EVAL,
example_shapes=reader_example_shapes,
batch_size=BATCH_SIZE,
shuffle_cache_size=SHUFFLE_CACHE_SIZE,
params=reader_params)
# Instantiate the neural network estimator
nn = tf.estimator.Estimator(model_fn=model_fn,
model_dir=args.model_path,
params={"learning_rate": 0.001},
config=tf.estimator.RunConfig())
# Hooks for validation summaries
val_summary_hook = tf.contrib.training.SummaryAtEndHook(
os.path.join(args.model_path, 'eval'))
step_cnt_hook = tf.train.StepCounterHook(every_n_steps=EVAL_EVERY_N_STEPS,
output_dir=args.model_path)
print('Starting training...')
try:
for _ in range(MAX_STEPS // EVAL_EVERY_N_STEPS):
nn.train(input_fn=train_input_fn,
hooks=[train_qinit_hook, step_cnt_hook],
steps=EVAL_EVERY_N_STEPS)
if args.run_validation:
results_val = nn.evaluate(
input_fn=val_input_fn,
hooks=[val_qinit_hook, val_summary_hook],
steps=EVAL_STEPS)
print('Step = {}; val loss = {:.5f};'.format(
results_val['global_step'], results_val['loss']))
except KeyboardInterrupt:
pass
print('Stopping now.')
export_dir = nn.export_savedmodel(
export_dir_base=args.model_path,
serving_input_receiver_fn=reader.serving_input_receiver_fn(
reader_example_shapes))
print('Model saved to {}.'.format(export_dir))
def tune_train(args):
np.random.seed(42)
tf.set_random_seed(42)
print('Setting up...')
# Parse csv files for file names
patch_filenames = pd.read_csv(args.train_csv,
dtype=object,
keep_default_na=False,
na_values=[]).as_matrix()
subj_filenames = pd.read_csv(args.val_csv,
dtype=object,
keep_default_na=False,
na_values=[]).as_matrix()
app_json = get_config_for_app()
val_filenames = []
for row in subj_filenames:
if row[4] == '1':
val_filenames.append(row)
# Set up a data reader to handle the file i/o.
reader_params = {
'n_examples': 16,
'example_size': [1, 64, 64],
'extract_examples': True
}
num_channels = app_json['num_channels']
reader_example_shapes = {
'features': {
'x': reader_params['example_size'] + [
num_channels,
]
},
'labels': {
'y': reader_params['example_size']
}
}
# module_name = 'contributions.applications.AL_framework.applications.app' + str(app_json['id']) + '.readers.'
#
# if app_json['reader_type'] == "Patch":
# module_name = module_name + 'patch_reader'
# elif app_json['reader_type'] == "Slice":
# module_name = module_name + 'slice_reader'
# elif app_json['reader_type'] == "Stack":
# module_name = module_name + 'stack_reader'
# else:
# print("Unsupported reader type: please specify a new one")
# return
# mod = import_module(module_name)
mod = import_module('readers.tune_reader')
read_fn = vars(mod)['read_fn']
reader = Reader(read_fn, {
'features': {
'x': tf.float32
},
'labels': {
'y': tf.int32
}
})
# Get input functions and queue initialisation hooks for training and
# validation data
batch_size = app_json['batch_size']
train_input_fn, train_qinit_hook = reader.get_inputs(
file_references=patch_filenames,
mode=tf.estimator.ModeKeys.TRAIN,
example_shapes=reader_example_shapes,
batch_size=batch_size,
shuffle_cache_size=SHUFFLE_CACHE_SIZE,
params=reader_params)
val_input_fn, val_qinit_hook = reader.get_inputs(
file_references=val_filenames,
mode=tf.estimator.ModeKeys.EVAL,
example_shapes=reader_example_shapes,
batch_size=batch_size,
shuffle_cache_size=SHUFFLE_CACHE_SIZE,
params=reader_params)
# Instantiate the neural network estimator
nn = tf.estimator.Estimator(model_fn=model_fn,
model_dir=args.model_path,
params={"learning_rate": 0.001},
config=tf.estimator.RunConfig())
# Hooks for validation summaries
val_summary_hook = tf.contrib.training.SummaryAtEndHook(
os.path.join(args.model_path, 'eval'))
step_cnt_hook = tf.train.StepCounterHook(every_n_steps=EVAL_EVERY_N_STEPS,
output_dir=args.model_path)
print('Starting tuning...')
max_steps = app_json['max_steps']
try:
for _ in range(max_steps // EVAL_EVERY_N_STEPS):
nn.train(input_fn=train_input_fn,
hooks=[train_qinit_hook, step_cnt_hook],
steps=EVAL_EVERY_N_STEPS)
if args.run_validation:
results_val = nn.evaluate(
input_fn=val_input_fn,
hooks=[val_qinit_hook, val_summary_hook],
steps=EVAL_STEPS)
print('Step = {}; val loss = {:.5f};'.format(
results_val['global_step'], results_val['loss']))
except KeyboardInterrupt:
pass
print('Stopping now.')
export_dir = nn.export_savedmodel(
export_dir_base=args.model_path,
serving_input_receiver_fn=reader.serving_input_receiver_fn(
reader_example_shapes))
print('Model saved to {}.'.format(export_dir))
app_json['model_status'] = 2
write_app_config(app_json)
print('Updated model status in model config')
def train(args):
np.random.seed(42)
tf.set_random_seed(42)
print('Setting up...')
# Parse csv files for file names
all_filenames = pd.read_csv(
args.data_csv,
dtype=object,
keep_default_na=False,
na_values=[]).as_matrix()
train_filenames = all_filenames[:100]
val_filenames = all_filenames[100:]
# Set up a data reader to handle the file i/o.
reader_params = {'n_examples': 8,
'example_size': [32, 128, 128],
'extract_examples': True}
reader_example_shapes = {'features': {'x': reader_params['example_size'] + [NUM_CHANNELS, ]}}
reader = Reader(read_fn, {'features': {'x': tf.float32}})
# Get input functions and queue initialisation hooks for training and
# validation data
train_input_fn, train_qinit_hook = reader.get_inputs(
file_references=train_filenames,
mode=tf.estimator.ModeKeys.TRAIN,
example_shapes=reader_example_shapes,
batch_size=BATCH_SIZE,
shuffle_cache_size=SHUFFLE_CACHE_SIZE,
params=reader_params)
val_input_fn, val_qinit_hook = reader.get_inputs(
file_references=val_filenames,
mode=tf.estimator.ModeKeys.EVAL,
example_shapes=reader_example_shapes,
batch_size=BATCH_SIZE,
shuffle_cache_size=SHUFFLE_CACHE_SIZE,
params=reader_params)
# Instantiate the neural network estimator
nn = tf.estimator.Estimator(
model_fn=model_fn,
model_dir=args.model_path,
params={'learning_rate': 0.01, 'upsampling_factor': UPSAMPLING_FACTOR},
config=tf.estimator.RunConfig())
# Hooks for validation summaries
val_summary_hook = tf.contrib.training.SummaryAtEndHook(
os.path.join(args.model_path, 'eval'))
step_cnt_hook = tf.train.StepCounterHook(
every_n_steps=EVAL_EVERY_N_STEPS,
output_dir=args.model_path)
print('Starting training...')
try:
for _ in range(MAX_STEPS // EVAL_EVERY_N_STEPS):
nn.train(
input_fn=train_input_fn,
hooks=[train_qinit_hook, step_cnt_hook],
steps=EVAL_EVERY_N_STEPS)
if args.run_validation:
results_val = nn.evaluate(
input_fn=val_input_fn,
hooks=[val_qinit_hook, val_summary_hook],
steps=EVAL_STEPS)
print('Step = {}; val loss = {:.5f};'.format(
results_val['global_step'], results_val['loss']))
except KeyboardInterrupt:
pass
print('Stopping now.')
export_dir = nn.export_savedmodel(
export_dir_base=args.model_path,
serving_input_receiver_fn=reader.serving_input_receiver_fn(reader_example_shapes))
print('Model saved to {}.'.format(export_dir))
def train(args):
np.random.seed(42)
tf.set_random_seed(42)
print('Setting up...')
# Parse csv files for file names
target_df = pd.read_csv('/data1/users/adoyle/IBIS/ibis_t1_qc.csv', dtype=object, keep_default_na=False)
all_filenames = target_df.iloc[:, 0].tolist()
all_labels = target_df.iloc[:, 1].tolist()
skf = StratifiedKFold(n_splits=10)
for train_filenames, val_filenames in skf.split(all_filenames, all_labels):
# Set up a data reader to handle the file i/o.
reader_params = {'n_examples': 2,
'example_size': [160, 256, 224],
'extract_examples': True}
reader_example_shapes = {'features': {'x': reader_params['example_size'] + [NUM_CHANNELS]},
'labels': {'y': [1]}}
reader = Reader(read_fn,
{'features': {'x': tf.float32},
'labels': {'y': tf.int32}})
# Get input functions and queue initialisation hooks for training and
# validation data
train_input_fn, train_qinit_hook = reader.get_inputs(
file_references=train_filenames,
mode=tf.estimator.ModeKeys.TRAIN,
example_shapes=reader_example_shapes,
batch_size=BATCH_SIZE,
shuffle_cache_size=SHUFFLE_CACHE_SIZE,
params=reader_params)
val_input_fn, val_qinit_hook = reader.get_inputs(
file_references=val_filenames,
mode=tf.estimator.ModeKeys.EVAL,
example_shapes=reader_example_shapes,
batch_size=BATCH_SIZE,
shuffle_cache_size=SHUFFLE_CACHE_SIZE,
params=reader_params)
# Instantiate the neural network estimator
nn = tf.estimator.Estimator(
model_fn=model_fn,
model_dir=args.model_path,
params={"learning_rate": 0.001},
config=tf.estimator.RunConfig())
# Hooks for validation summaries
val_summary_hook = tf.contrib.training.SummaryAtEndHook(
os.path.join(args.model_path, 'eval'))
step_cnt_hook = tf.train.StepCounterHook(every_n_steps=EVAL_EVERY_N_STEPS,
output_dir=args.model_path)
print('Starting training...')
try:
for _ in range(MAX_STEPS // EVAL_EVERY_N_STEPS):
nn.train(
input_fn=train_input_fn,
hooks=[train_qinit_hook, step_cnt_hook],
steps=EVAL_EVERY_N_STEPS)
if args.run_validation:
results_val = nn.evaluate(
input_fn=val_input_fn,
hooks=[val_qinit_hook, val_summary_hook],
steps=EVAL_STEPS)
print('Step = {}; val loss = {:.5f};'.format(
results_val['global_step'],
results_val['loss']))
except KeyboardInterrupt:
pass
# When exporting we set the expected input shape to be arbitrary.
export_dir = nn.export_savedmodel(
export_dir_base=args.model_path,
serving_input_receiver_fn=reader.serving_input_receiver_fn(
{'features': {'x': [None, None, None, NUM_CHANNELS]},
'labels': {'y': [1]}}))
print('Model saved to {}.'.format(export_dir))
def train(args):
np.random.seed(8)
tf.set_random_seed(8)
print('Setting Up...')
# Read Training-Fold.csv
train_filenames = pd.read_csv(args.train_csv,
dtype=object,
keep_default_na=False,
na_values=[]).values
# Read Validation-Fold.csv
val_filenames = pd.read_csv(args.val_csv,
dtype=object,
keep_default_na=False,
na_values=[]).values
# Set DLTK Reader Parameters (No. of Patches, Patch Size)
reader_params = {
'n_patches': NUM_PATCHES,
'patch_size': [PATCH_Z, PATCH_XY, PATCH_XY], # Target Patch Size
'extract_patches': True
} # Enable Training Mode Patch Extraction
# Set Patch Dimensions
reader_patch_shapes = {
'features': {
'x': reader_params['patch_size'] + [
NUM_CHANNELS,
]
},
'labels': {
'y': reader_params['patch_size']
}
}
# Initiate Data Reader + Patch Extraction
reader = Reader(read_fn, {
'features': {
'x': tf.float32
},
'labels': {
'y': tf.int32
}
})
# Create Input Functions + Queue Initialisation Hooks for Training/Validation Data
train_input_fn, train_qinit_hook = reader.get_inputs(
file_references=train_filenames,
mode=tf.estimator.ModeKeys.TRAIN,
example_shapes=reader_patch_shapes,
batch_size=BATCH_SIZE,
shuffle_cache_size=SHUFFLE_CACHE_SIZE,
prefetch_cache_size=PREFETCH_CACHE_SIZE,
params=reader_params)
val_input_fn, val_qinit_hook = reader.get_inputs(
file_references=val_filenames,
mode=tf.estimator.ModeKeys.EVAL,
example_shapes=reader_patch_shapes,
batch_size=BATCH_SIZE,
shuffle_cache_size=SHUFFLE_CACHE_SIZE,
prefetch_cache_size=PREFETCH_CACHE_SIZE,
params=reader_params)
# Instantiate Neural Network Estimator
nn = tf.estimator.Estimator(model_fn=model_fn,
model_dir=args.model_path,
config=tf.estimator.RunConfig())
# Hooks for Validation Summaries
val_summary_hook = tf.contrib.training.SummaryAtEndHook(
os.path.join(args.model_path, 'eval'))
step_cnt_hook = tf.train.StepCounterHook(every_n_steps=EVAL_EVERY_N_STEPS,
output_dir=args.model_path)
print('Begin Training...')
try:
for _ in range(MAX_STEPS // EVAL_EVERY_N_STEPS):
nn.train(input_fn=train_input_fn,
hooks=[train_qinit_hook, step_cnt_hook],
steps=EVAL_EVERY_N_STEPS)
if args.run_validation:
results_val = nn.evaluate(
input_fn=val_input_fn,
hooks=[val_qinit_hook, val_summary_hook],
steps=EVAL_STEPS)
EPOCH_DISPLAY = int(
int(results_val['global_step']) /
(TRAIN_SIZE / BATCH_SIZE))
print('Epoch = {}; Step = {} / ValLoss = {:.5f};'.format(
EPOCH_DISPLAY, results_val['global_step'],
results_val['loss']))
dim = args.model_path + 'Step{}ValLoss{:.5f}'.format(
results_val['global_step'], results_val['loss'])
export_dir = nn.export_savedmodel(
export_dir_base=dim,
serving_input_receiver_fn=reader.serving_input_receiver_fn(
reader_patch_shapes))
print('Model saved to {}.'.format(export_dir))
count_steps.append(results_val['global_step'])
count_loss.append(results_val['loss'])
except KeyboardInterrupt:
pass
# Arbitrary Input Shape during Export
export_dir = nn.export_savedmodel(
export_dir_base=args.model_path,
serving_input_receiver_fn=reader.serving_input_receiver_fn(
reader_patch_shapes))
print('Model saved to {}.'.format(export_dir))
step_Loss = pd.DataFrame(list(zip(count_steps, count_loss)),
columns=['steps', 'val_loss'])
step_Loss.to_csv("RU_Validation_Loss.csv", encoding='utf-8', index=False)
def train(args):
np.random.seed(42)
tf.set_random_seed(42)
print('Setting up...')
# Parse csv files for file names
all_filenames = pd.read_csv(args.data_csv,
dtype=object,
keep_default_na=False,
na_values=[]).as_matrix()
# 3300
train_filenames = all_filenames[:3300]
val_filenames = all_filenames[3300:]
# Set up a data reader to handle the file i/o.
reader_params = {
'n_examples': 1,
'example_size': RESIZE_SIZE,
'extract_examples': True
}
reader_example_shapes = {
'features': {
'x': reader_params['example_size'] + [NUM_CHANNELS]
},
'labels': {
'y': []
}
}
reader = Reader(read_fn, {
'features': {
'x': tf.float32
},
'labels': {
'y': tf.int32
}
})
# Get input functions and queue initialisation hooks for training and
# validation data
train_input_fn, train_qinit_hook = reader.get_inputs(
file_references=train_filenames,
mode=tf.estimator.ModeKeys.TRAIN,
example_shapes=reader_example_shapes,
batch_size=BATCH_SIZE,
shuffle_cache_size=SHUFFLE_CACHE_SIZE,
params=reader_params)
val_input_fn, val_qinit_hook = reader.get_inputs(
file_references=val_filenames,
mode=tf.estimator.ModeKeys.EVAL,
example_shapes=reader_example_shapes,
batch_size=BATCH_SIZE,
shuffle_cache_size=SHUFFLE_CACHE_SIZE,
params=reader_params)
# Instantiate the neural network estimator
config = tf.ConfigProto()
config.gpu_options.allow_growth = True
nn = tf.estimator.Estimator(
model_fn=model_fn,
model_dir=args.model_path,
params={"learning_rate": 1e-3},
config=tf.estimator.RunConfig(session_config=config))
# Hooks for validation summaries
val_summary_hook = tf.contrib.training.SummaryAtEndHook(
os.path.join(args.model_path, 'eval'))
train_summary_hook = tf.contrib.training.SummaryAtEndHook(
os.path.join(args.model_path, 'train'))
# train_summary_hook = tf.train.SummarySaverHook(save_steps=1,
# output_dir=os.path.join(args.model_path, 'train'),
# summary_op=tf.summary.)
step_cnt_hook = tf.train.StepCounterHook(every_n_steps=EVAL_EVERY_N_STEPS,
output_dir=args.model_path)
print('Starting training...')
try:
for _ in tqdm(range(MAX_STEPS // EVAL_EVERY_N_STEPS)):
nn.train(
input_fn=train_input_fn,
hooks=[train_qinit_hook, step_cnt_hook, train_summary_hook],
steps=EVAL_EVERY_N_STEPS)
if args.run_validation:
results_val = nn.evaluate(
input_fn=val_input_fn,
hooks=[val_qinit_hook, val_summary_hook],
steps=EVAL_STEPS)
print('Step = {}; val loss = {:.5f}; val acc = {:.5f}'.format(
results_val['global_step'], results_val['loss'],
results_val['accuracy_val']))
except KeyboardInterrupt:
pass
# When exporting we set the expected input shape to be arbitrary.
export_dir = nn.export_savedmodel(
export_dir_base=args.model_path,
serving_input_receiver_fn=reader.serving_input_receiver_fn({
'features': {
'x': [None, None, None, NUM_CHANNELS]
},
'labels': {
'y': [1]
}
}))
print('Model saved to {}.'.format(export_dir))
def train(args):
np.random.seed(42)
tf.set_random_seed(42)
print('Setting up...')
# Parse csv files for file names
all_filenames = pd.read_csv(
args.data_csv,
dtype=object,
keep_default_na=False,
na_values=[]).as_matrix()
train_filenames = all_filenames
# Set up a data reader to handle the file i/o.
reader_params = {'n_examples': 10,
'example_size': [4, 224, 224],
'extract_examples': True}
reader_example_shapes = {'labels': [4, 64, 64, 1],
'features': {'noise': [1, 1, 1, 100]}}
reader = Reader(read_fn, {'features': {'noise': tf.float32},
'labels': tf.float32})
# Get input functions and queue initialisation hooks for data
train_input_fn, train_qinit_hook = reader.get_inputs(
file_references=train_filenames,
mode=tf.estimator.ModeKeys.TRAIN,
example_shapes=reader_example_shapes,
batch_size=BATCH_SIZE,
params=reader_params)
# See TFGAN's `train.py` for a description of the generator and
# discriminator API.
def generator_fn(generator_inputs):
"""Generator function to build fake data samples. It creates a network
given input features (e.g. from a dltk.io.abstract_reader). Further,
custom Tensorboard summary ops can be added. For additional
information, please refer to https://www.tensorflow.org/versions/master/api_docs/python/tf/contrib/gan/estimator/GANEstimator.
Args:
generator_inputs (tf.Tensor): Noise input to generate samples from.
Returns:
tf.Tensor: Generated data samples
"""
gen = dcgan_generator_3d(
inputs=generator_inputs['noise'],
mode=tf.estimator.ModeKeys.TRAIN)
gen = gen['gen']
gen = tf.nn.tanh(gen)
return gen
def discriminator_fn(data, conditioning):
"""Discriminator function to discriminate real and fake data. It creates
a network given input features (e.g. from a dltk.io.abstract_reader).
Further, custom Tensorboard summary ops can be added. For additional
information, please refer to https://www.tensorflow.org/versions/master/api_docs/python/tf/contrib/gan/estimator/GANEstimator.
Args:
generator_inputs (tf.Tensor): Noise input to generate samples from.
Returns:
tf.Tensor: Generated data samples
"""
tf.summary.image('data', data[:, 0])
disc = dcgan_discriminator_3d(
inputs=data,
mode=tf.estimator.ModeKeys.TRAIN)
return disc['logits']
# get input tensors from queue
features, labels = train_input_fn()
# build generator
with tf.variable_scope('generator'):
gen = generator_fn(features)
# build discriminator on fake data
with tf.variable_scope('discriminator'):
disc_fake = discriminator_fn(gen, None)
# build discriminator on real data, reusing the previously created variables
with tf.variable_scope('discriminator', reuse=True):
disc_real = discriminator_fn(labels, None)
# building an LSGAN loss for the real examples
d_loss_real = tf.losses.mean_squared_error(
disc_real, tf.ones_like(disc_real))
# calculating a pseudo accuracy for the discriminator detecting a real
# sample and logging that
d_pred_real = tf.cast(tf.greater(disc_real, 0.5), tf.float32)
_, d_acc_real = tf.metrics.accuracy(tf.ones_like(disc_real), d_pred_real)
tf.summary.scalar('disc/real_acc', d_acc_real)
# building an LSGAN loss for the fake examples
d_loss_fake = tf.losses.mean_squared_error(
disc_fake, tf.zeros_like(disc_fake))
# calculating a pseudo accuracy for the discriminator detecting a fake
# sample and logging that
d_pred_fake = tf.cast(tf.greater(disc_fake, 0.5), tf.float32)
_, d_acc_fake = tf.metrics.accuracy(tf.zeros_like(disc_fake), d_pred_fake)
tf.summary.scalar('disc/fake_acc', d_acc_fake)
# building an LSGAN loss for the generator
g_loss = tf.losses.mean_squared_error(
disc_fake, tf.ones_like(disc_fake))
tf.summary.scalar('loss/gen', g_loss)
# combining the discriminator losses
d_loss = d_loss_fake + d_loss_real
tf.summary.scalar('loss/disc', d_loss)
# getting the list of discriminator variables
d_vars = tf.get_collection(tf.GraphKeys.TRAINABLE_VARIABLES,
'discriminator')
# building the discriminator optimizer
d_opt = tf.train.AdamOptimizer(
0.001, 0.5, epsilon=1e-5).minimize(d_loss, var_list=d_vars)
# getting the list of generator variables
g_vars = tf.get_collection(tf.GraphKeys.TRAINABLE_VARIABLES,
'generator')
# building the generator optimizer
g_opt = tf.train.AdamOptimizer(
0.001, 0.5, epsilon=1e-5).minimize(g_loss, var_list=g_vars)
# getting a variable to hold the global step
global_step = tf.train.get_or_create_global_step()
# build op to increment the global step - important for TensorBoard logging
inc_step = global_step.assign_add(1)
# build the training session.
# NOTE: we are not using a tf.estimator here, because they prevent some
# flexibility in the training procedure
s = tf.train.MonitoredTrainingSession(checkpoint_dir=args.model_path,
save_summaries_steps=100,
save_summaries_secs=None,
hooks=[train_qinit_hook])
# build dummy logging string
log = 'Step {} with Loss D: {}, Loss G: {}, Acc Real: {} Acc Fake: {}'
# start training
print('Starting training...')
loss_d = 0
loss_g = 0
try:
for step in range(MAX_STEPS):
# if discriminator is too good, only train generator
if not loss_g > 3 * loss_d:
s.run(d_opt)
# if generator is too good, only train discriminator
if not loss_d > 3 * loss_g:
s.run(g_opt)
# increment global step for logging hooks
s.run(inc_step)
# get statistics for training scheduling
loss_d, loss_g, acc_d, acc_g = s.run(
[d_loss, g_loss, d_acc_real, d_acc_fake])
# print stats for information
if step % SAVE_SUMMARY_STEPS == 0:
print(log.format(step, loss_d, loss_g, acc_d, acc_g))
except KeyboardInterrupt:
pass
print('Stopping now.')
def train(args):
np.random.seed(42)
tf.set_random_seed(42)
print('Setting up...')
# Parse csv files for file names
all_filenames = pd.read_csv(args.data_csv,
dtype=object,
keep_default_na=False,
na_values=[]).as_matrix()
train_filenames = all_filenames
# Set up a data reader to handle the file i/o.
reader_params = {
'n_examples': 10,
'example_size': [4, 224, 224],
'extract_examples': True
}
reader_example_shapes = {
'labels': [4, 64, 64, 1],
'features': {
'noise': [1, 1, 1, 100]
}
}
reader = Reader(read_fn, {
'features': {
'noise': tf.float32
},
'labels': tf.float32
})
# Get input functions and queue initialisation hooks for data
train_input_fn, train_qinit_hook = reader.get_inputs(
file_references=train_filenames,
mode=tf.estimator.ModeKeys.TRAIN,
example_shapes=reader_example_shapes,
batch_size=BATCH_SIZE,
params=reader_params)
# See TFGAN's `train.py` for a description of the generator and
# discriminator API.
def generator_fn(generator_inputs):
"""Generator function to build fake data samples. It creates a network
given input features (e.g. from a dltk.io.abstract_reader). Further,
custom Tensorboard summary ops can be added. For additional
information, please refer to https://www.tensorflow.org/versions/master/api_docs/python/tf/contrib/gan/estimator/GANEstimator.
Args:
generator_inputs (tf.Tensor): Noise input to generate samples from.
Returns:
tf.Tensor: Generated data samples
"""
gen = dcgan_generator_3d(inputs=generator_inputs['noise'],
mode=tf.estimator.ModeKeys.TRAIN)
gen = gen['gen']
gen = tf.nn.tanh(gen)
return gen
def discriminator_fn(data, conditioning):
"""Discriminator function to discriminate real and fake data. It creates
a network given input features (e.g. from a dltk.io.abstract_reader).
Further, custom Tensorboard summary ops can be added. For additional
information, please refer to https://www.tensorflow.org/versions/master/api_docs/python/tf/contrib/gan/estimator/GANEstimator.
Args:
generator_inputs (tf.Tensor): Noise input to generate samples from.
Returns:
tf.Tensor: Generated data samples
"""
tf.summary.image('data', data[:, 0])
disc = dcgan_discriminator_3d(inputs=data,
mode=tf.estimator.ModeKeys.TRAIN)
return disc['logits']
# get input tensors from queue
features, labels = train_input_fn()
# build generator
with tf.variable_scope('generator'):
gen = generator_fn(features)
# build discriminator on fake data
with tf.variable_scope('discriminator'):
disc_fake = discriminator_fn(gen, None)
# build discriminator on real data, reusing the previously created variables
with tf.variable_scope('discriminator', reuse=True):
disc_real = discriminator_fn(labels, None)
# building an LSGAN loss for the real examples
d_loss_real = tf.losses.mean_squared_error(disc_real,
tf.ones_like(disc_real))
# calculating a pseudo accuracy for the discriminator detecting a real
# sample and logging that
d_pred_real = tf.cast(tf.greater(disc_real, 0.5), tf.float32)
_, d_acc_real = tf.metrics.accuracy(tf.ones_like(disc_real), d_pred_real)
tf.summary.scalar('disc/real_acc', d_acc_real)
# building an LSGAN loss for the fake examples
d_loss_fake = tf.losses.mean_squared_error(disc_fake,
tf.zeros_like(disc_fake))
# calculating a pseudo accuracy for the discriminator detecting a fake
# sample and logging that
d_pred_fake = tf.cast(tf.greater(disc_fake, 0.5), tf.float32)
_, d_acc_fake = tf.metrics.accuracy(tf.zeros_like(disc_fake), d_pred_fake)
tf.summary.scalar('disc/fake_acc', d_acc_fake)
# building an LSGAN loss for the generator
g_loss = tf.losses.mean_squared_error(disc_fake, tf.ones_like(disc_fake))
tf.summary.scalar('loss/gen', g_loss)
# combining the discriminator losses
d_loss = d_loss_fake + d_loss_real
tf.summary.scalar('loss/disc', d_loss)
# getting the list of discriminator variables
d_vars = tf.get_collection(tf.GraphKeys.TRAINABLE_VARIABLES,
'discriminator')
# building the discriminator optimizer
d_opt = tf.train.AdamOptimizer(0.001, 0.5,
epsilon=1e-5).minimize(d_loss,
var_list=d_vars)
# getting the list of generator variables
g_vars = tf.get_collection(tf.GraphKeys.TRAINABLE_VARIABLES, 'generator')
# building the generator optimizer
g_opt = tf.train.AdamOptimizer(0.001, 0.5,
epsilon=1e-5).minimize(g_loss,
var_list=g_vars)
# getting a variable to hold the global step
global_step = tf.train.get_or_create_global_step()
# build op to increment the global step - important for TensorBoard logging
inc_step = global_step.assign_add(1)
# build the training session.
# NOTE: we are not using a tf.estimator here, because they prevent some
# flexibility in the training procedure
s = tf.train.MonitoredTrainingSession(checkpoint_dir=args.model_path,
save_summaries_steps=100,
save_summaries_secs=None,
hooks=[train_qinit_hook])
# build dummy logging string
log = 'Step {} with Loss D: {}, Loss G: {}, Acc Real: {} Acc Fake: {}'
# start training
print('Starting training...')
loss_d = 0
loss_g = 0
try:
for step in range(MAX_STEPS):
# if discriminator is too good, only train generator
if not loss_g > 3 * loss_d:
s.run(d_opt)
# if generator is too good, only train discriminator
if not loss_d > 3 * loss_g:
s.run(g_opt)
# increment global step for logging hooks
s.run(inc_step)
# get statistics for training scheduling
loss_d, loss_g, acc_d, acc_g = s.run(
[d_loss, g_loss, d_acc_real, d_acc_fake])
# print stats for information
if step % SAVE_SUMMARY_STEPS == 0:
print(log.format(step, loss_d, loss_g, acc_d, acc_g))
except KeyboardInterrupt:
pass
print('Stopping now.')
def predict(args):
# Read in the csv with the file names you would want to predict on
file_names = pd.read_csv(args.csv,
dtype=object,
keep_default_na=False,
na_values=[]).as_matrix()
# We trained on the first 4 subjects, so we predict on the rest
# file_names = file_names[-N_VALIDATION_SUBJECTS:]
config = tf.ConfigProto()
config.gpu_options.allow_growth = True
nn = tf.estimator.Estimator(
model_fn=model_fn,
model_dir=args.model_path,
params={"learning_rate": 1e-3},
config=tf.estimator.RunConfig(session_config=config))
reader_params = {
'n_examples': 1,
'example_size': RESIZE_SIZE,
'extract_examples': True
}
reader_example_shapes = {
'features': {
'x': reader_params['example_size'] + [1]
},
'labels': {
'y': []
}
}
reader = Reader(read_fn, {
'features': {
'x': tf.float32
},
'labels': {
'y': tf.int32
}
})
input_fn, qinit_hook = reader.get_inputs(
file_references=file_names,
mode=tf.estimator.ModeKeys.EVAL,
example_shapes=reader_example_shapes,
batch_size=1,
shuffle_cache_size=1,
params=reader_params)
features_to_save = []
labels_to_save = []
for i in tqdm(nn.predict(input_fn, hooks=[qinit_hook])):
features_to_save.append(i['features'])
labels_to_save.append(i['y_'])
np.savetxt('embeds.csv',
np.array(features_to_save),
delimiter=',',
newline='\n')
np.savetxt('labels.csv',
np.array(labels_to_save),
delimiter=',',
newline='\n')
def train(args):
np.random.seed(42)
tf.set_random_seed(42)
print('Setting up...')
# Parse csv files for file names
train_df = pd.read_csv(args.train_csv)
val_df = pd.read_csv(args.val_csv)
# Set up a data reader to handle the file i/o.
reader_params = {
'n_examples': 2,
'example_size': [121, 145, 121],
'extract_examples': False
}
reader_example_shapes = {
'features': {
'x': reader_params['example_size'] + [NUM_CHANNELS]
},
'labels': {
'y': [1]
}
}
reader = Reader(read_fn, {
'features': {
'x': tf.float32
},
'labels': {
'y': tf.float32
}
})
# Get input functions and queue initialisation hooks for training and
# validation data
train_input_fn, train_qinit_hook = reader.get_inputs(
file_references=train_df,
mode=tf.estimator.ModeKeys.TRAIN,
example_shapes=reader_example_shapes,
batch_size=BATCH_SIZE,
shuffle_cache_size=SHUFFLE_CACHE_SIZE,
params=reader_params)
val_input_fn, val_qinit_hook = reader.get_inputs(
file_references=val_df,
mode=tf.estimator.ModeKeys.EVAL,
example_shapes=reader_example_shapes,
batch_size=BATCH_SIZE,
shuffle_cache_size=SHUFFLE_CACHE_SIZE,
params=reader_params)
# Instantiate the neural network estimator
nn = tf.estimator.Estimator(model_fn=model_fn,
model_dir=args.model_path,
params={"learning_rate": 0.001},
config=tf.estimator.RunConfig())
# Hooks for validation summaries
val_summary_hook = tf.contrib.training.SummaryAtEndHook(
os.path.join(args.model_path, 'eval'))
step_cnt_hook = tf.train.StepCounterHook(every_n_steps=EVAL_EVERY_N_STEPS,
output_dir=args.model_path)
print('Starting training...')
best_model_path = os.path.join(args.model_path, 'best')
best_val_loss = None
try:
for _ in range(MAX_STEPS // EVAL_EVERY_N_STEPS):
print('Training epoch {}/{}'.format(
_, MAX_STEPS // EVAL_EVERY_N_STEPS))
nn.train(input_fn=train_input_fn,
hooks=[train_qinit_hook, step_cnt_hook],
steps=EVAL_EVERY_N_STEPS)
if args.run_validation:
results_val = nn.evaluate(
input_fn=val_input_fn,
hooks=[val_qinit_hook, val_summary_hook],
steps=EVAL_STEPS)
print('Step = {}; val loss = {:.5f};'.format(
results_val['global_step'], results_val['loss']))
if best_val_loss is None or results_val['loss'] < best_val_loss:
os.system('rm -rf {}/{}'.format(best_model_path, '*'))
export_dir = nn.export_savedmodel(
export_dir_base=os.path.join(args.model_path, 'best'),
serving_input_receiver_fn=reader.
serving_input_receiver_fn({
'features': {
'x': [None, None, None, NUM_CHANNELS]
},
'labels': {
'y': [1]
}
}))
print('Best Model saved to {}.'.format(export_dir))
best_val_loss = results_val['loss']
except KeyboardInterrupt:
pass
# When exporting we set the expected input shape to be arbitrary.
export_dir = nn.export_savedmodel(
export_dir_base=args.model_path,
serving_input_receiver_fn=reader.serving_input_receiver_fn({
'features': {
'x': [None, None, None, NUM_CHANNELS]
},
'labels': {
'y': [1]
}
}))
print('Model saved to {}.'.format(export_dir))