def _infer_network_outputs(*, sess, restored_model, num_of_anchors, anchors,
orig_image_width, orig_image_height,
model_image_width, model_image_height, img_np,
verbose):
start = time.time()
boxes = []
prob_class = []
for yolo_head_idx in range(len(restored_model.output)):
yolo_head = restored_model.output[yolo_head_idx]
yolo_head_shape = K.shape(yolo_head)
yolo_head_num_of_cols, yolo_head_num_of_rows = yolo_head_shape[
2], yolo_head_shape[1]
curr_yolo_head = K.reshape(yolo_head, [
-1, yolo_head_num_of_cols, yolo_head_num_of_rows, num_of_anchors,
NUM_OF_BOX_PARAMS + NUM_OF_CLASSES
])
grid = construct_grid(yolo_head_shape[1], yolo_head_shape[2])
grid = K.cast(grid, dtype=K.dtype(curr_yolo_head))
grid_size = K.cast([yolo_head_num_of_cols, yolo_head_num_of_rows],
dtype=K.dtype(curr_yolo_head))
curr_boxes_xy = (K.sigmoid(curr_yolo_head[..., :2]) + grid) / grid_size
curr_boxes_wh = K.exp(curr_yolo_head[...,
2:4]) * anchors[yolo_head_idx]
curr_prob_obj = K.sigmoid(curr_yolo_head[..., 4:5])
curr_prob_class = K.sigmoid(curr_yolo_head[..., 5:])
curr_prob_detected_class = curr_prob_obj * curr_prob_class
boxes.append(
get_corrected_boxes(box_width=curr_boxes_wh[..., 0:1],
box_height=curr_boxes_wh[..., 1:2],
box_x=curr_boxes_xy[..., 0:1],
box_y=curr_boxes_xy[..., 1:2],
orig_image_shape=(orig_image_width,
orig_image_height),
model_image_shape=(model_image_width,
model_image_height)))
curr_prob_detected_class = K.reshape(curr_prob_detected_class,
[-1, NUM_OF_CLASSES])
prob_class.append(curr_prob_detected_class)
prob_class = K.concatenate(prob_class, axis=0)
boxes = K.concatenate(boxes, axis=0)
out_tensors = [
boxes,
prob_class,
]
if verbose:
print(f'Took {time.time() - start} seconds to construct network.')
start = time.time()
sess_out = sess.run(out_tensors,
feed_dict={
restored_model.input: img_np,
K.learning_phase(): 0
})
if verbose:
print(
f'Took {time.time() - start} seconds to infer outputs in session.')
boxes, out_boxes_classes = sess_out
return boxes, out_boxes_classes
def custom_loss(y_true, y_pred):
# define a grid of offsets
# [[[ 0. 0.]]
# [[ 1. 0.]]
# [[ 0. 1.]]
# [[ 1. 1.]]]
grid = np.array([[[float(x), float(y)]] * nb_boxes for y in range(grid_h)
for x in range(grid_w)])
# first three values are classes : cat, rat, and none.
# However yolo doesn't predict none as a class, none is everything else and is just not predicted
# so I don't use it in the loss
y_true_class = y_true[..., 0:2]
y_pred_class = y_pred[..., 0:2]
# reshape array as a list of grid / grid cells / boxes / of 5 elements
pred_boxes = K.reshape(y_pred[..., 3:], (-1, grid_w * grid_h, nb_boxes, 5))
true_boxes = K.reshape(y_true[..., 3:], (-1, grid_w * grid_h, nb_boxes, 5))
# sum coordinates of center of boxes with cell offsets.
# as pred boxes are limited to 0 to 1 range, pred x,y + offset is limited to predicting elements inside a cell
y_pred_xy = pred_boxes[..., 0:2] + K.variable(grid)
# w and h predicted are 0 to 1 with 1 being image size
y_pred_wh = pred_boxes[..., 2:4]
# probability that there is something to predict here
y_pred_conf = pred_boxes[..., 4]
# same as predicate except that we don't need to add an offset, coordinate are already between 0 and cell count
y_true_xy = true_boxes[..., 0:2]
# with and height
y_true_wh = true_boxes[..., 2:4]
# probability that there is something in that cell. 0 or 1 here as it's a certitude.
y_true_conf = true_boxes[..., 4]
clss_loss = K.sum(K.square(y_true_class - y_pred_class), axis=-1)
xy_loss = K.sum(K.sum(K.square(y_true_xy - y_pred_xy), axis=-1) *
y_true_conf,
axis=-1)
wh_loss = K.sum(
K.sum(K.square(K.sqrt(y_true_wh) - K.sqrt(y_pred_wh)), axis=-1) *
y_true_conf,
axis=-1)
# when we add the confidence the box prediction lower in quality but we gain the estimation of the quality of the box
# however the training is a bit unstable
# compute the intersection of all boxes at once (the IOU)
intersect_wh = K.maximum(K.zeros_like(y_pred_wh),
(y_pred_wh + y_true_wh) / 2 -
K.square(y_pred_xy - y_true_xy))
intersect_area = intersect_wh[..., 0] * intersect_wh[..., 1]
true_area = y_true_wh[..., 0] * y_true_wh[..., 1]
pred_area = y_pred_wh[..., 0] * y_pred_wh[..., 1]
union_area = pred_area + true_area - intersect_area
iou = intersect_area / union_area
conf_loss = K.sum(K.square(y_true_conf * iou - y_pred_conf), axis=-1)
# final loss function
d = xy_loss + wh_loss + conf_loss + clss_loss
if False:
d = tf.Print(d, [d], "loss")
d = tf.Print(d, [xy_loss], "xy_loss")
d = tf.Print(d, [wh_loss], "wh_loss")
d = tf.Print(d, [clss_loss], "clss_loss")
d = tf.Print(d, [conf_loss], "conf_loss")
return d
def call(self, inputs, **kwargs):
"""
Creates the layer as a Keras graph
Notes:
This does not add self loops to the adjacency matrix.
Args:
inputs (list): list of inputs with 4 items:
node features (size b x N x F),
sparse graph adjacency matrix (size N x N),
where N is the number of nodes in the graph,
F is the dimensionality of node features
M is the number of output nodes
"""
X = inputs[0] # Node features (1 x N x F)
A_sparse = inputs[1] # Adjacency matrix (1 x N x N)
if not isinstance(A_sparse, tf.SparseTensor):
raise TypeError("A is not sparse")
# Get undirected graph edges (E x 2)
A_indices = A_sparse.indices
batch_dim, n_nodes, _ = K.int_shape(X)
if batch_dim != 1:
raise ValueError(
"Currently full-batch methods only support a batch dimension of one"
)
else:
# Remove singleton batch dimension
X = K.squeeze(X, 0)
outputs = []
for head in range(self.attn_heads):
kernel = self.kernels[head] # W in the paper (F x F')
attention_kernel = self.attn_kernels[
head] # Attention kernel a in the paper (2F' x 1)
# Compute inputs to attention network
features = K.dot(X, kernel) # (N x F')
# Compute feature combinations
# Note: [[a_1], [a_2]]^T [[Wh_i], [Wh_j]] = [a_1]^T [Wh_i] + [a_2]^T [Wh_j]
attn_for_self = K.dot(
features, attention_kernel[0]) # (N x 1), [a_1]^T [Wh_i]
attn_for_neighs = K.dot(
features, attention_kernel[1]) # (N x 1), [a_2]^T [Wh_j]
# Create sparse attention vector (All non-zero values of the matrix)
sparse_attn_self = tf.gather(K.reshape(attn_for_self, [-1]),
A_indices[:, 0],
axis=0)
sparse_attn_neighs = tf.gather(K.reshape(attn_for_neighs, [-1]),
A_indices[:, 1],
axis=0)
attn_values = sparse_attn_self + sparse_attn_neighs
# Add nonlinearity
attn_values = LeakyReLU(alpha=0.2)(attn_values)
# Apply dropout to features and attention coefficients
dropout_feat = Dropout(self.in_dropout_rate)(features) # (N x F')
dropout_attn = Dropout(self.attn_dropout_rate)(
attn_values) # (N x N)
# Convert to sparse matrix
sparse_attn = tf.sparse.SparseTensor(
A_indices, values=dropout_attn, dense_shape=[n_nodes, n_nodes])
# Apply softmax to get attention coefficients
sparse_attn = tf.sparse.softmax(
sparse_attn) # (N x N), Eq. 3 of the paper
# Linear combination with neighbors' features [YT: see Eq. 4]
node_features = tf.sparse.sparse_dense_matmul(
sparse_attn, dropout_feat) # (N x F')
if self.use_bias:
node_features = K.bias_add(node_features, self.biases[head])
# Add output of attention head to final output
outputs.append(node_features)
# Aggregate the heads' output according to the reduction method
if self.attn_heads_reduction == "concat":
output = K.concatenate(outputs) # (N x KF')
else:
output = K.mean(K.stack(outputs), axis=0) # N x F')
output = self.activation(output)
# Add batch dimension back if we removed it
if batch_dim == 1:
output = K.expand_dims(output, 0)
return output
def __call__(self, shape, dtype=None, partition_info=None):
w = self.base_initializer(shape=shape, dtype=dtype)
u = K.random_uniform(shape=tuple([1, shape[-1]]), dtype=dtype)
w_bar, _u, sigma = spectral_normalization(w, u, self.niter_spectral)
w_bar = bjorck_normalization(w_bar, self.niter_bjorck)
return K.reshape(w_bar, shape)
def call(self, x, mask=None): a = KB.permute_dimensions(x, (1,0,2)) a = KB.reshape(a, (x.shape[1] *x.shape[0], x.shape[2])) a = ifft(a) a = KB.reshape(a, (x.shape[1], x.shape[0], x.shape[2])) return KB.permute_dimensions(a, (1,0,2))
def yolo_head(feats, anchors, num_classes):
"""Convert final layer features to bounding box parameters.
Parameters
----------
feats : tensor
Final convolutional layer features.
anchors : array-like
Anchor box widths and heights.
num_classes : int
Number of target classes.
Returns
-------
box_xy : tensor
x, y box predictions adjusted by spatial location in conv layer.
box_wh : tensor
w, h box predictions adjusted by anchors and conv spatial resolution.
box_conf : tensor
Probability estimate for whether each box contains any object.
box_class_pred : tensor
Probability distribution estimate for each box over class labels.
"""
num_anchors = len(anchors)
# Reshape to batch, height, width, num_anchors, box_params.
anchors_tensor = K.reshape(K.variable(anchors), [1, 1, 1, num_anchors, 2])
# Static implementation for fixed models.
# TODO: Remove or add option for static implementation.
# _, conv_height, conv_width, _ = K.int_shape(feats)
# conv_dims = K.variable([conv_width, conv_height])
# Dynamic implementation of conv dims for fully convolutional model.
conv_dims = K.shape(feats)[1:3] # assuming channels last
# In YOLO the height index is the inner most iteration.
conv_height_index = K.arange(0, stop=conv_dims[0])
conv_width_index = K.arange(0, stop=conv_dims[1])
conv_height_index = K.tile(conv_height_index, [conv_dims[1]])
# TODO: Repeat_elements and tf.split doesn't support dynamic splits.
# conv_width_index = K.repeat_elements(conv_width_index, conv_dims[1], axis=0)
conv_width_index = K.tile(K.expand_dims(conv_width_index, 0),
[conv_dims[0], 1])
conv_width_index = K.flatten(K.transpose(conv_width_index))
conv_index = K.transpose(K.stack([conv_height_index, conv_width_index]))
conv_index = K.reshape(conv_index, [1, conv_dims[0], conv_dims[1], 1, 2])
conv_index = K.cast(conv_index, K.dtype(feats))
feats = K.reshape(
feats, [-1, conv_dims[0], conv_dims[1], num_anchors, num_classes + 5])
conv_dims = K.cast(K.reshape(conv_dims, [1, 1, 1, 1, 2]), K.dtype(feats))
# Static generation of conv_index:
# conv_index = np.array([_ for _ in np.ndindex(conv_width, conv_height)])
# conv_index = conv_index[:, [1, 0]] # swap columns for YOLO ordering.
# conv_index = K.variable(
# conv_index.reshape(1, conv_height, conv_width, 1, 2))
# feats = Reshape(
# (conv_dims[0], conv_dims[1], num_anchors, num_classes + 5))(feats)
box_xy = K.sigmoid(feats[..., :2])
box_wh = K.exp(feats[..., 2:4])
box_confidence = K.sigmoid(feats[..., 4:5])
box_class_probs = K.softmax(feats[..., 5:])
# Adjust preditions to each spatial grid point and anchor size.
# Note: YOLO iterates over height index before width index.
box_xy = (box_xy + conv_index) / conv_dims
box_wh = box_wh * anchors_tensor / conv_dims
return box_xy, box_wh, box_confidence, box_class_probs
def gather_nd_reshape(t, indices, final_shape):
h = tf.gather_nd(t, indices)
return K.reshape(h, final_shape)
def yolo_loss(inputs, num_anchors):
ignore_thresh = .5 # Порог вероятности обнаружения объекта
num_layers = num_anchors // 3 # Подсчитываем количество анкоров на каждом уровне сетки
y_pred = inputs[:num_layers] # Из входных данных выцепляем посчитанные моделью значения
y_true = inputs[num_layers:] # Из входных данных выцепляем эталонные значения
anchor_mask = [[6, 7, 8], [3, 4, 5], [0, 1, 2]] # Задаем маску анкоров для каждого уровня сеток
# Получаем размерность входного изображения ( (13 х 13) * 32 = (416 х 416)) и приводим к типу элемента y_true[0]
input_shape = K.cast(K.shape(y_pred[0])[1:3] * 32, K.dtype(y_true[0]))
# Получаем двумерный массив, соответствующий размерностям сеток ((13, 13), (26, 26), (52, 52))
grid_shapes = [K.cast(K.shape(y_pred[l])[1:3], K.dtype(y_true[0])) for l in range(num_layers)]
loss = 0 # Значение ошибки
# Считываем количество элементов
m = K.shape(y_pred[0])[0] # Размер пакета
batch_size = K.cast(m, K.dtype(y_pred[0])) # Преобразуем к типу y_pred[0]
for l in range(num_layers): # Пробегаем по всем трем уровням сеток
# Получаем маску для сетки l-го уровня по вероятности определения объекта (5-ый параметр в списке общих параметров).
# В массиве object_mask будут значения, которые соответствуют только вероятности обнаружения объекта
object_mask = y_true[l][..., 4:5] # Вернется набор данных вида ([0][0][0][0]...[1]...[0])
# Получаем аналогичную выборку для сетки l-го уровня с OHE (где записана позиция нашего класса)
# В массиве true_class будут значения, которые соответствуют только OHE представлению класса для данного уровня анкоров
true_class = y_true[l][..., 5:] # Вернется набор данных вида ([0][0][0][0]...[1]...[0])
num_sub_anchors = len(anchors[anchor_mask[l]]) # Получаем количество анкоров для отдельного уровян сетки (3)
# Решейпим анкоры отдельного уровня сетки и записываем в переменную anchors_tensor
anchors_tensor = K.reshape(K.constant(anchors[anchor_mask[l]]), [1, 1, 1, num_sub_anchors, 2])
# Создаем двумерный массив grid со значениями [[[0, 0] , [0, 1] , [0, 2] , ... , [0, k]],
# [[1, 0] , [1, 1] , [1, 2] , ... , [1 ,k]],
# ...
# [[k, 0] , [k, 1] , [k, 2] , ... , [k, k]]]
# где k - размерность сетки. Массив хранит индексы ячеек сетки
grid_shape = K.shape(y_pred[l])[1:3] # Получаем ширину и высоту сетки
grid_y = K.tile(K.reshape(K.arange(0, stop=grid_shape[0]), [-1, 1, 1, 1]),[1, grid_shape[1], 1, 1]) # Создаем вертикальную линию
grid_x = K.tile(K.reshape(K.arange(0, stop=grid_shape[1]), [1, -1, 1, 1]),[grid_shape[0], 1, 1, 1]) # Создаем горизонтальную линию
grid = K.concatenate([grid_x, grid_y]) # Объединяем
grid = K.cast(grid, K.dtype(y_pred[l])) # Приводим к типу y_pred[l]
# Решейпим y_pred[l]
feats = K.reshape(y_pred[l], [-1, grid_shape[0], grid_shape[1], num_sub_anchors, num_classes + 5])
# Считаем ошибку в определении координат центра объекта
# Получаем координаты центра объекта из спредиктенного значения
pred_xy = (K.sigmoid(feats[..., :2]) + grid) / K.cast(grid_shape[::-1], K.dtype(feats))
# Производим обратные вычисления для оригинальных значений из y_true для координат центра объекта
true_xy = y_true[l][..., :2] * grid_shapes[l][::-1] - grid # Реальные координаты центра bounding_box
box_loss_scale = 2 - y_true[l][...,2:3] * y_true[l][...,3:4] # чем больше бокс, тем меньше ошибка
# binary_crossentropy для истинного значения и спредиктенного (obect_mask для подсчета только требуемого значения)
xy_loss = object_mask * box_loss_scale * K.binary_crossentropy(true_xy, feats[...,0:2], from_logits=True)
# Считаем ошибку в определении координат ширины и высоты
# Получаем значения ширины и высоты изображения из спредиктенного значения
pred_wh = K.exp(feats[..., 2:4]) * anchors_tensor / K.cast(input_shape[::-1], K.dtype(feats))
# Производим обратные вычисления для оригинальных значений из y_true для ширины и высоты объекта
true_wh = K.log(y_true[l][..., 2:4] / anchors[anchor_mask[l]] * input_shape[::-1])
# Оставляем значение высоты и ширины только у тех элементов, где object_mask = 1
true_wh = K.switch(object_mask, true_wh, K.zeros_like(true_wh))
# Считаем значение ошибки в определении высоты и ширины
wh_loss = object_mask * box_loss_scale * 0.5 * K.square(true_wh-feats[...,2:4])
# Объединяем значения в один массив
pred_box = K.concatenate([pred_xy, pred_wh])
# Считаем ошибку в определении обнаружения какого-либо класса
# Для этого вначале надо отсечь все найденные объекты, вероятность которых меньше установленного значения ignore_thresh
# Определяем массив, который будет хранить данные о неподходящих значениях
ignore_mask = tf.TensorArray(K.dtype(y_true[0]), size=1, dynamic_size=True)
object_mask_bool = K.cast(object_mask, 'bool') # Приводим тип object_mask к типу 'bool'
# Функция, определяющая данные, которые требуется игнорировать
# Пробегаем по всем элементам пакета (b<m)
# Получаем параметры реального bounding_box для текущей ячейки
# Считаем IoU реального и спредиктенного
# В зависимости от best_iou < ignore_thresh помечаем его как верно распознанный или неверено
def loop_body(
b,
ignore_mask
):
# в true_box запишутся первые 4 параметра (центр, высота и ширина объекта) того элемента, значение которого в object_mask_bool равно True
true_box = tf.boolean_mask(y_true[l][b,...,0:4], object_mask_bool[b,...,0])
# Подсчитываем iou для спредиктенной ограничивающей рамки (pred_box) и оригинальной (true_box)
iou = calc_iou(pred_box[b], true_box)
# Находим лучшую ограничивающую рамку
best_iou = K.max(iou, axis=-1)
# Записываем в ignore_mask true или false в зависимости от (best_iou < ignore_thresh)
ignore_mask = ignore_mask.write(b, K.cast(best_iou < ignore_thresh, K.dtype(true_box)))
return b+1, ignore_mask # Увеличиваем счетчик на единицу и возвращаем ignore_mask
# Пробегаем в цикле по всем элементам в пределах значения m (m = batch size)
_, ignore_mask = tf.while_loop(lambda b,*args: b<m, loop_body, [0, ignore_mask])
ignore_mask = ignore_mask.stack() # Приводим ignore_mask к тензору
ignore_mask = K.expand_dims(ignore_mask, -1) # Добавляем еще одну размерность в конце ignore_mask
# Считаем значение ошибки
# 1 компонент - для значений, которые были верно спредиктены
# 2 компонент - для значения, которые были неверно спредиктены
confidence_loss = (
object_mask * K.binary_crossentropy(object_mask, feats[...,4:5], from_logits=True) +
(1-object_mask) * K.binary_crossentropy(object_mask, feats[...,4:5], from_logits=True) * ignore_mask
)
# Считаем ошибку в определении класса объекта
class_loss = object_mask * K.binary_crossentropy(true_class, feats[...,5:], from_logits=True)
# Считаем суммарную ошибку
xy_loss = K.sum(xy_loss) / batch_size
wh_loss = K.sum(wh_loss) / batch_size
confidence_loss = K.sum(confidence_loss) / batch_size
class_loss = K.sum(class_loss) / batch_size
loss += xy_loss + wh_loss + confidence_loss + class_loss
return loss # Возвращаем значение ошибки
def hw_flatten(x):
# Input shape x: [BATCH, HEIGHT, WIDTH, CHANNELS]
# flat the feature volume across the tensor width and height
shape = K.int_shape(x)
return K.reshape(
x, [shape[0], -1, shape[-1]]) # return [BATCH, W*H, CHANNELS]
def call(self, x):
#print("!",x[0].shape,x[1].shape)
mat = x[0]
val = x[1]
con = val[:, :, :self.keepconst]
var = val[:, :, self.keepconst:]
mata = K.constant(self.getmata(self.gs))
matb = K.constant(self.getmatb(self.gs))
for i in range(self.iterations):
#print("val",val.shape)
valp = K.permute_dimensions(val, (0, 2, 1))
#print("valp",valp.shape)
vA = K.dot(valp, mata)
vB = K.dot(valp, matb)
#print("vA",vA.shape,"vB",vB.shape)
feat = K.permute_dimensions(vA, (0, 2, 1))
diff = K.permute_dimensions(vB - vA, (0, 2, 1))
#print("feat",feat.shape,"diff",diff.shape)
premlp = K.concatenate((feat, diff), axis=-1)
#print("premlp",premlp.shape)
postmlp = premlp[:, :, :self.param - self.keepconst]
#postmlp=self.mlp(premlp)
#print("postmlp",postmlp.shape)
ppmlp = K.permute_dimensions(postmlp, (0, 2, 1))
#print("ppmlp",ppmlp.shape)
res = K.reshape(
ppmlp, (-1, self.param - self.keepconst, self.gs, self.gs))
#print("res",res.shape)
#print("mat",mat.shape)
resp = K.permute_dimensions(res, (1, 0, 2, 3))
presmat = resp * mat #kinda wondering that this(resp*mat) actually works...tja it actually does not
#print("presmat",presmat.shape)
resmat = K.permute_dimensions(presmat, (1, 0, 2, 3))
#resmat=K.permute_dimensions(presmat,(1,2,0,3))
print("resmat", resmat.shape)
#exit()
summ = K.sum(resmat, axis=-1) #/msumtra
print("summ", summ.shape)
#print("summ",summ.shape)
#print("pre",summ.shape)
summ /= self.k
#print("post",summ.shape)
#exit()
#print("permuted from",summ.shape)
#var=K.permute_dimensions(summ,(2,0,1))
var = K.permute_dimensions(
summ, (0, 2, 1)
) #hopefully implemented this into the resmat permute, nope, not diffbar
#print("permuted to",var.shape)
print("var", var.shape)
#print("con",con.shape)
#exit()
if self.activate:
var = self.advrelu(var, self.activation)
val = K.concatenate((con, var), axis=-1)
#print("concatting",con.shape,var.shape,"=>",val.shape)
#print("val",val.shape)
continue
#print("returning",val.shape)
#exit()
return val
def train_lstm(n_symbols, embedding_matrix, config):
#(batch_size,max_group_nums,max_post_nums,max_seq_len)
max_group_nums = config.max_group_nums
max_post_nums = config.max_post_nums
max_seq_len = config.max_seq_len
main_input = Input(shape=(max_group_nums, max_post_nums, max_seq_len))
sub_input = Input(shape=(max_group_nums, social_feature_nums))
embedding_layer_main = Embedding(input_dim=n_symbols,
output_dim=config.embeddingSize,
weights=[embedding_matrix],
input_length=max_seq_len,
mask_zero=True)(main_input)
dropout_layer_1 = Dropout(config.dropoutKeepProb)(embedding_layer_main)
#shape==(batch_size,max_group_nums,max_post_nums,max_seq_len,embeddingSize)
bid_GRU_layer_1 = Bidirectional(
GRU(32,
activation="tanh",
recurrent_dropout=0.5,
return_sequences=True),
merge_mode='concat')(backend.reshape(
dropout_layer_1, shape=[-1, max_seq_len, config.embeddingSize]))
#shape==(batch_size*max_group_nums*max_post_nums,max_seq_len,64)
bn_layer_1 = BatchNormalization()(bid_GRU_layer_1)
attention_layer_1 = AttentionLayer()(
bn_layer_1) #(batch_size*max_group_nums*max_post_nums,64)
bid_GRU_layer_2 = Bidirectional(GRU(32,
activation='tanh',
dropout=0.5,
recurrent_dropout=0.5,
return_sequences=True),
merge_mode='concat')(backend.reshape(
attention_layer_1,
shape=[-1, max_post_nums, 64]))
#shape==(batch_size*max_group_nums,max_post_nums,64)
bn_layer_2 = BatchNormalization()(bid_GRU_layer_2)
attention_layer_2 = AttentionLayer()(
bn_layer_2) #(batch_size*max_group_nums,64)
bid_GRU_layer_3 = Bidirectional(GRU(32,
activation='tanh',
dropout=0.5,
recurrent_dropout=0.5,
return_sequences=True),
merge_mode='concat')(backend.reshape(
attention_layer_2,
shape=[-1, max_group_nums, 64]))
bn_layer_3 = BatchNormalization()(concatenate([bid_GRU_layer_3, sub_input],
axis=2))
attention_layer_3 = AttentionLayer()(
bn_layer_3) #(batch_size,64+social_feature_nums)
dense_layer_1 = Dense(64, activation="tanh")(
sub_input) #(batch_size,max_group_nums,64)
social_attention_layer = AttentionLayer()(dense_layer_1) #(batch_size,64)
merge_layer = concatenate([attention_layer_3, social_attention_layer],
axis=1) #(bathc_size,128+social_attention_layer)
dropout_layer = Dropout(config.dropoutKeepProb)(merge_layer)
output_layer = Dense(2, activation='softmax')(Dense(
32, activation="tanh")(dropout_layer))
model = Model([main_input, sub_input], output_layer)
model.compile(loss='categorical_crossentropy',
optimizer='adam',
metrics=['accuracy'])
return model
def sinkhorn_loss(y_true, y_pred):
y_true = K.cast(y_true, y_pred.dtype)
y_pred = K.reshape(y_pred, (-1,48,1))
y_true = K.reshape(y_true, (-1,48,1))
cc = tf.concat([y_true, y_pred], axis=2)
return K.mean( tf.map_fn(myfunc, cc), axis=(-1) )
def display_yolo(image, yolo_model, score_threshold, iou_threshold,\
train_batch_size =16, grid_h =8, grid_w =8, image_h =256, image_w =256, anchors =[0.57273, 0.677385, 1.87446, 2.06253, 3.33843, 5.47434, 7.88282, 3.52778, 9.77052, 9.16828],\
plot = False):
'''
Display predictions from YOLO model.
Parameters
----------
- file : string list : list of images path.
- yolo_model : YOLO model.
- score_threshold : threshold used for filtering predicted bounding boxes.
- iou_threshold : threshold used for non max suppression.
'''
# load image
input_image = image[:,:,::-1]
input_image = image / 255.
input_image = np.expand_dims(input_image, 0)
# prediction
y_pred = yolo_model.predict_on_batch(input_image)
# post prediction process
# grid coords tensor
coord_x = tf.cast(tf.reshape(tf.tile(tf.range(grid_w), [grid_h]), (1, grid_h, grid_w, 1, 1)), tf.float32)
coord_y = tf.transpose(coord_x, (0,2,1,3,4))
coords = tf.tile(tf.concat([coord_x,coord_y], -1), [train_batch_size, 1, 1, 5, 1])
dims = K.cast_to_floatx(K.int_shape(y_pred)[1:3])
dims = K.reshape(dims,(1,1,1,1,2))
# anchors tensor
anchors = np.array(anchors)
anchors = anchors.reshape(len(anchors) // 2, 2)
# pred_xy and pred_wh shape (m, grid_w, grid_h, Anchors, 2)
pred_xy = K.sigmoid(y_pred[:,:,:,:,0:2])
pred_xy = (pred_xy + coords)
pred_xy = pred_xy / dims
pred_wh = K.exp(y_pred[:,:,:,:,2:4])
pred_wh = (pred_wh * anchors)
pred_wh = pred_wh / dims
# pred_confidence
box_conf = K.sigmoid(y_pred[:,:,:,:,4:5])
# pred_class
box_class_prob = K.softmax(y_pred[:,:,:,:,5:])
# Reshape
pred_xy = pred_xy[0,...]
pred_wh = pred_wh[0,...]
box_conf = box_conf[0,...]
box_class_prob = box_class_prob[0,...]
# Convert box coords from x,y,w,h to x1,y1,x2,y2
box_xy1 = pred_xy - 0.5 * pred_wh
box_xy2 = pred_xy + 0.5 * pred_wh
boxes = K.concatenate((box_xy1, box_xy2), axis=-1)
# Filter boxes
box_scores = box_conf * box_class_prob
box_classes = K.argmax(box_scores, axis=-1) # best score index
box_class_scores = K.max(box_scores, axis=-1) # best score
prediction_mask = box_class_scores >= score_threshold
boxes = tf.boolean_mask(boxes, prediction_mask)
scores = tf.boolean_mask(box_class_scores, prediction_mask)
classes = tf.boolean_mask(box_classes, prediction_mask)
# Scale box to image shape
boxes = boxes * image_h
# Non Max Supression
selected_idx = tf.image.non_max_suppression(boxes, scores, 50, iou_threshold=iou_threshold)
boxes = K.gather(boxes, selected_idx)
scores = K.gather(scores, selected_idx)
classes = K.gather(classes, selected_idx)
if plot:
# Draw image
plt.figure(figsize=(2,2))
f, (ax1) = plt.subplots(1,1, figsize=(10, 10))
ax1.imshow(image[:,:,::-1])
count_detected = boxes.shape[0]
ax1.set_title('Detected objects count : {}'.format(count_detected))
for i in range(count_detected):
box = boxes[i,...]
x = box[0]
y = box[1]
w = box[2] - box[0]
h = box[3] - box[1]
classe = classes[i].numpy()
if classe == 0:
color = (0, 1, 0)
else:
color = (1, 0, 0)
rect = patches.Rectangle((x.numpy(), y.numpy()), w.numpy(), h.numpy(), linewidth = 3, edgecolor=color,facecolor='none')
ax1.add_patch(rect)
return boxes, scores, classes
def testBackEnd():
x = tf.zeros([3, 4], tf.int32)
print('x=', x)
x = tf.zeros((3, 4), tf.int32)
print('x=', x)
return
a = tf.constant([1, 2, 3, 4, 5, 6, 7, 8], dtype=tf.float32)
a = K.reshape(a, (4, 4))
print('a=', a)
a = tf.constant([[1, 2], [3, 4]], dtype=tf.float32)
#a = K.abs(-1)
print('a=', type(a), a) #a= tf.Tensor(1, shape=(), dtype=int32)
# a = a.numpy()
# print('a=',a)
# a = tf.zeros([0, 3])
# a = tf.concat([a, [[1, 2, 3], [5, 6, 8]]], axis=0)
# print('a=',type(a),a)
b = tf.constant([[1, 8], [2, 3]], dtype=tf.float32)
c = K.square(a - b)
print('c=', c)
d = K.sum(c, axis=0)
print('d=', d)
d = K.sum(c, axis=1)
print('d=', d)
d = K.sum(c, axis=[0, 1])
print('d=', d)
return
a = K.abs([-1, 0, 9, -10])
print('a=', a) #a= tf.Tensor([ 1 0 9 10], shape=(4,), dtype=int32)
a = K.abs(np.array([-1, 0, 9, -10]))
print('a=', a) #a= tf.Tensor([ 1 0 9 10], shape=(4,), dtype=int32)
a = K.all(np.array([-1, 0, 9, -10]), axis=0)
print('a=', a) #a= tf.Tensor(False, shape=(), dtype=bool)
a = K.all(np.array([[-1, -2, -1], [-1, 0, 9]]), axis=0) #x axis
print('a=', a) #a= tf.Tensor([ True False True], shape=(3,), dtype=bool)
a = K.all(np.array([[-1, -2, -1], [-1, 0, 9]]), axis=1) #y axis
print('a=', a) #a= tf.Tensor([ True False], shape=(2,), dtype=bool)
a = K.arange(1, 100, 10)
print(
'a=', a
) #a= tf.Tensor([ 1 11 21 31 41 51 61 71 81 91], shape=(10,), dtype=int32)
a = K.sum(np.array([-1, 0, 9, -10]))
print('a=', a) #a= tf.Tensor(-2, shape=(), dtype=int32)
a = K.square(np.array([-1, 0, 9, -10]))
print('a=', a) #a= tf.Tensor([ 1 0 81 100], shape=(4,), dtype=int32)
x = K.placeholder(shape=(2, 3))
y = K.placeholder(shape=(3, 4))
xy = K.dot(x, y)
shape = K.int_shape(xy)
print('xy=', xy) #xy= Tensor("MatMul:0", shape=(2, 4), dtype=float32)
print('xy shape=', shape) #xy shape= (2, 4)
kvar = K.eye(3)
#K.eval(kvar)
print('kvar=', kvar)
'''
array([[1., 0., 0.],
[0., 1., 0.],
[0., 0., 1.]], dtype=float32)>
'''
a = np.array([[1, 2], [3, 4]])
a = K.transpose(a)
print('a=', a)
'''
a= tf.Tensor(
[[1 3]
[2 4]], shape=(2, 2), dtype=int32)
'''
a = K.clip(np.array([-1, 0, 1, 2, 3, 4, 5]), min_value=0, max_value=3)
print('a=', a) #a= tf.Tensor([0 0 1 2 3 3 3], shape=(7,), dtype=int32)
def call(self, inputs, **kwargs):
input_shape = K.int_shape(inputs)
sequence_length, d_model = input_shape[-2:]
# output of the "sigmoid halting unit" (not the probability yet)
halting = K.sigmoid(
K.reshape(
K.bias_add(
K.dot(K.reshape(inputs, [-1, d_model]),
self.halting_kernel),
self.halting_biases,
data_format='channels_last'),
[-1, sequence_length]))
if self.zeros_like_halting is None:
self.initialize_control_tensors(halting)
# useful flags
step_is_active = K.greater(self.halt_budget, 0)
no_further_steps = K.less_equal(self.halt_budget - halting, 0)
# halting probability is equal to
# a. halting output if this isn't the last step (we have some budget)
# b. to remainder if it is,
# c. and zero for the steps that shouldn't be executed at all
# (out of budget for them)
halting_prob = K.switch(
step_is_active,
K.switch(
no_further_steps,
self.remainder,
halting),
self.zeros_like_halting)
self.active_steps += K.switch(
step_is_active,
self.ones_like_halting,
self.zeros_like_halting)
# We don't know which step is the last, so we keep updating
# expression for the loss with each call of the layer
self.ponder_cost = (
self.time_penalty_t * K.mean(self.remainder + self.active_steps))
# Updating "the remaining probability" and the halt budget
self.remainder = K.switch(
no_further_steps,
self.remainder,
self.remainder - halting)
self.halt_budget -= halting # OK to become negative
# If none of the inputs are active at this step, then instead
# of zeroing them out by multiplying to all-zeroes halting_prob,
# we can simply use a constant tensor of zeroes, which means that
# we won't even calculate the output of those steps, saving
# some real computational time.
if self.zeros_like_input is None:
self.zeros_like_input = K.zeros_like(
inputs, name='zeros_like_input')
# just because K.any(step_is_active) doesn't work in PlaidML
any_step_is_active = K.greater(
K.sum(K.cast(step_is_active, 'int32')), 0)
step_weighted_output = K.switch(
any_step_is_active,
K.expand_dims(halting_prob, -1) * inputs,
self.zeros_like_input)
if self.weighted_output is None:
self.weighted_output = step_weighted_output
else:
self.weighted_output += step_weighted_output
return [inputs, self.weighted_output]
def build_model(hidden_dim, max_seq_len, vocabulary_size):
## encoder Input and layers
encoder_in = Input((max_seq_len, ), dtype='int32', name='encoder_in')
ith_str = Input((1, ), dtype='int32', name='ith_str')
word = Input((1, ), dtype='int32', name='word')
OneHot = Lambda(lambda x: K.one_hot(x, vocabulary_size), name='OneHot')
## building encoder
encoder_in_and_word = Concatenate()([ith_str, word, encoder_in])
encoder_GRU = GRU(hidden_dim, return_state=True, return_sequences=True)
encoder_out, state = encoder_GRU(OneHot(encoder_in_and_word))
encoder_out_dup = RepeatVector(max_seq_len)(encoder_out[:, -1])
## decoder Input and layers
decoder_in = Input((max_seq_len, ), dtype='int32', name='decoder_in')
ith = Input((1, ), dtype='int32', name='ith')
decoder_GRU = GRU(hidden_dim, return_sequences=True, return_state=True)
decoder_Dense = Dense(vocabulary_size,
activation='softmax',
name='decoder_out')
## building decoder
ith_dup = RepeatVector(max_seq_len)(K.cast(ith, 'float'))
word_dup = K.reshape(RepeatVector(max_seq_len)(word), (-1, max_seq_len))
x = Concatenate()(
[ith_dup,
OneHot(word_dup),
OneHot(decoder_in), encoder_out_dup])
x, _ = decoder_GRU(x, initial_state=state)
decoder_out = decoder_Dense(x)
## get the specific word
gather = K.concatenate(
[K.reshape(tf.range(K.shape(decoder_out)[0]), (-1, 1)), ith])
specific_word = tf.gather_nd(decoder_out, gather)
specific_word = Lambda(tf.identity, name='word_out')(
specific_word
) # Add this layer because the name of tf.gather_nd is too ugly
model = Model([encoder_in, decoder_in, ith, ith_str, word],
[decoder_out, specific_word])
## building decoder model given encoder_out and states
decoder_in_one_word = Input((1, ),
dtype='int32',
name='decoder_in_one_word')
decoder_state_in = Input((hidden_dim, ), name='decoder_state_in')
encoder_out = Input((hidden_dim, ), name='decoder_encoder_out')
x = Concatenate()([
K.cast(ith, 'float')[:, tf.newaxis],
OneHot(word),
OneHot(decoder_in_one_word), encoder_out[:, tf.newaxis]
])
x, decoder_state = decoder_GRU(x, initial_state=decoder_state_in)
decoder_out = decoder_Dense(x)
decoder_model = Model(
[decoder_in_one_word, encoder_out, decoder_state_in, ith, word],
[decoder_out, decoder_state])
encoder_in = Input((None, ), dtype='int32')
encoder_in_and_word = Concatenate()([ith_str, word, encoder_in])
encoder_out, state = encoder_GRU(OneHot(encoder_in_and_word))
encoder_model = Model([encoder_in, ith_str, word], [encoder_out, state])
return model, encoder_model, decoder_model
def yolo_loss(args,
anchors,
num_classes,
rescore_confidence=False,
print_loss=False):
"""YOLO localization loss function.
Parameters
----------
yolo_output : tensor
Final convolutional layer features.
true_boxes : tensor
Ground truth boxes tensor with shape [batch, num_true_boxes, 5]
containing box x_center, y_center, width, height, and class.
detectors_mask : array
0/1 mask for detector positions where there is a matching ground truth.
matching_true_boxes : array
Corresponding ground truth boxes for positive detector positions.
Already adjusted for conv height and width.
anchors : tensor
Anchor boxes for model.
num_classes : int
Number of object classes.
rescore_confidence : bool, default=False
If true then set confidence target to IOU of best predicted box with
the closest matching ground truth box.
print_loss : bool, default=False
If True then use a tf.Print() to print the loss components.
Returns
-------
mean_loss : float
mean localization loss across minibatch
"""
(yolo_output, true_boxes, detectors_mask, matching_true_boxes) = args
num_anchors = len(anchors)
object_scale = 5
no_object_scale = 1
class_scale = 1
coordinates_scale = 1
pred_xy, pred_wh, pred_confidence, pred_class_prob = yolo_head(
yolo_output, anchors, num_classes)
# Unadjusted box predictions for loss.
# TODO: Remove extra computation shared with yolo_head.
yolo_output_shape = K.shape(yolo_output)
feats = K.reshape(yolo_output, [
-1, yolo_output_shape[1], yolo_output_shape[2], num_anchors,
num_classes + 5
])
pred_boxes = K.concatenate((K.sigmoid(feats[..., 0:2]), feats[..., 2:4]),
axis=-1)
# TODO: Adjust predictions by image width/height for non-square images?
# IOUs may be off due to different aspect ratio.
# Expand pred x,y,w,h to allow comparison with ground truth.
# batch, conv_height, conv_width, num_anchors, num_true_boxes, box_params
pred_xy = K.expand_dims(pred_xy, 4)
pred_wh = K.expand_dims(pred_wh, 4)
pred_wh_half = pred_wh / 2.
pred_mins = pred_xy - pred_wh_half
pred_maxes = pred_xy + pred_wh_half
true_boxes_shape = K.shape(true_boxes)
# batch, conv_height, conv_width, num_anchors, num_true_boxes, box_params
true_boxes = K.reshape(true_boxes, [
true_boxes_shape[0], 1, 1, 1, true_boxes_shape[1], true_boxes_shape[2]
])
true_xy = true_boxes[..., 0:2]
true_wh = true_boxes[..., 2:4]
# Find IOU of each predicted box with each ground truth box.
true_wh_half = true_wh / 2.
true_mins = true_xy - true_wh_half
true_maxes = true_xy + true_wh_half
intersect_mins = K.maximum(pred_mins, true_mins)
intersect_maxes = K.minimum(pred_maxes, true_maxes)
intersect_wh = K.maximum(intersect_maxes - intersect_mins, 0.)
intersect_areas = intersect_wh[..., 0] * intersect_wh[..., 1]
pred_areas = pred_wh[..., 0] * pred_wh[..., 1]
true_areas = true_wh[..., 0] * true_wh[..., 1]
union_areas = pred_areas + true_areas - intersect_areas
iou_scores = intersect_areas / union_areas
# Best IOUs for each location.
best_ious = K.max(iou_scores, axis=4) # Best IOU scores.
best_ious = K.expand_dims(best_ious)
# A detector has found an object if IOU > thresh for some true box.
object_detections = K.cast(best_ious > 0.6, K.dtype(best_ious))
# TODO: Darknet region training includes extra coordinate loss for early
# training steps to encourage predictions to match anchor priors.
# Determine confidence weights from object and no_object weights.
# NOTE: YOLO does not use binary cross-entropy here.
no_object_weights = (no_object_scale * (1 - object_detections) *
(1 - detectors_mask))
no_objects_loss = no_object_weights * K.square(-pred_confidence)
if rescore_confidence:
objects_loss = (object_scale * detectors_mask *
K.square(best_ious - pred_confidence))
else:
objects_loss = (object_scale * detectors_mask *
K.square(1 - pred_confidence))
confidence_loss = objects_loss + no_objects_loss
# Classification loss for matching detections.
# NOTE: YOLO does not use categorical cross-entropy loss here.
matching_classes = K.cast(matching_true_boxes[..., 4], 'int32')
matching_classes = K.one_hot(matching_classes, num_classes)
classification_loss = (class_scale * detectors_mask *
K.square(matching_classes - pred_class_prob))
# Coordinate loss for matching detection boxes.
matching_boxes = matching_true_boxes[..., 0:4]
coordinates_loss = (coordinates_scale * detectors_mask *
K.square(matching_boxes - pred_boxes))
confidence_loss_sum = K.sum(confidence_loss)
classification_loss_sum = K.sum(classification_loss)
coordinates_loss_sum = K.sum(coordinates_loss)
total_loss = 0.5 * (confidence_loss_sum + classification_loss_sum +
coordinates_loss_sum)
if print_loss:
total_loss = tf.Print(
total_loss, [
total_loss, confidence_loss_sum, classification_loss_sum,
coordinates_loss_sum
],
message='yolo_loss, conf_loss, class_loss, box_coord_loss:')
return total_loss
def custom_loss(target, output):
output **= (1 / hp.T)
output /= K.reshape(K.sum(output, axis=1), (-1, 1))
return (hp.T)**2 * K.categorical_crossentropy(target, output)
def echo_sample(inputs, clip=None, d_max=100, batch=100, multiplicative=False, echo_mc = False,
replace=False, fx_clip=None, plus_sx=True, calc_log=True, return_noise=False, **kwargs):
# kwargs unused
# inputs should be specified as list:
# [ f(X), s(X) ] with s(X) in log space if calc_log = True
# plus_sx =
# True if logsigmoid activation for s(X)
# False for softplus (equivalent)
if isinstance(inputs, list):
fx = inputs[0]
sx = inputs[-1]
else:
fx = inputs
# TO DO : CALC_LOG currently determines both whether to do log space calculations AND whether sx is a log
fx_shape = fx.get_shape()
sx_shape = sx.get_shape()
z_dim = K.int_shape(fx)[-1]
batch_size = batch
batch = K.shape(fx)[0]
if clip is None:
# clip is multiplied times s(x) to ensure that last sampled term:
# (clip^d_max)*f(x) < machine precision
max_fx = fx_clip if fx_clip is not None else 1.0
clip = (2**(-23)/max_fx)**(1.0/d_max)
# fx_clip can be used to restrict magnitude of f(x), not used in paper
# defaults to no clipping and M = 1 (e.g. with tanh activation for f(x))
if fx_clip is not None:
fx = K.clip(fx, -fx_clip, fx_clip)
if not calc_log:
sx = tf.multiply(clip,sx)
sx = tf.where(tf.abs(sx) < K.epsilon(), K.epsilon()*tf.sign(sx), sx)
#raise ValueError('calc_log=False is not supported; sx has to be log_sigmoid')
else:
# plus_sx based on activation for sx = s(x):
# True for log_sigmoid
# False for softplus
sx = tf.log(clip) + (-1*sx if not plus_sx else sx)
#if echo_mc is not None and echo_mc:
# use mean centered fx for noise
# fx = fx - K.mean(fx, axis = 0, keepdims = True)
if replace: # replace doesn't set batch size (using permute_neighbor_indices does)
sx = K.batch_flatten(sx) if len(sx_shape) > 2 else sx
fx = K.batch_flatten(fx) if len(fx_shape) > 2 else fx
inds = K.reshape(random_indices(batch, d_max), (-1, 1))
select_sx = gather_nd_reshape(sx, inds, (-1, d_max, z_dim))
select_fx = gather_nd_reshape(fx, inds, (-1, d_max, z_dim))
if len(sx_shape)>2:
select_sx = K.expand_dims(K.expand_dims(select_sx, 2), 2)
sx = K.expand_dims(K.expand_dims(sx, 1),1)
if len(fx_shape)>2:
select_fx = K.expand_dims(K.expand_dims(select_fx, 2), 2)
fx = K.expand_dims(K.expand_dims(fx, 1),1)
else:
# batch x batch x z_dim
# for all i, stack_sx[i, :, :] = sx
repeat = tf.multiply(tf.ones_like(tf.expand_dims(fx, 0)), tf.ones_like(tf.expand_dims(fx, 1)))
stack_fx = tf.multiply(fx, repeat)
stack_sx = tf.multiply(sx, repeat)
# select a set of dmax examples from original fx / sx for each batch entry
inds = indices_without_replacement(batch, d_max)
# Alterntive method: but note that permute_neighbor_indices sets the batch_size dimension != None
# this necessitates the use of fit_generator, e.g. in training to avoid 'remainder' batches if data_size % batch > 0
#inds = permute_neighbor_indices(batch_size, d_max, replace = replace)
select_sx = tf.gather_nd(stack_sx, inds)
select_fx = tf.gather_nd(stack_fx, inds)
if calc_log:
sx_echoes = tf.cumsum(select_sx, axis = 1, exclusive = True)
else:
sx_echoes = tf.cumprod(select_sx, axis = 1, exclusive = True)
# calculates S(x0)S(x1)...S(x_l)*f(x_(l+1))
sx_echoes = tf.exp(sx_echoes) if calc_log else sx_echoes
fx_sx_echoes = tf.multiply(select_fx, sx_echoes)
# performs the sum over dmax terms to calculate noise
noise = tf.reduce_sum(fx_sx_echoes, axis = 1)
sx = sx if not calc_log else tf.exp(sx)
if multiplicative: # log z according to echo
output = tf.exp(fx + tf.multiply(sx, noise))
else:
output = fx + tf.multiply(sx, noise)
return output if not return_noise else noise
def cat_acc(y_true, y_pred):
y_true = K.reshape(y_true, shape=(-1, 7, 37))
y_pred = K.reshape(y_pred, shape=(-1, 7, 37))
return K.mean(tf.keras.metrics.categorical_accuracy(y_true, y_pred))
def body(X, Kernel, n, f, c, kH, kW, w_out, i, j):
X_reshape = K.reshape(X[n, c, i:i + kH, j:j + kW], [1, -1])
a = K.sum(Kernel * X_reshape)
return a
def top_3_k(y_true, y_pred):
# Reshape into 2-d
y_true = K.reshape(y_true, (-1, 37))
y_pred = K.reshape(y_pred, (-1, 37))
return K.mean(
tf.keras.metrics.top_k_categorical_accuracy(y_true, y_pred, k=3))
def call(self, obj_vecs, pred_vecs, edges):
"""
Inputs:
- obj_vecs: FloatTensor of shape (B, O, D) giving vectors for all objects
- pred_vecs: FloatTensor of shape (B, T, D) giving vectors for all predicates
- edges: LongTensor of shape (B, T, 2) where edges[k] = [i, j] indicates the
presence of a triple [obj_vecs[i], pred_vecs[k], obj_vecs[j]]
Outputs:
- new_obj_vecs: FloatTensor of shape (B, O, D) giving new vectors for objects
- new_pred_vecs: FloatTensor of shape (B, T, D) giving new vectors for predicates
"""
O, T = K.int_shape(obj_vecs)[1], K.int_shape(pred_vecs)[1]
Din, H, Dout = self.input_dim, self.hidden_dim, self.output_dim
# Break apart indices for subjects and objects; these have shape (B, T,)
s_idx, o_idx = tf.split(edges, 2, axis=2) #shape =(B,T,1)
s_idx = K.reshape(s_idx, (-1, T)) #shape =(B,T)
o_idx = K.reshape(o_idx, (-1, T))
i = tf.meshgrid(tf.range(self.batch_size), indexing="ij")
i = K.reshape(i, (self.batch_size, 1))
i = tf.broadcast_to(i, (self.batch_size, T))
idx_s = tf.stack([i, s_idx], axis=-1)
idx_o = tf.stack([i, o_idx], axis=-1)
cur_s_vecs = tf.gather_nd(obj_vecs, idx_s)
cur_o_vecs = tf.gather_nd(obj_vecs, idx_o)
# Get current vectors for triples; shape is (B, T, 3 * Din)
# Pass through net1 to get new triple vecs; shape is (B, T, 2 * H + Dout)
cur_t_vecs = K.concatenate([cur_s_vecs, pred_vecs, cur_o_vecs], axis=2)
new_t_vecs = self.net1(cur_t_vecs)
# Break apart into new s, p, and o vecs; s and o vecs have shape (B, T, H) and
# p vecs have shape (B, T, Dout)
new_s_vecs = new_t_vecs[:, :, :H]
new_p_vecs = new_t_vecs[:, :, H:(H + Dout)]
new_o_vecs = new_t_vecs[:, :, (H + Dout):(2 * H + Dout)]
# Allocate space for pooled object vectors of shape (B, O, H)
pooled_obj_vecs = tf.zeros(shape=(self.batch_size, O, H))
shape = K.shape(pooled_obj_vecs)
# Use scatter_add to sum vectors for objects that appear in multiple triples;
# we first need to expand the indices to have shape (B, T, H)
s_idx = K.reshape(s_idx, (-1, T))
o_idx = K.reshape(o_idx, (-1, T))
i = tf.meshgrid(tf.range(self.batch_size), indexing="ij")
i = K.reshape(i, (self.batch_size, 1))
i = tf.broadcast_to(i, (self.batch_size, T))
idx_s = tf.stack([i, s_idx], axis=-1)
idx_o = tf.stack([i, o_idx], axis=-1)
pooled_obj_vecs = tf.scatter_nd(idx_s, new_s_vecs, shape=shape)
pooled_obj_vecs = tf.scatter_nd(idx_o, new_o_vecs,
shape=shape) # shape(B, O, H)
if self.pooling == 'avg':
# Figure out how many times each object has appeared, again using
# some scatter_add trickery.
obj_counts = tf.zeros(shape=(self.batch_size, O, H))
ones = tf.ones(shape=(self.batch_size, T, H))
obj_counts = tf.scatter_nd(idx_s, ones, shape=shape)
obj_counts = tf.scatter_nd(idx_o, ones, shape=shape)
# Divide the new object vectors by the number of times they
# appeared, but first clamp at 1 to avoid dividing by zero;
# objects that appear in no triples will have output vector 0
# so this will not affect them.
obj_counts = K.clip(obj_counts, min_value=1, max_value=None)
pooled_obj_vecs = pooled_obj_vecs / obj_counts
# Send pooled object vectors through net2 to get output object vectors, of shape (O, Dout)
new_obj_vecs = self.net2(pooled_obj_vecs)
return new_obj_vecs, new_p_vecs
def styleTransfer(cData, sData, tData):
print(" Building transfer model.")
contentTensor = K.variable(cData, dtype=tf.float64)
styleTensor = K.variable(sData, dtype=tf.float64)
genFlatten = K.placeholder(CONTENT_IMG_H * CONTENT_IMG_W * 3,
dtype=tf.float64)
genTensor = K.reshape(genFlatten, (1, CONTENT_IMG_H, CONTENT_IMG_W, 3))
inputTensor = K.concatenate([contentTensor, styleTensor, genTensor],
axis=0)
model = vgg19.VGG19(include_top=False,
weights='imagenet',
input_tensor=inputTensor)
outputDict = dict([(layer.name, layer.output) for layer in model.layers])
print(" VGG19 model loaded.")
loss = 0.0
styleLayerNames = [
"block1_conv1", "block2_conv1", "block3_conv1", "block4_conv1",
"block5_conv1"
]
contentLayerName = "block5_conv2"
print(" Calculating content loss.")
contentLayer = outputDict[contentLayerName]
contentOutput = contentLayer[0, :, :, :]
genOutput = contentLayer[2, :, :, :]
loss += CONTENT_WEIGHT * contentLoss(contentOutput, genOutput)
print(" Calculating style loss.")
styleLayerWeight = 1 / len(styleLayerNames)
for layerName in styleLayerNames:
styleLayer = outputDict[layerName]
originalStyleOutput = styleLayer[1, :, :, :]
genStyleOutput = styleLayer[2, :, :, :]
loss += STYLE_WEIGHT * styleLayerWeight * styleLoss(
originalStyleOutput, genStyleOutput)
loss = totalLoss(loss)
gradients = K.gradients(loss, genFlatten)[0]
loss_function = K.function([genFlatten], [loss, gradients])
gen_image_np = tData.flatten()
print(gradients)
print(loss)
print(gen_image_np.shape)
print(" Beginning transfer.")
for i in range(TRANSFER_ROUNDS):
print(" Step %d." % i)
#loss_val, grad_val = loss_function([gen_image_np])
#gen_image_np = gen_image_np + 0.1 * grad_val
gen_image_np, f, d = fmin_l_bfgs_b(func=loss_function,
x0=gen_image_np,
maxiter=100,
maxls=1200,
maxfun=5000000)
print(" Loss: %f." % f)
img = deprocessImage(gen_image_np)
saveFile = "./transfer_" + str(i) + ".jpg"
imageio.imwrite(saveFile,
img) #Uncomment when everything is working right.
print(" Image saved to \"%s\"." % saveFile)
print(" Transfer complete.")
def _process_sample(args):
_hm, _reg, _wh, _kps, _hm_hp, _hp_offset = args
_scores, _inds = tf.math.top_k(_hm, k=k, sorted=True)
_classes = K.cast(_inds % cat, 'float32')
_inds = K.cast(_inds / cat, 'int32')
_xs = K.cast(_inds % width, 'float32')
_ys = K.cast(K.cast(_inds / width, 'int32'), 'float32')
_wh = K.gather(_wh, _inds)
_reg = K.gather(_reg, _inds)
_kps = K.gather(_kps, _inds)
# shift keypoints by their center
_kps_x = _kps[:, ::2]
_kps_y = _kps[:, 1::2]
_kps_x = _kps_x + K.expand_dims(_xs, -1) # k x J
_kps_y = _kps_y + K.expand_dims(_ys, -1) # k x J
_kps = K.stack([_kps_x, _kps_y], -1) # k x J x 2
_xs = _xs + _reg[..., 0]
_ys = _ys + _reg[..., 1]
_x1 = _xs - _wh[..., 0] / 2
_y1 = _ys - _wh[..., 1] / 2
_x2 = _xs + _wh[..., 0] / 2
_y2 = _ys + _wh[..., 1] / 2
# snap center keypoints to the closest heatmap keypoint
def _process_channel(args):
__kps, __hm_hp = args
thresh = 0.1
__hm_scores, __hm_inds = tf.math.top_k(__hm_hp, k=k, sorted=True)
__hm_xs = K.cast(__hm_inds % width, 'float32')
__hm_ys = K.cast(K.cast(__hm_inds / width, 'int32'), 'float32')
__hp_offset = K.gather(_hp_offset, __hm_inds)
__hm_xs = __hm_xs + __hp_offset[..., 0]
__hm_ys = __hm_ys + __hp_offset[..., 1]
mask = K.cast(__hm_scores > thresh, 'float32')
__hm_scores = (1. - mask) * -1. + mask * __hm_scores
__hm_xs = (1. - mask) * -10000. + mask * __hm_xs
__hm_ys = (1. - mask) * -10000. + mask * __hm_ys
__hm_kps = K.stack([__hm_xs, __hm_ys], -1) # k x 2
__broadcast_hm_kps = K.expand_dims(__hm_kps, 1) # k x 1 x 2
__broadcast_kps = K.expand_dims(__kps, 0) # 1 x k x 2
dist = K.sqrt(
K.sum(K.pow(__broadcast_kps - __broadcast_hm_kps, 2),
2)) # k, k
min_dist = K.min(dist, 0)
min_ind = K.argmin(dist, 0)
__hm_scores = K.gather(__hm_scores, min_ind)
__hm_kps = K.gather(__hm_kps, min_ind)
mask = (K.cast(__hm_kps[..., 0] < _x1, 'float32') +
K.cast(__hm_kps[..., 0] > _x2, 'float32') +
K.cast(__hm_kps[..., 1] < _y1, 'float32') +
K.cast(__hm_kps[..., 1] > _y2, 'float32') +
K.cast(__hm_scores < thresh, 'float32') + K.cast(
min_dist > 0.3 *
(K.maximum(_wh[..., 0], _wh[..., 1])), 'float32'))
mask = K.expand_dims(mask, -1)
mask = K.cast(mask > 0, 'float32')
__kps = (1. - mask) * __hm_kps + mask * __kps
return __kps
_kps = K.permute_dimensions(_kps, (1, 0, 2)) # J x k x 2
_hm_hp = K.permute_dimensions(_hm_hp, (1, 0)) # J x -1
_kps = K.map_fn(_process_channel, [_kps, _hm_hp], dtype='float32')
_kps = K.reshape(K.permute_dimensions(_kps, (1, 2, 0)),
(k, -1)) # k x J * 2
# rescale to image coordinates
_x1 = output_stride * _x1
_y1 = output_stride * _y1
_x2 = output_stride * _x2
_y2 = output_stride * _y2
_kps = output_stride * _kps
_boxes = K.stack([_x1, _y1, _x2, _y2], -1)
_scores = K.expand_dims(_scores, -1)
_classes = K.expand_dims(_classes, -1)
_detection = K.concatenate([_boxes, _scores, _kps, _classes], -1)
return _detection
def call(self, x):
x = x[0]
#print("!x",x.shape)
gs = self.gs
k = self.k
param = self.param
C = self.numericalC
#print("gs",gs,"k",k,"param",param,"C",C)
for i in range(10):
t.print(
"!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!"
)
t.print("calling onetopk", self.metrik, output_stream=sys.stdout)
#exit()
mata = K.constant(self.getmata(gs))
matb = K.constant(self.getmatb(gs))
#print("mata",mata.shape)
#print("matb",matb.shape)
xp = K.permute_dimensions(x, (0, 2, 1))
#print("xp",xp.shape)
xa = K.dot(xp, mata)
xb = K.dot(xp, matb)
#print("xa",xa.shape,"xb",xb.shape)
#exit()
isval = xa[:, self.flag, :] * xb[:, self.flag, :]
#return isval,isval
#print("isval",isval.shape)
#exit()
ds = xa - xb
#print("ds",ds.shape)
dsp = K.permute_dimensions(ds, (0, 2, 1))
#print("dsp",dsp.shape)
dspsq = K.square(dsp)
#print("dspsq",dspsq.shape)
#print("self.metrik",self.metrik.shape)
delt = K.dot(dspsq, self.metrik)
#print("delt",delt.shape)
delt = K.reshape(
delt, (-1, self.gs * self.gs)) + (1 - isval) * self.emptyconst
d = K.reshape(delt, (-1, gs, gs))
#print("d",d.shape)
#return d,d
#####no self interactions
if self.self_interaction == False:
one = K.eye(gs)
#print("one",one.shape)
d += self.self_interaction_const * one
#####end no self interactions
v, _ = t.math.top_k(-d, k=k)
#print("v",v.shape)
#return v,v
vb = v[:, :, -1]
#print("vb",vb.shape)
vbs = K.reshape(vb, (-1, gs, 1))
#print("vbs",vbs.shape)
su = d + vbs #plus since top_k(-d)
#print("su",su.shape)
#map anything above 0 to 0 and anything below to 1, also map 0 to 1
#p(-x)=C*d_C(-x)
# =d(-C*x)
# =1-r(Cx-1)+r(Cx)
#experimentally:
# r(1-Cx)-r(-Cx)
rel = K.relu(1 - C * su) - K.relu(-C * su)
#print("rel",rel.shape)
rel = K.relu(rel) - K.relu(rel - 1)
#return rel,rel
dez1 = K.reshape(rel, (-1, self.gs * self.gs))
#print("dez1",dez1.shape)
dez2 = dez1 * isval
#print("dez2",dez2.shape)
rel = K.reshape(dez2, (-1, self.gs, self.gs))
print("rel", rel.shape)
numnei = K.sum(rel, axis=-1)
print("numnei", numnei.shape)
factor = self.k / (numnei + 0.00000000001)
#return K.concatenate((numnei,factor),axis=-1),factor#,factor#numnei,numnei
print("factor", factor.shape)
refactor = K.repeat(factor, self.gs)
print("refactor", refactor.shape)
refactor = K.permute_dimensions(refactor, (0, 2, 1))
#return refactor,refactor
rel = rel * refactor
print("rel", rel.shape)
#exit()
if self.free == 0: return rel, x
zero1 = K.zeros_like(x[:, :, 0])
zero1 = K.reshape(zero1, (-1, x.shape[1], 1))
#print("!",zero1.shape)
zerolis = []
for i in range(self.free):
zerolis.append(zero1)
zeros = K.concatenate(zerolis, axis=-1)
#print(zeros.shape)
return rel, K.concatenate((x, zeros), axis=-1)
def sparse_gather(y_pred, target_indices, task_name):
clf_h = Lambda(lambda x: K.reshape(x, (-1, K.int_shape(x)[-1])),
name=task_name + '_flatten')(y_pred)
return Lambda(lambda x: K.gather(x[0], K.cast(x[1], 'int32')),
name=task_name + '_gather')([clf_h, target_indices])
def call(self, x, **kwargs):
assert isinstance(x, list), 'SliceLayer input is not a list'
return x[0] * K.reshape(x[1], (-1, 1))
def call(self, inputs, q_mask=None, v_mask=None, a_mask=None):
"""实现多头注意力
q_mask: 对输入的query序列的mask。
主要是将输出结果的padding部分置0。
v_mask: 对输入的value序列的mask。
主要是防止attention读取到padding信息。
a_mask: 对attention矩阵的mask。
不同的attention mask对应不同的应用。
"""
q, k, v = inputs[:3]
if a_mask:
if len(inputs) == 3:
a_mask = 'history_only'
else:
a_mask = inputs[3]
if q_mask is not None:
if not hasattr(self, 'q_mask_layer'):
self.q_mask_layer = search_layer(q, q_mask)
q_mask = self.q_mask_layer.output_mask
if v_mask is not None:
if not hasattr(self, 'v_mask_layer'):
self.v_mask_layer = search_layer(v, v_mask)
v_mask = self.v_mask_layer.output_mask
# Pooling
if self.pool_size > 1:
is_self_attention = (q is k is v)
q_in_len = K.shape(q)[1]
q = sequence_masking(q, q_mask, 0)
q = divisible_temporal_padding(q, self.pool_size)
q = pool1d(q, self.pool_size, self.pool_size, pool_mode='avg')
if is_self_attention:
k = v = q
else:
k = sequence_masking(k, v_mask, 0)
k = divisible_temporal_padding(k, self.pool_size)
k = pool1d(k, self.pool_size, self.pool_size, pool_mode='avg')
v = sequence_masking(v, v_mask, 0)
v = divisible_temporal_padding(v, self.pool_size)
v = pool1d(v, self.pool_size, self.pool_size, pool_mode='avg')
if v_mask is not None:
v_mask = v_mask[:, ::self.pool_size]
if a_mask is not None and not is_string(a_mask):
a_mask = a_mask[..., ::self.pool_size, ::self.pool_size]
# 线性变换
qw = self.q_dense(q)
kw = self.k_dense(k)
vw = self.v_dense(v)
# 形状变换
qw = K.reshape(qw, (-1, K.shape(q)[1], self.heads, self.key_size))
kw = K.reshape(kw, (-1, K.shape(k)[1], self.heads, self.key_size))
vw = K.reshape(vw, (-1, K.shape(v)[1], self.heads, self.head_size))
# Attention
a = tf.einsum('bjhd,bkhd->bhjk', qw, kw)
# 相对位置编码
if self.max_relative_position is not None:
q_idxs = K.arange(0, K.shape(q)[1], dtype='int32')
q_idxs = K.expand_dims(q_idxs, 1)
v_idxs = K.arange(0, K.shape(v)[1], dtype='int32')
v_idxs = K.expand_dims(v_idxs, 0)
pos_ids = v_idxs - q_idxs
pos_ids = K.clip(pos_ids, -self.max_relative_position,
self.max_relative_position)
pos_ids = pos_ids + self.max_relative_position
pos_embeddings = K.gather(self.relative_embeddings, pos_ids)
a = a + tf.einsum('bjhd,jkd->bhjk', qw, pos_embeddings)
# Attention（续）
a = a / self.key_size ** 0.5
a = sequence_masking(a, v_mask, 1, -1)
if a_mask is not None:
if is_string(a_mask):
ones = K.ones_like(a[:1, :1])
a_mask = (ones - tf.linalg.band_part(ones, -1, 0)) * 1e12
a = a - a_mask
else:
a = a - (1 - a_mask) * 1e12
a = K.softmax(a)
# 完成输出
o = tf.einsum('bhjk,bkhd->bjhd', a, vw)
if self.max_relative_position is not None:
o = o + tf.einsum('bhjk,jkd->bjhd', a, pos_embeddings)
o = K.reshape(o, (-1, K.shape(o)[1], self.out_dim))
o = self.o_dense(o)
# 恢复长度
if self.pool_size > 1:
o = K.repeat_elements(o, self.pool_size, 1)[:, :q_in_len]
# 返回结果
o = sequence_masking(o, q_mask, 0)
return o
def call(self, x, mask=None):
assert (len(x) == 2)
img = x[0]
rois = x[1]
input_shape = K.shape(img)
outputs = []
for roi_idx in range(self.num_rois):
x = rois[0, roi_idx, 0]
y = rois[0, roi_idx, 1]
w = rois[0, roi_idx, 2]
h = rois[0, roi_idx, 3]
row_length = w / float(self.pool_size)
col_length = h / float(self.pool_size)
num_pool_regions = self.pool_size
# NOTE: the RoiPooling implementation differs between theano and tensorflow due to the lack of a resize op
# in theano. The theano implementation is much less efficient and leads to long compile times
if self.dim_ordering == 'th':
for jy in range(num_pool_regions):
for ix in range(num_pool_regions):
x1 = x + ix * row_length
x2 = x1 + row_length
y1 = y + jy * col_length
y2 = y1 + col_length
x1 = K.cast(x1, 'int32')
x2 = K.cast(x2, 'int32')
y1 = K.cast(y1, 'int32')
y2 = K.cast(y2, 'int32')
x2 = x1 + K.maximum(1, x2 - x1)
y2 = y1 + K.maximum(1, y2 - y1)
new_shape = [
input_shape[0], input_shape[1], y2 - y1, x2 - x1
]
x_crop = img[:, :, y1:y2, x1:x2]
xm = K.reshape(x_crop, new_shape)
pooled_val = K.max(xm, axis=(2, 3))
outputs.append(pooled_val)
elif self.dim_ordering == 'tf':
x = K.cast(x, 'int32')
y = K.cast(y, 'int32')
w = K.cast(w, 'int32')
h = K.cast(h, 'int32')
rs = tf.image.resize_images(img[:, y:y + h, x:x + w, :],
(self.pool_size, self.pool_size))
outputs.append(rs)
final_output = K.concatenate(outputs, axis=0)
final_output = K.reshape(final_output,
(1, self.num_rois, self.pool_size,
self.pool_size, self.nb_channels))
if self.dim_ordering == 'th':
final_output = K.permute_dimensions(final_output, (0, 1, 4, 2, 3))
else:
final_output = K.permute_dimensions(final_output, (0, 1, 2, 3, 4))
return final_output
