def get_states_b(self):
"""
Iterates through time/ sequence to get all hidden state
"""
all_hidden_states, all_memory_states = self.get_states_f()
# Reversing the hidden and memory state to get the final hidden and
# memory state
last_hidden_states = tf.reverse(
all_hidden_states, [True, False, False])[0, :, :]
last_memory_states = tf.reverse(
all_memory_states, [True, False, False])[0, :, :]
# For backward pass using the last hidden and memory of the forward
# pass
initial_hidden = tf.pack([last_hidden_states, last_memory_states])
# Getting all hidden state throuh time
all_hidden_memory_states = tf.scan(self.Lstm_b,
self.processed_input_rev,
initializer=initial_hidden,
name='states')
# Now reversing the states to keep those in original order
all_hidden_states = tf.reverse(all_hidden_memory_states[
:, 0, :, :], [True, False, False])
all_memory_states = tf.reverse(all_hidden_memory_states[
:, 1, :, :], [True, False, False])
return all_hidden_states, all_memory_states
def cummax(x, reverse=False, name=None):
"""Compute the cumulative maximum of the tensor `x` along `axis`. This
operation is similar to the more classic `cumsum`. Only support 1D Tensor
for now.
Args:
x: A `Tensor`. Must be one of the following types: `float32`, `float64`,
`int64`, `int32`, `uint8`, `uint16`, `int16`, `int8`, `complex64`,
`complex128`, `qint8`, `quint8`, `qint32`, `half`.
axis: A `Tensor` of type `int32` (default: 0).
reverse: A `bool` (default: False).
name: A name for the operation (optional).
Returns:
A `Tensor`. Has the same type as `x`.
"""
with ops.name_scope(name, "Cummax", [x]) as name:
x = ops.convert_to_tensor(x, name="x")
# Not very optimal: should directly integrate reverse into tf.scan.
if reverse:
x = tf.reverse(x, axis=[0])
# 'Accumlating' maximum: ensure it is always increasing.
cmax = tf.scan(lambda a, y: tf.maximum(a, y), x,
initializer=None, parallel_iterations=1,
back_prop=False, swap_memory=False)
if reverse:
cmax = tf.reverse(cmax, axis=[0])
return cmax
def forward(self, x, x_mask=None, context_layer=None):
assert not (self.reverse_alternation and x_mask == None)
# assert (context_layer == None or
# tf.shape(context_layer)[-1] == self.context_state_size)
def create_step_fun(gru):
def step_fn(prev_state, x):
gates_x2d, proposal_x2d = x[0], x[1]
new_state = gru.forward(prev_state,
gates_x=gates_x2d,
proposal_x=proposal_x2d)
if len(x) > 2:
mask = x[2]
new_state *= mask # batch x 1
# first couple of states of reversed encoder should be zero
# this is why we need to multiply by mask
# this way, when the reversed encoder reaches actual words
# the state will be zeros and not some accumulated garbage
return new_state
return step_fn
init_state = tf.zeros(shape=[self.batch_size, self.state_size],
dtype=tf.float32)
if x_mask != None:
x_mask_r = tf.reverse(x_mask, axis=[0])
x_mask_bwd = tf.expand_dims(x_mask_r, axis=[2]) #seqLen x batch x 1
for i, gru in enumerate(self.grus):
layer = RecurrentLayer(initial_state=init_state,
step_fn=create_step_fun(gru))
if context_layer == None:
x2 = x
else:
x2 = tf.concat([x, context_layer], axis=-1)
if not self.alternating:
left_to_right = True
else:
if self.reverse_alternation:
left_to_right = (i % 2 == 1)
else:
left_to_right = (i % 2 == 0)
if left_to_right:
# Recurrent state flows from left to right in this layer.
gates_x, proposal_x = gru.precompute_from_x(x2)
h = layer.forward((gates_x, proposal_x))
else:
# Recurrent state flows from right to left in this layer.
x2_reversed = tf.reverse(x2, axis=[0])
gates_x, proposal_x = gru.precompute_from_x(x2_reversed)
h_reversed = layer.forward((gates_x, proposal_x, x_mask_bwd))
h = tf.reverse(h_reversed, axis=[0])
# Compute the word states, which will become the input for the
# next layer (or the output of the stack if we're at the top).
if i == 0:
x = h
else:
x += h # Residual connection
return x
def sparse_transpose(sp_input):
transposed_indices = tf.reverse(tf.cast(sp_input.indices, tf.int32), [False, True])
transposed_values = sp_input.values
transposed_shape = tf.reverse(tf.cast(sp_input.shape, tf.int32), [True])
sp_output = tf.SparseTensor(tf.cast(transposed_indices, tf.int64), transposed_values, tf.cast(transposed_shape, tf.int64))
sp_output = tf.sparse_reorder(sp_output)
return sp_output
def embed_sequences(self, embed_sequence_batch):
"""Return sentence embeddings as a tensor with with shape
[batch_size, hidden_size * 2]
"""
forward_values = embed_sequence_batch.values
forward_mask = embed_sequence_batch.mask
backward_values = tf.reverse(forward_values, [False, True, False])
backward_mask = tf.reverse(forward_mask, [False, True])
# Initialize LSTMs
self._forward_lstm = LSTM(self.hidden_size, return_sequences=True)
self._backward_lstm = LSTM(self.hidden_size, return_sequences=True)
# Pass input through the LSTMs
# Shape: (batch_size, seq_length, hidden_size)
forward_seq = self._forward_lstm(forward_values, forward_mask)
forward_seq.set_shape((None, self.seq_length, self.hidden_size))
backward_seq = self._backward_lstm(backward_values, backward_mask)
backward_seq.set_shape((None, self.seq_length, self.hidden_size))
# Stitch the outputs together --> hidden states (for computing attention)
# Final dimension: (batch_size, seq_length, hidden_size * 2)
lstm_states = tf.concat(2, [forward_seq, tf.reverse(backward_seq, [False, True, False])])
self._hidden_states = SequenceBatch(lstm_states, forward_mask)
# Stitch the final outputs together --> sequence embedding
# Final dimension: (batch_size, hidden_size * 2)
seq_length = tf.shape(forward_values)[1]
forward_final = tf.slice(forward_seq, [0, seq_length - 1, 0], [-1, 1, self.hidden_size])
backward_final = tf.slice(backward_seq, [0, seq_length - 1, 0], [-1, 1, self.hidden_size])
return tf.squeeze(tf.concat(2, [forward_final, backward_final]), [1])
def ndlstm_base_dynamic(inputs, noutput, scope=None, reverse=False):
"""Run an LSTM, either forward or backward.
This is a 1D LSTM implementation using dynamic_rnn and
the TensorFlow LSTM op.
Args:
inputs: input sequence (length, batch_size, ninput)
noutput: depth of output
scope: optional scope name
reverse: run LSTM in reverse
Returns:
Output sequence (length, batch_size, noutput)
"""
with tf.variable_scope(scope, "SeqLstm", [inputs]):
# TODO(tmb) make batch size, sequence_length dynamic
# example: sequence_length = tf.shape(inputs)[0]
_, batch_size, _ = _shape(inputs)
lstm_cell = tf.nn.rnn_cell.BasicLSTMCell(noutput, state_is_tuple=False)
state = tf.zeros([batch_size, lstm_cell.state_size])
sequence_length = int(inputs.get_shape()[0])
sequence_lengths = tf.to_int64(tf.fill([batch_size], sequence_length))
if reverse:
inputs = tf.reverse(inputs, [True, False, False])
outputs, _ = tf.nn.dynamic_rnn(lstm_cell,
inputs,
sequence_lengths,
state,
time_major=True)
if reverse:
outputs = tf.reverse(outputs, [True, False, False])
return outputs
def random_transformation2(x, y, padding, phase_train, rnd_vflip=True, rnd_hflip=True, rnd_transpose=True, rnd_colour=False):
"""
Perform random crop, flip, transpose, hue, saturation, brightness, contrast.
Args:
x: [B, H, W, 3]
y: [B, T, H, W]
padding: int
phase_train: bool
"""
# Random image transformation layers.
phase_train_f = tf.to_float(phase_train)
x_shape = tf.shape(x)
y_shape = tf.shape(y)
num_ex = x_shape[0]
inp_height = x_shape[1]
inp_width = x_shape[2]
inp_depth_x = x_shape[3]
inp_depth_y = y_shape[3]
# Add padding
x_pad = tf.pad(x, [[0, 0], [padding, padding], [padding, padding], [0, 0]])
y_pad = tf.pad(y, [[0, 0], [padding, padding], [padding, padding], [0, 0]])
# Random crop
offset = tf.random_uniform([2], dtype='int32', maxval=padding * 2)
x_rand = tf.slice(x_pad, tf.pack([0, offset[0], offset[1], 0]),
tf.pack([-1, inp_height, inp_width, inp_depth_x]))
y_rand = tf.slice(y_pad, tf.pack([0, offset[0], offset[1], 0]),
tf.pack([-1, inp_height, inp_width, inp_depth_y]))
# Center slices (for inference)
x_ctr = tf.slice(x_pad, [0, padding, padding, 0],
tf.pack([-1, inp_height, inp_width, -1]))
y_ctr = tf.slice(y_pad, [0, padding, padding, 0],
tf.pack([-1, inp_height, inp_width, -1]))
# Random horizontal & vertical flip & transpose
rand_h = tf.random_uniform([1], 1.0 - float(rnd_hflip), 1.0)
rand_v = tf.random_uniform([1], 1.0 - float(rnd_vflip), 1.0)
mirror = tf.pack([1.0, rand_v[0], rand_h[0], 1.0]) < 0.5
x_rand = tf.reverse(x_rand, mirror)
y_rand = tf.reverse(y_rand, mirror)
rand_t = tf.random_uniform([1], 1.0 - float(rnd_transpose), 1.0)
do_tr = tf.cast(rand_t[0] < 0.5, 'int32')
x_rand = tf.transpose(x_rand, tf.pack([0, 1 + do_tr, 2 - do_tr, 3]))
y_rand = tf.transpose(y_rand, tf.pack([0, 1 + do_tr, 2 - do_tr, 3]))
# Random hue, saturation, brightness, contrast
if rnd_colour:
x_rand = random_hue(x_rand, 0.1)
x_rand = random_saturation(x_rand, 0.9, 1.1)
x_rand = tf.image.random_brightness(x_rand, 0.1)
x_rand = tf.image.random_contrast(x_rand, 0.9, 1.1)
x = (1.0 - phase_train_f) * x_ctr + phase_train_f * x_rand
y = (1.0 - phase_train_f) * y_ctr + phase_train_f * y_rand
return x, y
def image_distortions(image, distortions):
distort_left_right_random = distortions[0]
mirror = tf.less(tf.pack([1.0, distort_left_right_random, 1.0]), 0.5)
image = tf.reverse(image, mirror)
distort_up_down_random = distortions[1]
mirror = tf.less(tf.pack([distort_up_down_random, 1.0, 1.0]), 0.5)
image = tf.reverse(image, mirror)
return image
def _compute_rnn_outputs(self):
reversed_inputs = tf.reverse(self.inputs, [False, True, False])
reversed_resets = tf.reverse(self.resets, [False, True, False])
self._rv_lstm = LSTM(reversed_inputs, reversed_resets, self.training,
self.num_layers, self.hidden_layer_size,
self.init_scale, self.dropout_keep_prob)
outputs = tf.reverse(self._rv_lstm.outputs, [False, True, False])
return outputs
def reconstruct(self, inputs, samples=1, sample_static=False,
sample_dynamic=False, swap_static=False, swap_dynamic=False,
fix_static=False, fix_dynamic=False):
"""Reconstruct the given input sequences.
Args:
inputs: A batch of image sequences `x_{1:T}` of shape
`[batch_size, timesteps, height, width, channels]`.
samples: Number of samples to draw from the latent distributions.
sample_static: Boolean for whether or not to randomly sample the
static latent variable `f` from its prior distribution.
sample_dynamic: Boolean for whether or not to randomly sample the
dynamic latent variable `z_{1:T}` from its prior distribution.
swap_static: Boolean for whether or not to swap the encodings for
the static latent variable `f` between the examples.
swap_dynamic: Boolean for whether or not to swap the encodings for
the dynamic latent variable `z_{1:T}` between the examples.
fix_static: Boolean for whether or not to share the same random
sample of the static latent variable `f` from its prior across
all examples.
fix_dynamic: Boolean for whether or not to share the same random
sample of the dynamic latent variable `z_{1:T}` from its prior
across all examples.
Returns:
A batched Independent distribution wrapping a set of Normal
distributions over the pixels of the reconstruction of the input,
where the Independent distribution has event shape [height, width,
channels], batch shape [samples, batch_size, timesteps], and
sample shape [sample_shape, samples, batch_size, timesteps,
height, width, channels].
"""
batch_size = tf.shape(inputs)[-5]
length = len(tf.unstack(inputs, axis=-4)) # hack for graph mode
features = self.compressor(inputs) # (..., batch, timesteps, hidden)
if sample_static:
static_sample, _ = self.sample_static_prior(
samples, batch_size, fix_static)
else:
static_sample, _ = self.sample_static_posterior(features, samples)
if swap_static:
static_sample = tf.reverse(static_sample, axis=[1])
if sample_dynamic:
dynamic_sample, _ = self.sample_dynamic_prior(
samples, batch_size, length, fix_dynamic)
else:
dynamic_sample, _ = self.sample_dynamic_posterior(
features, samples, static_sample)
if swap_dynamic:
dynamic_sample = tf.reverse(dynamic_sample, axis=[1])
likelihood = self.decoder((dynamic_sample, static_sample))
return likelihood
def __call__(self, cell_output, scores, scores_state, ignore_mask):
# apply exponential moving average with interpolation gate weight
# to scores from previous time which are equal to probs at this point
# different from original NTM where it is applied after softmax
i_g = self.inter_gate(cell_output)
# scores limited by time
scores = tf.concat([i_g * scores[:, :-1] + (1 - i_g) * scores_state,
scores[:, -1:]], 1)
next_scores_state = scores
# create probabilities for attention
if self.sparse_attention:
probs = tf.contrib.sparsemax.sparsemax(scores)
else:
probs = tf.nn.softmax(scores)
if self.shift_weight is not None:
s_w = self.shift_weight(cell_output)
# we want to go back in time during convolution
conv_probs = tf.reverse(probs, axis=[1])
# preare probs for tf.nn.depthwise_conv2d
# [in_width, in_channels=batch]
conv_probs = tf.transpose(conv_probs, [1, 0])
# [batch=1, in_height=1, in_width=time+1, in_channels=batch]
conv_probs = conv_probs[tf.newaxis, tf.newaxis, :, :]
# [filter_height=1, filter_width=2*attn_shift_range+1,
# in_channels=batch, channel_multiplier=1]
conv_s_w = tf.transpose(s_w, [1, 0])
conv_s_w = conv_s_w[tf.newaxis, :, :, tf.newaxis]
# perform 1d convolution
# [batch=1, out_height=1, out_width=time+1, out_channels=batch]
conv_probs = tf.nn.depthwise_conv2d_native(conv_probs, conv_s_w,
[1, 1, 1, 1], 'SAME')
conv_probs = conv_probs[0, 0, :, :]
conv_probs = tf.transpose(conv_probs, [1, 0])
probs = tf.reverse(conv_probs, axis=[1])
# Sharpening
g_sh = self.gamma_sharp(cell_output)
powed_probs = tf.pow(probs, g_sh)
probs = powed_probs / (
tf.reduce_sum(powed_probs, 1, keepdims=True) + 1e-32)
# set probs for no intents and action_listens to zero
if ignore_mask is not None:
probs = tf.concat([tf.where(ignore_mask,
tf.zeros_like(probs[:, :-1]),
probs[:, :-1]),
probs[:, -1:]], 1)
return probs, next_scores_state
def discounted_return(reward, length, discount):
"""Discounted Monte-Carlo returns."""
timestep = tf.range(reward.shape[1].value)
mask = tf.cast(timestep[None, :] < length[:, None], tf.float32)
return_ = tf.reverse(tf.transpose(tf.scan(
lambda agg, cur: cur + discount * agg,
tf.transpose(tf.reverse(mask * reward, [1]), [1, 0]),
tf.zeros_like(reward[:, -1]), 1, False), [1, 0]), [1])
return tf.check_numerics(tf.stop_gradient(return_), 'return')
def ni_slice(sub_values, last_ind, axis=0): # TODO: Allow both to be negative indexed... ndims = len(shape(sub_values)) im1 = 0 + abs(last_ind) i = [[None, None]] * ndims i[axis] = [im1, None] am = [False] * ndims am[axis] = True sl = [slice(*ii) for ii in i] ti = tf.reverse(sub_values, am)[sl] return tf.reverse(ti, am)
def lambda_advantage(reward, value, length, discount):
"""Generalized Advantage Estimation."""
timestep = tf.range(reward.shape[1].value)
mask = tf.cast(timestep[None, :] < length[:, None], tf.float32)
next_value = tf.concat([value[:, 1:], tf.zeros_like(value[:, -1:])], 1)
delta = reward + discount * next_value - value
advantage = tf.reverse(tf.transpose(tf.scan(
lambda agg, cur: cur + discount * agg,
tf.transpose(tf.reverse(mask * delta, [1]), [1, 0]),
tf.zeros_like(delta[:, -1]), 1, False), [1, 0]), [1])
return tf.check_numerics(tf.stop_gradient(advantage), 'advantage')
def lambda_return(reward, value, length, discount, lambda_):
"""TD-lambda returns."""
timestep = tf.range(reward.shape[1].value)
mask = tf.cast(timestep[None, :] < length[:, None], tf.float32)
sequence = mask * reward + discount * value * (1 - lambda_)
discount = mask * discount * lambda_
sequence = tf.stack([sequence, discount], 2)
return_ = tf.reverse(tf.transpose(tf.scan(
lambda agg, cur: cur[0] + cur[1] * agg,
tf.transpose(tf.reverse(sequence, [1]), [1, 2, 0]),
tf.zeros_like(value[:, -1]), 1, False), [1, 0]), [1])
return tf.check_numerics(tf.stop_gradient(return_), 'return')
def fix_variables(self, sess, pretrained_model):
print('Fix Resnet V1 layers..')
with tf.variable_scope('Fix_Resnet_V1') as scope:
with tf.device("/cpu:0"):
# fix RGB to BGR
conv1_rgb = tf.get_variable("conv1_rgb", [7, 7, 3, 64], trainable=False)
restorer_fc = tf.train.Saver({self._resnet_scope + "/conv1/weights": conv1_rgb})
restorer_fc.restore(sess, pretrained_model)
sess.run(tf.assign(self._variables_to_fix['rpn_network/'+self._resnet_scope + '/conv1/weights:0'],
tf.reverse(conv1_rgb, [2])))
sess.run(tf.assign(self._variables_to_fix['rfcn_network/'+self._resnet_scope + '/conv1/weights:0'],
tf.reverse(conv1_rgb, [2])))
def image_mirroring(img, label):
"""
Randomly mirrors the images.
Args:
img: Training image to mirror.
label: Segmentation mask to mirror.
"""
distort_left_right_random = tf.random_uniform([1], 0, 1.0, dtype=tf.float32)[0]
mirror = tf.less(tf.stack([1.0, distort_left_right_random, 1.0]), 0.5)
mirror = tf.boolean_mask([0, 1, 2], mirror)
img = tf.reverse(img, mirror)
label = tf.reverse(label, mirror)
return img, label
def process_reals(x, lod, mirror_augment, drange_data, drange_net):
with tf.name_scope('ProcessReals'):
with tf.name_scope('DynamicRange'):
x = tf.cast(x, tf.float32)
x = misc.adjust_dynamic_range(x, drange_data, drange_net)
if mirror_augment:
with tf.name_scope('MirrorAugment'):
s = tf.shape(x)
mask = tf.random_uniform([s[0], 1, 1, 1], 0.0, 1.0)
mask = tf.tile(mask, [1, s[1], s[2], s[3]])
x = tf.where(mask < 0.5, x, tf.reverse(x, axis=[3]))
with tf.name_scope('FadeLOD'): # Smooth crossfade between consecutive levels-of-detail.
s = tf.shape(x)
y = tf.reshape(x, [-1, s[1], s[2]//2, 2, s[3]//2, 2])
y = tf.reduce_mean(y, axis=[3, 5], keepdims=True)
y = tf.tile(y, [1, 1, 1, 2, 1, 2])
y = tf.reshape(y, [-1, s[1], s[2], s[3]])
x = tfutil.lerp(x, y, lod - tf.floor(lod))
with tf.name_scope('UpscaleLOD'): # Upscale to match the expected input/output size of the networks.
s = tf.shape(x)
factor = tf.cast(2 ** tf.floor(lod), tf.int32)
x = tf.reshape(x, [-1, s[1], s[2], 1, s[3], 1])
x = tf.tile(x, [1, 1, 1, factor, 1, factor])
x = tf.reshape(x, [-1, s[1], s[2] * factor, s[3] * factor])
return x
def fastrcnn_training(self, image,
rcnn_labels, fg_rcnn_boxes, gt_boxes_per_fg,
rcnn_label_logits, fg_rcnn_box_logits):
"""
Args:
image (NCHW):
rcnn_labels (n): labels for each sampled targets
fg_rcnn_boxes (fg x 4): proposal boxes for each sampled foreground targets
gt_boxes_per_fg (fg x 4): matching gt boxes for each sampled foreground targets
rcnn_label_logits (n): label logits for each sampled targets
fg_rcnn_box_logits (fg x 4): box logits for each sampled foreground targets
"""
with tf.name_scope('fg_sample_patch_viz'):
fg_sampled_patches = crop_and_resize(
image, fg_rcnn_boxes,
tf.zeros(tf.shape(fg_rcnn_boxes)[0], dtype=tf.int32), 300)
fg_sampled_patches = tf.transpose(fg_sampled_patches, [0, 2, 3, 1])
fg_sampled_patches = tf.reverse(fg_sampled_patches, axis=[-1]) # BGR->RGB
tf.summary.image('viz', fg_sampled_patches, max_outputs=30)
encoded_boxes = encode_bbox_target(
gt_boxes_per_fg, fg_rcnn_boxes) * tf.constant(config.FASTRCNN_BBOX_REG_WEIGHTS)
fastrcnn_label_loss, fastrcnn_box_loss = fastrcnn_losses(
rcnn_labels, rcnn_label_logits,
encoded_boxes,
fg_rcnn_box_logits)
return fastrcnn_label_loss, fastrcnn_box_loss
def final_hidden(self, input_idxes):
rnn_hiddens = self.hiddens(input_idxes)
# execute our rnn using scan function
# extract final timestep's hidden
rnn_hiddens_reverse = tf.reverse(rnn_hiddens, [True,False,False])
rnn_final_hidden = rnn_hiddens_reverse[0,:,:]
return rnn_final_hidden
def test_flip(self):
shape = [3, 4, 5]
data = np.random.rand(*shape)
var = llo.variable(data, 'var')
tf_var = tf.Variable(data)
sess = tf.Session()
sess.run(tf_var.initializer)
out = age.flip(var, 1)
tf_out = tf.reverse(tf_var, [1])
fout = out.evaluate(dtype=np.dtype(float))
tf_fout = sess.run(tf_out)
self._array_close(tf_fout, fout)
var2 = llo.variable(data, 'var2')
zero = llo.derive(out, var2)
ex = llo.derive(out, var)
rej = zero.evaluate()
der = ex.evaluate()
tf_grad = tf.gradients(tf_fout, [tf_var])[0]
self.assertEqual(None, tf_grad)
data0 = np.zeros(shape, dtype=float)
data1 = np.ones(shape, dtype=float)
self._array_eq(data0, rej)
self._array_eq(data1, der)
def compute_states(self):
states_list = []
state = self.zero_state()
timesteps = xrange(self.esn.n_steps)
if self.backwards:
timesteps = reversed(timesteps)
for i in timesteps:
prev_state = state
state = (tf.matmul(self.esn.scaled_x[:, i, :], self.w_xh) +
tf.matmul(prev_state, self.w_hh, b_is_sparse=True))
state = (1 - self.leakage) * tf.tanh(state) + self.leakage * prev_state
states_list.append(state)
states = tf.reshape(tf.concat(
1, states_list), [-1, self.esn.n_steps, self.esn.n_hidden])
if self.backwards:
states = tf.reverse(states, [False, True, False])
return states
def _conv_2d(X, h, strides=[1,1,1,1]):
"""
Perform 2d convolution in tensorflow.
X will to be manipulated to be of shape [batch, height, width, ch],
and h to be of shape [height, width, ch, num]. This function does the
necessary reshaping before calling the conv2d function, and does the
reshaping on the output, returning Y of shape [batch, height, width]
"""
# Check the shape of X is what we expect
if len(X.shape) != 3:
raise ValueError('X needs to be of shape [batch, height, width] ' +
'for conv_2d')
# Check the shape of h is what we expect
if len(h.shape) != 2:
raise ValueError('Filter inputs must only have height and width ' +
'for conv_2d')
# Add in the unit dimensions for conv
X = tf.expand_dims(X, axis=-1)
h = tf.expand_dims(tf.expand_dims(h, axis=-1),axis=-1)
# Have to reverse h as tensorflow 2d conv is actually cross-correlation
h = tf.reverse(h, axis=[0,1])
Y = tf.nn.conv2d(X, h, strides=strides, padding='VALID')
# Remove the final dimension, returning a result of shape
# [batch, height, width]
Y = tf.squeeze(Y, axis=-1)
return Y
def model(voxels, transform_matrix, params, is_training):
"""Model transforming the 3D voxels into 2D projections.
Args:
voxels: A tensor of size [batch, depth, height, width, channel]
representing the input of projection layer (tf.float32).
transform_matrix: A tensor of size [batch, 16] representing
the flattened 4-by-4 matrix for transformation (tf.float32).
params: Model parameters (dict).
is_training: Set to True if while training (boolean).
Returns:
A transformed tensor (tf.float32)
"""
del is_training # Doesn't make a difference for projector
# Rearrangement (batch, z, y, x, channel) --> (batch, y, z, x, channel).
# By the standard, projection happens along z-axis but the voxels
# are stored in a different way. So we need to switch the y and z
# axis for transformation operation.
voxels = tf.transpose(voxels, [0, 2, 1, 3, 4])
z_near = params.focal_length
z_far = params.focal_length + params.focal_range
transformed_voxels = perspective_transform.transformer(
voxels, transform_matrix, [params.vox_size] * 3, z_near, z_far)
views = tf.reduce_max(transformed_voxels, [1])
views = tf.reverse(views, [1])
return views
def bisine_wahwah_wave(frequency): """Emit two sine waves with balance oscillating left and right.""" # # This is clearly intended to build on the bisine wave defined above, # so we can start by generating that. waves_a = bisine_wave(frequency) # # Then, by reversing axis 2, we swap the stereo channels. By mixing # this with `waves_a`, we'll be able to create the desired effect. waves_b = tf.reverse(waves_a, axis=[2]) # # Let's have the balance oscillate from left to right four times. iterations = 4 # # Now, we compute the balance for each sample: `ts` has values # in [0, 1] that indicate how much we should use `waves_a`. xs = tf.reshape(tf.range(_samples(), dtype=tf.float32), [1, _samples(), 1]) thetas = xs / _samples() * iterations ts = (tf.sin(math.pi * 2 * thetas) + 1) / 2 # # Finally, we can mix the two together, and we're done. wave = ts * waves_a + (1.0 - ts) * waves_b # # Alternately, we can make the effect more pronounced by exaggerating # the sample data. Let's emit both variations. exaggerated_wave = wave ** 3.0 return tf.concat([wave, exaggerated_wave], axis=0)
def preprocess_for_test(image, gt_boxes, gt_masks):
ih, iw = tf.shape(image)[0], tf.shape(image)[1]
## min size resizing
new_ih, new_iw = preprocess_utils._smallest_size_at_least(ih, iw, cfg.FLAGS.image_min_size)
image = tf.expand_dims(image, 0)
image = tf.image.resize_bilinear(image, [new_ih, new_iw], align_corners=False)
image = tf.squeeze(image, axis=[0])
gt_masks = tf.expand_dims(gt_masks, -1)
gt_masks = tf.cast(gt_masks, tf.float32)
gt_masks = tf.image.resize_nearest_neighbor(gt_masks, [new_ih, new_iw], align_corners=False)
gt_masks = tf.cast(gt_masks, tf.int32)
gt_masks = tf.squeeze(gt_masks, axis=[-1])
scale_ratio = tf.to_float(new_ih) / tf.to_float(ih)
gt_boxes = preprocess_utils.resize_gt_boxes(gt_boxes, scale_ratio)
## zero mean image
image = tf.cast(image, tf.float32)
image = image / 256.0
image = (image - 0.5) * 2.0
image = tf.expand_dims(image, axis=0)
## rgb to bgr
image = tf.reverse(image, axis=[-1])
return image, gt_boxes, gt_masks
def get_test_iterator(src_dataset, src_vocab_table, batch_size, config):
src_eos_id = tf.cast(src_vocab_table.lookup(tf.constant(config.eos)), tf.int32)
src_dataset = src_dataset.map(lambda src: tf.string_split([src]).values)
src_dataset = src_dataset.map(lambda src: src[:config.src_max_len])
src_dataset = src_dataset.map(
lambda src: tf.cast(src_vocab_table.lookup(src), tf.int32))
if config.reverse_src:
src_dataset = src_dataset.map(lambda src: tf.reverse(src, axis=[0]))
src_dataset = src_dataset.map(lambda src: (src, tf.size(src)))
def batching_func(x):
return x.padded_batch(
config.batch_size,
padded_shapes=(tf.TensorShape([None]),
tf.TensorShape([])),
padding_values=(src_eos_id,
0))
batched_dataset = batching_func(src_dataset)
batched_iter = batched_dataset.make_initializable_iterator()
src_ids, src_seq_len = batched_iter.get_next()
return BatchedInput(
initializer=batched_iter.initializer,
source=src_ids,
target_input=None,
target_output=None,
source_sequence_length=src_seq_len,
target_sequence_length=None)
def forward_backward(self, obs_prob_seq):
"""
runs forward backward algorithm on observation sequence
Arguments
---------
- obs_seq : matrix of size N by S, where N is number of timesteps and
S is the number of states
Returns
-------
- forward : matrix of size N by S representing
the forward probability of each state at each time step
- backward : matrix of size N by S representing
the backward probability of each state at each time step
- posterior : matrix of size N by S representing
the posterior probability of each state at each time step
"""
obs_prob_list_for = tf.split(0, self.N, obs_prob_seq)
with tf.name_scope('forward_belief_propagation'):
# forward belief propagation
self._forward(obs_prob_list_for)
obs_prob_seq_rev = tf.reverse(obs_prob_seq, [True, False])
obs_prob_list_back = tf.split(0, self.N, obs_prob_seq_rev)
with tf.name_scope('backward_belief_propagation'):
# backward belief propagation
self._backward(obs_prob_list_back)
def _rnd_img_transformation(x, y_gt, padding, phase_train):
"""
Perform random crop, flip, transpose, hue, saturation, brightness, contrast.
"""
# Random image transformation layers.
phase_train_f = tf.to_float(phase_train)
x_shape = tf.shape(x)
num_ex = x_shape[0]
full_height = x_shape[1]
full_width = x_shape[2]
inp_height = full_height - 2 * padding
inp_width = full_width - 2 * padding
inp_depth = x_shape[3]
# Random crop
offset = tf.random_uniform([2], dtype='int32', maxval=padding * 2)
x_rand = tf.slice(x, tf.pack([0, offset[0], offset[1], 0]),
tf.pack([-1, inp_height, inp_width, inp_depth]))
y_rand = tf.slice(y_gt, tf.pack([0, 0, offset[0], offset[1]]),
tf.pack([-1, -1, inp_height, inp_width]))
# Center slices (for inference)
x_ctr = tf.slice(x, [0, padding, padding, 0],
tf.pack([-1, inp_height, inp_width, -1]))
y_ctr = tf.slice(y_gt, [0, 0, padding, padding],
tf.pack([-1, -1, inp_height, inp_width]))
# Random horizontal & vertical flip & transpose
rand = tf.random_uniform([3], 0, 1.0)
mirror_x = tf.pack([1.0, rand[0], rand[1], 1.0]) < 0.5
mirror_y = tf.pack([1.0, 1.0, rand[0], rand[1]]) < 0.5
x_rand = tf.reverse(x_rand, mirror_x)
y_rand = tf.reverse(y_rand, mirror_y)
do_tr = tf.cast(rand[2] > 0.5, 'int32')
x_rand = tf.transpose(x_rand, tf.pack([0, 1 + do_tr, 2 - do_tr, 3]))
y_rand = tf.transpose(y_rand, tf.pack([0, 1, 2 + do_tr, 3 - do_tr]))
# Random hue, saturation, brightness, contrast
# x_rand = img.random_hue(x_rand, 0.5)
# x_rand = img.random_saturation(x_rand, 0.5, 2.0)
# x_rand = tf.image.random_brightness(x_rand, 0.5)
# x_rand = tf.image.random_contrast(x_rand, 0.5, 2.0)
x = (1.0 - phase_train_f) * x_ctr + phase_train_f * x_rand
y_gt = (1.0 - phase_train_f) * y_ctr + phase_train_f * y_rand
return x, y_gt
def reverse(x, axis=-1):
'''Apply [::-1] to appropriate axis'''
ndim = len(x.get_shape())
dims = [False] * ndim
if axis < 0:
axis = axis % ndim
dims[axis] = True
return tf.reverse(x, dims)
(source_int_text,
target_int_text), (source_vocab_to_int,
target_vocab_to_int), _ = helper.load_preprocess()
max_source_sentence_length = max(
[len(sentence) for sentence in source_int_text])
train_graph = tf.Graph()
with train_graph.as_default():
input_data, targets, lr, keep_prob = model_inputs()
sequence_length = tf.placeholder_with_default(max_source_sentence_length,
None,
name='sequence_length')
input_shape = tf.shape(input_data)
train_logits, inference_logits = seq2seq_model(
tf.reverse(input_data, [-1]), targets, keep_prob,
batch_size, sequence_length, len(source_vocab_to_int),
len(target_vocab_to_int), encoding_embedding_size,
decoding_embedding_size, rnn_size, num_layers, target_vocab_to_int)
tf.identity(inference_logits, 'logits')
with tf.name_scope("optimization"):
# Loss function
cost = tf.contrib.seq2seq.sequence_loss(
train_logits, targets, tf.ones([input_shape[0], sequence_length]))
# Optimizer
optimizer = tf.train.AdamOptimizer(lr)
# Gradient Clipping
gradients = optimizer.compute_gradients(cost)
def ae_transformer_internal(inputs,
targets,
target_space,
hparams,
cache=None,
predict_mask=1.0,
means=None,
ema_count=None,
ema_means=None):
"""AE Transformer, main step used for training."""
# Summaries break with the do_refine cond, turn them off in that case.
global _DO_SUMMARIES
if hparams.do_refine:
_DO_SUMMARIES = False
# Prepare.
batch_size = common_layers.shape_list(inputs)[0]
targets = tf.reshape(targets, [batch_size, -1, 1, hparams.hidden_size])
# Encoder.
if inputs is not None:
inputs = common_layers.flatten4d3d(inputs)
inputs, ed = encode(inputs, target_space, hparams, "input_enc")
else:
ed = None
# Autoencoding.
losses = {"extra": tf.constant(0.0), "latent_pred": tf.constant(0.0)}
if hparams.do_ae:
max_targets_len_from_inputs = tf.concat([inputs, inputs], axis=1)
targets, _ = common_layers.pad_to_same_length(
targets, max_targets_len_from_inputs,
final_length_divisible_by=2**hparams.num_compress_steps)
targets_c = compress(targets, inputs, False, hparams, "compress")
if hparams.mode != tf.estimator.ModeKeys.PREDICT:
# Compress and bottleneck.
latents_dense, latents_discrete, extra_loss, _ = bottleneck(
targets_c, hparams, 2 * 2048, "vc", means, ema_count, ema_means)
if _DO_SUMMARIES:
tf.summary.histogram("b0", tf.reshape(latents_discrete[:, 0, :], [-1]))
pc = common_layers.inverse_exp_decay(hparams.startup_steps)
pc = pc if hparams.mode == tf.estimator.ModeKeys.TRAIN else 1.0
cond = tf.less(tf.random_uniform([batch_size]), pc)
latents_dense = tf.where(cond, latents_dense, targets_c)
# TODO(lukaszkaiser): return extra losses batchwise, multiply before mean.
losses["extra"] = extra_loss * tf.reduce_mean(tf.to_float(cond))
# Extra loss predicting latent code from input. Discrete only.
if hparams.bottleneck_kind not in ["dense", "vae"]:
latents_pred = decode_transformer(
tf.stop_gradient(inputs), tf.stop_gradient(ed),
tf.stop_gradient(latents_dense), hparams, "extra")
_, latent_pred_loss = ae_latent_softmax(
latents_pred, latents_discrete, hparams)
losses["latent_pred"] = tf.reduce_mean(
latent_pred_loss * 0.5 * tf.to_float(cond))
else:
inputs_c = decode_transformer(inputs, ed, targets_c, hparams, "dec_c")
losses["latent_pred"] = tf.reduce_mean((inputs_c - targets_c)**2) * 20
def bn_inputs():
with tf.variable_scope(tf.get_variable_scope(), reuse=True):
bn, _, _, _ = bottleneck(inputs_c, hparams, 2 * 2048, "vc", means,
ema_count, ema_means)
return bn
pbn = 0.8 if hparams.mode == tf.estimator.ModeKeys.TRAIN else 1.0
inputs_c = tf.cond(tf.less(tf.random_uniform([]), pbn),
bn_inputs, lambda: inputs_c)
ptc = 1.0 - common_layers.inverse_lin_decay(200000) * 0.5
ptc = ptc if hparams.mode == tf.estimator.ModeKeys.TRAIN else 1.0
latents_dense = tf.where(tf.less(tf.random_uniform([batch_size]), ptc),
latents_dense, inputs_c)
else:
if hparams.bottleneck_kind in ["dense", "vae"]:
inputs_c = decode_transformer(inputs, ed, targets_c, hparams, "dec_c")
latents_dense, _, _, _ = bottleneck(inputs_c, hparams, 2 * 2048, "vc",
means, ema_count, ema_means)
else:
latent_len = common_layers.shape_list(targets_c)[1]
_, _, _, embed = bottleneck(targets_c, hparams, 2 * 2048, "vc", means,
ema_count, ema_means)
latents_dense = tf.zeros_like(targets_c[:, :latent_len, :, :])
if cache is None:
cache = ae_latent_sample(latents_dense, inputs, ed, embed, 8, hparams)
latents_dense = embed(cache)
# Postprocess.
d = latents_dense
pos = tf.get_variable("pos", [1, 1000, 1, hparams.hidden_size])
pos = pos[:, :common_layers.shape_list(latents_dense)[1] + 1, :, :]
latents_dense = tf.pad(latents_dense,
[[0, 0], [1, 0], [0, 0], [0, 0]]) + pos
# Masking.
if hparams.do_mask:
masking = common_layers.inverse_lin_decay(hparams.mask_startup_steps)
masking *= common_layers.inverse_exp_decay(
hparams.mask_startup_steps // 4) # Not much at start.
if not hparams.do_refine:
masking -= tf.random_uniform([]) * hparams.unmasked_percentage
masking = tf.minimum(tf.maximum(masking, 0.0), 1.0)
if hparams.mode == tf.estimator.ModeKeys.PREDICT:
masking = predict_mask
mask = tf.less(masking, tf.random_uniform(
common_layers.shape_list(targets)[:-1]))
mask = tf.expand_dims(tf.to_float(mask), 3)
for i in xrange(hparams.num_compress_steps):
j = hparams.num_compress_steps - i - 1
d = residual_conv(d, 1, (3, 1), hparams, "decompress_rc_%d" % j)
if hparams.do_attend_decompress:
d = attend(d, inputs, hparams, "decompress_attend_%d" % j)
d = decompress_step(d, hparams, i > 0, False, "decompress_%d" % j)
targets = mask * targets + (1.0 - mask) * d
targets = tf.concat([tf.reverse(latents_dense, [1]), targets], axis=1)
res = decode_transformer(inputs, ed, targets, hparams, "decoder")
if hparams.do_ae:
res = res[:, common_layers.shape_list(latents_dense)[1]:, :, :]
if hparams.do_mask and hparams.do_refine:
def refine_res():
# return residual_conv(res, 1, (5, 1), hparams, "refine")
r, _ = encode(tf.squeeze(res, axis=[2]),
target_space, hparams, "refine_enc")
return tf.expand_dims(r, axis=2)
masked_batches = tf.reduce_sum(mask, axis=[1, 2, 3])
all_masked = tf.less(masked_batches, 0.1)
res = tf.where(all_masked, refine_res(), res)
# We'll start training the extra model of latents after mask_startup_steps.
latent_time = tf.less(hparams.mask_startup_steps,
tf.to_int32(tf.train.get_global_step()))
losses["latent_pred"] *= tf.to_float(latent_time)
return res, losses, cache
def rot90(images):
return tf.transpose(tf.reverse(images, [2]), [0, 2, 1, 3])
def _parse_eval_(image): image = (image*1.0+1.0)*127.5 image = tf.reverse(image,[-1]) image = tf.clip_by_value(image,0,255) return image
def build_bert_inputs(example):
"""Convert example <Tensor [30, 70]> into bert inputs."""
k_size = FLAGS.k_size
CLS_ID = tf.constant([101], dtype=tf.int64) # pylint: disable=invalid-name
SEP_ID = tf.constant([102], dtype=tf.int64) # pylint: disable=invalid-name
max_len = tf.constant([FLAGS.max_para_length])
context_size = tf.constant([FLAGS.context_size])
intermediate_examples_tensor = tf.reduce_sum(tf.abs(example), 1)
examples_zero_vector = tf.zeros(shape=(1, 1), dtype=tf.int64)
examples_bool_mask = tf.squeeze(
tf.not_equal(intermediate_examples_tensor, examples_zero_vector))
paragraph_len = tf.reduce_sum(tf.cast(examples_bool_mask, tf.int32))
start = tf.random.uniform([1],
0,
tf.reshape(paragraph_len, []) -
tf.reshape(context_size, []) + 1,
dtype=tf.int32)
# Slice the document into the before, after and context.
# Discard the zero padding.
sizes = tf.squeeze(
tf.concat([[
start, context_size, paragraph_len - context_size - start,
max_len - paragraph_len
]], 0))
before, context, after, _ = tf.split(example, sizes, axis=0)
# Gather the context removing zero padding at end of sentences.
non_zeros = tf.where(tf.not_equal(context, tf.zeros_like(context)))
context_gathered = tf.gather_nd(context, non_zeros)
# Flip before so we select the 4 sentences closest to target
before = tf.reverse(before, axis=[0])
# pad both to longer than needed
paddings = tf.constant([[0, 8], [0, 0]])
before = tf.pad(before, paddings)
after = tf.pad(after, paddings)
# Extend targets to 3 sentences
# pad both
before_minus_one = before[1:][:k_size]
before_minus_two = before[2:][:k_size]
after_plus_one = after[1:][:k_size]
after_plus_two = after[2:][:k_size]
before = before[:k_size]
after = after[:k_size]
before = tf.concat([before_minus_two, before_minus_one, before], axis=1)
after = tf.concat([after, after_plus_one, after_plus_two], axis=1)
############################################################################
# These 8 sentences are the 8 surrounding targets. Some are padding.
targets = tf.concat([before, after], axis=0)
# Remove the padding from the sourrounding sentences
# Eg. if context starts at beginning of paragraph, before is all padding
intermediate_tensor = tf.reduce_sum(tf.abs(targets), 1)
zero_vector = tf.zeros(shape=(1, 1), dtype=tf.int64)
bool_mask = tf.squeeze(tf.not_equal(intermediate_tensor, zero_vector))
bool_mask.set_shape([None])
targets = tf.boolean_mask(targets, bool_mask)
# Randomly select 4 targets
# We will also select the label_types for each selected target
indices = tf.range(0, limit=tf.shape(targets)[0], dtype=tf.int32)
shuffled_indices = tf.random.shuffle(indices)[:k_size]
targets = tf.gather(targets, shuffled_indices)
if k_size == 4:
full_labels = tf.concat([tf.range(3, -1, -1), tf.range(4, 8)], axis=0)
elif k_size == 3:
full_labels = tf.concat([tf.range(2, -1, -1), tf.range(3, 6)], axis=0)
elif k_size == 2:
full_labels = tf.concat([tf.range(1, -1, -1), tf.range(2, 4)], axis=0)
elif k_size == 1:
full_labels = tf.concat([tf.range(0, -1, -1), tf.range(1, 2)], axis=0)
label_types = tf.boolean_mask(full_labels, bool_mask)
label_types = tf.gather(label_types, shuffled_indices)
# create inputs
bert_inputs = []
input_masks = []
segment_ids = []
# make context
ctx_segment_id = tf.concat([
tf.zeros_like(CLS_ID, dtype=tf.int64),
tf.zeros_like(context_gathered),
tf.zeros_like(SEP_ID, dtype=tf.int64)
],
axis=0)
ctx_segment_id = pad_and_cut(ctx_segment_id, FLAGS.max_seq_length)
segment_ids.append(ctx_segment_id)
new_ctx_input = tf.concat([CLS_ID, context_gathered, SEP_ID], axis=0)
ctx_input_mask = tf.ones_like(new_ctx_input)
ctx_input_mask = pad_and_cut(ctx_input_mask, FLAGS.max_seq_length)
input_masks.append(ctx_input_mask)
padded_new_ctx_input = pad_and_cut(new_ctx_input, FLAGS.max_seq_length)
bert_inputs.append(padded_new_ctx_input)
for i in range(k_size):
target_non_zero = tf.where(
tf.not_equal(targets[i], tf.zeros_like(targets[i])))
targets_stripped = tf.gather_nd(targets[i], target_non_zero)
if FLAGS.include_context:
segment_id = tf.concat([
tf.zeros_like(CLS_ID, dtype=tf.int64),
tf.zeros_like(context_gathered),
tf.zeros_like(SEP_ID, dtype=tf.int64),
tf.ones_like(targets_stripped),
tf.ones_like(SEP_ID, dtype=tf.int64)
],
axis=0)
else:
segment_id = tf.concat([
tf.zeros_like(CLS_ID, dtype=tf.int64),
tf.zeros_like(targets_stripped),
tf.zeros_like(SEP_ID, dtype=tf.int64)
],
axis=0)
segment_id = pad_and_cut(segment_id, FLAGS.max_seq_length)
segment_ids.append(segment_id)
if FLAGS.include_context:
new_input = tf.concat(
[CLS_ID, context_gathered, SEP_ID, targets_stripped, SEP_ID],
axis=0)
else:
new_input = tf.concat([CLS_ID, targets_stripped, SEP_ID], axis=0)
input_mask = tf.ones_like(new_input)
input_mask = pad_and_cut(input_mask, FLAGS.max_seq_length)
input_masks.append(input_mask)
padded_new_input = pad_and_cut(new_input, FLAGS.max_seq_length)
bert_inputs.append(padded_new_input)
bert_inputs = tf.stack(bert_inputs, axis=0)
input_masks = tf.stack(input_masks, axis=0)
segment_ids = tf.stack(segment_ids, axis=0)
out = Outputs_And_Context(bert_inputs, input_masks, segment_ids,
label_types, context_gathered)
return out
def tf_flip_slices_2d(dl, ur, ul):
flip_dl = tf.reverse(dl, axis=[1])
flip_ur = tf.reverse(ur, axis=[2])
flip_ul = tf.reverse(ul, axis=[1, 2])
return flip_dl, flip_ur, flip_ul
def tf_flip_slices_1d(l):
return tf.reverse(l, axis=[1])
learning_rate = 0.001
learning_rate_decay = 0.9
min_learning_rate = 0.0001
keep_probability = 0.5
tf.reset_default_graph()
session = tf.InteractiveSession()
inputs, targets, lr, keep_prob = model_inputs()
sequence_length = tf.placeholder_with_default(25, None, name='sequence_length')
input_shape = tf.shape(inputs)
training_predictions, test_predictions = seq2seq_model(
tf.reverse(inputs, [-1]), targets, keep_prob, batch_size, sequence_length,
len(answerswords2int), len(questionswords2int), encoding_embedding_size,
decoding_embedding_size, rnn_size, num_layers, questionswords2int)
with tf.name_scope("optimization"):
loss_error = tf.contrib.seq2seq.sequence_loss(
training_predictions, targets,
tf.ones([input_shape[0], sequence_length]))
optimizer = tf.train.AdamOptimizer(learning_rate)
gradients = optimizer.compute_gradients(loss_error)
clipped_gradients = [(tf.clip_by_value(grad_tensor, -5.,
5.), grad_variable)
for grad_tensor, grad_variable in gradients
if grad_tensor is not None]
optimizer_gradient_clipping = optimizer.apply_gradients(clipped_gradients)
def trainSurvivalNet(mat_file_path, n_hidden, num_steps, num_shuffles,
penaltyLambdaArray, alphaArray, prefix):
""" This function is to train SurvivalNet with Tensorflow.
:type mat_file_path: string
:param mat_file_path: path to the file that stores data in .mat format
:type n_hidden: integer
:param n_hidden: number of hidden nodes in a layer
:type num_steps: integer
:param num_steps: number of iterations to run
:type num_shuffles: integer
:param num_shuffles: number of shuffles to run
:type penaltyLambdaArray: np.float32 array
:param penaltyLambdaArray: array of lambda (regularization parameters) to train to model
:type alphaArray: np.float32 array
:param alphaArray: array of alpha (balancing factor between L1 and L2 in elastic net) to train the model
:type prefix: string
:param prefix: prefix of output file that stores all results
"""
p = os.path.join(os.getcwd(), mat_file_path)
Brain_C = sio.loadmat(p)
data = Brain_C['Integ_X']
censor = np.asarray([c[0] for c in Brain_C['Censored']])
survival = np.asarray([t[0] for t in Brain_C['Survival']])
T = np.asarray([t[0] for t in Brain_C['Survival']])
O = 1 - np.asarray([c[0] for c in Brain_C['Censored']])
X = Brain_C['Integ_X']
#Use the whole dataset fotr pretraining
pretrain_set = X
#foldsize denotes th amount of data used for testing. The same amount
#of data is used for model selection. The rest is used for training.
fold_size = int(len(X) / 10)
train_set = {}
test_set = {}
final_set = {}
#caclulate the risk group for every patient i: patients who die after i
sa = SurvivalAnalysis()
train_set['X'], train_set['T'], train_set['O'], train_set[
'A'] = sa.calc_at_risk(X[0:fold_size * 6, ], T[0:fold_size * 6],
O[0:fold_size * 6])
test_set['X'], test_set['T'], test_set['O'], test_set[
'A'] = sa.calc_at_risk(X[fold_size * 6:fold_size * 8, ],
T[fold_size * 6:fold_size * 8],
O[fold_size * 6:fold_size * 8])
final_set['X'], final_set['T'], final_set['O'], final_set[
'A'] = sa.calc_at_risk(X[fold_size * 8:, ], T[fold_size * 8:],
O[fold_size * 8:])
## initialization
n_obs = train_set['X'].shape[0] # 302
n_in = train_set['X'].shape[1] # 201
test_obs = test_set['X'].shape[0] # 64
test_in = test_set['X'].shape[1] # 201
n_out = 1
#### tensorflow implementation
def cumsum(x, observations):
x = tf.reshape(x, (1, observations))
values = tf.split(1, x.get_shape()[1], x)
out = []
prev = tf.zeros_like(values[0])
for val in values:
s = prev + val
out.append(s)
prev = s
cumsum = tf.concat(1, out)
cumsum = tf.reshape(cumsum, (observations, 1))
return cumsum
with tf.device('/gpu:1'):
## dropout
keep_prob = tf.placeholder(tf.float32)
## penaltyLambda
penaltyLambda = tf.placeholder(tf.float32)
## alpha
alpha = tf.placeholder(tf.float32)
## data
input = tf.placeholder(tf.float32, [n_obs, n_in])
at_risk = tf.placeholder(tf.int32, [
n_obs,
])
observed = tf.placeholder(tf.float32, [
n_obs,
])
# testing data
test_input = tf.placeholder(tf.float32, [test_obs, test_in])
prediction_at_risk = tf.placeholder(tf.int32, [
test_obs,
])
prediction_observed = tf.placeholder(tf.float32, [
test_obs,
])
## layer_1
w_1 = tf.Variable(
tf.truncated_normal([n_in, n_hidden], dtype=tf.float32) / 20)
output_layer1 = tf.nn.relu(tf.matmul(input, w_1))
output_layer1_drop = tf.nn.dropout(output_layer1, keep_prob)
prediciton_layer1 = tf.nn.relu(tf.matmul(test_input, w_1))
## layer_2
w_2 = tf.Variable(
tf.truncated_normal([n_hidden, n_hidden], dtype=tf.float32) / 20)
output_layer2 = tf.nn.relu(tf.matmul(output_layer1_drop, w_2))
output_layer2_drop = tf.nn.dropout(output_layer2, keep_prob)
prediciton_layer2 = tf.nn.relu(tf.matmul(prediciton_layer1, w_2))
## layer_3
w_3 = tf.Variable(
tf.truncated_normal([n_hidden, n_hidden], dtype=tf.float32) / 20)
output_layer3 = tf.nn.relu(tf.matmul(output_layer2_drop, w_3))
output_layer3_drop = tf.nn.dropout(output_layer3, keep_prob)
prediciton_layer3 = tf.nn.relu(tf.matmul(prediciton_layer2, w_3))
## layer_4
w_4 = tf.Variable(
tf.truncated_normal([n_hidden, n_hidden], dtype=tf.float32) / 20)
output_layer4 = tf.nn.relu(tf.matmul(output_layer3_drop, w_4))
output_layer4_drop = tf.nn.dropout(output_layer4, keep_prob)
prediciton_layer4 = tf.nn.relu(tf.matmul(prediciton_layer3, w_4))
# layer_5
w_5 = tf.Variable(
tf.truncated_normal([n_hidden, n_hidden], dtype=tf.float32) / 20)
output_layer5 = tf.nn.relu(tf.matmul(output_layer4_drop, w_5))
output_layer5_drop = tf.nn.dropout(output_layer5, keep_prob)
prediciton_layer5 = tf.nn.relu(tf.matmul(prediciton_layer4, w_5))
## output layer
w_6 = tf.Variable(
tf.truncated_normal([n_hidden, n_out], dtype=tf.float32) / 20)
output = tf.matmul(output_layer5_drop, w_6)
prediction_output = tf.matmul(prediciton_layer5, w_6)
exp = tf.reverse(tf.exp(output), dims=[True, False])
partial_sum_a = cumsum(exp, n_obs)
partial_sum = tf.reverse(partial_sum_a, dims=[True, False]) + 1
log_at_risk = tf.log(
tf.gather(partial_sum, tf.reshape(at_risk, [-1])) + 1e-50)
diff = tf.sub(output, log_at_risk)
times = tf.reshape(diff, [-1]) * observed
cost = -(tf.reduce_sum(times)) + alpha * tf.reduce_sum(
penaltyLambda * tf.nn.l2_loss(w_6)) + alpha * tf.reduce_sum(
penaltyLambda * tf.nn.l2_loss(w_5)) + alpha * tf.reduce_sum(
penaltyLambda *
tf.nn.l2_loss(w_4)) + alpha * tf.reduce_sum(
penaltyLambda *
tf.nn.l2_loss(w_3)) + alpha * tf.reduce_sum(
penaltyLambda *
tf.nn.l2_loss(w_3)) + alpha * tf.reduce_sum(
penaltyLambda *
tf.nn.l2_loss(w_2)) + alpha * tf.reduce_sum(
penaltyLambda * tf.nn.l2_loss(w_1)
) + (1 - alpha) * tf.reduce_sum(
penaltyLambda *
tf.abs(w_6)) + (1 - alpha) * tf.reduce_sum(
penaltyLambda * tf.abs(w_5)
) + (1 - alpha) * tf.reduce_sum(
penaltyLambda * tf.abs(w_4)
) + (1 - alpha) * tf.reduce_sum(
penaltyLambda * tf.abs(w_3)
) + (1 - alpha) * tf.reduce_sum(
penaltyLambda * tf.abs(w_3)
) + (1 - alpha) * tf.reduce_sum(
penaltyLambda * tf.abs(w_2)) + (
1 - alpha) * tf.reduce_sum(
penaltyLambda * tf.abs(w_1))
weightSize = tf.nn.l2_loss(w_1) + tf.nn.l2_loss(w_2) + tf.nn.l2_loss(
w_3) + tf.nn.l2_loss(w_4) + tf.nn.l2_loss(w_5) + tf.nn.l2_loss(w_6)
### prediction
prediction_exp = tf.reverse(tf.exp(prediction_output),
dims=[True, False])
prediction_partial_sum_a = cumsum(prediction_exp, test_obs)
prediction_partial_sum = tf.reverse(prediction_partial_sum_a,
dims=[True, False]) + 1
prediction_log_at_risk = tf.log(
tf.gather(prediction_partial_sum,
tf.reshape(prediction_at_risk, [-1])) + 1e-50)
prediction_diff = tf.sub(prediction_output, prediction_log_at_risk)
prediction_times = tf.reshape(prediction_diff,
[-1]) * prediction_observed
prediction_cost = -(tf.reduce_sum(prediction_times))
global_step = tf.Variable(0, trainable=False)
starter_learning_rate = 0.0001
learning_rate = tf.train.exponential_decay(starter_learning_rate,
global_step,
100000,
0.989,
staircase=True)
# optimizer
optimizer = tf.train.GradientDescentOptimizer(learning_rate).minimize(
cost)
for alphaArrayIndex in range(len(alphaArray)):
print("alpha: " + str(alphaArray[alphaArrayIndex]))
for penaltyLambdaIndex in range(len(penaltyLambdaArray)):
print("lambda: " + str(penaltyLambdaArray[penaltyLambdaIndex]))
targetFile = prefix + ".lambda." + str(
penaltyLambdaArray[penaltyLambdaIndex]) + ".alpha." + str(
alphaArray[alphaArrayIndex]) + ".txt"
target = open(targetFile, "w")
finalTestingAcc = np.zeros(num_shuffles)
testingAcc = np.zeros(num_shuffles)
bestAccInOneShuffle = np.zeros(num_steps)
session = tf.InteractiveSession()
header = prefix + ".lambda." + str(
penaltyLambdaArray[penaltyLambdaIndex]) + ".alpha." + str(
alphaArray[alphaArrayIndex])
for shuffle in range(num_shuffles):
outputFile = header + "." + str(shuffle) + ".txt"
outputFH = open(outputFile, "w")
outputFH.write("trainCost" + "\t" + "testCost" + "\t" +
"trainCIndex" + "\t" + "testCIndex" + "\t" +
"weightSize" + "\n")
tf.initialize_all_variables().run()
index = np.arange(data.shape[0])
random.shuffle(index)
X = X[index, :]
O = O[index]
T = T[index]
fold_size = int(len(X) / 10)
train_set = {}
test_set = {}
final_set = {}
sa = SurvivalAnalysis()
train_set['X'], train_set['T'], train_set['O'], train_set[
'A'] = sa.calc_at_risk(X[0:fold_size * 6, ],
T[0:fold_size * 6],
O[0:fold_size * 6])
test_set['X'], test_set['T'], test_set['O'], test_set[
'A'] = sa.calc_at_risk(X[fold_size * 6:fold_size * 8, ],
T[fold_size * 6:fold_size * 8],
O[fold_size * 6:fold_size * 8])
final_set['X'], final_set['T'], final_set['O'], final_set[
'A'] = sa.calc_at_risk(X[fold_size * 8:fold_size * 10, ],
T[fold_size * 8:fold_size * 10],
O[fold_size * 8:fold_size * 10])
number_of_range = 0
sum_of_test_c_index = np.zeros(15)
for step in range(num_steps):
feed_dict = {
input: train_set['X'],
at_risk: train_set['A'],
observed: train_set['O'],
test_input: test_set['X'],
prediction_at_risk: test_set['A'],
prediction_observed: test_set['O'],
keep_prob: 1,
penaltyLambda: penaltyLambdaArray[penaltyLambdaIndex],
alpha: alphaArray[alphaArrayIndex]
}
timesV, _, test_outputV, outputV, costV, expV, partialV, logV, diffV, w1V, costTestV, weightSizeV = session.run(
[
times, optimizer, prediction_output, output, cost,
exp, partial_sum, log_at_risk, diff, w_1,
prediction_cost, weightSize
],
feed_dict=feed_dict)
train_c_index = _naive_concordance_index(
train_set['T'], -outputV, train_set['O'])
test_c_index = _naive_concordance_index(
test_set['T'], -test_outputV, test_set['O'])
bestAccInOneShuffle[step] = test_c_index
outputFH.write(
str(costV) + "\t" + str(costTestV) + "\t" +
str(train_c_index) + "\t" + str(test_c_index) + "\t" +
str(weightSizeV) + "\n")
if (step % 10 == 0):
print("step: " + str(step) + ", cost: " + str(costV))
print("train cIndex: " + str(train_c_index) +
", test cIndex: " + str(test_c_index))
if (step == num_steps - 1):
print("best result: " +
str(np.max(bestAccInOneShuffle)))
feed_dict = {
input: train_set['X'],
at_risk: train_set['A'],
observed: train_set['O'],
test_input: final_set['X'],
keep_prob: 1,
penaltyLambda:
penaltyLambdaArray[penaltyLambdaIndex],
alpha: alphaArray[alphaArrayIndex]
}
final_outputV = session.run(prediction_output,
feed_dict=feed_dict)
final_c_index = _naive_concordance_index(
final_set['T'], -final_outputV, final_set['O'])
finalTestingAcc[shuffle] = final_c_index
testingAcc[shuffle] = test_c_index
outputFH.close()
target.write("final mean: " + str(np.mean(finalTestingAcc)) + "\n")
target.write("final sd: " + str(np.std(finalTestingAcc)) + "\n")
target.write("---\n")
target.write("validation mean: " + str(np.mean(testingAcc)) + "\n")
target.write("validation sd: " + str(np.std(testingAcc)) + "\n")
target.close()
def tf_discount_rewards(tf_r): # tf_r ~ [game_steps,1]
discount_f = lambda a, v: a * gamma + v
tf_r_reverse = tf.scan(discount_f, tf.reverse(tf_r, [True, False]))
tf_discounted_r = tf.reverse(tf_r_reverse, [True, False])
return tf_discounted_r
def pixel_wise_softmax(output_map):
exponential_map = tf.exp(output_map)
evidence = tf.add(exponential_map,
tf.reverse(exponential_map, [False, False, False, True]))
return tf.div(exponential_map, evidence, name="pixel_wise_softmax")
# Defining a session
tf.reset_default_graph()
session = tf.InteractiveSession()
# Loading the model inputs
inputs, targets, lr, keep_prob = model_inputs()
# Setting the sequence length
sequence_length = tf.placeholder_with_default(25, None, name = 'sequence_length')
# Getting the shape of the inputs tensor
input_shape = tf.shape(inputs)
# Getting the training and test predictions
training_predictions, test_predictions = seq2seq_model(tf.reverse(inputs, [-1]),
targets,
keep_prob,
batch_size,
sequence_length,
len(answerswords2int),
len(questionswords2int),
encoding_embedding_size,
decoding_embedding_size,
rnn_size,
num_layers,
questionswords2int)
# Setting up the Loss Error, the Optimizer and Gradient Clipping
with tf.name_scope("optimization"):
loss_error = tf.contrib.seq2seq.sequence_loss(training_predictions,