#!/usr/bin/env python
# coding: utf-8
from tensorflow.keras.models import Model
from tensorflow.keras.layers import GlobalAveragePooling2D, Conv2D, Dense, Concatenate, Flatten, Input
from .abe_models import ABE
from ..layers import L2Normalisation
from ..layers.horde_layers import CKOP, PKOB
from .extractors import get_extractor, get_preprocess_method
def KOrderModel(extractor_name,
embedding_sizes,
high_order_dims,
ho_trainable=False,
end_layer=None):
model = get_extractor(extractor_name, end_layer=end_layer)
inputs = model.input
x = model.output
max_order = len(high_order_dims)
output_list = [x]
# Add all high-order approximation layers:
for k, order_dim in enumerate(high_order_dims, start=2):
x_ho = CKOP(output_dim=order_dim, name='CKOP_' + str(k), ho_trainable=ho_trainable)([x] * k)
output_list.append(x_ho)
# Add pooling and embedding layers:
for k in range(len(output_list)):
output_list[k] = GlobalAveragePooling2D(name='GAP_' + extractor_name + '_O' + str(k + 1))(output_list[k])
if embedding_sizes[k] > 0:
output_list[k] = Dense(embedding_sizes[k], use_bias=False)(output_list[k])
output_list[k] = L2Normalisation(name='L2_' + extractor_name + '_O' + str(k + 1))(output_list[k])
return Model(inputs=inputs, outputs=output_list, name=extractor_name + '_O' + str(max_order)), get_preprocess_method(extractor_name)
def CascadedKOrder(extractor_name,
embedding_sizes,
high_order_dims,
ho_trainable=True,
end_layer=None):
model = get_extractor(extractor_name, end_layer=end_layer)
inputs = model.input
x = model.output
max_order = len(high_order_dims)
output_list = [x]
# Add all high-order approximation layers:
for k, order_dim in enumerate(high_order_dims, start=2):
only_project_second = False if k == 2 else True
x_ho = PKOB(order_dim,
only_project_second=only_project_second,
ho_trainable=ho_trainable)([output_list[-1], x])
output_list.append(x_ho)
# Add pooling and embedding layers:
for k in range(len(output_list)):
output_list[k] = GlobalAveragePooling2D(name='GAP_' + extractor_name + '_O' + str(k + 1))(output_list[k])
if ho_trainable:
output_list[k] = Dense(embedding_sizes[k],
use_bias=False,
name='Proj_' + extractor_name + '_O' + str(k + 1))(output_list[k])
elif k == 0:
output_list[k] = Dense(embedding_sizes[k],
use_bias=False,
name='Proj_' + extractor_name + '_O' + str(k + 1))(output_list[k])
output_list[k] = L2Normalisation(name='L2_' + extractor_name + '_O' + str(k + 1))(output_list[k])
return Model(inputs=inputs, outputs=output_list, name=extractor_name + '_O' + str(max_order)), get_preprocess_method(extractor_name)
def CascadedABE(embedding_size,
high_order_dims,
features_reduction=256,
ho_trainable=True,
n_head=8):
model, preprocess_method = ABE(embedding_size[0], n_head=8)
inp = model.input
multi_head_out = [model.get_layer(name='inception_5b/output').get_output_at(k) for k in range(n_head)]
concat = Concatenate()(multi_head_out) # Nx H x W x n_ensemble*1024
if features_reduction is not None:
concat = Conv2D(filters=features_reduction,
kernel_size=(1, 1),
use_bias=False)(concat)
output_list = [concat]
# Add all high-order approximation layers:
for k, order_dim in enumerate(high_order_dims, start=2):
only_project_second = False if k == 2 else True
x_ho = PKOB(order_dim,
only_project_second=only_project_second,
ho_trainable=ho_trainable)([output_list[-1], concat])
output_list.append(x_ho)
# Add pooling and embedding layers:
for k in range(1, len(output_list)):
output_list[k] = GlobalAveragePooling2D(name='GAP_O' + str(k + 1))(output_list[k])
if ho_trainable:
output_list[k] = Dense(embedding_size[k],
use_bias=False,
name='Proj_O' + str(k + 1))(output_list[k])
output_list[k] = L2Normalisation(name='L2_O' + str(k + 1))(output_list[k])
# Finally we replace the first order by the true model:
output_list[0] = model.get_layer(name='ABE'+str(n_head)).output
return Model(inp, output_list, name='ABE'+str(n_head)+'_O'+str(len(embedding_size))), preprocess_method
