229 lines
10 KiB
Python
229 lines
10 KiB
Python
|
# Preferred settings for this model is:
|
||
|
|
# Training epochs = 80
|
||
|
|
# Crop Size = 224
|
||
|
|
# Learning Rate = 0.001
|
||
|
|
# Under advanced learning rate options:
|
||
|
|
# Step Size = 10.0
|
||
|
|
# Gamma = 0.96
|
||
|
|
|
||
|
|
# The auxillary branches as spcified in the original googlenet V1 model do exist in this implementation of
|
||
|
|
# googlenet but it is not used. To use it, be sure to check self.is_training to ensure that it is only used
|
||
|
|
# during training.
|
||
|
|
|
||
|
|
from model import Tower
|
||
|
|
from utils import model_property
|
||
|
|
import tensorflow as tf
|
||
|
|
import utils as digits
|
||
|
|
|
||
|
|
|
||
|
|
class UserModel(Tower):
|
||
|
|
|
||
|
|
all_inception_settings = {
|
||
|
|
'3a': [[64], [96, 128], [16, 32], [32]],
|
||
|
|
'3b': [[128], [128, 192], [32, 96], [64]],
|
||
|
|
'4a': [[192], [96, 208], [16, 48], [64]],
|
||
|
|
'4b': [[160], [112, 224], [24, 64], [64]],
|
||
|
|
'4c': [[128], [128, 256], [24, 64], [64]],
|
||
|
|
'4d': [[112], [144, 288], [32, 64], [64]],
|
||
|
|
'4e': [[256], [160, 320], [32, 128], [128]],
|
||
|
|
'5a': [[256], [160, 320], [32, 128], [128]],
|
||
|
|
'5b': [[384], [192, 384], [48, 128], [128]]
|
||
|
|
}
|
||
|
|
|
||
|
|
@model_property
|
||
|
|
def inference(self):
|
||
|
|
# rescale to proper form, really we expect 224 x 224 x 1 in HWC form
|
||
|
|
model = tf.reshape(self.x, shape=[-1, self.input_shape[0], self.input_shape[1], self.input_shape[2]])
|
||
|
|
|
||
|
|
conv_7x7_2s_weight, conv_7x7_2s_bias = self.create_conv_vars([7, 7, self.input_shape[2], 64], 'conv_7x7_2s')
|
||
|
|
model = self.conv_layer_with_relu(model, conv_7x7_2s_weight, conv_7x7_2s_bias, 2)
|
||
|
|
|
||
|
|
model = self.max_pool(model, 3, 2)
|
||
|
|
|
||
|
|
# model = tf.nn.local_response_normalization(model)
|
||
|
|
|
||
|
|
conv_1x1_vs_weight, conv_1x1_vs_bias = self.create_conv_vars([1, 1, 64, 64], 'conv_1x1_vs')
|
||
|
|
model = self.conv_layer_with_relu(model, conv_1x1_vs_weight, conv_1x1_vs_bias, 1, 'VALID')
|
||
|
|
|
||
|
|
conv_3x3_1s_weight, conv_3x3_1s_bias = self.create_conv_vars([3, 3, 64, 192], 'conv_3x3_1s')
|
||
|
|
model = self.conv_layer_with_relu(model, conv_3x3_1s_weight, conv_3x3_1s_bias, 1)
|
||
|
|
|
||
|
|
# model = tf.nn.local_response_normalization(model)
|
||
|
|
|
||
|
|
model = self.max_pool(model, 3, 2)
|
||
|
|
|
||
|
|
inception_settings_3a = InceptionSettings(192, UserModel.all_inception_settings['3a'])
|
||
|
|
model = self.inception(model, inception_settings_3a, '3a')
|
||
|
|
|
||
|
|
inception_settings_3b = InceptionSettings(256, UserModel.all_inception_settings['3b'])
|
||
|
|
model = self.inception(model, inception_settings_3b, '3b')
|
||
|
|
|
||
|
|
model = self.max_pool(model, 3, 2)
|
||
|
|
|
||
|
|
inception_settings_4a = InceptionSettings(480, UserModel.all_inception_settings['4a'])
|
||
|
|
model = self.inception(model, inception_settings_4a, '4a')
|
||
|
|
|
||
|
|
# first auxiliary branch for making training faster
|
||
|
|
# aux_branch_1 = self.auxiliary_classifier(model, 512, "aux_1")
|
||
|
|
|
||
|
|
inception_settings_4b = InceptionSettings(512, UserModel.all_inception_settings['4b'])
|
||
|
|
model = self.inception(model, inception_settings_4b, '4b')
|
||
|
|
|
||
|
|
inception_settings_4c = InceptionSettings(512, UserModel.all_inception_settings['4c'])
|
||
|
|
model = self.inception(model, inception_settings_4c, '4c')
|
||
|
|
|
||
|
|
inception_settings_4d = InceptionSettings(512, UserModel.all_inception_settings['4d'])
|
||
|
|
model = self.inception(model, inception_settings_4d, '4d')
|
||
|
|
|
||
|
|
# second auxiliary branch for making training faster
|
||
|
|
# aux_branch_2 = self.auxiliary_classifier(model, 528, "aux_2")
|
||
|
|
|
||
|
|
inception_settings_4e = InceptionSettings(528, UserModel.all_inception_settings['4e'])
|
||
|
|
model = self.inception(model, inception_settings_4e, '4e')
|
||
|
|
|
||
|
|
model = self.max_pool(model, 3, 2)
|
||
|
|
|
||
|
|
inception_settings_5a = InceptionSettings(832, UserModel.all_inception_settings['5a'])
|
||
|
|
model = self.inception(model, inception_settings_5a, '5a')
|
||
|
|
|
||
|
|
inception_settings_5b = InceptionSettings(832, UserModel.all_inception_settings['5b'])
|
||
|
|
model = self.inception(model, inception_settings_5b, '5b')
|
||
|
|
|
||
|
|
model = self.avg_pool(model, 7, 1, 'VALID')
|
||
|
|
|
||
|
|
fc_weight, fc_bias = self.create_fc_vars([1024, self.nclasses], 'fc')
|
||
|
|
model = self.fully_connect(model, fc_weight, fc_bias)
|
||
|
|
|
||
|
|
# if self.is_training:
|
||
|
|
# return [aux_branch_1, aux_branch_2, model]
|
||
|
|
|
||
|
|
return model
|
||
|
|
|
||
|
|
@model_property
|
||
|
|
def loss(self):
|
||
|
|
model = self.inference
|
||
|
|
loss = digits.classification_loss(model, self.y)
|
||
|
|
accuracy = digits.classification_accuracy(model, self.y)
|
||
|
|
self.summaries.append(tf.summary.scalar(accuracy.op.name, accuracy))
|
||
|
|
return loss
|
||
|
|
|
||
|
|
def inception(self, model, inception_setting, layer_name):
|
||
|
|
weights, biases = self.create_inception_variables(inception_setting, layer_name)
|
||
|
|
conv_1x1 = self.conv_layer_with_relu(model, weights['conv_1x1_1'], biases['conv_1x1_1'], 1)
|
||
|
|
|
||
|
|
conv_3x3 = self.conv_layer_with_relu(model, weights['conv_1x1_2'], biases['conv_1x1_2'], 1)
|
||
|
|
conv_3x3 = self.conv_layer_with_relu(conv_3x3, weights['conv_3x3'], biases['conv_3x3'], 1)
|
||
|
|
|
||
|
|
conv_5x5 = self.conv_layer_with_relu(model, weights['conv_1x1_3'], biases['conv_1x1_3'], 1)
|
||
|
|
conv_5x5 = self.conv_layer_with_relu(conv_5x5, weights['conv_5x5'], biases['conv_5x5'], 1)
|
||
|
|
|
||
|
|
conv_pool = self.max_pool(model, 3, 1)
|
||
|
|
conv_pool = self.conv_layer_with_relu(conv_pool, weights['conv_pool'], biases['conv_pool'], 1)
|
||
|
|
|
||
|
|
final_model = tf.concat([conv_1x1, conv_3x3, conv_5x5, conv_pool], 3)
|
||
|
|
|
||
|
|
return final_model
|
||
|
|
|
||
|
|
def create_inception_variables(self, inception_setting, layer_name):
|
||
|
|
model_dim = inception_setting.model_dim
|
||
|
|
conv_1x1_1_w, conv_1x1_1_b = self.create_conv_vars([1, 1, model_dim, inception_setting.conv_1x1_1_layers],
|
||
|
|
layer_name + '-conv_1x1_1')
|
||
|
|
conv_1x1_2_w, conv_1x1_2_b = self.create_conv_vars([1, 1, model_dim, inception_setting.conv_1x1_2_layers],
|
||
|
|
layer_name + '-conv_1x1_2')
|
||
|
|
conv_1x1_3_w, conv_1x1_3_b = self.create_conv_vars([1, 1, model_dim, inception_setting.conv_1x1_3_layers],
|
||
|
|
layer_name + '-conv_1x1_3')
|
||
|
|
conv_3x3_w, conv_3x3_b = self.create_conv_vars([3, 3, inception_setting.conv_1x1_2_layers,
|
||
|
|
inception_setting.conv_3x3_layers],
|
||
|
|
layer_name + '-conv_3x3')
|
||
|
|
conv_5x5_w, conv_5x5_b = self.create_conv_vars([5, 5, inception_setting.conv_1x1_3_layers,
|
||
|
|
inception_setting.conv_5x5_layers],
|
||
|
|
layer_name + '-conv_5x5')
|
||
|
|
conv_pool_w, conv_pool_b = self.create_conv_vars([1, 1, model_dim, inception_setting.conv_pool_layers],
|
||
|
|
layer_name + '-conv_pool')
|
||
|
|
|
||
|
|
weights = {
|
||
|
|
'conv_1x1_1': conv_1x1_1_w,
|
||
|
|
'conv_1x1_2': conv_1x1_2_w,
|
||
|
|
'conv_1x1_3': conv_1x1_3_w,
|
||
|
|
'conv_3x3': conv_3x3_w,
|
||
|
|
'conv_5x5': conv_5x5_w,
|
||
|
|
'conv_pool': conv_pool_w
|
||
|
|
}
|
||
|
|
|
||
|
|
biases = {
|
||
|
|
'conv_1x1_1': conv_1x1_1_b,
|
||
|
|
'conv_1x1_2': conv_1x1_2_b,
|
||
|
|
'conv_1x1_3': conv_1x1_3_b,
|
||
|
|
'conv_3x3': conv_3x3_b,
|
||
|
|
'conv_5x5': conv_5x5_b,
|
||
|
|
'conv_pool': conv_pool_b
|
||
|
|
}
|
||
|
|
|
||
|
|
return weights, biases
|
||
|
|
|
||
|
|
def auxiliary_classifier(self, model, input_size, name):
|
||
|
|
aux_classifier = self.avg_pool(model, 5, 3, 'VALID')
|
||
|
|
|
||
|
|
conv_weight, conv_bias = self.create_conv_vars([1, 1, input_size, input_size], name + '-conv_1x1')
|
||
|
|
aux_classifier = self.conv_layer_with_relu(aux_classifier, conv_weight, conv_bias, 1)
|
||
|
|
|
||
|
|
fc_weight, fc_bias = self.create_fc_vars([4*4*input_size, self.nclasses], name + '-fc')
|
||
|
|
aux_classifier = self.fully_connect(aux_classifier, fc_weight, fc_bias)
|
||
|
|
|
||
|
|
aux_classifier = tf.nn.dropout(aux_classifier, 0.7)
|
||
|
|
|
||
|
|
return aux_classifier
|
||
|
|
|
||
|
|
def conv_layer_with_relu(self, model, weights, biases, stride_size, padding='SAME'):
|
||
|
|
new_model = tf.nn.conv2d(model, weights, strides=[1, stride_size, stride_size, 1], padding=padding)
|
||
|
|
new_model = tf.nn.bias_add(new_model, biases)
|
||
|
|
new_model = tf.nn.relu(new_model)
|
||
|
|
return new_model
|
||
|
|
|
||
|
|
def max_pool(self, model, kernal_size, stride_size, padding='SAME'):
|
||
|
|
new_model = tf.nn.max_pool(model, ksize=[1, kernal_size, kernal_size, 1],
|
||
|
|
strides=[1, stride_size, stride_size, 1], padding=padding)
|
||
|
|
return new_model
|
||
|
|
|
||
|
|
def avg_pool(self, model, kernal_size, stride_size, padding='SAME'):
|
||
|
|
new_model = tf.nn.avg_pool(model, ksize=[1, kernal_size, kernal_size, 1],
|
||
|
|
strides=[1, stride_size, stride_size, 1], padding=padding)
|
||
|
|
return new_model
|
||
|
|
|
||
|
|
def fully_connect(self, model, weights, biases):
|
||
|
|
fc_model = tf.reshape(model, [-1, weights.get_shape().as_list()[0]])
|
||
|
|
fc_model = tf.matmul(fc_model, weights)
|
||
|
|
fc_model = tf.add(fc_model, biases)
|
||
|
|
fc_model = tf.nn.relu(fc_model)
|
||
|
|
return fc_model
|
||
|
|
|
||
|
|
def create_conv_vars(self, size, name):
|
||
|
|
weight = self.create_weight(size, name + '_W')
|
||
|
|
bias = self.create_bias(size[3], name + '_b')
|
||
|
|
return weight, bias
|
||
|
|
|
||
|
|
def create_fc_vars(self, size, name):
|
||
|
|
weight = self.create_weight(size, name + '_W')
|
||
|
|
bias = self.create_bias(size[1], name + '_b')
|
||
|
|
return weight, bias
|
||
|
|
|
||
|
|
def create_weight(self, size, name):
|
||
|
|
weight = tf.get_variable(name, size, initializer=tf.contrib.layers.xavier_initializer())
|
||
|
|
return weight
|
||
|
|
|
||
|
|
def create_bias(self, size, name):
|
||
|
|
bias = tf.get_variable(name, [size], initializer=tf.constant_initializer(0.2))
|
||
|
|
return bias
|
||
|
|
|
||
|
|
|
||
|
|
class InceptionSettings():
|
||
|
|
|
||
|
|
def __init__(self, model_dim, inception_settings):
|
||
|
|
self.model_dim = model_dim
|
||
|
|
self.conv_1x1_1_layers = inception_settings[0][0]
|
||
|
|
self.conv_1x1_2_layers = inception_settings[1][0]
|
||
|
|
self.conv_1x1_3_layers = inception_settings[2][0]
|
||
|
|
self.conv_3x3_layers = inception_settings[1][1]
|
||
|
|
self.conv_5x5_layers = inception_settings[2][1]
|
||
|
|
self.conv_pool_layers = inception_settings[3][0]
|