202 lines
9.8 KiB
Python
202 lines
9.8 KiB
Python
|
import tensorflow as tf
|
||
|
|
import numpy as np
|
||
|
|
from sklearn.utils import shuffle
|
||
|
|
|
||
|
|
def stl10(path):
|
||
|
|
labels=['avion', 'oiseau', 'voiture', 'chat', 'cerf', 'chien', 'cheval', 'singe', 'bateau', 'camion']
|
||
|
|
train_images=np.fromfile(path+"/train_X.bin", dtype=np.uint8).reshape(-1, 3, 96, 96).transpose(0, 2, 3, 1)
|
||
|
|
train_labels=np.eye(10)[np.fromfile(path+"/train_y.bin", dtype=np.uint8)-1]
|
||
|
|
train_images, train_labels=shuffle(train_images, train_labels)
|
||
|
|
test_images=np.fromfile(path+"/test_X.bin", dtype=np.uint8).reshape(-1, 3, 96, 96).transpose(0, 2, 3, 1)
|
||
|
|
test_labels=np.eye(10)[np.fromfile(path+"/test_y.bin", dtype=np.uint8)-1]
|
||
|
|
return labels, train_images, train_labels, test_images, test_labels
|
||
|
|
|
||
|
|
def convolution(input, taille_noyau, nbr_noyau, stride, b_norm, f_activation, training):
|
||
|
|
w_filtre=tf.Variable(tf.random.truncated_normal(shape=(taille_noyau, taille_noyau, int(input.get_shape()[-1]), nbr_noyau)))
|
||
|
|
b_filtre=np.zeros(nbr_noyau)
|
||
|
|
result=tf.nn.conv2d(input, w_filtre, strides=[1, stride, stride, 1], padding='SAME')+b_filtre
|
||
|
|
if b_norm is True:
|
||
|
|
result=tf.layers.batch_normalization(result, training=training)
|
||
|
|
if f_activation is not None:
|
||
|
|
result=f_activation(result)
|
||
|
|
return result
|
||
|
|
|
||
|
|
def fc(input, nbr_neurone, b_norm, f_activation, training):
|
||
|
|
w=tf.Variable(tf.random.truncated_normal(shape=(int(input.get_shape()[-1]), nbr_neurone), dtype=tf.float32))
|
||
|
|
b=tf.Variable(np.zeros(shape=(nbr_neurone)), dtype=tf.float32)
|
||
|
|
result=tf.matmul(input, w)+b
|
||
|
|
if b_norm is True:
|
||
|
|
result=tf.layers.batch_normalization(result, training=training)
|
||
|
|
if f_activation is not None:
|
||
|
|
result=f_activation(result)
|
||
|
|
return result
|
||
|
|
|
||
|
|
def b_resnet_1(input, kernel, nbr_cc, reduce, training, dropout=None):
|
||
|
|
if reduce is True:
|
||
|
|
stride=2
|
||
|
|
result2=convolution(input, 1, nbr_cc[-1], stride, True, tf.nn.relu, training)
|
||
|
|
else:
|
||
|
|
stride=1
|
||
|
|
if nbr_cc[-1]!=int(input.get_shape()[-1]):
|
||
|
|
result2=convolution(input, 1, nbr_cc[-1], stride, True, tf.nn.relu, training)
|
||
|
|
else:
|
||
|
|
result2=input
|
||
|
|
result=input
|
||
|
|
for shift in range(len(kernel)-1):
|
||
|
|
result=convolution(result, kernel[shift], nbr_cc[shift], stride, True, tf.nn.relu, training)
|
||
|
|
stride=1
|
||
|
|
result=convolution(result, kernel[len(kernel)-1], nbr_cc[len(kernel)-1], stride, True, None, training)
|
||
|
|
result=result+result2
|
||
|
|
result=tf.nn.relu(result)
|
||
|
|
if dropout is not None:
|
||
|
|
result=tf.layers.dropout(result, dropout)
|
||
|
|
return result
|
||
|
|
|
||
|
|
def b_resnet_1M(input, kernel, nbr_cc, reduce, training, dropout=None):
|
||
|
|
if reduce is True:
|
||
|
|
result=tf.nn.max_pool(input, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding='SAME')
|
||
|
|
result2=tf.nn.max_pool(input, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding='SAME')
|
||
|
|
if nbr_cc[-1]!=int(input.get_shape()[-1]):
|
||
|
|
result2=convolution(result2, 1, nbr_cc[-1], 1, True, tf.nn.relu, training)
|
||
|
|
else:
|
||
|
|
result=input
|
||
|
|
if nbr_cc[-1]!=int(input.get_shape()[-1]):
|
||
|
|
result2=convolution(input, 1, nbr_cc[-1], 1, True, tf.nn.relu, training)
|
||
|
|
else:
|
||
|
|
result2=input
|
||
|
|
for shift in range(len(kernel)-1):
|
||
|
|
result=convolution(result, kernel[shift], nbr_cc[shift], 1, True, tf.nn.relu, training)
|
||
|
|
result=convolution(result, kernel[len(kernel)-1], nbr_cc[len(kernel)-1], 1, True, None, training)
|
||
|
|
result=result+result2
|
||
|
|
result=tf.nn.relu(result)
|
||
|
|
if dropout is not None:
|
||
|
|
result=tf.layers.dropout(result, dropout)
|
||
|
|
return result
|
||
|
|
|
||
|
|
def b_resnet_2(input, kernel, nbr_cc, reduce, training, dropout=None):
|
||
|
|
if reduce is True:
|
||
|
|
stride=2
|
||
|
|
result2=convolution(input, 1, nbr_cc[-1], stride, True, tf.nn.relu, training)
|
||
|
|
else:
|
||
|
|
stride=1
|
||
|
|
if nbr_cc[-1]!=int(input.get_shape()[-1]):
|
||
|
|
result2=convolution(input, 1, nbr_cc[-1], stride, True, tf.nn.relu, training)
|
||
|
|
else:
|
||
|
|
result2=input
|
||
|
|
result=input
|
||
|
|
for shift in range(len(kernel)-1):
|
||
|
|
result=convolution(result, kernel[shift], nbr_cc[shift], stride, True, tf.nn.relu, training)
|
||
|
|
stride=1
|
||
|
|
result=convolution(result, kernel[len(kernel)-1], nbr_cc[len(kernel)-1], stride, False, None, training)
|
||
|
|
result=result+result2
|
||
|
|
result=tf.layers.batch_normalization(result, training=training)
|
||
|
|
result=tf.nn.relu(result)
|
||
|
|
if dropout is not None:
|
||
|
|
result=tf.layers.dropout(result, dropout)
|
||
|
|
return result
|
||
|
|
|
||
|
|
def b_resnet_2M(input, kernel, nbr_cc, reduce, training, dropout=None):
|
||
|
|
if reduce is True:
|
||
|
|
result=tf.nn.max_pool(input, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding='SAME')
|
||
|
|
result2=tf.nn.max_pool(input, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding='SAME')
|
||
|
|
if nbr_cc[-1]!=int(input.get_shape()[-1]):
|
||
|
|
result2=convolution(result2, 1, nbr_cc[-1], 1, True, tf.nn.relu, training)
|
||
|
|
else:
|
||
|
|
result=input
|
||
|
|
if nbr_cc[-1]!=int(input.get_shape()[-1]):
|
||
|
|
result2=convolution(input, 1, nbr_cc[-1], 1, True, tf.nn.relu, training)
|
||
|
|
else:
|
||
|
|
result2=input
|
||
|
|
for shift in range(len(kernel)-1):
|
||
|
|
result=convolution(result, kernel[shift], nbr_cc[shift], 1, True, tf.nn.relu, training)
|
||
|
|
result=convolution(result, kernel[len(kernel)-1], nbr_cc[len(kernel)-1], 1, False, None, training)
|
||
|
|
result=result+result2
|
||
|
|
result=tf.layers.batch_normalization(result, training=training)
|
||
|
|
result=tf.nn.relu(result)
|
||
|
|
if dropout is not None:
|
||
|
|
result=tf.layers.dropout(result, dropout)
|
||
|
|
return result
|
||
|
|
|
||
|
|
def b_resnet_3(input, kernel, nbr_cc, reduce, training, dropout=None):
|
||
|
|
if reduce is True:
|
||
|
|
stride=2
|
||
|
|
result2=convolution(input, 1, nbr_cc[-1], stride, True, tf.nn.relu, training)
|
||
|
|
else:
|
||
|
|
if nbr_cc[-1]!=int(input.get_shape()[-1]):
|
||
|
|
result2=convolution(input, 1, nbr_cc[-1], 1, True, tf.nn.relu, training)
|
||
|
|
else:
|
||
|
|
result2=input
|
||
|
|
stride=1
|
||
|
|
result=input
|
||
|
|
for shift in range(len(kernel)-1):
|
||
|
|
result=convolution(result, kernel[shift], nbr_cc[shift], stride, True, tf.nn.relu, training)
|
||
|
|
stride=1
|
||
|
|
shift=len(kernel)-1
|
||
|
|
result=convolution(result, kernel[shift], nbr_cc[shift], stride, True, None, training)
|
||
|
|
result=result+result2
|
||
|
|
if dropout is not None:
|
||
|
|
result=tf.layers.dropout(result, dropout)
|
||
|
|
return result
|
||
|
|
|
||
|
|
def b_resnet_3M(input, kernel, nbr_cc, reduce, training, dropout=None):
|
||
|
|
if reduce is True:
|
||
|
|
result =tf.nn.max_pool(input, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding='SAME')
|
||
|
|
result2=tf.nn.max_pool(input, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding='SAME')
|
||
|
|
if nbr_cc[-1]!=int(input.get_shape()[-1]):
|
||
|
|
result2=convolution(result2, 1, nbr_cc[-1], 1, True, tf.nn.relu, training)
|
||
|
|
else:
|
||
|
|
result=convolution(input, kernel[0], nbr_cc[0], 1, True, tf.nn.relu, training)
|
||
|
|
if nbr_cc[-1]!=int(input.get_shape()[-1]):
|
||
|
|
result2=convolution(input, 1, nbr_cc[-1], 1, True, tf.nn.relu, training)
|
||
|
|
else:
|
||
|
|
result2=input
|
||
|
|
for shift in range(1, len(kernel)):
|
||
|
|
result=convolution(result, kernel[shift], nbr_cc[shift], 1, True, tf.nn.relu, training)
|
||
|
|
result=result+result2
|
||
|
|
if dropout is not None:
|
||
|
|
result=tf.layers.dropout(result, dropout)
|
||
|
|
return result
|
||
|
|
|
||
|
|
def resnet(nbr_classes, b_resnet, learning_rate):
|
||
|
|
ph_images=tf.placeholder(shape=(None, 96, 96, 3), dtype=tf.float32)
|
||
|
|
ph_labels=tf.placeholder(shape=(None, nbr_classes), dtype=tf.float32)
|
||
|
|
ph_is_training=tf.placeholder_with_default(False, (), name='is_training')
|
||
|
|
|
||
|
|
#result=convolution(ph_images, 7, 64, 2, True, tf.nn.relu, ph_is_training)
|
||
|
|
result=convolution(ph_images, 5, 64, 1, True, tf.nn.relu, ph_is_training)
|
||
|
|
result=tf.nn.max_pool(result, ksize=[1, 3, 3, 1], strides=[1, 2, 2, 1], padding='SAME')
|
||
|
|
|
||
|
|
result=b_resnet(result, [1, 3, 1], [64, 64, 256], False, ph_is_training, None)
|
||
|
|
result=b_resnet(result, [1, 3, 1], [64, 64, 256], False, ph_is_training, None)
|
||
|
|
result=b_resnet(result, [1, 3, 1], [64, 64, 256], False, ph_is_training, None)
|
||
|
|
|
||
|
|
result=b_resnet(result, [1, 3, 1], [128, 128, 512], True, ph_is_training, None)
|
||
|
|
result=b_resnet(result, [1, 3, 1], [128, 128, 512], False, ph_is_training, None)
|
||
|
|
result=b_resnet(result, [1, 3, 1], [128, 128, 512], False, ph_is_training, None)
|
||
|
|
result=b_resnet(result, [1, 3, 1], [128, 128, 512], False, ph_is_training, None)
|
||
|
|
|
||
|
|
result=b_resnet(result, [1, 3, 1], [256, 256, 1024], True, ph_is_training, None)
|
||
|
|
result=b_resnet(result, [1, 3, 1], [256, 256, 1024], False, ph_is_training, None)
|
||
|
|
result=b_resnet(result, [1, 3, 1], [256, 256, 1024], False, ph_is_training, None)
|
||
|
|
result=b_resnet(result, [1, 3, 1], [256, 256, 1024], False, ph_is_training, None)
|
||
|
|
result=b_resnet(result, [1, 3, 1], [256, 256, 1024], False, ph_is_training, None)
|
||
|
|
result=b_resnet(result, [1, 3, 1], [256, 256, 1024], False, ph_is_training, None)
|
||
|
|
|
||
|
|
result=b_resnet(result, [1, 3, 1], [512, 512, 2048], True, ph_is_training, None)
|
||
|
|
result=b_resnet(result, [1, 3, 1], [512, 512, 2048], False, ph_is_training, None)
|
||
|
|
result=b_resnet(result, [1, 3, 1], [512, 512, 2048], False, ph_is_training, None)
|
||
|
|
taille=result.get_shape()[1]
|
||
|
|
result=tf.nn.avg_pool(result, ksize=[1, taille, taille, 1], strides=[1, 1, 1, 1], padding='SAME')
|
||
|
|
|
||
|
|
result=tf.contrib.layers.flatten(result)
|
||
|
|
result=fc(result, nbr_classes, False, None, ph_is_training)
|
||
|
|
socs=tf.nn.softmax(result)
|
||
|
|
|
||
|
|
loss=tf.nn.softmax_cross_entropy_with_logits_v2(labels=ph_labels, logits=result)
|
||
|
|
extra_update_ops=tf.get_collection(tf.GraphKeys.UPDATE_OPS)
|
||
|
|
with tf.control_dependencies(extra_update_ops):
|
||
|
|
train=tf.train.AdamOptimizer(learning_rate).minimize(loss)
|
||
|
|
accuracy=tf.reduce_mean(tf.cast(tf.equal(tf.argmax(socs, 1), tf.argmax(ph_labels, 1)), tf.float32))
|
||
|
|
|
||
|
|
return ph_images, ph_labels, ph_is_training, socs, train, accuracy, tf.train.Saver()
|