def predict(model, params):
"""
Compute the prediction result of the given model
"""
model_args = {arg.name: arg for arg in model.arguments}
context = model_args['context']
entity_ids_mask = model_args['entity_ids_mask']
entity_condition = greater(entity_ids_mask, 0, name='condidion')
# Get all the enities in the paragraph via gather operator, which will create a new dynamic sequence axis
entities_all = sequence.gather(entity_condition,
entity_condition,
name='entities_all')
# The generated dynamic axis has the same length as the input enity id sequence,
# so we asign it as the entity id's dynamic axis.
entity_ids = input(shape=(params.entity_dim),
is_sparse=True,
dynamic_axes=entities_all.dynamic_axes,
name='entity_ids')
wordvocab_dim = params.vocab_dim
answers = sequence.scatter(sequence.gather(model.outputs[-1],
entity_condition),
entities_all,
name='Final_Ans')
entity_id_matrix = ops.slice(ops.reshape(entity_ids, params.entity_dim),
-1, 1, params.entity_dim)
expand_pred = sequence.reduce_sum(element_times(answers, entity_id_matrix))
pred_max = ops.hardmax(expand_pred, name='pred_max')
return pred_max
def loss(model, params):
"""
Compute the loss and accuracy of the model output
"""
model_args = {arg.name: arg for arg in model.arguments}
context = model_args['context']
entity_ids_mask = model_args['entity_ids_mask']
entity_condition = greater(entity_ids_mask, 0, name='condidion')
# Get all the enities in the paragraph via gather operator, which will create a new dynamic sequence axis
entities_all = sequence.gather(entity_condition,
entity_condition,
name='entities_all')
# The generated dynamic axis has the same length as the input enity id sequence,
# so we asign it as the entity id's dynamic axis.
entity_ids = input(shape=(params.entity_dim),
is_sparse=True,
dynamic_axes=entities_all.dynamic_axes,
name='entity_ids')
wordvocab_dim = params.vocab_dim
labels_raw = input(shape=(1, ),
is_sparse=False,
dynamic_axes=context.dynamic_axes,
name='labels')
answers = sequence.scatter(sequence.gather(model.outputs[-1],
entity_condition),
entities_all,
name='Final_Ans')
labels = sequence.scatter(sequence.gather(labels_raw, entity_condition),
entities_all,
name='EntityLabels')
entity_id_matrix = ops.reshape(entity_ids, params.entity_dim)
expand_pred = sequence.reduce_sum(element_times(answers, entity_id_matrix))
expand_label = ops.greater_equal(
sequence.reduce_sum(element_times(labels, entity_id_matrix)), 1)
predictions_and_stop_probabilities = []
for step in range(int((len(model.outputs) - 1) / 2)):
predictions_and_stop_probabilities += [(model.outputs[step * 2],
model.outputs[step * 2 + 1])]
loss_value = contractive_reward(labels_raw,
predictions_and_stop_probabilities)
accuracy = accuracy_func(expand_pred, expand_label, name='accuracy')
apply_loss = combine([loss_value, answers, labels, accuracy], name='Loss')
return apply_loss
def KitModel(weight_file=None):
global __weights_dict
__weights_dict = load_weights(weight_file)
imgLow = cntk.input_variable((
227,
227,
3,
), name='imgLow')
conv1 = convolution(imgLow,
strides=(4, 4),
auto_padding=[False, False, False],
dilation=(1, ),
groups=1,
name='conv1')
relu1 = layers.Activation(activation=ops.relu, name='relu1')(conv1)
pool1 = pooling(relu1,
pooling_type=0,
pooling_window_shape=(3, 3),
strides=(2, 2),
auto_padding=[False, False, False],
ceil_out_dim=True)
norm1 = lrn(pool1,
k=1,
n=3,
alpha=9.999999747378752e-05,
beta=0.75,
name='norm1')
conv2 = convolution(norm1,
strides=(1, 1),
auto_padding=[False, True, True],
dilation=(1, ),
groups=2,
name='conv2')
relu2 = layers.Activation(activation=ops.relu, name='relu2')(conv2)
pool2 = pooling(relu2,
pooling_type=0,
pooling_window_shape=(3, 3),
strides=(2, 2),
auto_padding=[False, False, False],
ceil_out_dim=True)
norm2 = lrn(pool2,
k=1,
n=3,
alpha=9.999999747378752e-05,
beta=0.75,
name='norm2')
conv3 = convolution(norm2,
strides=(1, 1),
auto_padding=[False, True, True],
dilation=(1, ),
groups=1,
name='conv3')
relu3 = layers.Activation(activation=ops.relu, name='relu3')(conv3)
conv4 = convolution(relu3,
strides=(1, 1),
auto_padding=[False, True, True],
dilation=(1, ),
groups=2,
name='conv4')
relu4 = layers.Activation(activation=ops.relu, name='relu4')(conv4)
conv5 = convolution(relu4,
strides=(1, 1),
auto_padding=[False, True, True],
dilation=(1, ),
groups=2,
name='conv5')
relu5 = layers.Activation(activation=ops.relu, name='relu5')(conv5)
pool5 = pooling(relu5,
pooling_type=0,
pooling_window_shape=(3, 3),
strides=(2, 2),
auto_padding=[False, False, False],
ceil_out_dim=True)
fc6_0 = ops.reshape(pool5, (-1, ), name='fc6_0')
fc6_1 = dense(fc6_0, name='fc6_1')
relu6 = layers.Activation(activation=ops.relu, name='relu6')(fc6_1)
fc7_0 = ops.reshape(relu6, (-1, ), name='fc7_0')
fc7_1 = dense(fc7_0, name='fc7_1')
relu7 = layers.Activation(activation=ops.relu, name='relu7')(fc7_1)
fc8_VividColor_0 = ops.reshape(relu7, (-1, ), name='fc8_VividColor_0')
fc8_Light_0 = ops.reshape(relu7, (-1, ), name='fc8_Light_0')
fc8_DoF_0 = ops.reshape(relu7, (-1, ), name='fc8_DoF_0')
fc8_Symmetry_0 = ops.reshape(relu7, (-1, ), name='fc8_Symmetry_0')
fc8_MotionBlur_0 = ops.reshape(relu7, (-1, ), name='fc8_MotionBlur_0')
fc8_Repetition_0 = ops.reshape(relu7, (-1, ), name='fc8_Repetition_0')
fc8_RuleOfThirds_0 = ops.reshape(relu7, (-1, ), name='fc8_RuleOfThirds_0')
fc8_ColorHarmony_0 = ops.reshape(relu7, (-1, ), name='fc8_ColorHarmony_0')
fc8_Object_0 = ops.reshape(relu7, (-1, ), name='fc8_Object_0')
fc8_Content_0 = ops.reshape(relu7, (-1, ), name='fc8_Content_0')
fc8new_0 = ops.reshape(relu7, (-1, ), name='fc8new_0')
fc8_BalancingElement_0 = ops.reshape(relu7, (-1, ),
name='fc8_BalancingElement_0')
fc8_VividColor_1 = dense(fc8_VividColor_0, name='fc8_VividColor_1')
fc8_Light_1 = dense(fc8_Light_0, name='fc8_Light_1')
fc8_DoF_1 = dense(fc8_DoF_0, name='fc8_DoF_1')
fc8_Symmetry_1 = dense(fc8_Symmetry_0, name='fc8_Symmetry_1')
fc8_MotionBlur_1 = dense(fc8_MotionBlur_0, name='fc8_MotionBlur_1')
fc8_Repetition_1 = dense(fc8_Repetition_0, name='fc8_Repetition_1')
fc8_RuleOfThirds_1 = dense(fc8_RuleOfThirds_0, name='fc8_RuleOfThirds_1')
fc8_ColorHarmony_1 = dense(fc8_ColorHarmony_0, name='fc8_ColorHarmony_1')
fc8_Object_1 = dense(fc8_Object_0, name='fc8_Object_1')
fc8_Content_1 = dense(fc8_Content_0, name='fc8_Content_1')
fc8new_1 = dense(fc8new_0, name='fc8new_1')
fc8_BalancingElement_1 = dense(fc8_BalancingElement_0,
name='fc8_BalancingElement_1')
relu8_VividColor = layers.Activation(
activation=ops.relu, name='relu8_VividColor')(fc8_VividColor_1)
relu8_Light = layers.Activation(activation=ops.relu,
name='relu8_Light')(fc8_Light_1)
relu8_DoF = layers.Activation(activation=ops.relu,
name='relu8_DoF')(fc8_DoF_1)
relu8_Symmetry = layers.Activation(activation=ops.relu,
name='relu8_Symmetry')(fc8_Symmetry_1)
relu8_MotionBlur = layers.Activation(
activation=ops.relu, name='relu8_MotionBlur')(fc8_MotionBlur_1)
relu8_Repetition = layers.Activation(
activation=ops.relu, name='relu8_Repetition')(fc8_Repetition_1)
relu8_RuleOfThirds = layers.Activation(
activation=ops.relu, name='relu8_RuleOfThirds')(fc8_RuleOfThirds_1)
relu8_ColorHarmony = layers.Activation(
activation=ops.relu, name='relu8_ColorHarmony')(fc8_ColorHarmony_1)
relu8_Object = layers.Activation(activation=ops.relu,
name='relu8_Object')(fc8_Object_1)
relu8_Content = layers.Activation(activation=ops.relu,
name='relu8_Content')(fc8_Content_1)
relu8new = layers.Activation(activation=ops.relu,
name='relu8new')(fc8new_1)
relu8_BalancingElement = layers.Activation(
activation=ops.relu,
name='relu8_BalancingElement')(fc8_BalancingElement_1)
fc9_VividColor_0 = ops.reshape(relu8_VividColor, (-1, ),
name='fc9_VividColor_0')
fc9_Light_0 = ops.reshape(relu8_Light, (-1, ), name='fc9_Light_0')
fc9_DoF_0 = ops.reshape(relu8_DoF, (-1, ), name='fc9_DoF_0')
fc9_Symmetry_0 = ops.reshape(relu8_Symmetry, (-1, ), name='fc9_Symmetry_0')
fc9_MotionBlur_0 = ops.reshape(relu8_MotionBlur, (-1, ),
name='fc9_MotionBlur_0')
fc9_Repetition_0 = ops.reshape(relu8_Repetition, (-1, ),
name='fc9_Repetition_0')
fc9_RuleOfThirds_0 = ops.reshape(relu8_RuleOfThirds, (-1, ),
name='fc9_RuleOfThirds_0')
fc9_ColorHarmony_0 = ops.reshape(relu8_ColorHarmony, (-1, ),
name='fc9_ColorHarmony_0')
fc9_Object_0 = ops.reshape(relu8_Object, (-1, ), name='fc9_Object_0')
fc9_Content_0 = ops.reshape(relu8_Content, (-1, ), name='fc9_Content_0')
fc9_BalancingElement_0 = ops.reshape(relu8_BalancingElement, (-1, ),
name='fc9_BalancingElement_0')
Concat9 = cntk.splice(relu8new,
relu8_BalancingElement,
relu8_ColorHarmony,
relu8_Content,
relu8_DoF,
relu8_Light,
relu8_MotionBlur,
relu8_Object,
relu8_Repetition,
relu8_RuleOfThirds,
relu8_Symmetry,
relu8_VividColor,
axis=2,
name='Concat9')
fc9_VividColor_1 = dense(fc9_VividColor_0, name='fc9_VividColor_1')
fc9_Light_1 = dense(fc9_Light_0, name='fc9_Light_1')
fc9_DoF_1 = dense(fc9_DoF_0, name='fc9_DoF_1')
fc9_Symmetry_1 = dense(fc9_Symmetry_0, name='fc9_Symmetry_1')
fc9_MotionBlur_1 = dense(fc9_MotionBlur_0, name='fc9_MotionBlur_1')
fc9_Repetition_1 = dense(fc9_Repetition_0, name='fc9_Repetition_1')
fc9_RuleOfThirds_1 = dense(fc9_RuleOfThirds_0, name='fc9_RuleOfThirds_1')
fc9_ColorHarmony_1 = dense(fc9_ColorHarmony_0, name='fc9_ColorHarmony_1')
fc9_Object_1 = dense(fc9_Object_0, name='fc9_Object_1')
fc9_Content_1 = dense(fc9_Content_0, name='fc9_Content_1')
fc9_BalancingElement_1 = dense(fc9_BalancingElement_0,
name='fc9_BalancingElement_1')
fc10_merge_0 = ops.reshape(Concat9, (-1, ), name='fc10_merge_0')
fc10_merge_1 = dense(fc10_merge_0, name='fc10_merge_1')
relu10_merge = layers.Activation(activation=ops.relu,
name='relu10_merge')(fc10_merge_1)
fc11_score_0 = ops.reshape(relu10_merge, (-1, ), name='fc11_score_0')
fc11_score_1 = dense(fc11_score_0, name='fc11_score_1')
return fc9_VividColor_1, fc9_Symmetry_1, fc9_Content_1, fc9_RuleOfThirds_1, fc9_ColorHarmony_1, fc9_MotionBlur_1, fc11_score_1, fc9_BalancingElement_1, fc9_Light_1, fc9_DoF_1, fc9_Repetition_1, fc9_Object_1