def build_model(self):
# declaring the placeholders for our extracted image feature vectors, our caption, and our mask
# (describes how long our caption is with an array of 0/1 values of length `maxlen`
img = tf.placeholder(tf.float32, [self.batch_size, self.dim_in])
self.img=img
caption_placeholder = tf.placeholder(tf.int32, [self.batch_size, self.n_lstm_steps])
self.caption_placeholder=caption_placeholder
mask = tf.placeholder(tf.float32, [self.batch_size, self.n_lstm_steps])
self.mask=mask
self.output_placeholder = tf.placeholder(tf.int32, [self.batch_size, self.n_lstm_steps])
network_weights = self._initialize_weights()
# getting an initial LSTM embedding from our image_imbedding
image_embedding = tf.matmul(img, self.img_embedding) + self.img_embedding_bias
flat_caption_placeholder=tf.reshape(caption_placeholder,[-1])
#leverage one-hot sparsity to lookup embeddings fast
embedded_input,KLD_loss=self._get_word_embedding([network_weights['variational_encoding'],network_weights['biases_variational_encoding']],network_weights['input_meaning'],flat_caption_placeholder,logit=True)
KLD_loss=tf.multiply(KLD_loss,tf.reshape(mask,[-1,1]))
KLD_loss=tf.reduce_sum(KLD_loss)*0
word_embeddings=tf.matmul(embedded_input,self.word_embedding)+self.embedding_bias
word_embeddings=tf.reshape(word_embeddings,[self.batch_size,self.n_lstm_steps,-1])
#initialize lstm state
state = self.lstm.zero_state(self.batch_size, dtype=tf.float32)
rnn_output=[]
with tf.variable_scope("RNN"):
# unroll lstm
for i in range(self.n_lstm_steps):
if i > 0:
# if this isn’t the first iteration of our LSTM we need to get the word_embedding corresponding
# to the (i-1)th word in our caption
current_embedding = word_embeddings[:,i-1,:]
else:
#if this is the first iteration of our LSTM we utilize the embedded image as our input
current_embedding = image_embedding
if i > 0:
# allows us to reuse the LSTM tensor variable on each iteration
tf.get_variable_scope().reuse_variables()
out, state = self.lstm(current_embedding, state)
# if i>0:
rnn_output.append(tf.expand_dims(out,1))
#perform classification of output
rnn_output=tf.concat(rnn_output,axis=1)
rnn_output=tf.reshape(rnn_output,[self.batch_size*(self.n_lstm_steps),-1])
encoded_output=tf.matmul(rnn_output,self.word_encoding)+self.word_encoding_bias
encoded_output=tf.reshape(tf.square(encoded_output),[self.batch_size*self.n_lstm_steps,-1])[:,1:]
#get loss
# normed_embedding= tf.nn.l2_normalize(encoded_output, dim=-1)
# normed_target=tf.nn.l2_normalize(embedded_input,dim=-1)
# cos_sim=tf.multiply(normed_embedding,normed_target)[:,1:]
# cos_sim=(tf.reduce_sum(cos_sim,axis=-1))
# cos_sim=tf.reshape(cos_sim,[self.batch_size,-1])
# cos_sim=tf.reduce_sum(cos_sim[:,1:]*mask[:,1:])
# cos_sim=cos_sim/tf.reduce_sum(mask[:,1:])
# self.exp_loss=tf.reduce_sum((-cos_sim))
# # self.exp_loss=tf.reduce_sum(xentropy)/float(self.batch_size)
# total_loss = tf.reduce_sum(-(cos_sim))
# mse=tf.reduce_sum(tf.reshape(tf.square(encoded_output-embedded_input),[self.batch_size*self.n_lstm_steps,1]),axis=-1)[:,1:]*(mask[:,1:])
# mse=tf.reduce_sum(mse)/tf.reduce_sum(mask[:,1:])
with tf.variable_scope('D',reuse=True) as scope:
total_loss+=tf.reduce_mean(-tf.log(self.discriminator.discriminate(encoded_output,train=False)))
self.D2=self.discriminator.discriminate(tf.stop_gradients(encoded_output),train=True)
#average over timeseries length
# total_loss=tf.reduce_sum(masked_xentropy)/tf.reduce_sum(mask[:,1:])
self.print_loss=total_loss
total_loss+=KLD_loss/tf.reduce_sum(mask)
return total_loss, img, caption_placeholder, mask
def metastep_graph(inp):
meta_train_x, meta_train_y, meta_val_x, meta_val_y = inp
meta_train_loss_list = []
meta_val_loss_list = []
weights = self._weights
meta_train_output = self._contruct_forward(meta_train_x,
weights,
reuse=False,
norm=norm,
is_train=self._is_train)
# Meta train loss: Calculate gradient
meta_train_loss = self._loss_fn(meta_train_y, meta_train_output)
meta_train_loss = tf.reduce_mean(meta_train_loss)
meta_train_loss_list.append(meta_train_loss)
grads = dict(
zip(weights.keys(),
tf.gradients(meta_train_loss, list(weights.values()))))
new_weights = dict(
zip(weights.keys(), [
weights[key] - self._alpha * grads[key]
for key in weights.keys()
]))
if self._avoid_second_derivative:
new_weights = tf.stop_gradients(new_weights)
meta_val_output = self._contruct_forward(meta_val_x,
new_weights,
reuse=True,
norm=norm,
is_train=self._is_train)
# Meta val loss: Calculate loss (meta step)
meta_val_loss = self._loss_fn(meta_val_y, meta_val_output)
meta_val_loss = tf.reduce_mean(meta_val_loss)
meta_val_loss_list.append(meta_val_loss)
# If perform multiple updates
for _ in range(self._num_updates - 1):
meta_train_output = self._contruct_forward(
meta_train_x,
new_weights,
reuse=True,
norm=norm,
is_train=self._is_train)
meta_train_loss = self._loss_fn(meta_train_y,
meta_train_output)
meta_train_loss = tf.reduce_mean(meta_train_loss)
meta_train_loss_list.append(meta_train_loss)
grads = dict(
zip(
new_weights.keys(),
tf.gradients(meta_train_loss,
list(new_weights.values()))))
new_weights = dict(
zip(new_weights.keys(), [
new_weights[key] - self._alpha * grads[key]
for key in new_weights.keys()
]))
if self._avoid_second_derivative:
new_weights = tf.stop_gradients(new_weights)
meta_val_output = self._contruct_forward(
meta_val_x,
new_weights,
reuse=True,
norm=norm,
is_train=self._is_train)
meta_val_loss = self._loss_fn(meta_val_y, meta_val_output)
meta_val_loss = tf.reduce_mean(meta_val_loss)
meta_val_loss_list.append(meta_val_loss)
return [
meta_train_loss_list, meta_val_loss_list, meta_train_output,
meta_val_output
]
def get_standard_loss(inputs,
outputs,
is_chief=True,
verbose=False,
loss_base_name="net",
epsilon=1e-8,
num_cells=None,
**kwargs):
""" Compute the loss function for standard object detection (final stage).
Args:
inputs: A dictionnary of inputs
outputs: A dictionnary of outputs
is_chief: Adds additional summaries iff is_chief is True
verbose: verbosity level
base_name: Prefix for all the loss colelctions to be added in the graph
epsilon: for avoiding overflow error
num_cells: 2D array number of cells in teh grid, used to normalize the centers distance
Kwargs:
centers_localization_loss_weight: Weights for the localization losses of centers. defaults to 1
scales_localization_loss_weight: Weights for the localization losses of log scales. defaults to 1
confidence_loss_weight: Weights for the confidence loss. defaults to 5
noobj_confidence_loss_weight: Weights for the confidence loss (empty cells). defautls to 1
classification_loss_weight: weights for the counting loss. defaults to 1
"""
(centers_localization_loss_weight, scales_localization_loss_weight,
confidence_loss_weight, noobj_confidence_loss_weight) = get_defaults(kwargs, [
'centers_localization_loss_weight', 'scales_localization_loss_weight',
'confidence_loss_weight', 'noobj_confidence_loss_weight'], verbose=verbose)
assert num_cells is not None
# obj_i_mask: (batch, num_cells, num_cells, 1, num_gt), indicates presence of a box in a cell
obj_i_mask = inputs['obj_i_mask_bbs']
## Split coordinates
# pred_bbs: 4 * (batch, num_cells, num_cells, num_preds, 1)
# true_bbs: 4 * (batch, 1, 1, 1, num_gt)
with tf.name_scope('coordinates'):
pred_bbs = tf.split(outputs['bounding_boxes'], 4, axis=-1)
true_bbs = tf.split(tf.expand_dims(tf.expand_dims(
tf.transpose(inputs['bounding_boxes'], (0, 2, 1)), axis=1), axis=1), 4, axis=-2)
## Compute target value for the assignment reward and the confidences
# target_confs: (batch, num_cells, num_cells, num_preds, num_gt)
with tf.name_scope('compute_target_confidence'):
target_confs = get_iou(true_bbs, pred_bbs, epsilon=epsilon)
target_confs = tf.stop_gradient(target_confs)
# assignment_rewards: (batch, num_cells, num_cells, num_preds, num_gt)
with tf.name_scope('compute_assignment_reward'):
# TODO rename
assignment_rewards = tf.stop_gradient(target_confs)
## Create obj mask mapping ground-truth to predictors
# obj_ij_mask: (batch, num_cells, num_cells, num_preds, num_gt, 1)
with tf.name_scope('assign_predictors'):
best_reward = tf.reduce_max(assignment_rewards, axis=-2, keepdims=True)
obj_ij_mask = tf.to_float(tf.greater_equal(assignment_rewards, best_reward))
obj_ij_mask *= obj_i_mask
obj_ij_mask = tf.expand_dims(obj_ij_mask, axis=-1)
obj_ij_mask = tf.stop_gradient(obj_ij_mask)
## Localization loss
with tf.name_scope('localization_loss'):
# true_mins, true_maxs: (batch, num_cells, num_cells, num_preds, num_gt, 2)
true_mins = tf.stack(true_bbs[:2], axis=-1)
true_maxs = tf.stack(true_bbs[2:], axis=-1)
# centers
with tf.name_scope('xy_loss'):
centers_diffs = tf.expand_dims(outputs['shifted_centers'], axis=-2) - num_cells * (true_mins + true_maxs) / 2
centers_localization_loss = tf.losses.compute_weighted_loss(
centers_diffs**2,
weights=centers_localization_loss_weight * obj_ij_mask,
reduction=tf.losses.Reduction.SUM_BY_NONZERO_WEIGHTS)
# scales
with tf.name_scope('wh_loss'):
scales_diff = tf.expand_dims(outputs['log_scales'], axis=-2) - tf.log(tf.maximum(epsilon, true_maxs - true_mins))
scales_localization_loss = tf.losses.compute_weighted_loss(
scales_diff**2,
weights=scales_localization_loss_weight * obj_ij_mask,
reduction=tf.losses.Reduction.SUM_BY_NONZERO_WEIGHTS)
## Confidence loss
with tf.name_scope('conf_loss'):
# Best predictor in non-empty cells
with tf.name_scope('non_empty'):
# confs_diffs: (batch, num_cells, num_cells, num_preds, num_gt)
obj_mask = tf.squeeze(obj_ij_mask, axis=-1)
confs_diffs = target_confs - outputs["confidence_scores"]
confidence_loss_obj = tf.losses.compute_weighted_loss(
confs_diffs**2,
weights=confidence_loss_weight * obj_mask,
reduction=tf.losses.Reduction.SUM_BY_NONZERO_WEIGHTS)
# Predictors in empty cells
with tf.name_scope('empty'):
# noobj_mask: (batch, num_cells, num_cells, 1, 1)
noobj_mask = 1. - tf.minimum(1., tf.reduce_sum(obj_i_mask, axis=-1, keepdims=True))
confidence_loss_noobj = tf.losses.compute_weighted_loss(
outputs["confidence_scores"]**2,
weights=noobj_confidence_loss_weight * noobj_mask,
reduction=tf.losses.Reduction.SUM_BY_NONZERO_WEIGHTS)
## Classification loss
if 'classification_probs' in outputs:
assert 'class_labels' in inputs
with tf.name_scope('classification_loss'):
classification_loss_weight = get_defaults(kwargs, ['classification_loss_weight'], verbose=verbose)[0]
# labels: (batch, 1, 1, 1, num_gt, num_classes)
labels = inputs['class_labels'] # (batch, num_gt, num_classes)
labels = tf.expand_dims(labels, axis=1)
labels = tf.expand_dims(labels, axis=1)
labels = tf.stop_gradients(tf.to_float(tf.expand_dims(labels, axis=1)))
# logits: (batch, num_cells, num_cells, num_preds, 1, num_classes)
logits = outputs['classification_probs']
logits = tf.expand_dims(logits, axis=4)
# classification loss
class_diffs = labels - logits
classification_loss = tf.losses.compute_weighted_loss(
class_diffs**2,
weights=classification_loss_weight * obj_ij_mask,
reduction=tf.losses.Reduction.SUM_BY_NONZERO_WEIGHTS)
else:
classification_loss = 0.
## Add informative summaries
if is_chief:
is_assigned_predictor = tf.to_float(tf.reduce_sum(obj_ij_mask, axis=-2) > 0.)
outputs["target_bounding_boxes"] = outputs["bounding_boxes"] * is_assigned_predictor
return [('%s_centers_localization_loss' % loss_base_name, centers_localization_loss),
('%s_scales_localization_loss' % loss_base_name, scales_localization_loss),
('%s_confidence_obj_loss' % loss_base_name, confidence_loss_obj),
('%s_confidence_noobj_loss' % loss_base_name, confidence_loss_noobj),
('%s_classification_loss' % loss_base_name, classification_loss)]
def _build_graph(self, dim_input, dim_output, norm):
def model_summary():
model_vars = tf.trainable_variables()
slim.model_analyzer.analyze_vars(model_vars, print_info=True)
learning_x, learning_y, meta_x, meta_y = [
self._learning_x, self._learning_y, self._meta_x, self._meta_y
]
learning_loss_list = []
meta_loss_list = []
weights = self._weights
learning_output = self.construct_forward(learning_x,
weights,
reuse=False,
norm=norm,
is_train=self._is_train)
# Meta train loss: Calculate gradient
learning_loss = self._loss_fn(learning_y, learning_output)
learning_loss = tf.reduce_mean(learning_loss)
learning_loss_list.append(learning_loss)
grads = dict(
zip(weights.keys(),
tf.gradients(learning_loss, list(weights.values()))))
# learning rate
self.learning_rate_op = tf.maximum(
self._min_alpha,
tf.train.exponential_decay(self._alpha,
self.alpha_step,
self._alpha_decay_step,
self._alpha_decay_rate,
staircase=True))
self.learning_train_op = tf.train.AdamOptimizer(
self.learning_rate_op).minimize(learning_loss)
if self.dic_agent_conf['GRADIENT_CLIP']:
for key in grads.keys():
grads[key] = tf.clip_by_value(
grads[key], -1 * self.dic_agent_conf['CLIP_SIZE'],
self.dic_agent_conf['CLIP_SIZE'])
self._learning_grads = grads
new_weights = dict(
zip(weights.keys(), [
weights[key] - self.learning_rate_op * grads[key]
for key in weights.keys()
]))
if self._avoid_second_derivative:
new_weights = tf.stop_gradients(new_weights)
meta_output = self.construct_forward(meta_x,
new_weights,
reuse=True,
norm=norm,
is_train=self._is_train)
# Meta val loss: Calculate loss (meta step)
meta_loss = self._loss_fn(meta_y, meta_output)
meta_loss = tf.reduce_mean(meta_loss)
meta_loss_list.append(meta_loss)
# If perform multiple updates
for _ in range(self._num_updates - 1):
learning_output = self.construct_forward(learning_x,
new_weights,
reuse=True,
norm=norm,
is_train=self._is_train)
learning_loss = self._loss_fn(learning_y, learning_output)
learning_loss = tf.reduce_mean(learning_loss)
learning_loss_list.append(learning_loss)
grads = dict(
zip(new_weights.keys(),
tf.gradients(learning_loss, list(new_weights.values()))))
new_weights = dict(
zip(new_weights.keys(), [
new_weights[key] - self.learning_rate_op * grads[key]
for key in new_weights.keys()
]))
if self._avoid_second_derivative:
new_weights = tf.stop_gradients(new_weights)
meta_output = self.construct_forward(meta_x,
new_weights,
reuse=True,
norm=norm,
is_train=self._is_train)
meta_loss = self._loss_fn(meta_y, meta_output)
meta_loss = tf.reduce_mean(meta_loss)
meta_loss_list.append(meta_loss)
self._new_weights = new_weights
# output
self._learning_output = learning_output
self._meta_output = meta_output
# Loss
learning_loss = tf.reduce_mean(learning_loss_list[-1])
meta_loss = tf.reduce_mean(meta_loss_list[-1])
self._learning_loss = learning_loss
self._meta_loss = meta_loss
model_summary()
state, reward, done = tf_env_step(env, action)
state.set_shape(initial_state_shape)
next_value = critic(next_state)
return next_state, done, action_probs, value, next_value, rewards
def compute_loss(action_probs: tf.Tensor, value: tf.Tensor, gamma: float
next_value: tf.Tensor
, reward: tf.Tensor,
done: tf.Tensor) -> tf.Tensor:
advs = tf.stop_gradients(values - reward + tf.cast(gamma, tf.float32) *
next_value * (1 - done))
loss = -tf.math.reduce_sum(action_log_probs * advs)
return loss
huber_loss = tf.keras.losses.Huber(reduction=tf.keras.losses.Reduction.SUM)
class Policy(tf.keras.Model):
def __init__(self, n_actions: int):
super(Policy, self).__init__(name='Policy')
self.fc1 = Dense(6, activation='relu')
self.fc2 = Dense(n_actions)
def call(self, inputs: tf.Tensor):
x = self.fc1(inputs)
action_logits = self.fc2(x)
def compute_infinite_nature_loss(generated_rgbd, gt_rgbd, discriminator,
mu_logvar):
"""Computes loss between a generated RGBD sequence and the ground truth.
Lambda terms are the default values used during the original submission.
Args:
generated_rgbd: [B, T, H, W, 4] A batch of T-length RGBD sequences
produced by the generator. Ranges from (0, 1)
gt_rgbd: [B, T, H, W, 4] The ground truth sequence from a video.
Ranges from (0, 1)
discriminator: a discriminator which accepts an [B, H, W, D] tensor
and runs a discriminator on multiple scales and returns
a list of (features, logit) for each scale.
mu_logvar: ([B, 128], [B, 128]) A tuple of mu, log-variance features
parameterizing the Gaussian used to sample the variational noise.
Returns:
A dictionary of losses. total_loss refers to the final
loss used to optimize the generator and total_disc_loss refers to the
loss used by the discriminator.
"""
_, _, height, width, _ = tf.shape(generated_rgbd).as_list()
gen_flatten = tf.reshape(generated_rgbd, [-1, height, width, 4])
gt_flatten = tf.reshape(gt_rgbd, [-1, height, width, 4])
# discriminator returns:
# [(scale_1_feats, scale_1_logits), (scale_2_feats, scale_2_logits), ...]
disc_on_generated = discriminator(gen_flatten)
generated_features = [f[0] for f in disc_on_generated]
generated_logits = [f[1] for f in disc_on_generated]
disc_on_real = discriminator(gt_flatten)
real_features = [f[0] for f in disc_on_real]
real_logits = [f[1] for f in disc_on_real]
disc_loss, _, _ = compute_discriminator_loss(
real_logits, tf.stop_gradients(generated_logits))
fool_d_loss = compute_adversarial_loss(generated_logits)
feature_matching_loss = compute_feature_matching(real_features,
generated_features)
kld_loss = compute_kld_loss(mu_logvar[0], mu_logvar[1])
rgbd_loss = tf.reduce_mean(tf.abs(generated_rgbd - gt_rgbd))
perceptual_loss = compute_perceptual_loss(generated_rgbd * 255.,
gt_rgbd * 255.)
loss_dict = {}
loss_dict["disc_loss"] = disc_loss
loss_dict["adversarial_loss"] = fool_d_loss
loss_dict["feature_matching_loss"] = feature_matching_loss
loss_dict["kld_loss"] = kld_loss
loss_dict["perceptual_loss"] = perceptual_loss
loss_dict["reconstruction_loss"] = rgbd_loss
total_loss = (1e-2 * perceptual_loss + 10.0 * feature_matching_loss +
0.05 * kld_loss + 1.5 * fool_d_loss + 0.5 * rgbd_loss)
total_disc_loss = 1.5 * disc_loss
loss_dict["total_loss"] = total_loss
loss_dict["total_disc_loss"] = total_disc_loss
return loss_dict
def _build_model(self):
param_dict = self._params
stack = self.conv2D_layer(self._X,
filter_size=self._filters,
n_filters=48,
relu=self._relu,
elu=self._elu,
batch_norm=True,
layer_name='ConvLayer',
param_dict=param_dict)
combine_layers = []
layer_sizes = []
for i, depth in enumerate(self._layer_depths):
dense_block = self._dense_block(
stack,
depth=depth,
growth=self._growth,
param_dict=param_dict,
block_name='DenseBlock{}'.format(i),
filter_size=self._filters)
print(dense_block.shape)
layer_sizes.append(int(dense_block.shape[2]))
if self._short_residual:
bypass = self.conv2D_layer(stack,
filter_size=1,
n_filters=int(
dense_block.shape[-1]),
relu=False,
batch_norm=True,
dropout=False,
layer_name='Bypass{}'.format(i),
param_dict=param_dict,
strides=(1, 1),
padding='SAME')
combine = dense_block + bypass
else:
combine = tf.concat([stack, dense_block], axis=-1)
combine_layers.append(combine)
stack = self._transition_down(
combine,
param_dict=param_dict,
block_name='TransitionDown{}'.format(i))
print(stack.shape)
combine_layers = combine_layers[::-1]
layer_sizes = layer_sizes[::-1]
layer_depths = self._layer_depths[::-1]
stack = self._dense_block(stack,
depth=self._bottleneck,
growth=16,
param_dict=param_dict,
block_name='BottleNeck',
filter_size=self._filters)
print(stack.shape)
# volumes = self._regression_block(stack,
# param_dict=param_dict,
# n_reg=self._n_classes,
# block_name='RegeressionBlock')
for i, (combine, layer_size, depth) in enumerate(
zip(combine_layers, layer_sizes, layer_depths)):
padding = 'SAME' if 2 * int(
stack.shape[2]) == layer_size else 'VALID'
transition_up = self._transition_up(
stack,
output_shape=(layer_size, layer_size),
param_dict=param_dict,
batch_size=self._batch_size,
block_name='TransitionUp{}'.format(i),
padding=padding)
print(transition_up.shape)
if self._long_residual:
bypass = self.conv2D_layer(combine,
filter_size=1,
n_filters=int(
transition_up.shape[-1]),
relu=False,
batch_norm=True,
dropout=False,
layer_name='BypassUp{}'.format(i),
param_dict=param_dict,
strides=(1, 1),
padding='SAME')
combine_up = bypass + transition_up
else:
combine_up = tf.concat([combine, transition_up], axis=-1)
stack = self._dense_block(combine_up,
depth=depth,
growth=self._growth,
param_dict=param_dict,
block_name='DenseBlockUp{}'.format(i),
filter_size=self._filters)
print(stack.shape)
self._logits = self.conv2D_layer(stack,
filter_size=self._output_filters,
n_filters=self._n_classes,
dropout=True,
relu=self._relu,
elu=self._elu,
batch_norm=False,
layer_name='OutputConv',
param_dict=param_dict)
self._soft_prob = tf.nn.softmax(self._logits, dim=-1)
# self._dice_loss = self.soft_dice_loss(self._logits,
# self._Y,
# index=-1,
# weights=self._class_weights)
output_vis = self.visualize_tensor(self._soft_prob)
gt_vis = self.visualize_tensor(self._Y)
self._entropy_loss = self.weighted_softmax_cross_entropy_with_logits(
self._logits, self._Y, self._class_weights, axis=-1)
learning_rate = tf.train.exponential_decay(5e-4,
self._params['global_step'],
1000,
0.90,
staircase=True)
self._cost = self._entropy_loss
self._cost_summary = tf.summary.scalar('Cost', self._cost)
self._train_op = tf.train.AdamOptimizer(
learning_rate=learning_rate, beta1=0.99,
beta2=0.995).minimize(self._cost,
global_step=self._params['global_step'])
self._predictions = tf.stop_gradients(tf.argmax(self._logits, -1))
self._output_summary = tf.summary.image('Output', output_vis)
self._input_summary = tf.summary.image('Ground Truth', gt_vis)
self._image_summary = tf.summary.merge(
[self._output_summary, self._input_summary])
folder_dir = os.path.split(self._saved)[0]
folder_dir = os.path.split(folder_dir)[0]
self.log_dir = os.path.join(folder_dir, 'logs')
if os.path.exists(self.log_dir):
shutil.rmtree(self.log_dir)
os.makedirs(self.log_dir)
else:
os.makedirs(self.log_dir)
self._summary_writer = tf.summary.FileWriter(
self.log_dir, graph=tf.get_default_graph())
################
#Reinforcement Learning nodes
self.optimizer = tf.train.AdamOptimizer(learning_rate=learning_rate,
beta1=0.99,
beta2=0.995)
self.cross_entropy_loss = tf.nn.softmax_cross_entropy_with_logits(
logits=self._logits, labels=self._predictions)
self.pg_loss = tf.reduce_mean(self.cross_entropy_loss)
self.gradients = self.optimizer.compute_gradients(self.pg_loss)
self.discounted_rewards = self.stop_gradients(
self.soft_dice_loss(logits=self._logits, labels=self._Y) -
self._threshold)
for i, (grad, var) in enumerate(self.gradients):
if grad is not None:
self.gradients[i] = (grad * self.discounted_rewards, var)
self._reinforce_op = self.optimizer.apply_gradients(self.gradients)