# Copyright (C) 2016 Ross Wightman. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
# ==============================================================================
""" The Inception v4 network.
Implementation of Inception V4, Inception-Resnet-V1, Inception-Resnet-V2
based on https://arxiv.org/abs/1602.07261
"""
from __future__ import absolute_import
from __future__ import division
from __future__ import print_function
import tensorflow as tf
from tensorflow.contrib.framework import arg_scope
from tensorflow.contrib import layers
def _block_stem(net, endpoints, scope='Stem'):
# Stem shared by inception-v4 and inception-resnet-v2 (resnet-v1 uses simpler _stem below)
# NOTE observe endpoints of first 3 layers
with arg_scope([layers.conv2d, layers.max_pool2d, layers.avg_pool2d], padding='VALID'):
with tf.variable_scope(scope):
# 299 x 299 x 3
net = layers.conv2d(net, 32, [3, 3], stride=2, scope='Conv1_3x3/2')
endpoints[scope + '/Conv1'] = net
# 149 x 149 x 32
net = layers.conv2d(net, 32, [3, 3], scope='Conv2_3x3')
endpoints[scope + '/Conv2'] = net
# 147 x 147 x 32
net = layers.conv2d(net, 64, [3, 3], padding='SAME', scope='Conv3_3x3')
endpoints[scope + '/Conv3'] = net
# 147 x 147 x 64
with tf.variable_scope('Br1A_Pool'):
br1a = layers.max_pool2d(net, [3, 3], stride=2, scope='Pool1_3x3/2')
with tf.variable_scope('Br1B_3x3'):
br1b = layers.conv2d(net, 96, [3, 3], stride=2, scope='Conv4_3x3/2')
net = tf.concat(3, [br1a, br1b], name='Concat1')
endpoints[scope + '/Concat1'] = net
# 73 x 73 x 160
with tf.variable_scope('Br2A_3x3'):
br2a = layers.conv2d(net, 64, [1, 1], padding='SAME', scope='Conv5_1x1')
br2a = layers.conv2d(br2a, 96, [3, 3], scope='Conv6_3x3')
with tf.variable_scope('Br2B_7x7x3'):
br2b = layers.conv2d(net, 64, [1, 1], padding='SAME', scope='Conv5_1x1')
br2b = layers.conv2d(br2b, 64, [7, 1], padding='SAME', scope='Conv6_7x1')
br2b = layers.conv2d(br2b, 64, [1, 7], padding='SAME', scope='Conv7_1x7')
br2b = layers.conv2d(br2b, 96, [3, 3], scope='Conv8_3x3')
net = tf.concat(3, [br2a, br2b], name='Concat2')
endpoints[scope + '/Concat2'] = net
# 71 x 71 x 192
with tf.variable_scope('Br3A_3x3'):
br3a = layers.conv2d(net, 192, [3, 3], stride=2, scope='Conv9_3x3/2')
with tf.variable_scope('Br3B_Pool'):
br3b = layers.max_pool2d(net, [3, 3], stride=2, scope='Pool2_3x3/2')
net = tf.concat(3, [br3a, br3b], name='Concat3')
endpoints[scope + '/Concat3'] = net
print('%s output shape: %s' % (scope, net.get_shape()))
# 35x35x384
return net
def _block_a(net, scope='BlockA'):
# 35 x 35 x 384 grid
# default padding = SAME
# default stride = 1
with tf.variable_scope(scope):
with tf.variable_scope('Br1_Pool'):
br1 = layers.avg_pool2d(net, [3, 3], scope='Pool1_3x3')
br1 = layers.conv2d(br1, 96, [1, 1], scope='Conv1_1x1')
with tf.variable_scope('Br2_1x1'):
br2 = layers.conv2d(net, 96, [1, 1], scope='Conv1_1x1')
with tf.variable_scope('Br3_3x3'):
br3 = layers.conv2d(net, 64, [1, 1], scope='Conv1_1x1')
br3 = layers.conv2d(br3, 96, [3, 3], scope='Conv2_3x3')
with tf.variable_scope('Br4_3x3Dbl'):
br4 = layers.conv2d(net, 64, [1, 1], scope='Conv1_1x1')
br4 = layers.conv2d(br4, 96, [3, 3], scope='Conv2_3x3')
br4 = layers.conv2d(br4, 96, [3, 3], scope='Conv3_3x3')
net = tf.concat(3, [br1, br2, br3, br4], name='Concat1')
# 35 x 35 x 384
return net
def _block_a_reduce(net, endpoints, k=192, l=224, m=256, n=384, scope='BlockReduceA'):
# 35 x 35 -> 17 x 17 reduce
# inception-v4: k=192, l=224, m=256, n=384
# inception-resnet-v1: k=192, l=192, m=256, n=384
# inception-resnet-v2: k=256, l=256, m=384, n=384
# default padding = VALID
# default stride = 1
with arg_scope([layers.conv2d, layers.max_pool2d, layers.avg_pool2d], padding='VALID'):
with tf.variable_scope(scope):
with tf.variable_scope('Br1_Pool'):
br1 = layers.max_pool2d(net, [3, 3], stride=2, scope='Pool1_3x3/2')
# 17 x 17 x input
with tf.variable_scope('Br2_3x3'):
br2 = layers.conv2d(net, n, [3, 3], stride=2, scope='Conv1_3x3/2')
# 17 x 17 x n
with tf.variable_scope('Br3_3x3Dbl'):
br3 = layers.conv2d(net, k, [1, 1], padding='SAME', scope='Conv1_1x1')
br3 = layers.conv2d(br3, l, [3, 3], padding='SAME', scope='Conv2_3x3')
br3 = layers.conv2d(br3, m, [3, 3], stride=2, scope='Conv3_3x3/2')
# 17 x 17 x m
net = tf.concat(3, [br1, br2, br3], name='Concat1')
# 17 x 17 x input + n + m
# 1024 for v4 (384 + 384 + 256)
# 896 for res-v1 (256 + 384 +256)
# 1152 for res-v2 (384 + 384 + 384)
endpoints[scope] = net
print('%s output shape: %s' % (scope, net.get_shape()))
return net
def _block_b(net, scope='BlockB'):
# 17 x 17 x 1024 grid
# default padding = SAME
# default stride = 1
with tf.variable_scope(scope):
with tf.variable_scope('Br1_Pool'):
br1 = layers.avg_pool2d(net, [3, 3], scope='Pool1_3x3')
br1 = layers.conv2d(br1, 128, [1, 1], scope='Conv1_1x1')
with tf.variable_scope('Br2_1x1'):
br2 = layers.conv2d(net, 384, [1, 1], scope='Conv1_1x1')
with tf.variable_scope('Br3_7x7'):
br3 = layers.conv2d(net, 192, [1, 1], scope='Conv1_1x1')
br3 = layers.conv2d(br3, 224, [1, 7], scope='Conv2_1x7')
br3 = layers.conv2d(br3, 256, [7, 1], scope='Conv3_7x1')
with tf.variable_scope('Br4_7x7Dbl'):
br4 = layers.conv2d(net, 192, [1, 1], scope='Conv1_1x1')
br4 = layers.conv2d(br4, 192, [1, 7], scope='Conv2_1x7')
br4 = layers.conv2d(br4, 224, [7, 1], scope='Conv3_7x1')
br4 = layers.conv2d(br4, 224, [1, 7], scope='Conv4_1x7')
br4 = layers.conv2d(br4, 256, [7, 1], scope='Conv5_7x1')
net = tf.concat(3, [br1, br2, br3, br4], name='Concat1')
# 17 x 17 x 1024
return net
def _block_b_reduce(net, endpoints, scope='BlockReduceB'):
# 17 x 17 -> 8 x 8 reduce
with arg_scope([layers.conv2d, layers.max_pool2d, layers.avg_pool2d], padding='VALID'):
with tf.variable_scope(scope):
with tf.variable_scope('Br1_Pool'):
br1 = layers.max_pool2d(net, [3, 3], stride=2, scope='Pool1_3x3/2')
with tf.variable_scope('Br2_3x3'):
br2 = layers.conv2d(net, 192, [1, 1], padding='SAME', scope='Conv1_1x1')
br2 = layers.conv2d(br2, 192, [3, 3], stride=2, scope='Conv2_3x3/2')
with tf.variable_scope('Br3_7x7x3'):
br3 = layers.conv2d(net, 256, [1, 1], padding='SAME', scope='Conv1_1x1')
br3 = layers.conv2d(br3, 256, [1, 7], padding='SAME', scope='Conv2_1x7')
br3 = layers.conv2d(br3, 320, [7, 1], padding='SAME', scope='Conv3_7x1')
br3 = layers.conv2d(br3, 320, [3, 3], stride=2, scope='Conv4_3x3/2')
net = tf.concat(3, [br1, br2, br3], name='Concat1')
endpoints[scope] = net
print('%s output shape: %s' % (scope, net.get_shape()))
return net
def _block_c(net, scope='BlockC'):
# 8 x 8 x 1536 grid
# default padding = SAME
# default stride = 1
with tf.variable_scope(scope):
with tf.variable_scope('Br1_Pool'):
br1 = layers.avg_pool2d(net, [3, 3], scope='Pool1_3x3')
br1 = layers.conv2d(br1, 256, [1, 1], scope='Conv1_1x1')
with tf.variable_scope('Br2_1x1'):
br2 = layers.conv2d(net, 256, [1, 1], scope='Conv1_1x1')
with tf.variable_scope('Br3_3x3'):
br3 = layers.conv2d(net, 384, [1, 1], scope='Conv1_1x1')
br3a = layers.conv2d(br3, 256, [1, 3], scope='Conv2_1x3')
br3b = layers.conv2d(br3, 256, [3, 1], scope='Conv3_3x1')
with tf.variable_scope('Br4_7x7Dbl'):
br4 = layers.conv2d(net, 384, [1, 1], scope='Conv1_1x1')
br4 = layers.conv2d(br4, 448, [1, 7], scope='Conv2_1x7')
br4 = layers.conv2d(br4, 512, [7, 1], scope='Conv3_7x1')
br4a = layers.conv2d(br4, 256, [1, 7], scope='Conv4a_1x7')
br4b = layers.conv2d(br4, 256, [7, 1], scope='Conv4b_7x1')
net = tf.concat(3, [br1, br2, br3a, br3b, br4a, br4b], name='Concat1')
# 8 x 8 x 1536
return net
def _block_stem_res(net, endpoints, scope='Stem'):
# Simpler _stem for inception-resnet-v1 network
# NOTE observe endpoints of first 3 layers
# default padding = VALID
# default stride = 1
with arg_scope([layers.conv2d, layers.max_pool2d, layers.avg_pool2d], padding='VALID'):
with tf.variable_scope(scope):
# 299 x 299 x 3
net = layers.conv2d(net, 32, [3, 3], stride=2, scope='Conv1_3x3/2')
endpoints[scope + '/Conv1'] = net
# 149 x 149 x 32
net = layers.conv2d(net, 32, [3, 3], scope='Conv2_3x3')
endpoints[scope + '/Conv2'] = net
# 147 x 147 x 32
net = layers.conv2d(net, 64, [3, 3], padding='SAME', scope='Conv3_3x3')
endpoints[scope + '/Conv3'] = net
# 147 x 147 x 64
net = layers.max_pool2d(net, [3, 3], stride=2, scope='Pool1_3x3/2')
# 73 x 73 x 64
net = layers.conv2d(net, 80, [1, 1], padding='SAME', scope='Conv4_1x1')
# 73 x 73 x 80
net = layers.conv2d(net, 192, [3, 3], scope='Conv5_3x3')
# 71 x 71 x 192
net = layers.conv2d(net, 256, [3, 3], stride=2, scope='Conv6_3x3/2')
# 35 x 35 x 256
endpoints[scope] = net
print('%s output shape: %s' % (scope, net.get_shape()))
return net
def _block_a_res(net, ver=2, res_scale=None, scope='BlockA', activation_fn=tf.nn.relu):
# 35x35 grid
# configure branch filter numbers
br3_num = 32
if ver == 1:
br3_inc = 0
else:
br3_inc = 16
# default padding = SAME
# default stride = 1
with tf.variable_scope(scope):
shortcut = tf.identity(net, name='Shortcut')
if res_scale:
shortcut = tf.mul(shortcut, res_scale) # scale residual
with tf.variable_scope('Br1_1x1'):
br1 = layers.conv2d(net, 32, [1, 1], scope='Conv1_1x1')
with tf.variable_scope('Br2_3x3'):
br2 = layers.conv2d(net, 32, [1, 1], scope='Conv1_1x1')
br2 = layers.conv2d(br2, 32, [3, 3], scope='Conv2_3x3')
with tf.variable_scope('Br3_3x3Dbl'):
br3 = layers.conv2d(net, br3_num, [1, 1], scope='Conv1_1x1')
br3 = layers.conv2d(br3, br3_num + 1*br3_inc, [3, 3], scope='Conv2_3x3')
br3 = layers.conv2d(br3, br3_num + 2*br3_inc, [3, 3], scope='Conv3_3x3')
net = tf.concat(3, [br1, br2, br3], name='Concat1')
net = layers.conv2d(net, shortcut.get_shape()[-1], [1, 1], activation_fn=None, scope='Conv4_1x1')
net = activation_fn(tf.add(shortcut, net, name='Sum1'))
# 35 x 35 x 256 res-v1, 384 res-v2
return net
def _block_b_res(net, ver=2, res_scale=None, scope='BlockB', activation_fn=tf.nn.relu):
# 17 x 17 grid
# configure branch filter numbers
if ver == 1:
br1_num = 128
br2_num = 128
br2_inc = 0
else:
br1_num = 192
br2_num = 128
br2_inc = 32
# default padding = SAME
# default stride = 1
with tf.variable_scope(scope):
shortcut = tf.identity(net, name='Shortcut')
if res_scale:
shortcut = tf.mul(shortcut, res_scale) # scale residual
with tf.variable_scope('Br1_1x1'):
br1 = layers.conv2d(net, br1_num, [1, 1], scope='Conv1_1x1')
with tf.variable_scope('Br2_7x7'):
br2 = layers.conv2d(net, br2_num, [1, 1], scope='Conv1_1x1')
br2 = layers.conv2d(br2, br2_num + 1*br2_inc, [1, 7], scope='Conv2_1x7')
br2 = layers.conv2d(br2, br2_num + 2*br2_inc, [7, 1], scope='Conv3_7x1')
net = tf.concat(3, [br1, br2], name='Concat1')
net = layers.conv2d(net, shortcut.get_shape()[-1], [1, 1], activation_fn=None, scope='Conv4_1x1')
# 17 x 17 x 896 res-v1, 1152 res-v2. Typo in paper, 1152, not 1154
net = activation_fn(tf.add(shortcut, net, name='Sum1'))
return net
def _block_b_reduce_res(net, endpoints, ver=2, scope='BlockReduceB'):
# 17 x 17 -> 8 x 8 reduce
# configure branch filter numbers
br3_num = 256
br4_num = 256
if ver == 1:
br3_inc = 0
br4_inc = 0
else:
br3_inc = 32
br4_inc = 32
with arg_scope([layers.conv2d, layers.max_pool2d, layers.avg_pool2d], padding='VALID'):
with tf.variable_scope(scope):
with tf.variable_scope('Br1_Pool'):
br1 = layers.max_pool2d(net, [3, 3], stride=2, scope='Pool1_3x3/2')
with tf.variable_scope('Br2_3x3'):
br2 = layers.conv2d(net, 256, [1, 1], padding='SAME', scope='Conv1_1x1')
br2 = layers.conv2d(br2, 384, [3, 3], stride=2, scope='Conv2_3x3/2')
with tf.variable_scope('Br3_3x3'):
br3 = layers.conv2d(net, br3_num, [1, 1], padding='SAME', scope='Conv1_1x1')
br3 = layers.conv2d(br3, br3_num + br3_inc, [3, 3], stride=2, scope='Conv2_3x3/2')
with tf.variable_scope('Br4_3x3Dbl'):
br4 = layers.conv2d(net, br4_num, [1, 1], padding='SAME', scope='Conv1_1x1')
br4 = layers.conv2d(br4, br4_num + 1*br4_inc, [3, 3], padding='SAME', scope='Conv2_3x3')
br4 = layers.conv2d(br4, br4_num + 2*br4_inc, [3, 3], stride=2, scope='Conv3_3x3/2')
net = tf.concat(3, [br1, br2, br3, br4], name='Concat1')
# 8 x 8 x 1792 v1, 2144 v2 (paper indicates 2048 but only get this if we use a v1 config for this block)
endpoints[scope] = net
print('%s output shape: %s' % (scope, net.get_shape()))
return net
def _block_c_res(net, ver=2, res_scale=None, scope='BlockC', activation_fn=tf.nn.relu):
# 8 x 8 grid
# configure branch filter numbers
br2_num = 192
if ver == 1:
br2_inc = 0
else:
br2_inc = 32
# default padding = SAME
# default stride = 1
with tf.variable_scope(scope):
shortcut = tf.identity(net, name='Shortcut')
if res_scale:
shortcut = tf.mul(shortcut, res_scale) # scale residual
with tf.variable_scope('Br1_1x1'):
br1 = layers.conv2d(net, 192, [1, 1], scope='Conv1_1x1')
with tf.variable_scope('Br2_3x3'):
br2 = layers.conv2d(net, br2_num, [1, 1], scope='Conv1_1x1')
br2 = layers.conv2d(br2, br2_num + 1*br2_inc, [1, 3], scope='Conv2_1x3')
br2 = layers.conv2d(br2, br2_num + 2*br2_inc, [3, 1], scope='Conv3_3x1')
net = tf.concat(3, [br1, br2], name='Concat1')
net = layers.conv2d(net, shortcut.get_shape()[-1], [1, 1], activation_fn=None, scope='Conv4_1x1')
# 1792 res-1, 2144 (2048?) res-2
net = activation_fn(tf.add(shortcut, net, name='Sum1'))
return net
def _block_output(net, endpoints, num_classes=1000, dropout_keep_prob=0.5, scope='Output'):
with tf.variable_scope(scope):
# 8 x 8 x 1536
shape = net.get_shape()
net = layers.avg_pool2d(net, shape[1:3], padding='VALID', scope='Pool1_Global')
endpoints['Output/Pool1'] = net
# 1 x 1 x 1536
net = layers.dropout(net, dropout_keep_prob)
net = layers.flatten(net)
# 1536
net = layers.fully_connected(net, num_classes, activation_fn=None, scope='Logits')
# num classes
endpoints['Logits'] = net
return net
def _stack(net, endpoints, fn=None, count=1, **kwargs):
scope = kwargs.pop('scope')
for i in range(count):
block_scope = '%s%d' % (scope, (i+1))
kwargs['scope'] = block_scope
net = fn(net, **kwargs)
endpoints[block_scope] = net
print('%s output shape: %s' % (scope, net.get_shape()))
return net
def _build_inception_v4(
inputs,
stack_counts=[4, 7, 3],
dropout_keep_prob=0.8,
num_classes=1000,
is_training=True,
scope=''):
"""Inception v4 from http://arxiv.org/abs/
Args:
inputs: a tensor of size [batch_size, height, width, channels].
dropout_keep_prob: dropout keep_prob.
num_classes: number of predicted classes.
is_training: whether is training or not.
scope: Optional scope for op_scope.
Returns:
a list containing 'logits' Tensors and a dict of Endpoints.
"""
# endpoints will collect relevant activations for external use, for example, summaries or losses.
endpoints = {}
name_scope_net = tf.name_scope(scope, 'Inception_v4', [inputs])
arg_scope_train = arg_scope([layers.batch_norm, layers.dropout], is_training=is_training)
arg_scope_conv = arg_scope([layers.conv2d, layers.max_pool2d, layers.avg_pool2d], stride=1, padding='SAME')
with name_scope_net, arg_scope_train, arg_scope_conv:
net = _block_stem(inputs, endpoints)
# 35 x 35 x 384
with tf.variable_scope('Scale1'):
net = _stack(net, endpoints, fn=_block_a, count=stack_counts[0], scope='BlockA')
# 35 x 35 x 384
with tf.variable_scope('Scale2'):
net = _block_a_reduce(net, endpoints)
# 17 x 17 x 1024
net = _stack(net, endpoints, fn=_block_b, count=stack_counts[1], scope='BlockB')
# 17 x 17 x 1024
with tf.variable_scope('Scale3'):
net = _block_b_reduce(net, endpoints)
# 8 x 8 x 1536
net = _stack(net, endpoints, fn=_block_c, count=stack_counts[2], scope='BlockC')
# 8 x 8 x 1536
logits = _block_output(net, endpoints, num_classes, dropout_keep_prob, scope='Output')
endpoints['Predictions'] = tf.nn.softmax(logits, name='Predictions')
return logits, endpoints
def _build_inception_resnet(
inputs,
stack_counts=[5, 10, 5],
ver=2,
res_scale=None,
activation_fn=tf.nn.relu,
dropout_keep_prob=0.8,
num_classes=1000,
is_training=True,
scope=''):
"""Inception v4 from http://arxiv.org/abs/
Args:
inputs: a tensor of size [batch_size, height, width, channels].
dropout_keep_prob: dropout keep_prob.
num_classes: number of predicted classes.
is_training: whether is training or not.
scope: Optional scope for op_scope.
Returns:
'logits' tensor
'endpoints' dict
"""
# endpoints will collect relevant activations for external use, for example, summaries or losses.
assert ver == 1 or ver == 2
endpoints = {}
network_name = 'inception_resnet_v%d' % ver
print("Building %s" % network_name)
name_scope_net = tf.name_scope(scope, network_name, [inputs])
arg_scope_train = arg_scope([layers.batch_norm, layers.dropout], is_training=is_training)
arg_scope_conv = arg_scope([layers.conv2d, layers.max_pool2d, layers.avg_pool2d], stride=1, padding='SAME')
with name_scope_net, arg_scope_train, arg_scope_conv:
net = _block_stem_res(inputs, endpoints) if ver == 1 else _block_stem(inputs, endpoints)
# 35 x 35 x 384 (v2)
with tf.variable_scope('Scale1'):
net = _stack(
net, endpoints, fn=_block_a_res, count=stack_counts[0], scope='BlockA',
ver=ver, res_scale=res_scale, activation_fn=activation_fn)
# 35 x 35 x 384
with tf.variable_scope('Scale2'):
k, l, m, n = (192, 192, 256, 384) if ver == 1 else (256, 256, 384, 384)
net = _block_a_reduce(net, endpoints, k=k, l=l, m=m, n=n)
# 17 x 17 x 896 v1, 1152 v2
net = _stack(
net, endpoints, fn=_block_b_res, count=stack_counts[1], scope='BlockB',
ver=ver, res_scale=res_scale, activation_fn=activation_fn)
# 17 x 17 x 896 v1, 1152 v2
with tf.variable_scope('Scale3'):
net = _block_b_reduce_res(net, endpoints, ver=ver)
# 8 x 8 x 1792 v1, 2144 v2
net = _stack(
net, endpoints, fn=_block_c_res, count=stack_counts[2], scope='BlockC',
ver=ver, res_scale=res_scale, activation_fn=activation_fn)
# 8 x 8 x 1792 v1, 2144 v2
logits = _block_output(net, endpoints, num_classes, dropout_keep_prob, 'Output')
endpoints['Predictions'] = tf.nn.softmax(logits, name='Predictions')
return logits, endpoints
def params_inception(version=4, residual=False):
params = {
'version': version,
'residual': residual,
'dropout_keep_prob': 0.8,
'residual_scale': 0.67,
}
return params
def inception_arg_scope(
weight_decay=0.00004,
use_batch_norm=True,
batch_norm_decay=0.9997,
batch_norm_epsilon=0.001,
):
# Parameters for BatchNorm.
batch_norm_params = {
# Decay for the moving averages.
'decay': batch_norm_decay,
# epsilon to prevent 0s in variance.
'epsilon': batch_norm_epsilon,
}
if use_batch_norm:
normalizer_fn = layers.batch_norm
normalizer_params = batch_norm_params
else:
normalizer_fn = None
normalizer_params = {}
# Set weight_decay for weights in Conv and FC layers.
l2_regularizer = layers.l2_regularizer(weight_decay)
activation_fn = tf.nn.relu # tf.nn.elu
arg_scope_weights = arg_scope(
[layers.conv2d, layers.fully_connected],
weights_initializer=layers.variance_scaling_initializer(factor=1.0),
weights_regularizer=l2_regularizer
)
arg_scope_conv = arg_scope(
[layers.conv2d],
activation_fn=activation_fn,
normalizer_fn=normalizer_fn,
normalizer_params=normalizer_params
)
with arg_scope_weights, arg_scope_conv as arg_sc:
return arg_sc
def build_inception(
inputs,
num_classes=1000,
params=params_inception(),
is_training=True,
scope=''
):
"""Build Inception v4 architectures.
See here for reference: http://arxiv.org/pdf/1602.07261v1.pdf
"""
residual = params['residual']
version = params['version']
dropout_keep_prob = params['dropout_keep_prob']
residual_scale = params['residual_scale']
with arg_scope(inception_arg_scope()) as sc:
# fetch the activation_fn from scope for direct use in build fn
key = getattr(layers.conv2d, '_key_op', str(layers.conv2d))
activation_fn = sc[key]['activation_fn']
if residual:
assert version == 1 or version == 2
logits, endpoints = _build_inception_resnet(
inputs,
num_classes=num_classes,
ver=version,
res_scale=residual_scale,
activation_fn=activation_fn, # activation_fn used directly in res blocks
dropout_keep_prob=dropout_keep_prob,
is_training=is_training,
scope=scope)
else:
assert version == 4
logits, endpoints = _build_inception_v4(
inputs,
num_classes=num_classes,
dropout_keep_prob=dropout_keep_prob,
is_training=is_training,
scope=scope)
return logits, endpoints
