def correct_boxes(box_xy, box_wh, input_shape, image_shape):
'''Get corrected boxes'''
box_yx = box_xy[..., ::-1]
box_hw = box_wh[..., ::-1]
input_shape = K.cast(input_shape, K.dtype(box_yx))
image_shape = K.cast(image_shape, K.dtype(box_yx))
new_shape = K.round(image_shape * K.min(input_shape / image_shape))
offset = (input_shape - new_shape) / 2. / input_shape
scale = input_shape / new_shape
box_yx = (box_yx - offset) * scale
box_hw *= scale
box_mins = box_yx - (box_hw / 2.)
box_maxes = box_yx + (box_hw / 2.)
boxes = K.concatenate([
box_mins[..., 0:1], # y_min
box_mins[..., 1:2], # x_min
box_maxes[..., 0:1], # y_max
box_maxes[..., 1:2] # x_max
])
# Scale boxes back to original image shape.
boxes *= K.concatenate([image_shape, image_shape])
return boxes
def mask_aware_max(x):
mask = K.not_equal(K.sum(K.abs(x), axis=2, keepdims=True), 0)
mask = K.cast(mask, 'float32')
vecmin = K.min(x, axis=1, keepdims=True)
xstar = x + (vecmin * (1 - mask)) # setting masked values to the min value
return K.max(xstar, axis=1, keepdims=False)
def _batch_hard_triplet_loss(self, y_true: Tensor, pairwise_dist: Tensor) -> Tensor:
mask_anchor_positive = self._get_anchor_positive_triplet_mask(y_true, pairwise_dist)
anchor_positive_dist = mask_anchor_positive * pairwise_dist
hardest_positive_dist = K.max(anchor_positive_dist, axis=1, keepdims=True)
mask_anchor_negative = self._get_anchor_negative_triplet_mask(y_true, pairwise_dist)
anchor_negative_dist = mask_anchor_negative * pairwise_dist
mask_anchor_negative = self._get_semihard_anchor_negative_triplet_mask(anchor_negative_dist,
hardest_positive_dist,
mask_anchor_negative)
max_anchor_negative_dist = K.max(pairwise_dist, axis=1, keepdims=True)
anchor_negative_dist = pairwise_dist + max_anchor_negative_dist * (1.0 - mask_anchor_negative)
hardest_negative_dist = K.min(anchor_negative_dist, axis=1, keepdims=True)
triplet_loss = K.clip(hardest_positive_dist - hardest_negative_dist + self.margin, 0.0, None)
triplet_loss = K.mean(triplet_loss)
return triplet_loss
def starGAN_train(self, D_lr, G_lr, lamda_gp, lamda_cls, lamda_rec):
x_real = Input(shape=self.image_size)
label_real = Input(shape=(self.n_class,))
label_fake = Input(shape=(self.n_class,))
label_real_matrix = Input(shape=(self.image_size[0],self.image_size[1],self.n_class))
label_fake_matrix = Input(shape=(self.image_size[0],self.image_size[1],self.n_class))
x_fake = self.generator([x_real, label_fake_matrix])
# loss for discriminator
d_out_src_real, d_out_cls_real = self.discriminator(x_real)
d_loss_real = -K.mean(d_out_src_real)
d_loss_cls = K.mean(K.categorical_crossentropy(label_real, d_out_cls_real))
# cal acc
label_sub = d_out_cls_real - label_real
c1 = 1 + K.min(label_sub, axis=1) # label为1的最小置信度
c2 = K.max(label_sub, axis=1) # label为0的最大置信度
d_acc = K.mean(K.cast(K.greater(c1 - c2, 0), K.floatx())) # 如果label为1的最小置信度大于label为0的最大置信度，则正确，否则错误
# label_pred = K.cast(K.greater(K.clip(d_out_cls_real, 0, 1), 0.5), K.floatx())
# d_acc = 1 - K.mean(K.clip(K.sum(K.abs(label_real - label_pred), axis=1), 0, 1))
d_out_src_fake, d_out_cls_fake = self.discriminator(x_fake)
d_loss_fake = K.mean(d_out_src_fake)
# gradient penalty
e = K.placeholder(shape=(None, 1, 1, 1))
x_mixed = Input(shape=self.image_size, tensor=e * x_real + (1 - e) * x_fake)
x_mixed_gradient = K.gradients(self.discriminator(x_mixed), [x_mixed])[0]
x_mixed_gradient_norm = K.sqrt(K.sum(K.square(x_mixed_gradient), axis=[1, 2, 3])) # not norm in batch_size
gradient_penalty = K.mean(K.square(x_mixed_gradient_norm - 1))
d_loss = d_loss_real + d_loss_fake + lamda_gp * gradient_penalty + lamda_cls * d_loss_cls
d_training_updates = RMSprop(lr=D_lr).get_updates(d_loss, self.discriminator.trainable_weights)
D_train = K.function([x_real, label_real, label_real_matrix, label_fake, label_fake_matrix, e], [d_loss, d_acc], d_training_updates)
# loss for generator
x_rec = self.generator([x_fake, label_real_matrix])
g_out_src_fake, g_out_cls_fake = self.discriminator(x_fake)
g_loss_fake = -K.mean(g_out_src_fake)
g_loss_rec = K.mean(K.abs(x_real - x_rec))
g_loss_cls = K.mean(K.categorical_crossentropy(label_fake, g_out_cls_fake))
g_loss = g_loss_fake + lamda_rec * g_loss_rec + lamda_cls * g_loss_cls
g_training_updates = RMSprop(lr=G_lr).get_updates(g_loss, self.generator.trainable_weights)
G_train = K.function([x_real, label_real, label_real_matrix, label_fake, label_fake_matrix], [g_loss], g_training_updates)
return D_train, G_train
def loss(y_true, y_pred):
global output_feature;
y_pred_positive=y_pred[:,:output_feature]
y_pred_label=y_pred[:,output_feature:2*output_feature]
y_pred_negative1=y_pred[:,2*output_feature:3*output_feature]
y_pred_negative2=y_pred[:,3*output_feature:4*output_feature]
y_pred_negative3=y_pred[:,4*output_feature:5*output_feature]
y_pred_negative4=y_pred[:,5*output_feature:6*output_feature]
y_pred_negative5=y_pred[:,6*output_feature:7*output_feature]
y_pred_negative6=y_pred[:,7*output_feature:8*output_feature]
l=K.min(y_true-y_true, axis=1)
return K.maximum(cosine_proximity(y_pred_positive, y_pred_negative1)-cosine_proximity(y_pred_positive, y_pred_label)+0.3, 0.)\
+K.maximum(cosine_proximity(y_pred_positive, y_pred_negative2)-cosine_proximity(y_pred_positive, y_pred_label)+0.3, 0.)\
+K.maximum(cosine_proximity(y_pred_positive, y_pred_negative3)-cosine_proximity(y_pred_positive, y_pred_label)+0.3, 0.)\
+K.maximum(cosine_proximity(y_pred_positive, y_pred_negative4)-cosine_proximity(y_pred_positive, y_pred_label)+0.3, 0.)\
+K.maximum(cosine_proximity(y_pred_positive, y_pred_negative5)-cosine_proximity(y_pred_positive, y_pred_label)+0.3, 0.)\
+K.maximum(cosine_proximity(y_pred_positive, y_pred_negative6)-cosine_proximity(y_pred_positive, y_pred_label)+0.3, 0.)\
+K.maximum(0.98-cosine_proximity(y_pred_positive, y_pred_label), 0.)+l
def step(self, input_energy_t, states, return_logZ=True):
# not in the following `prev_target_val` has shape = (B, F)
# where B = batch_size, F = output feature dim
# Note: `i` is of float32, due to the behavior of `K.rnn`
prev_target_val, i, chain_energy = states[:3]
t = K.cast(i[0, 0], dtype='int32')
if len(states) > 3:
if K.backend() == 'theano':
m = states[3][:, t:(t + 2)]
else:
m = K.tf.slice(states[3], [0, t], [-1, 2])
input_energy_t = input_energy_t * K.expand_dims(m[:, 0])
chain_energy = chain_energy * K.expand_dims(
K.expand_dims(m[:, 0] * m[:, 1])) # (1, F, F)*(B, 1, 1) -> (B, F, F)
if return_logZ:
energy = chain_energy + K.expand_dims(input_energy_t - prev_target_val,
2) # shapes: (1, B, F) + (B, F, 1) -> (B, F, F)
new_target_val = K.logsumexp(-energy, 1) # shapes: (B, F)
return new_target_val, [new_target_val, i + 1]
else:
energy = chain_energy + K.expand_dims(input_energy_t + prev_target_val, 2)
min_energy = K.min(energy, 1)
argmin_table = K.cast(K.argmin(energy, 1), K.floatx()) # cast for tf-version `K.rnn`
return argmin_table, [min_energy, i + 1]
def triplet_loss(y_true,
y_pred,
margin=input_margin,
P=P_param,
K=K_param,
output_dim=input_output_dim):
embeddings = Keras.reshape(y_pred, (-1, output_dim))
loss = Keras.variable(0, dtype='float32')
for i in range(P):
for a in range(K):
pred_anchor = embeddings[i * K + a]
hard_pos = Keras.max(
dist(pred_anchor, embeddings[i * K:(i + 1) * K]))
hard_neg = Keras.min(
dist(
pred_anchor,
Keras.concatenate(
[embeddings[0:i * K], embeddings[(i + 1) * K:]], 0)))
if margin == None:
loss += log1p(hard_pos - hard_neg)
else:
loss += Keras.maximum(margin + hard_pos - hard_neg, 0.0)
return loss
def ranking_loss_with_margin(y_pred, y_true):
"""
Using this loss trains the model to give scores to all correct elements in y_true that are
higher than all scores it gives to incorrect elements in y_true, plus a margin.
For example, let ``y_true = [0, 0, 1, 1, 0]``, and let ``y_pred = [-1, 1, 2, 0, -2]``. We will
find the lowest score assigned to correct elements in ``y_true`` (``0`` in this case), and the
highest score assigned to incorrect elements in ``y_true`` (``1`` in this case). We will then
compute a hinge loss given these values: ``K.maximum(0.0, 1 + 1 - 0)``.
Note that the way we do this uses ``K.max()`` and ``K.min()`` over the elements in ``y_true``,
which means that if you have a lot of values in here, you'll only get gradients backpropping
through two of them (the ones on the margin). This could be an inefficient use of your
computation time. Think carefully about the data that you're using with this loss function.
Because of the way masking works with Keras loss functions, also, you need to be sure that any
masked elements in ``y_pred`` have very negative values before they get passed into this loss
function.
"""
correct_elements = y_pred + (1.0 - y_true) * VERY_LARGE_NUMBER
lowest_scoring_correct = K.min(correct_elements, axis=-1)
incorrect_elements = y_pred + y_true * VERY_NEGATIVE_NUMBER
highest_scoring_incorrect = K.max(incorrect_elements, axis=-1)
return K.mean(K.maximum(0.0, 1.0 + highest_scoring_incorrect - lowest_scoring_correct))
def call(self, x, mask=None):
J = self.num_component
# Initialize loss as zero, we'll add terms to it later
loss = backend.variable(0.)
# The mean vector = [mu_0, mu_{1,..., J}]
means = backend.concatenate([backend.variable([0.]), self.means], axis=0)
# Convert back to 1 / variance = precision
precision = backend.exp(self.gammas)
# Get the mixing proportions back from exp(tau)
# Employ the log_sum_exp trick to prevent over/underflow
# log \sum exp(x_i) = a + log \sum exp(x_i - a)
# a = min x_i
a = backend.min(self.rhos)
mix_prop = backend.exp(self.rhos - a)
mix_prop = (1 - self.pi_0) * mix_prop / backend.sum(mix_prop)
mix_prop = backend.concatenate([backend.variable([self.pi_0]), mix_prop], axis=0)
# Compute the negative log likelihood of the weights w.r.t to the mixture model
loss += sum([self.nll(weights, mix_prop, means, precision) for weights in self.arr_net_weight])
# Gamma hyper-prior parameters - on the zero mean gaussian
(alpha_0, beta_0) = (5000.0, 2.0)
negative_log_prob = (1 - alpha_0) * backend.gather(self.gammas, [0]) + beta_0 * backend.gather(precision, [0])
loss += backend.sum(negative_log_prob)
# Gamma hyper-prior parameters - on the rest of the gaussians
alpha, beta = (250, 0.1)
index = np.arange(1, J)
negative_log_prob = (1 - alpha) * backend.gather(self.gammas, index) + beta * backend.gather(precision, index)
loss += backend.sum(negative_log_prob)
return loss
def msml_loss(y_true, y_pred):
global SN
global PN
feat_num = SN*PN # images num
y_pred = K.l2_normalize(y_pred,axis=1)
feat1 = K.tile(K.expand_dims(y_pred,axis = 0),[feat_num,1,1])
feat2 = K.tile(K.expand_dims(y_pred,axis = 1),[1,feat_num,1])
delta = feat1 - feat2
dis_mat = K.sum(K.square(delta),axis = 2) + K.epsilon() # Avoid gradients becoming NAN
dis_mat = K.sqrt(dis_mat)
positive = dis_mat[0:SN,0:SN]
negetive = dis_mat[0:SN,SN:]
for i in range(1,PN):
positive = tf.concat([positive,dis_mat[i*SN:(i+1)*SN,i*SN:(i+1)*SN]],axis = 0)
if i != PN-1:
negs = tf.concat([dis_mat[i*SN:(i+1)*SN,0:i*SN],dis_mat[i*SN:(i+1)*SN, (i+1)*SN:]],axis = 1)
else:
negs = tf.concat(dis_mat[i*SN:(i+1)*SN, 0:i*SN],axis = 0)
negetive = tf.concat([negetive,negs],axis = 0)
positive = K.max(positive)
negetive = K.min(negetive)
a1 = 0.6
loss = K.mean(K.maximum(0.0,positive-negetive+a1))
return loss
def yolo_correct_boxes(box_xy, box_wh, input_shape, image_shape):
'''Get corrected boxes'''
box_yx = box_xy[..., ::-1]
box_hw = box_wh[..., ::-1]
input_shape = K.cast(input_shape, K.dtype(box_yx))
image_shape = K.cast(image_shape, K.dtype(box_yx))
new_shape = K.round(image_shape * K.min(input_shape / image_shape))
offset = (input_shape - new_shape) / 2. / input_shape
scale = input_shape / new_shape
box_yx = (box_yx - offset) * scale
box_hw *= scale
box_mins = box_yx - (box_hw / 2.)
box_maxes = box_yx + (box_hw / 2.)
boxes = K.concatenate([
box_mins[..., 0:1], # y_min
box_mins[..., 1:2], # x_min
box_maxes[..., 0:1], # y_max
box_maxes[..., 1:2] # x_max
])
# Scale boxes back to original image shape.
boxes *= K.concatenate([image_shape, image_shape])
return boxes
def triplet_hard_loss(y_true, y_pred):
global SN
global PN
feat_num = SN*PN # images num
y_pred = K.l2_normalize(y_pred,axis=1)
feat1 = K.tile(K.expand_dims(y_pred,axis = 0),[feat_num,1,1])
feat2 = K.tile(K.expand_dims(y_pred,axis = 1),[1,feat_num,1])
delta = feat1 - feat2
dis_mat = K.sum(K.square(delta),axis = 2)
dis_mat = K.sqrt(dis_mat) + 1e-8 #1e-8 is not necessary
positive = dis_mat[0:SN,0:SN]
negetive = dis_mat[0:SN,SN:]
for i in range(1,PN):
positive = tf.concat([positive,dis_mat[i*SN:(i+1)*SN,i*SN:(i+1)*SN]],axis = 0)
if i != PN-1:
negs = tf.concat([dis_mat[i*SN:(i+1)*SN,0:i*SN],dis_mat[i*SN:(i+1)*SN, (i+1)*SN:]],axis = 1)
else:
negs = tf.concat(dis_mat[i*SN:(i+1)*SN, 0:i*SN],axis = 0)
negetive = tf.concat([negetive,negs],axis = 0)
positive = K.max(positive,axis=1)
negetive = K.min(negetive,axis=1)
a1 = 0.6
loss = K.mean(K.maximum(0.0,positive-negetive+a1))
return loss
def yolo_correct_boxes(box_xy, box_wh, input_shape, image_shape):
'''Get corrected boxes'''
'''还原原始图像真实物理大小
@param box_xy 预测值,相对全图大小偏移量,0-1之间,shape=>(batch_size,height,width,num_anchors,2)
@param box_wh 预测值,相对全图大小比例值， shape=>(batch_size,height,width,num_anchors,2)
@param input_shape 模型输入尺寸,val=>(416,416)
@param image_shape 图像真实尺寸,shape=>(?,?)
@return boxes
'''
#归一化数据状态
box_yx = box_xy[..., ::-1]
box_hw = box_wh[..., ::-1]
input_shape = K.cast(input_shape, K.dtype(box_yx)) #模型输入大小
image_shape = K.cast(image_shape, K.dtype(box_yx)) #图像真实大小
new_shape = K.round(image_shape *
K.min(input_shape / image_shape)) #原始图像等比缩放后的shape
offset = (input_shape -
new_shape) / 2. / input_shape #原始图像等比缩放后与模型输入shape的边偏移量(归一化)
scale = input_shape / new_shape #图像缩放倍数
#真实数据复原,box_yx-offset为get_random_data的box[:, [0,2]] = box[:, [0,2]]*scale + dx的逆向操作
box_yx = (box_yx - offset) * scale
box_hw *= scale #真实数据复原
box_mins = box_yx - (box_hw / 2.)
box_maxes = box_yx + (box_hw / 2.)
boxes = K.concatenate([
box_mins[..., 0:1], # y_min
box_mins[..., 1:2], # x_min
box_maxes[..., 0:1], # y_max
box_maxes[..., 1:2] # x_max
])
# Scale boxes back to original image shape.
#还原真实物理尺寸
boxes *= K.concatenate([image_shape, image_shape])
return boxes
def yolo_trans_boxes(box_xy, box_wh, input_shape, image_shape):
# 获取到box的x,y,w,h信息
box_yx = box_xy[..., ::-1]
box_hw = box_wh[..., ::-1]
# 进行shape的转换
input_shape = K.cast(input_shape, K.dtype(box_yx))
image_shape = K.cast(image_shape, K.dtype(box_yx))
new_shape = K.round(image_shape * K.min(input_shape / image_shape))
# 获取box的实际位置信息
offset = (input_shape - new_shape) / 2. / input_shape
scale = input_shape / new_shape
box_yx = (box_yx - offset) * scale
box_hw *= scale
# box转换
box_mins = box_yx - (box_hw / 2.)
box_maxes = box_yx + (box_hw / 2.)
boxes = K.concatenate([
box_mins[..., 0:1], box_mins[..., 1:2], box_maxes[..., 0:1],
box_maxes[..., 1:2]
])
boxes *= K.concatenate([image_shape, image_shape])
return boxes
def yolo_correct_boxes(box_xy, box_wh, input_shape, image_shape):
box_yx = box_xy[..., ::-1]
box_hw = box_wh[..., ::-1]
input_shape = K.cast(input_shape, K.dtype(box_yx))
image_shape = K.cast(image_shape, K.dtype(box_yx))
new_shape = K.round(image_shape * K.min(input_shape/image_shape))
offset = (input_shape-new_shape)/2./input_shape
scale = input_shape/new_shape
box_yx = (box_yx - offset) * scale
box_hw *= scale
box_mins = box_yx - (box_hw / 2.)
box_maxes = box_yx + (box_hw / 2.)
boxes = K.concatenate([
box_mins[..., 0:1],
box_mins[..., 1:2],
box_maxes[..., 0:1],
box_maxes[..., 1:2]
])
boxes *= K.concatenate([image_shape, image_shape])
return boxes
def min(t, axis=None, keepdims=False):
"""
min Minimum values of tensor, element-wise
Parameters
----------
t : backend.Tensor
axis : int, optional
The dimension to min over, or if None min over all
dimensions. By default None
keepdims : bool, optional
If `keepdims` is `False`, the rank of the tensor is reduced
by 1. If `keepdims` is `True`, the reduced dimension is retained
with length 1., by default False
Returns
-------
backend.Tensor
Min of t
"""
if isinstance(t, np.ndarray):
return t.sum(axis=axis, keepdims=keepdims)
return K.min(t, axis, keepdims)
def _batch_hard_triplet_loss(self, y_true: Tensor,
pairwise_dist: Tensor) -> Tensor:
mask_anchor_positive = self._get_anchor_positive_triplet_mask(
y_true, pairwise_dist)
anchor_positive_dist = mask_anchor_positive * pairwise_dist
hardest_positive_dist = K.max(anchor_positive_dist,
axis=1,
keepdims=True)
mask_anchor_negative = self._get_anchor_negative_triplet_mask(
y_true, pairwise_dist)
anchor_negative_dist = mask_anchor_negative * pairwise_dist
mask_anchor_negative = self._get_semihard_anchor_negative_triplet_mask(
anchor_negative_dist, hardest_positive_dist, mask_anchor_negative)
max_anchor_negative_dist = K.max(pairwise_dist, axis=1, keepdims=True)
anchor_negative_dist = pairwise_dist + max_anchor_negative_dist * (
1.0 - mask_anchor_negative)
hardest_negative_dist = K.min(anchor_negative_dist,
axis=1,
keepdims=True)
triplet_loss = K.clip(
hardest_positive_dist - hardest_negative_dist + self.margin, 0.0,
None)
triplet_loss = K.mean(triplet_loss)
return triplet_loss
def yolo_correct_boxes(box_xy, box_wh, input_shape, image_shape):
'''
通过上面得到的系数,计算真实框的位置.
:param box_xy:
:param box_wh:
:param input_shape:
:param image_shape:
:return:
'''
'''Get corrected boxes'''
box_yx = box_xy[..., ::-1] # 中心坐标 ,下面一行是长宽.
box_hw = box_wh[..., ::-1] #先逆序
input_shape = K.cast(input_shape, K.dtype(box_yx))
image_shape = K.cast(image_shape, K.dtype(box_yx))
new_shape = K.round(image_shape * K.min(input_shape / image_shape))
offset = (input_shape - new_shape) / 2. / input_shape
scale = input_shape / new_shape #new_shape 是输出的图片大小,也就是label数据集里面图片大小.
box_yx = (box_yx - offset) * scale #image_shape 是在new_shape外面多一层0
#box_yx表示0-1之间,的图片中心坐标.在图片整体的百分比坐标.
#得到的box_yx 是在new_shape 图片大小意义下,中心点百分比坐标.
box_hw *= scale
box_mins = box_yx - (box_hw / 2.)
box_maxes = box_yx + (box_hw / 2.)
boxes = K.concatenate([ #默认拼接最后一列.
box_mins[..., 0:1], # y_min
box_mins[..., 1:2], # x_min
box_maxes[..., 0:1], # y_max
box_maxes[..., 1:2] # x_max
])
# Scale boxes back to original image shape.
boxes *= K.concatenate([image_shape, image_shape])
return boxes
def huber_loss(y_true, y_pred):
d = 0.15
x = K.abs(y_true - y_pred)
d_t = d * K.ones_like(x)
quad = K.min(K.stack([x, d_t], axis=-1), axis=-1)
return (0.5 * K.square(quad) + d * (x - quad))
def categorical_accuracy_per_sequence(y_true, y_pred):
return K.mean(K.min(K.equal(K.argmax(y_true, axis=-1), K.argmax(y_pred, axis=-1)), axis=-1))
def active(y_true, y_pred):
return K.max(y_pred) - K.min(y_pred)
def DirectedSumMin1(distance, x, y, N):
x2, y2 = _repeat_nxn_matrix(x,y,N)
d = K.sum(distance(x2,y2), axis=-1) # [batch, N, N]
sum_x = K.sum(K.min(d, axis=2), axis=1) # [batch]
return K.mean(sum_x)
def max_min_1d(X):
max_ = K.max(X, axis=1)
min_ = K.min(X, axis=1)
return K.concatenate([max_, min_], axis=1)
def _all_accuracy(y_true, y_pred):
acc_all = K.cast(K.equal(y_true, K.round(y_pred)), 'int8')
acc_batch = K.min(acc_all, axis=[1, 2, 3, 4, 5])
acc = K.mean(K.cast(acc_batch, 'float32'), axis=-1)
return acc
def call(self, inputs):
if self.data_format == 'channels_last':
pooled = K.min(inputs, axis=[1, 2])
else:
pooled = K.min(inputs, axis=[2, 3])
return pooled
def metrics_pred_min(y_true, y_pred):
return K.min(y_pred)
def act_min(y_true, y_pred):
return K.min(y_pred)
def exact_accruacy(y_true, y_pred):
return K.min(K.cast(K.equal(y_true, y_pred), K.floatx()), axis=-1)
def metrics_pred_min(y_true, y_pred):
return K.min(y_pred)
fcsv = csv.reader(f)
for i in range(10):
next(fcsv)
a = next(fcsv)
pxl = a[1]
pxl = np.fromstring(pxl, dtype=float, sep=' ').reshape((1, 48, 48, 1))
input_img = model.input
val_proba = model.predict(pxl)
pred = val_proba.argmax(axis=-1)
target = K.mean(model.output[:, pred])
grads = K.gradients(target, input_img)[0]
grads = (grads - K.mean(grads)) / K.std(grads)
grads = (grads - K.min(grads)) / (K.max(grads) - K.min(grads))
fn = K.function([input_img, K.learning_phase()], [grads])
heatmap = fn([pxl, False])
heatmap = np.array(heatmap).reshape(48, 48)
thres = 0.5
see = pxl.reshape(48, 48)
see[np.where(heatmap <= thres)] = np.mean(see)
plt.figure()
plt.imshow(heatmap, cmap=plt.cm.jet)
plt.colorbar()
plt.tight_layout()
fig = plt.gcf()
def min_error(y_true, y_pred):
return K.min(K.abs(y_true - y_pred))
def __call__(self, w):
inc = K.minimum(w, K.epsilon()) * self.rate
pos = w - K.min(inc, axis=self.axis, keepdims=True)
abs_sum = K.sum(K.abs(pos), axis=self.axis, keepdims=True)
desired = self.rate + (1 - self.rate) * abs_sum
return pos * desired / (K.maximum(K.epsilon(), abs_sum))
def ctn(X):
max_ = K.max(X, axis=1)
min_ = K.min(X, axis=1)
mean_ = K.mean(X, axis=1)
return K.concatenate([mean_, max_, min_], axis=1)
def call(self, x):
"""Build the actual logic."""
# (None, n_features, n_particles) -> (None, n_features, n_particles)
if self.debug:
x = K.variable(
np.array([[[
229.46118164, 132.46817017, 26.43243217, 13.2313776,
5.75571156
],
[
-195.08522034, -113.19028473, -22.73009872,
-10.31623554, -4.25184822
],
[
-114.19178772, -65.08143616, -12.34527397,
-8.04754353, -2.59461427
],
[
-39.42618179, -22.36474037, -5.44153976,
-1.97019398, -2.88409066
]],
[[129., 135., 26., 15., 7.],
[-105., -114., -20., -10., -6.],
[-100., -60., -10., -8., -1.],
[-32., -20., -5., -1., -2.]]]))
weight_index = 0
metric_index = 0
out_features = []
metric3 = K.variable(np.array([0., 1., 1., 1.]))
# Our input is of the form
# (b,f,p)
# -> (batch_size, features, particles)
# Let's build a few helpful matrices
# All the individual dimensions
# bp
Es = x[:, 0, :]
Xs = x[:, 1, :]
Ys = x[:, 2, :]
Zs = x[:, 3, :]
# A1s = x[:,4,:]
# A2s = x[:,5,:]
# Element wise square of x
# bfp
x2 = pow(x, 2)
# Mass^2 and transverse momentum^2
# bp
Ms = x2[:, 0, :] - x2[:, 1, :] - x2[:, 2, :] - x2[:, 3, :]
Pts = K.abs(K.sqrt(x2[:, 1, :] + x2[:, 2, :]))
#Pts = K.map_fn( lambda x:K.switch(K.less(x,self.t[0,0]),0,1), Es)
if self.ms:
out_features.append(Ms)
if self.es:
out_features.append(Es)
if self.pts:
out_features.append(Pts)
# out_features.append(A1s)
# out_features.append(A2s)
for i in range(self.n_train_es):
out_features.append(
theano.tensor.tensordot(Es,
self.w[weight_index, :, :],
axes=[1, 0]))
weight_index += 1
for i in range(self.n_train_ms):
out_features.append(
theano.tensor.tensordot(Ms,
self.w[weight_index, :, :],
axes=[1, 0]))
weight_index += 1
for i in range(self.n_train_pts):
out_features.append(
theano.tensor.tensordot(Pts,
self.w[weight_index, :, :],
axes=[1, 0]))
weight_index += 1
for i in range(self.n_train_a1s):
out_features.append(
theano.tensor.tensordot(A1s,
self.w[weight_index, :, :],
axes=[1, 0]))
weight_index += 1
for i in range(self.n_train_a2s):
out_features.append(
theano.tensor.tensordot(A2s,
self.w[weight_index, :, :],
axes=[1, 0]))
weight_index += 1
# Helper tensor for building sums/differences
# magic1: 111 000 000
# 000 111 000
# 000 000 111
#
# magic2: 100 100 100
# 010 010 010
# 001 001 001
magic1 = K.repeat_elements(K.eye(self.n_particles), self.n_particles,
1)
magic2 = K.tile(K.eye(self.n_particles), [1, self.n_particles])
magic_diff = magic1 - magic2
magic_sum = magic1 - magic2
# Build d_ij^2 = (k_i - k_j)^mu (k_i - k_j)^nu eta_mu_nu
# b f p p'
# x * magic gives b f p^2, reshape to b f p p'
d2_ij = K.reshape(K.expand_dims(K.dot(x, magic_diff), -1),
(x.shape[0], x.shape[1], x.shape[2], x.shape[2]))
# elements squared
d2_ij = K.pow(d2_ij, 2)
# fold with the metric
# b f p p' * f = b p p'
for i in range(self.n_train_dijs):
if self.debug:
print("metric:")
print(K.eval(self.m[metric_index, :].shape))
print(K.eval(self.m[metric_index, :]))
if self.debug:
print("d2_ij:")
print(K.eval(d2_ij.shape))
print(K.eval(d2_ij))
if self.train_metric:
metric = self.m[metric_index, :]
metric_index += 1
else:
metric = K.variable(np.array([-1., 1., 1., 1.]))
m_d2_ij = theano.tensor.tensordot(d2_ij, metric, axes=[1, 0])
out_features.append(
K.sum(theano.tensor.tensordot(m_d2_ij,
self.w[weight_index, :, :],
axes=[1, 0]),
axis=2))
weight_index += 1
if self.debug:
print("Done with d2_ijs")
# Build d_ij^2 = (k_i - k_j)^mu (k_i - k_j)^nu eta_mu_nu
# b f p p'
# x * magic gives b f p^2, reshape to b f p p'
d2_3d_ij = K.reshape(K.expand_dims(K.dot(x, magic_diff), -1),
(x.shape[0], x.shape[1], x.shape[2], x.shape[2]))
# elements squared
d2_3d_ij = K.pow(d2_3d_ij, 2)
# fold with the metric
# b f p p' * f = b p p'
d2_3d_ij = theano.tensor.tensordot(d2_3d_ij, metric3, axes=[1, 0])
for i in range(self.n_train_dijs_3d):
out_features.append(
K.sum(theano.tensor.tensordot(d2_3d_ij,
self.w[weight_index, :, :],
axes=[1, 0]),
axis=2))
weight_index += 1
# Build m_ij^2 = (k_i + k_j)^mu (k_i + k_j)^nu eta_mu_nu
# b f p p'
# x * magic gives b f p^2, reshape to b f p p'
m2_ij = K.reshape(K.expand_dims(K.dot(x, magic_sum), -1),
(x.shape[0], x.shape[1], x.shape[2], x.shape[2]))
# elements squared
m2_ij = K.pow(m2_ij, 2)
# fold with the metric
# b f p p' * f = b p p'
for i in range(self.n_train_mijs):
if self.train_metric:
metric = self.m[metric_index, :]
metric_index += 1
else:
metric = K.variable(np.array([-1., 1., 1., 1.]))
m_m2_ij = theano.tensor.tensordot(m2_ij, metric, axes=[1, 0])
out_features.append(
K.min(theano.tensor.tensordot(m_m2_ij,
self.w[weight_index, :, :],
axes=[1, 0]),
axis=2))
weight_index += 1
if self.debug:
print("done with m2_ij")
# Build m_ij^2 = (k_i + k_j)^mu (k_i + k_j)^nu (only 3metric)
# b f p p'
# x * magic gives b f p^2, reshape to b f p p'
m2_3d_ij = K.reshape(K.expand_dims(K.dot(x, magic_sum), -1),
(x.shape[0], x.shape[1], x.shape[2], x.shape[2]))
# elements squared
m2_3d_ij = K.pow(m2_3d_ij, 2)
# fold with the metric
# b f p p' * f = b p p'
m2_3d_ij = theano.tensor.tensordot(m2_3d_ij, metric3, axes=[1, 0])
for i in range(self.n_train_mijs_3d):
out_features.append(
K.sum(theano.tensor.tensordot(m2_3d_ij,
self.w[weight_index, :, :],
axes=[1, 0]),
axis=2))
weight_index += 1
# Build cos_ij = m_ij^2 / 2 E_i E_j
Ei = K.expand_dims(Es, -1) # bpN
Ej = K.expand_dims(Es, -2) # bNp
Eij = Ei * Ej # bpp
Eij = K.clip(Eij, 0.0001, 100000000.)
ratio_m2ij_Eij = m2_ij * pow(Eij, -1)
onemat = K.ones((self.n_particles, self.n_particles))
onemat = K.expand_dims(onemat, 0)
cos_ij = onemat + ratio_m2ij_Eij
for i in range(self.n_train_cosijs):
out_features.append(
K.sum(theano.tensor.tensordot(cos_ij,
self.w[weight_index, :, :],
axes=[1, 0]),
axis=2))
weight_index += 1
# if self.debug:
# print ("Eij:")
# print (K.eval(Eij))
# print (K.eval(Eij.shape))
# print ("sum_ij:")
# print (K.eval(sum_ij))
# print (K.eval(sum_ij.shape))
# print ("ratio:")
# print (K.eval(ratio_dij_Eij))
# print (K.eval(ratio_dij_Eij.shape))
# print ("cos:")
# print (K.eval(cos))
# print (K.eval(cos.shape))
# TODO: Also enable these..
Pts_over_lead = Pts / K.repeat_elements(
K.expand_dims(Pts[:, 0], axis=-1), self.n_particles, -1)
Es_over_lead = Es / K.repeat_elements(K.expand_dims(Es[:, 0], axis=-1),
self.n_particles, -1)
results = K.stack(out_features, axis=1)
if self.debug:
print("results:")
print(K.eval(results))
print(K.eval(results.shape))
if self.debug:
sys.exit()
return results
def call(self, inputs, **kwargs):
return K.min(inputs, axis=1)
def f1_min(y_true, y_pred):
f1 = base_f1(y_true, y_pred)
return K.min(f1)
def DirectedSumMin2(distance, x, y, N):
x2, y2 = _repeat_nxn_matrix(x,y,N)
d = K.sum(distance(x2,y2), axis=-1) # [batch, N, N]
sum_y = K.sum(K.min(d, axis=1), axis=1) # [batch] --- mind the axis
return K.mean(sum_y)
def loss_hardnet(x, anchor_swap = False, anchor_ave = False,\
margin = 100.0, batch_reduce = 'min', loss_type = "triplet_margin"):
"""HardNet margin loss - calculates loss based on distance matrix based on positive distance and closest negative distance.
"""
anchor, positive = x
#print("anchor_shape, pos shape")
#assert anchor.shape == positive.shape, "Input sizes between positive and negative must be equal."
assert len(anchor.shape) == 2, "Inputd must be a 2D matrix."
eps = 1e-8
dist_matrix = distance_matrix_vector(anchor, positive) + eps
eye = tf.cast(tf.diag(tf.fill(tf.shape(dist_matrix[0]), 1)),
dist_matrix.dtype)
# steps to filter out same patches that occur in distance matrix as negatives
pos1 = tf.linalg.tensor_diag_part(dist_matrix)
dist_without_min_on_diag = dist_matrix + eye * 10
#con = tf.constant([0.008], dtype=dist_without_min_on_diag.dtype)
#mask = (tf.cast(tf.math.greater_equal(dist_without_min_on_diag, con), dist_without_min_on_diag.dtype)-1)*-1
#mask = (dist_without_min_on_diag.ge(0.008).float()-1)*-1
#mask = tf.cast(mask, dist_without_min_on_diag.dtype)*10
#dist_without_min_on_diag = dist_without_min_on_diag+mask
if batch_reduce == 'min':
min_neg = K.min(dist_without_min_on_diag, 1)[0]
if anchor_swap:
min_neg2 = K.min(dist_without_min_on_diag, 0)[0]
min_neg = K.min(min_neg, min_neg2)
# if False:
# dist_matrix_a = distance_matrix_vector(anchor, anchor)+ eps
# dist_matrix_p = distance_matrix_vector(positive,positive)+eps
# dist_without_min_on_diag_a = dist_matrix_a+eye*10
# dist_without_min_on_diag_p = dist_matrix_p+eye*10
# min_neg_a = torch.min(dist_without_min_on_diag_a,1)[0]
# min_neg_p = torch.t(torch.min(dist_without_min_on_diag_p,0)[0])
# min_neg_3 = torch.min(min_neg_p,min_neg_a)
# min_neg = torch.min(min_neg,min_neg_3)
# print (min_neg_a)
# print (min_neg_p)
# print (min_neg_3)
# print (min_neg)
min_neg = min_neg
pos = pos1
#print("min_neg, pos", K.eval(min_neg), K.eval(pos))
# elif batch_reduce == 'average':
# pos = pos1.repeat(anchor.size(0)).view(-1,1).squeeze(0)
# min_neg = dist_without_min_on_diag.view(-1,1)
# if anchor_swap:
# min_neg2 = torch.t(dist_without_min_on_diag).contiguous().view(-1,1)
# min_neg = torch.min(min_neg,min_neg2)
# min_neg = min_neg.squeeze(0)
# elif batch_reduce == 'random':
# idxs = torch.autograd.Variable(torch.randperm(anchor.size()[0]).long()).cuda()
# min_neg = dist_without_min_on_diag.gather(1,idxs.view(-1,1))
# if anchor_swap:
# min_neg2 = torch.t(dist_without_min_on_diag).gather(1,idxs.view(-1,1))
# min_neg = torch.min(min_neg,min_neg2)
# min_neg = torch.t(min_neg).squeeze(0)
# pos = pos1
# else:
# print ('Unknown batch reduce mode. Try min, average or random')
# sys.exit(1)
if loss_type == "triplet_margin":
loss = K.clip(margin + pos - min_neg, 0.0, 1e10)
# elif loss_type == 'softmax':
# exp_pos = torch.exp(2.0 - pos);
# exp_den = exp_pos + torch.exp(2.0 - min_neg) + eps;
# loss = - torch.log( exp_pos / exp_den )
# elif loss_type == 'contrastive':
# loss = torch.clamp(margin - min_neg, min=0.0) + pos;
# else:
# print ('Unknown loss type. Try triplet_margin, softmax or contrastive')
# sys.exit(1)
loss = K.expand_dims(loss, axis=1) #K.mean(loss)
return loss
def min_pred(y_true, y_pred):
return K.min(y_pred)
def content_loss(base, style, combination):
# Changes from equation 7 (Pg# 5)
G = style / (base + 1e-04)
G_clamped = K.max(K.min(G, g_max), g_min) # Clamping values
Fm = base * G_clamped
return K.sum(K.square(combination - Fm))