# https://github.com/tensorflow/models/blob/master/research/resnet/resnet_model.py
#
# Copyright 2016 The TensorFlow Authors. 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
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
# ==============================================================================
"""ResNet model.
Related papers:
https://arxiv.org/pdf/1603.05027v2.pdf
https://arxiv.org/pdf/1512.03385v1.pdf
https://arxiv.org/pdf/1605.07146v1.pdf
"""
from collections import namedtuple
import tensorflow as tf
import six
from model import MLPnGPU
from model import Conv2DnGPU
from model import LinearnGPU
from model import LayerNorm
HParams = namedtuple('HParams',
'batch_size, num_classes, min_lrn_rate, lrn_rate, '
'num_residual_units, use_bottleneck, weight_decay_rate, '
'relu_leakiness, momentum')
class ResNetTF(MLPnGPU):
"""ResNet model."""
def __init__(self, batch_size=None, name=None, **kwargs):
self.global_step = tf.contrib.framework.get_or_create_global_step()
self.hps = HParams(batch_size=batch_size,
num_classes=10,
min_lrn_rate=0.0001,
lrn_rate=0.1,
num_residual_units=5,
# num_residual_units=4,
use_bottleneck=False,
weight_decay_rate=0.0002,
relu_leakiness=0.1,
momentum=.9)
self.layers = []
self.layer_idx = 0
self.init_layers = True
self.decay_cost = None
self.training = None
self.device_name = None
def set_training(self, training):
super(ResNetTF, self).set_training(training)
self.training = training
def fprop(self, x):
self.layer_idx = 0
with tf.variable_scope('Resnet'):
logits, probs = self._build_model(x)
self.init_layers = False
states = {'logits': logits, 'probs': probs}
return states
def _stride_arr(self, stride):
"""Map a stride scalar to the stride array for tf.nn.conv2d."""
return [1, stride, stride, 1]
def _build_model(self, x):
"""Build the core model within the graph."""
with tf.variable_scope('init'):
x = self._conv('init_conv', x, 3, x.shape[3], 16,
self._stride_arr(1))
strides = [1, 2, 2]
activate_before_residual = [True, False, False]
if self.hps.use_bottleneck:
res_func = self._bottleneck_residual
filters = [16, 64, 128, 256]
else:
res_func = self._residual
filters = [16, 16, 32, 64]
# Uncomment the following codes to use w28-10 wide residual
# network.
# It is more memory efficient than very deep residual network and
# has
# comparably good performance.
# https://arxiv.org/pdf/1605.07146v1.pdf
# filters = [16, 160, 320, 640]
# Update hps.num_residual_units to 4
with tf.variable_scope('unit_1_0'):
x = res_func(x, filters[0], filters[1],
self._stride_arr(strides[0]),
activate_before_residual[0])
for i in six.moves.range(1, self.hps.num_residual_units):
with tf.variable_scope('unit_1_%d' % i):
x = res_func(x, filters[1], filters[1],
self._stride_arr(1), False)
with tf.variable_scope('unit_2_0'):
x = res_func(x, filters[1], filters[2],
self._stride_arr(strides[1]),
activate_before_residual[1])
for i in six.moves.range(1, self.hps.num_residual_units):
with tf.variable_scope('unit_2_%d' % i):
x = res_func(x, filters[2], filters[2],
self._stride_arr(1), False)
with tf.variable_scope('unit_3_0'):
x = res_func(x, filters[2], filters[3],
self._stride_arr(strides[2]),
activate_before_residual[2])
for i in six.moves.range(1, self.hps.num_residual_units):
with tf.variable_scope('unit_3_%d' % i):
x = res_func(x, filters[3], filters[3],
self._stride_arr(1), False)
with tf.variable_scope('unit_last'):
x = self._layer_norm('final_bn', x)
x = self._relu(x, self.hps.relu_leakiness)
x = self._global_avg_pool(x)
with tf.variable_scope('logit'):
logits = self._fully_connected(x, self.hps.num_classes)
predictions = tf.nn.softmax(logits)
return logits, predictions
def build_cost(self, labels, logits):
"""
Build the graph for cost from the logits if logits are provided.
If predictions are provided, logits are extracted from the operation.
"""
op = logits.op
if "softmax" in str(op).lower():
logits, = op.inputs
with tf.variable_scope('costs'):
xent = tf.nn.softmax_cross_entropy_with_logits(
logits=logits, labels=labels)
cost = tf.reduce_mean(xent, name='xent')
cost += self._decay()
cost = cost
return cost
def build_train_op_from_cost(self, cost):
"""Build training specific ops for the graph."""
self.lrn_rate = tf.constant(self.hps.lrn_rate, tf.float32,
name='learning_rate')
self.momentum = tf.constant(self.hps.momentum, tf.float32,
name='momentum')
trainable_variables = tf.trainable_variables()
grads = tf.gradients(cost, trainable_variables)
devs = set([v.device for v in trainable_variables])
assert len(devs) == 1, ('There should be no trainable variables'
' on any device other than the last GPU.')
optimizer = tf.train.MomentumOptimizer(self.lrn_rate, self.momentum)
gv_pairs = zip(grads, trainable_variables)
gv_pairs = [gv for gv in gv_pairs if gv[0] is not None]
devs = set([gv[1].device for gv in gv_pairs])
assert len(devs) == 1, ('There should be no gradients wrt'
' vars on other GPUs.')
apply_op = optimizer.apply_gradients(
gv_pairs,
global_step=self.global_step, name='train_step')
train_ops = [apply_op]
train_op = tf.group(*train_ops)
return train_op
def _layer_norm(self, name, x):
"""Layer normalization."""
if self.init_layers:
bn = LayerNorm()
bn.name = name
self.layers += [bn]
else:
bn = self.layers[self.layer_idx]
self.layer_idx += 1
bn.device_name = self.device_name
bn.set_training(self.training)
x = bn.fprop(x)
return x
def _residual(self, x, in_filter, out_filter, stride,
activate_before_residual=False):
"""Residual unit with 2 sub layers."""
if activate_before_residual:
with tf.variable_scope('shared_activation'):
x = self._layer_norm('init_bn', x)
x = self._relu(x, self.hps.relu_leakiness)
orig_x = x
else:
with tf.variable_scope('residual_only_activation'):
orig_x = x
x = self._layer_norm('init_bn', x)
x = self._relu(x, self.hps.relu_leakiness)
with tf.variable_scope('sub1'):
x = self._conv('conv1', x, 3, in_filter, out_filter, stride)
with tf.variable_scope('sub2'):
x = self._layer_norm('bn2', x)
x = self._relu(x, self.hps.relu_leakiness)
x = self._conv('conv2', x, 3, out_filter, out_filter, [1, 1, 1, 1])
with tf.variable_scope('sub_add'):
if in_filter != out_filter:
orig_x = tf.nn.avg_pool(orig_x, stride, stride, 'VALID')
orig_x = tf.pad(
orig_x, [[0, 0], [0, 0], [0, 0],
[(out_filter - in_filter) // 2,
(out_filter - in_filter) // 2]])
x += orig_x
return x
def _bottleneck_residual(self, x, in_filter, out_filter, stride,
activate_before_residual=False):
"""Bottleneck residual unit with 3 sub layers."""
if activate_before_residual:
with tf.variable_scope('common_bn_relu'):
x = self._layer_norm('init_bn', x)
x = self._relu(x, self.hps.relu_leakiness)
orig_x = x
else:
with tf.variable_scope('residual_bn_relu'):
orig_x = x
x = self._layer_norm('init_bn', x)
x = self._relu(x, self.hps.relu_leakiness)
with tf.variable_scope('sub1'):
x = self._conv('conv1', x, 1, in_filter, out_filter / 4, stride)
with tf.variable_scope('sub2'):
x = self._layer_norm('bn2', x)
x = self._relu(x, self.hps.relu_leakiness)
x = self._conv('conv2', x, 3, out_filter / 4,
out_filter / 4, [1, 1, 1, 1])
with tf.variable_scope('sub3'):
x = self._layer_norm('bn3', x)
x = self._relu(x, self.hps.relu_leakiness)
x = self._conv('conv3', x, 1, out_filter /
4, out_filter, [1, 1, 1, 1])
with tf.variable_scope('sub_add'):
if in_filter != out_filter:
orig_x = self._conv('project', orig_x, 1,
in_filter, out_filter, stride)
x += orig_x
return x
def _decay(self):
"""L2 weight decay loss."""
if self.decay_cost is not None:
return self.decay_cost
costs = []
if self.device_name is None:
for var in tf.trainable_variables():
if var.op.name.find(r'DW') > 0:
costs.append(tf.nn.l2_loss(var))
else:
for layer in self.layers:
for var in layer.params_device[self.device_name].values():
if (isinstance(var, tf.Variable) and
var.op.name.find(r'DW') > 0):
costs.append(tf.nn.l2_loss(var))
self.decay_cost = tf.multiply(self.hps.weight_decay_rate,
tf.add_n(costs))
return self.decay_cost
def _conv(self, name, x, filter_size, in_filters, out_filters, strides):
"""Convolution."""
if self.init_layers:
conv = Conv2DnGPU(out_filters,
(filter_size, filter_size),
strides[1:3], 'SAME', w_name='DW')
conv.name = name
self.layers += [conv]
else:
conv = self.layers[self.layer_idx]
self.layer_idx += 1
conv.device_name = self.device_name
conv.set_training(self.training)
return conv.fprop(x)
def _relu(self, x, leakiness=0.0):
"""Relu, with optional leaky support."""
return tf.where(tf.less(x, 0.0), leakiness * x, x, name='leaky_relu')
def _fully_connected(self, x, out_dim):
"""FullyConnected layer for final output."""
if self.init_layers:
fc = LinearnGPU(out_dim, w_name='DW')
fc.name = 'logits'
self.layers += [fc]
else:
fc = self.layers[self.layer_idx]
self.layer_idx += 1
fc.device_name = self.device_name
fc.set_training(self.training)
return fc.fprop(x)
def _global_avg_pool(self, x):
assert x.get_shape().ndims == 4
return tf.reduce_mean(x, [1, 2])
