import tensorflow as tf import numpy as np from sklearn.utils import shuffle from sklearn.model_selection import train_test_split import cv2 import os import numpy as np from PIL import ImageFont, ImageDraw, Image taille_batch=100 nbr_entrainement=100 nbr=2*42 def modif_image(image, seuil=1): b=np.random.normal(0, 1, (28, 28)) a=image.copy() a[b>seuil]=255 a[b<-seuil]=0 return a def convolution(input, taille_noyau, nbr_noyau, stride, b_norm=False, f_activation=None, training=False): 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=False, f_activation=None, training=False): 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 ia(nbr_classes, size, couche, learning_rate=1E-3): ph_images=tf.placeholder(shape=(None, size, size, couche), dtype=tf.float32, name='entree') 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, 3, 64, 1, True, tf.nn.relu, ph_is_training) result=tf.layers.dropout(result, 0.3, training=ph_is_training) result=convolution(result, 3, 128, 1, True, tf.nn.relu, ph_is_training) result=tf.layers.dropout(result, 0.4, training=ph_is_training) result=tf.nn.max_pool(result, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding='SAME') result=tf.contrib.layers.flatten(result) result=fc(result, 128, True, tf.nn.relu, ph_is_training) result=tf.layers.dropout(result, 0.5, training=ph_is_training) result=fc(result, nbr_classes) socs=tf.nn.softmax(result, name="sortie") 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() tab_images=[] tab_labels=[] for dir in ["/usr/share/fonts/truetype/ubuntu-font-family/", "/usr/share/fonts/truetype/freefont/"]: for root, dirs, files in os.walk(dir): for file in files: if file.endswith("ttf"): print(root+"/"+file) for i in range(1, 10): for cpt in range(nbr): image=Image.new("L", (28, 28)) draw=ImageDraw.Draw(image) font=ImageFont.truetype(root+"/"+file, np.random.randint(26, 32)) text="{:d}".format(i) draw.text((np.random.randint(1, 10), np.random.randint(-4, 0)), text, font=font, fill=(255)) image=np.array(image).reshape(28, 28, 1) tab_images.append(image) tab_labels.append(np.eye(10)[i]) image_m=modif_image(image, 1.05+np.random.rand()) tab_images.append(image_m) tab_labels.append(np.eye(10)[i]) image=np.zeros((28, 28, 1)) for cpt in range(3*nbr): image_m=modif_image(image, 1.05+np.random.rand()) tab_images.append(image_m) tab_labels.append(np.eye(10)[0]) tab_images=np.array(tab_images) tab_labels=np.array(tab_labels) tab_images=tab_images/255 tab_images, tab_labels=shuffle(tab_images, tab_labels) if False: # Changer en True si vous voulez voir les images générées for i in range(len(tab_images)): cv2.imshow('chiffre', tab_images[i].reshape(28, 28, 1)) print(tab_labels[i], np.argmax(tab_labels[i])) if cv2.waitKey()&0xFF==ord('q'): break print("Nbr:", len(tab_images)) train_images, test_images, train_labels, test_labels=train_test_split(tab_images, tab_labels, test_size=0.10) images, labels, is_training, sortie, train, accuracy, saver=ia(10, 28, 1) with tf.Session() as s: s.run(tf.global_variables_initializer()) tab_train=[] tab_test=[] for id_entrainement in np.arange(nbr_entrainement): print("> Entrainement", id_entrainement) for batch in np.arange(0, len(train_images), taille_batch): s.run(train, feed_dict={ images: train_images[batch:batch+taille_batch], labels: train_labels[batch:batch+taille_batch], is_training: True }) print(" entrainement OK") tab_accuracy_train=[] for batch in np.arange(0, len(train_images), taille_batch): p=s.run(accuracy, feed_dict={ images: train_images[batch:batch+taille_batch], labels: train_labels[batch:batch+taille_batch], is_training: True }) tab_accuracy_train.append(p) print(" train:", np.mean(tab_accuracy_train)) tab_accuracy_test=[] for batch in np.arange(0, len(test_images), taille_batch): p=s.run(accuracy, feed_dict={ images: test_images[batch:batch+taille_batch], labels: test_labels[batch:batch+taille_batch], is_training: True }) tab_accuracy_test.append(p) print(" test :", np.mean(tab_accuracy_test)) tab_train.append(1-np.mean(tab_accuracy_train)) tab_test.append(1-np.mean(tab_accuracy_test)) saver.save(s, './mon_modele/modele')