def infogan_comp():
from infogan import InfoGAN
STYLE_DIM = 2
LABEL_DIM = 10
RAND_DIM = 88
IMG_SHAPE = (28, 28, 1)
FIX_STD = True
model = InfoGAN(RAND_DIM, STYLE_DIM, LABEL_DIM, IMG_SHAPE, FIX_STD)
model.load_weights("./models/infogan/model.ckpt").expect_partial()
img_label = np.arange(0, 10).astype(np.int32).repeat(10, axis=0)
noise = tf.repeat(tf.random.normal((1, model.rand_dim)),
len(img_label),
axis=0)
def plot(noise, img_label, img_style, n):
img_label = tf.convert_to_tensor(img_label, dtype=tf.int32)
img_style = tf.convert_to_tensor(img_style, dtype=tf.float32)
imgs = model.g.call([noise, img_label, img_style],
training=False).numpy()
plt.clf()
nc, nr = 10, 10
plt.figure(0, (nc * 2, nr * 2))
for c in range(nc):
for r in range(nr):
i = r * nc + c
plt.subplot(nc, nr, i + 1)
plt.imshow(imgs[i], cmap="gray_r")
plt.axis("off")
plt.text(23, 26, int(r), fontsize=23)
plt.tight_layout()
model_name = model.__class__.__name__.lower()
dir_ = "visual/{}".format(model_name)
os.makedirs(dir_, exist_ok=True)
path = dir_ + "/style{}.png".format(n)
plt.savefig(path)
img_style = np.concatenate(
[np.linspace(-model.style_scale, model.style_scale, 10)] * 10).reshape(
(100, 1)).astype(np.float32)
plot(noise, img_label,
np.concatenate((img_style, np.zeros_like(img_style)), axis=1), 1)
plot(noise, img_label,
np.concatenate((np.zeros_like(img_style), img_style), axis=1), 2)
def _gaussian_kernel(self, pos):
'''
Generates all gaussian kernels needed for the loss function
'''
# Get vector of spatial indices
idxs = np.indices((self.dim_len[0], self.dim_len[1]))
idxs = idxs.reshape(2, -1)
idxs = tf.convert_to_tensor(idxs)
# Each value of the kernel is computed as exp(dist^2/sigma)
# where dist is the distance between position (i, j) and the center
pos_rep = tf.repeat(tf.expand_dims(pos, 1),
self.dim_len[0] * self.dim_len[1],
axis=1)
dists = tf.math.reduce_euclidean_norm(tf.cast(idxs - pos_rep,
tf.float64),
axis=0)
gaussian = tf.math.exp(-tf.math.square(dists) / self.sigma)
return gaussian
def _spatial_transform(image,
rotation=5.0,
shear=2.0,
hzoom=8.0,
wzoom=8.0,
hshift=8.0,
wshift=8.0):
ydim = tf.gather(tf.shape(image), 0)
xdim = tf.gather(tf.shape(image), 1)
xxdim = xdim % 2
yxdim = ydim % 2
# random rotation, shear, zoom and shift
rotation = rotation * tf.random.normal([1], dtype='float32')
shear = shear * tf.random.normal([1], dtype='float32')
hzoom = 1.0 + tf.random.normal([1], dtype='float32') / hzoom
wzoom = 1.0 + tf.random.normal([1], dtype='float32') / wzoom
hshift = hshift * tf.random.normal([1], dtype='float32')
wshift = wshift * tf.random.normal([1], dtype='float32')
m = _get_transform_matrix(rotation, shear, hzoom, wzoom, hshift, wshift)
# origin pixels
y = tf.repeat(tf.range(ydim // 2, -ydim // 2, -1), xdim)
x = tf.tile(tf.range(-xdim // 2, xdim // 2), [ydim])
z = tf.ones([ydim * xdim], dtype='int32')
idx = tf.stack([y, x, z])
# destination pixels
idx2 = tf.matmul(m, tf.cast(idx, dtype='float32'))
idx2 = tf.cast(idx2, dtype='int32')
# clip to origin pixels range
idx2y = tf.clip_by_value(idx2[0, ], -ydim // 2 + yxdim + 1, ydim // 2)
idx2x = tf.clip_by_value(idx2[1, ], -xdim // 2 + xxdim + 1, xdim // 2)
idx2 = tf.stack([idx2y, idx2x, idx2[2, ]])
# apply destinations pixels to image
idx3 = tf.stack([ydim // 2 - idx2[0, ], xdim // 2 - 1 + idx2[1, ]])
d = tf.gather_nd(image, tf.transpose(idx3))
image = tf.reshape(d, [ydim, xdim, 3])
return image
def draft_decoder(self,
input_ids,
enc_output,
beam_size,
length_penalty,
temperature,
top_p,
top_k,
batch_size,
training=False):
"""
Inference call, builds a draft output_sequence auto-regressively
"""
start_ids = tf.repeat(config.CLS_ID, repeats=batch_size)
input_ids = tfa.seq2seq.tile_batch(input_ids, multiplier=beam_size)
enc_output = tfa.seq2seq.tile_batch(enc_output, multiplier=beam_size)
#start_ids = tf.cast(start_ids, dtype=tf.int64)
def perform_beam_search(dec_input):
return query_decoder(self,
enc_output,
input_ids,
dec_input,
batch_size,
temperature,
top_p,
top_k,
training=training)
predicted_beam_search_op, _, _, attention_dist = beam_search(
perform_beam_search,
initial_ids=start_ids,
beam_size=beam_size,
decode_length=config.target_seq_length,
vocab_size=config.target_vocab_size,
alpha=length_penalty,
stop_early=False,
eos_id=config.SEP_ID)
predicted_output_sequence = predicted_beam_search_op[:, 0, :]
return predicted_output_sequence, attention_dist
def grad(dx):
# Gradients with respect to off grid signal
fft_dx = scale_and_fft_on_image_volume(dx,
scaling_coef,
grid_size,
im_size,
norm,
im_rank=im_rank)
dx_dy = kbinterp(fft_dx, om, interpob)
if grad_traj:
# Gradients with respect to trajectory locations
r = [
tf.linspace(-im_size[i] / 2, im_size[i] / 2 - 1,
im_size[i]) for i in range(im_rank)
]
# This wont work for multicoil case as the dimension for dx is `batch_size x coil x Nx x Ny`
grid_r = tf.cast(tf.meshgrid(*r, indexing='ij'),
dx.dtype)[None, ...]
ifft_dxr = scale_and_fft_on_image_volume(tf.math.conj(dx) *
grid_r,
scaling_coef,
grid_size,
im_size,
norm,
im_rank=im_rank,
do_ifft=True)
# Do this when handling batches
ifft_dxr = tf.reshape(ifft_dxr,
shape=(-1, 1, *ifft_dxr.shape[2:]))
inufft_dxr = kbinterp(ifft_dxr,
tf.repeat(om, im_rank, axis=0),
interpob,
conj=True)
# Unbatch back the data
inufft_dxr = tf.reshape(inufft_dxr,
shape=(-1, im_rank,
*inufft_dxr.shape[2:]))
dx_dom = tf.cast(1j * y * inufft_dxr, om.dtype)
# dx_dom = tf.math.reduce_sum(dx_dom, axis=1)[None, :]
else:
dx_dom = None
return dx_dy, dx_dom
def si_weight(states, actions, n_agents):
data = tf.concat([states, actions], axis=-1)
all_head_key, all_head_agents, all_head_action = [[] for _ in range(3)]
with tf.variable_scope('adv_hyer'):
for i in range(num_kernel):
all_head_key.append(
tf.layers.dense(states,
1,
kernel_initializer=w_initializer,
bias_initializer=b_initializer,
name='all_head_key-{}'.format(i)))
all_head_agents.append(
tf.layers.dense(states,
n_agents,
kernel_initializer=w_initializer,
bias_initializer=b_initializer,
name='all_head_agents-{}'.format(i)))
all_head_action.append(
tf.layers.dense(data,
n_agents,
kernel_initializer=w_initializer,
bias_initializer=b_initializer,
name='all_head_action-{}'.format(i)))
head_attend_weights = []
for curr_head_key, curr_head_agents, curr_head_action in zip(
all_head_key, all_head_agents, all_head_action):
x_key = tf.repeat(tf.abs(curr_head_key), n_agents) + 1e-10
x_key = tf.reshape(x_key, shape=[-1, n_agents])
x_agents = tf.sigmoid(curr_head_agents)
x_action = tf.sigmoid(curr_head_action)
weights = x_key * x_agents * x_action
head_attend_weights.append(weights)
head_attend = tf.stack(head_attend_weights, axis=1)
head_attend = tf.reshape(head_attend, shape=(-1, num_kernel, n_agents))
head_attend = tf.reduce_sum(head_attend, axis=1, keepdims=False)
return head_attend
def update_output(self, input_latent, input_labels, img_width=1600):
input_latent = tf.convert_to_tensor(input_latent, dtype=tf.float32)
input_latent = tf.repeat(input_latent, self.num_img, axis=0)
input_labels = tf.convert_to_tensor(input_labels, dtype=tf.float32)
img_per_batch = 12
self.output = []
for i in range(self.num_img // img_per_batch):
index_start = i * img_per_batch
index_end = min((i + 1) * img_per_batch, self.num_img)
self.output.extend(
self.model([
input_latent[index_start:index_end],
input_labels[index_start:index_end]
]))
self.output = (np.asarray(self.output) * 0.5) + 0.5
self.img_size = self.output.shape[1]
imgs = []
for i in range(self.output.shape[0]):
imgs.append(
tf.keras.preprocessing.image.array_to_img(self.output[i]))
self.output_img = Image.new(
'L', (self.num_img // self.num_rows_chars * self.img_size,
self.img_size * self.num_rows_chars))
for i in range(len(imgs)):
self.output_img.paste(
imgs[i],
(self.img_size * (i % (self.num_img // self.num_rows_chars)),
(i // (self.num_img // self.num_rows_chars)) * self.img_size))
img_height = int(self.output_img.size[1] * img_width /
self.output_img.size[0])
self.output_img = self.output_img.resize((img_width, img_height),
resample=Image.NEAREST)
q_pix = QPixmap.fromImage(ImageQt.ImageQt(self.output_img))
self.output_label.setPixmap(q_pix)
def shear_img(image, dim, n_channels, shear_factor = 7.0):
'''
Shears an image by `shear_factor` degrees
Args:
image: array of one squared image of size (dim, dim, n_channels)
n_channels: (int)
shear_factor: (int or float) the shear factor in degrees
Returns:
An sheared image
'''
xdim = dim%2
shr = shear_factor * tf.ones([1], dtype='float32')
shear = 3.141593 * shr / 180.
def get_3x3_mat(lst):
return tf.reshape(tf.concat([lst],axis=0), [3,3])
# shear matrix
c2 = tf.math.cos(shear)
s2 = tf.math.sin(shear)
one = tf.constant([1],dtype='float32')
zero = tf.constant([0],dtype='float32')
m = get_3x3_mat([one, s2, zero,
zero, c2, zero,
zero, zero, one])
# list destination pixel indices
x = tf.repeat(tf.range(dim//2, -dim//2,-1), dim)
y = tf.tile(tf.range(-dim//2, dim//2), [dim])
z = tf.ones([dim*dim], dtype='int32')
idx = tf.stack( [x,y,z] )
# rotate destination pixels onto origin pixels
idx2 = K.dot(m, tf.cast(idx, dtype='float32'))
idx2 = K.cast(idx2, dtype='int32')
idx2 = K.clip(idx2, -dim//2+xdim+1, dim//2)
# find origin pixel values
idx3 = tf.stack([dim//2-idx2[0,], dim//2-1+idx2[1,]])
d = tf.gather_nd(image, tf.transpose(idx3))
return tf.reshape(d,[dim, dim, n_channels])
def test_get_box_as_dotted_lines(self):
corners = tf.repeat(tf.expand_dims(tf.expand_dims(tf.range(
0, 8, dtype=tf.float32),
axis=0),
axis=2),
3,
axis=2)
corners = tf.concat([corners, corners + 10.], axis=0)
num_of_points_per_line = 100
points = box_utils.get_box_as_dotted_lines(
corners, num_of_points_per_line=num_of_points_per_line)
expected_points = np.repeat(np.linspace(
0., 1., num_of_points_per_line)[:, np.newaxis],
repeats=3,
axis=1)
self.assertAllEqual(points.shape, [2, 12 * num_of_points_per_line, 3])
self.assertAllClose(points[0, :num_of_points_per_line, :],
expected_points)
self.assertAllClose(points[1, :num_of_points_per_line, :],
expected_points + 10.)
def _replace_tensor_values(_data_to_edit, _new_value, _replace_idx):
#function performs _data_to_edit[_replace_idx] = _new_value
#acting as a work around for eager tensors
if (tf.size(_replace_idx) == 1):
_replace_value = _new_value
else:
_replace_value = tf.repeat(_new_value, tf.size(_replace_idx))
#find the location of index's to keep constant
_keep_idx = tf.where(
tf.equal(
tf.constant([0]),
tf.cast(tf.scatter_nd(
_replace_idx, tf.ones(tf.size(_replace_idx)),
tf.expand_dims(tf.size(_data_to_edit), axis=0)),
dtype=tf.int32)))
_keep_value = tf.gather_nd(_data_to_edit, _keep_idx)
return tf.dynamic_stitch([
tf.cast(tf.squeeze(_replace_idx), dtype=tf.int32),
tf.cast(tf.squeeze(_keep_idx), dtype=tf.int32)
], [_replace_value, _keep_value])
def random_walk_layer(B, adj_list, degree_matrix, num_steps, num_sims):
"""Computes omega using random walks
Omega has shape (batch size, k, num sims)
"""
P = tf.tile(B, (num_sims, ))
Ps = [P]
for i in range(num_steps):
x = tf.random.uniform((num_sims, ))
x = tf.repeat(x, B.shape[0])
degrees = tf.gather(degree_matrix, P)
neighbor_indices = tf.cast(tf.floor(x * tf.cast(degrees, 'float32')),
'int32')
adj_indices = tf.transpose(tf.stack([P, neighbor_indices]))
P_ = tf.gather_nd(adj_list, adj_indices)
Ps.append(P_)
P = P_
return tf.reshape(tf.transpose(tf.stack(Ps)),
(B.shape[0], num_steps + 1, num_sims))
def function_with_pending_points(x: TensorType) -> TensorType:
# stuff below is quite tricky, and thus deserves an elaborate comment
# we receive unknown number N of batches to evaluate
# and need to collect batch of B new points
# so the shape of `x` is [N, B, D]
# we want to add P pending points to each batch
# so that acquisition actually receives N batches of shape [P+B, D] each
# therefore here we prepend each batch with all pending points
# resulting a shape [N, P+B, D]
# we do that by repeating pending points N times and concatenating with x
# pending points are 2D, we need 3D and repeat along first axis
expanded = tf.expand_dims(pending_points, axis=0)
pending_points_repeated = tf.repeat(expanded,
[tf.shape(x)[0]],
axis=0)
all_points = tf.concat([pending_points_repeated, x],
axis=1)
return cast(AcquisitionFunction,
self._acquisition_function)(all_points)
def rbf_neighborhood(datacube):
datacube_reshape = tf.reshape(datacube,
(datacube.shape[0], datacube.shape[1] *
datacube.shape[2], datacube.shape[3]))
centers = datacube[:,
int((datacube.shape[1] - 1) / 2),
int((datacube.shape[2] - 1) / 2), :]
centers_repmat = tf.repeat(centers[:, tf.newaxis, :],
datacube_reshape.shape[1],
axis=1)
dists = tf.math.pow(
tf.math.reduce_euclidean_norm(datacube_reshape - centers_repmat,
axis=2), 2)
betas = (1 / (tf.math.reduce_mean(dists)))**1
betas = matlib.repmat(betas, dists.shape[0], dists.shape[1])
return np.squeeze(betas)
def test_DotProductAttention(self):
"""
query and key are vectors of same length
Returns:
"""
query_size = key_size = 2
queries, keys = tf.random.normal(shape=(2, 1, query_size)), tf.ones(
(2, 10, key_size))
# The two value matrices in the `values` minibatch are identical
values = tf.repeat(tf.reshape(tf.range(40, dtype=tf.float32),
shape=(1, 10, 4)),
repeats=2,
axis=0)
valid_lens = tf.constant([2, 6])
attention = DotProductAttention(
# num_hiddens=8,
dropout=0.1)
output = attention(queries, keys, values, valid_lens, training=False)
# output.shape
self.assertTrue(output.shape == [2, 1, 4])
def costs_decorated_to_logits(cls, symbol_costs_decorated, questions):
symbol_costs_valid = tf.ragged.boolean_mask(
symbol_costs_decorated['costs'], symbol_costs_decorated['valid'])
questions_valid = tf.ragged.boolean_mask(
questions, symbol_costs_decorated['valid'])
logits_valid = cls.costs_to_logits(symbol_costs_valid, questions_valid)
index_mask = tf.repeat(symbol_costs_decorated['valid'],
questions.row_lengths())
indices_all = tf.range(questions.row_splits[-1])
indices_valid = indices_all[index_mask]
indices_valid = tf.expand_dims(indices_valid, axis=1)
# tf.debugging.assert_equal(indices_valid.shape[:-1], logits_valid.flat_values.shape)
# We assign the value 0 to the invalid logits. `scatter_nd` defaults the output values to 0.
logits_values = tf.scatter_nd(
indices_valid, logits_valid.flat_values,
tf.reshape(questions.row_splits[-1], (1, )))
logits = tf.RaggedTensor.from_row_splits(logits_values,
questions.row_splits)
tf.debugging.assert_equal(logits.row_splits, questions.row_splits)
return logits
def propagating(self,
x,
mask=None,
pos=None,
past=None,
training=False,
softmax=True):
p1, p2 = (tf.unstack(past, axis=1) if past is not None else [None] *
self.param['n_layer']), []
t1 = pos if pos is not None else tf.repeat(
[past.shape[3]], x.shape[0], 0) if past is not None else None
x1 = self.embedding.propagating(x, None, t1, training)
for i1 in range(self.param['n_layer']):
x1, a1 = self.encoder[i1].propagating(x1, mask, p1[i1], training)
p2.append(a1)
x1, h1 = self.norm(x1), tf.stack(p2, 1)
x2 = tf.matmul(x1, self.embedding.emb, transpose_b=True)
return tf.nn.softmax(x2) if softmax else x2, x1, h1
def add_selfloops_edge(edge_index,
num_nodes,
edge_weight=None,
fill_weight=1.0):
diagnal_edge_index = tf.reshape(
tf.repeat(tf.range(num_nodes, dtype=intx()), 2), [num_nodes, 2])
updated_edge_index = tf.concat([edge_index, diagnal_edge_index],
axis=0)
if edge_weight:
diagnal_edge_weight = tf.cast(tf.fill([num_nodes], fill_weight),
dtype=floatx())
updated_edge_weight = tf.concat([edge_weight, diagnal_edge_weight],
axis=0)
else:
updated_edge_weight = None
return updated_edge_index, updated_edge_weight
def add_self_loops_indices(indices, n_nodes=None):
"""
Given the indices of a square SparseTensor, adds the diagonal entries (i, i)
and returns the reordered indices.
:param indices: Tensor of rank 2, the indices to a SparseTensor.
:param n_nodes: the size of the n_nodes x n_nodes SparseTensor indexed by
the indices. If `None`, n_nodes is calculated as the maximum entry in the
indices plus 1.
:return: Tensor of rank 2, the indices to a SparseTensor.
"""
n_nodes = tf.reduce_max(indices) + 1 if n_nodes is None else n_nodes
row, col = indices[..., 0], indices[..., 1]
mask = tf.ensure_shape(row != col, row.shape)
sl_indices = tf.range(n_nodes, dtype=row.dtype)[:, None]
sl_indices = tf.repeat(sl_indices, 2, -1)
indices = tf.concat((indices[mask], sl_indices), 0)
dummy_values = tf.ones_like(indices[:, 0])
indices, _ = gen_sparse_ops.sparse_reorder(indices, dummy_values,
(n_nodes, n_nodes))
return indices
def get_vector_elements_equalities_matrix_op(vector_op):
"""
Given a vector_op, return a square matrix such that element (i,j) is 1 if
vector_op[i] == vector_op[j] and 0 otherwise
:param vector_op: 1D tensor of ints
:return: 2D matrix of ints
"""
elements_count_op = tf.shape(vector_op)[0]
# Unroll vector so that each element can be matched with each other element
vector_repeated_elements_wise = tf.repeat(vector_op, repeats=elements_count_op)
vector_repeated_vector_wise = tf.tile(vector_op, multiples=[elements_count_op])
# Compute equalities, cast booleans to floats
equalities_vector_op = tf.cast(vector_repeated_elements_wise == vector_repeated_vector_wise, tf.float32)
# Reshape vector to square matrix
return tf.reshape(equalities_vector_op, shape=(elements_count_op, elements_count_op))
def call(self, x):
batch_size = tf.shape(x)[0]
# Expand the class token batch number of times.
class_token = tf.repeat(self.cls, repeats=batch_size, axis=0)
# Concat the input with the trainable class token.
x = tf.concat([class_token, x], axis=1)
# Apply attention to x.
x = self.layer_norm1(x)
x, viz_weights = self.attention(query=x[:, 0:1],
key=x,
value=x,
return_attention_scores=True)
class_token = class_token + x
class_token = self.layer_norm2(class_token)
class_token = self.flatten(class_token)
class_token = self.layer_norm3(class_token)
class_token = class_token + self.mlp(class_token)
# Build the logits
logits = self.dense(class_token)
return logits, tf.squeeze(viz_weights)[..., 1:]
def histmatch(x, y, nbins=256):
x_shape = tf.shape(x)
y_shape = tf.shape(y)
ch = x_shape[-1]
size = x_shape[0] * x_shape[1]
x_flat = tf.transpose(tf.reshape(x, [-1, ch]), [1, 0])
y_flat = tf.transpose(tf.reshape(y, [-1, ch]), [1, 0])
x_min = tf.reduce_min(x_flat, -1)[..., None]
x_max = tf.reduce_max(x_flat, -1)[..., None]
y_min = tf.reduce_min(y_flat, -1)[..., None]
y_max = tf.reduce_max(y_flat, -1)[..., None]
x_flat = (x_flat - x_min) / (x_max - x_min)
y_flat = (y_flat - y_min) / (y_max - y_min)
x_bins = tf.minimum(tf.cast(x_flat * tf.cast(nbins, tf.float32), tf.int32),
nbins - 1)
x_hist = tf.math.bincount(x_bins, minlength=nbins, axis=-1)
x_cdf = tf.cumsum(x_hist, -1)
x_cdf = tf.cast(x_cdf, tf.float32)
x_cdf = x_cdf / x_cdf[:, -1, None]
y_bins = tf.minimum(tf.cast(y_flat * tf.cast(nbins, tf.float32), tf.int32),
nbins - 1)
y_hist = tf.math.bincount(y_bins, minlength=nbins, axis=-1)
y_cdf = tf.cumsum(y_hist, -1)
y_cdf = tf.cast(y_cdf, tf.float32)
y_cdf = y_cdf / y_cdf[:, -1, None]
x_match = x_cdf[:, :, None]
y_match = y_cdf[:, None, :]
where = (y_match >= x_match)
amax = tf.argmax(where, -1)
indices = tf.stack([tf.repeat(tf.range(ch)[..., None], size, -1), x_bins],
-1)
corr = tf.gather_nd(amax, indices)
corr = tf.cast(corr, tf.float32) / tf.cast(nbins, tf.float32)
corr = corr * (y_max - y_min) + y_min
corr = tf.reshape(tf.transpose(corr, [1, 0]), x_shape)
return corr
def get_delta_errors_fixed_size(y_true, all_but_one_auxiliary_outputs,
error_with_all_features, loss_function,
log_transform, math_ops):
n_out = math_ops.int_shape(all_but_one_auxiliary_outputs[0])[1]
n_feats = len(all_but_one_auxiliary_outputs)
x = tf.stack(all_but_one_auxiliary_outputs,
axis=1) #(batch, n_feats, n_out)
y_true_tile = tf.reshape(tf.repeat(y_true, n_feats, axis=1),
(-1, n_feats, n_out)) # (batch, n_feats, n_out)
error_without_one_feature = safe_evaluate_custom_loss_function(
loss_function, y_true_tile, x, math_ops) # (batch, n_feats, n_out)
error_with_all_features = tf.reshape(error_with_all_features,
(-1, n_out)) # (batch, n_out)
delta_errors = math_ops.maximum(
error_without_one_feature - error_with_all_features,
math_ops.epsilon())
# delta_errors = error_without_one_feature - error_with_all_features
# delta_errors = delta_errors - math_ops.reduce_min(delta_errors, axis=1) + math_ops.epsilon()
if log_transform:
delta_errors = math_ops.log(1 + delta_errors)
# ---
# delta_errors = []
# for all_but_one_auxiliary_output in all_but_one_auxiliary_outputs:
# error_without_one_feature = safe_evaluate_custom_loss_function(
# loss_function, y_true, all_but_one_auxiliary_output, math_ops
# )
# # The error without the feature is an indicator as to how potent the left-out feature is as a predictor.
# delta_error = math_ops.maximum(error_without_one_feature - error_with_all_features, math_ops.epsilon())
# if log_transform:
# delta_error = math_ops.log(1 + delta_error)
# delta_errors.append(delta_error)
# delta_errors = math_ops.stack(delta_errors, axis=-1)
return delta_errors
def transform(image, inv_mat, image_shape):
'''Used in random_rotate
Adapted from https://www.kaggle.com/xiejialun/gridmask-data-augmentation-with-tensorflow/notebook'''
h, w, c = image_shape
cx, cy = w // 2, h // 2
new_xs = tf.repeat(tf.range(-cx, cx, 1), h)
new_ys = tf.tile(tf.range(-cy, cy, 1), [w])
new_zs = tf.ones([h * w], dtype=tf.int32)
old_coords = tf.matmul(
inv_mat, tf.cast(tf.stack([new_xs, new_ys, new_zs]), tf.float32))
old_coords_x, old_coords_y = tf.round(old_coords[0, :] +
w // 2), tf.round(old_coords[1, :] +
h // 2)
clip_mask_x = tf.logical_or(old_coords_x < 0, old_coords_x > w - 1)
clip_mask_y = tf.logical_or(old_coords_y < 0, old_coords_y > h - 1)
clip_mask = tf.logical_or(clip_mask_x, clip_mask_y)
old_coords_x = tf.boolean_mask(old_coords_x, tf.logical_not(clip_mask))
old_coords_y = tf.boolean_mask(old_coords_y, tf.logical_not(clip_mask))
new_coords_x = tf.boolean_mask(new_xs + cx, tf.logical_not(clip_mask))
new_coords_y = tf.boolean_mask(new_ys + cy, tf.logical_not(clip_mask))
old_coords = tf.cast(tf.stack([old_coords_y, old_coords_x]), tf.int32)
new_coords = tf.cast(tf.stack([new_coords_y, new_coords_x]), tf.int64)
rotated_image_values = tf.gather_nd(image, tf.transpose(old_coords))
rotated_image_channel = list()
for i in range(c):
vals = rotated_image_values[:, i]
sparse_channel = tf.SparseTensor(tf.transpose(new_coords), vals,
[h, w])
rotated_image_channel.append(
tf.sparse.to_dense(sparse_channel,
default_value=0,
validate_indices=False))
return tf.transpose(tf.stack(rotated_image_channel), [1, 2, 0])
def _train_step(self, times, sequences, optimizer):
with tf.GradientTape() as tape:
params = self.encoder((times, sequences), training=True)
times_list = tf.expand_dims(times.values, axis=-1)
params_list = tf.repeat(params, times.row_lengths(), axis=0)
latents = self.trajectory((times_list, params_list))
reconstructed_sequences_seq = self.decoder(latents, training=True)
reconstructed_sequences = tf.RaggedTensor.from_row_splits(
reconstructed_sequences_seq, row_splits=times.row_splits)
loss, reconstruction_error, regularisation = self._loss(
sequences, reconstructed_sequences, params)
heart_rates = tf.exp(params[:, 0]) * 60
# update model weights
variables = self.trainable_variables
grads = tape.gradient(loss, variables)
optimizer.apply_gradients(zip(grads, variables))
# logging
self._train_metrics['loss'](loss)
self._train_metrics['reconstruction_error'](reconstruction_error)
self._train_metrics['regularisation'](regularisation)
self._train_metrics['heart_rate'](heart_rates)
self._train_metrics['regularisation_strength'](self._sigma_squared)
# reconstruction_stddev = tf.reduce_mean(tf.math.reduce_std(reconstructed_sequences, axis=1, keepdims=False).values)
means = tf.expand_dims(tf.reduce_mean(reconstructed_sequences, axis=1),
axis=1)
reconstruction_var = tf.reduce_mean(
tf.reduce_mean((reconstructed_sequences - means)**2, axis=1))
reconstruction_stddev = tf.sqrt(reconstruction_var)
self._train_metrics['reconstruction_stddev'](reconstruction_stddev)
# update regularisation weight
update_rate = 0.99
self._sigma_squared.assign(update_rate * self._sigma_squared +
(1 - update_rate) *
tf.reduce_mean(reconstruction_error))
def rotate_img(image, dim, n_channels, degrees):
'''
Rotates an image by `degrees` degrees
Args:
image: array of one squared image of size (dim, dim, n_channels)
n_channels: (int)
degrees: (int or float) the angle of the rotation in degrees
Returns:
An image rotated by an angle of `degrees`
'''
xdim = dim%2
degrees = degrees * tf.ones([1], dtype='float32')
rotation = 3.141593 * degrees / 180.
def get_3x3_mat(lst):
return tf.reshape(tf.concat([lst],axis=0), [3,3])
c1 = tf.math.cos(rotation)
s1 = tf.math.sin(rotation)
one = tf.constant([1],dtype='float32')
zero = tf.constant([0],dtype='float32')
m = get_3x3_mat([c1, s1, zero,
-s1, c1, zero,
zero, zero, one])
# list destination pixel indices
x = tf.repeat(tf.range(dim//2, -dim//2,-1), dim)
y = tf.tile(tf.range(-dim//2, dim//2), [dim])
z = tf.ones([dim*dim], dtype='int32')
idx = tf.stack( [x,y,z] )
# rotate destination pixels onto origin pixels
idx2 = K.dot(m, tf.cast(idx, dtype='float32'))
idx2 = K.cast(idx2, dtype='int32')
idx2 = K.clip(idx2, -dim//2+xdim+1, dim//2)
# find origin pixel values
idx3 = tf.stack([dim//2-idx2[0,], dim//2-1+idx2[1,]])
d = tf.gather_nd(image, tf.transpose(idx3))
return tf.reshape(d,[dim, dim, n_channels])
def cem_planner(state, num_actions, horizon, proposals, topk, iterations,
imagine, objective):
dtype = prec.global_policy().compute_dtype
B, P = list(state.values())[0].shape[0], proposals
H, A = horizon, num_actions
flat_state = {k: tf.repeat(v, P, 0) for k, v in state.items()}
mean = tf.zeros((B, H, A), dtype)
std = tf.ones((B, H, A), dtype)
for _ in range(iterations):
proposals = tf.random.normal((B, P, H, A), dtype=dtype)
proposals = proposals * std[:, None] + mean[:, None]
proposals = tf.clip_by_value(proposals, -1, 1)
flat_proposals = tf.reshape(proposals, (B * P, H, A))
states = imagine(flat_proposals, flat_state)
scores = objective(states)
scores = tf.reshape(tf.reduce_sum(scores, -1), (B, P))
_, indices = tf.math.top_k(scores, topk, sorted=False)
best = tf.gather(proposals, indices, axis=1, batch_dims=1)
mean, var = tf.nn.moments(best, 1)
std = tf.sqrt(var + 1e-6)
return mean[:, 0, :]
def add_noise(audio, noises=None, scale=0.5):
if noises is not None:
# Create a random tensor of the same size as audio ranging from
# 0 to the number of noise stream samples that we have.
tf_rnd = tf.random.uniform((tf.shape(audio)[0], ),
0,
noises.shape[0],
dtype=tf.int32)
noise = tf.gather(noises, tf_rnd, axis=0)
# Get the amplitude proportion between the audio and the noise
prop = tf.math.reduce_max(audio, axis=1) / tf.math.reduce_max(
noise, axis=1)
prop = tf.repeat(tf.expand_dims(prop, axis=1),
tf.shape(audio)[1],
axis=1)
# Adding the rescaled noise to audio
audio = audio + noise * prop * scale
return audio
def body(i, batch_size, outputs, encoder_masks, encoder_hidden_states,
durations_gt, max_durations):
repeats = durations_gt[i]
real_length = tf.reduce_sum(repeats)
pad_size = max_durations - real_length
masks = tf.sequence_mask([real_length],
max_durations,
dtype=tf.int32)
repeat_encoder_hidden_states = tf.repeat(encoder_hidden_states[i],
repeats=repeats,
axis=0)
repeat_encoder_hidden_states = tf.expand_dims(
tf.pad(repeat_encoder_hidden_states, [[0, pad_size], [0, 0]]),
0) # [1, max_durations, hidden_size]
outputs = tf.concat([outputs, repeat_encoder_hidden_states],
axis=0)
encoder_masks = tf.concat([encoder_masks, masks], axis=0)
return [
i + 1, batch_size, outputs, encoder_masks,
encoder_hidden_states, durations_gt, max_durations
]
def broadcast_wff_and_mask(wff: Formula, mask: Formula) -> Formula:
"""Broadcast the wff to include all vars in mask; put the vars of the mask in the first axes"""
wff = wff._copy()
# 1. broadcast wff with vars that are in the mask but not yet in the formula
mask_vars_not_in_wff = [
var for var in mask.free_vars if var not in wff.free_vars
]
for var in mask_vars_not_in_wff:
new_idx = len(wff.free_vars)
wff.tensor = tf.expand_dims(wff.tensor, axis=new_idx)
wff.tensor = tf.repeat(wff.tensor,
mask._get_dim_of_free_var(var),
axis=new_idx)
wff.free_vars.append(var)
# 2. transpose wff so that the masked vars on the first axes
vars_not_in_mask = [
var for var in wff.free_vars if var not in mask.free_vars
]
wff = transpose_free_vars(wff,
new_var_order=mask.free_vars + vars_not_in_mask)
return wff
def init_qstar_mean(self, m, batch_index, batch_num):
"""Initialize the q0 with mean."""
# use q0=avg(m) (or q0=0)
# batch_shape = ksb.shape(batch_num)
q = tf.math.segment_mean(m, batch_index) # (batch,feat)
# r0
qt = tf.repeat(q, batch_num, axis=0) # (batch*num,feat)
et = self.f_et(m, qt) # (batch*num,)
# get at = exp(et)/sum(et) with sum(et)=norm
at = ksb.exp(et - self.get_scale_per_sample(et, batch_index,
batch_num)) # (batch*num,)
norm = self.get_norm(at, batch_index, batch_num) # (batch*num,)
at = norm * at # (batch*num,) x (batch*num,)
# calculate rt
at = ksb.expand_dims(at, axis=1) # (batch*num,1)
rt = m * at # (batch*num,feat) x (batch*num,1)
rt = tf.math.segment_sum(rt, batch_index) # (batch,feat)
# [q0,r0]
qstar = ksb.concatenate([q, rt], axis=1) # (batch,2*feat)
qstar = ksb.expand_dims(qstar, axis=1) # (batch,1,2*feat)
return qstar