Initial commit

Tensorflow/tutoriel38/README.md

@@ -0,0 +1,5 @@
+# Tutoriel tensorflow
+## Vecteur embedding, TripletLoss, FaceNet
+
+La vidéo de ce tutoriel est disponible à l'adresse suivante: https://www.youtube.com/watch?v=x3PbnSNV5n0
+

Tensorflow/tutoriel38/common.py

@@ -0,0 +1,27 @@
+import numpy as np
+from tensorflow.keras import layers, models
+import tensorflow as tf
+import io
+
+def write_labels_embs(model, ds, file_embeddings, file_labels):
+    embeddings=model.predict(ds)
+    np.savetxt(file_embeddings, embeddings, delimiter='\t')
+    if file_labels is not None:
+        fichier=io.open(file_labels, 'w', encoding='utf-8')
+        for images, labels in ds:
+            [fichier.write("{:d}\n".format(x)) for x in labels]
+        fichier.close()
+
+def model_embedding(nbr_cc, embeddings_size):
+    entree=layers.Input(shape=(28, 28, 1), dtype=tf.float32)
+
+    result=layers.Conv2D(nbr_cc, 3, activation='relu', padding='same')(entree)
+    result=layers.MaxPool2D()(result)
+    result=layers.Conv2D(nbr_cc, 3, activation='relu', padding='same')(result)
+    result=layers.MaxPool2D()(result)
+    result=layers.Flatten()(result)
+    result=layers.Dense(embeddings_size, activation=None)(result)
+    embeddings=layers.Lambda(lambda x: tf.math.l2_normalize(x, axis=1))(result)
+
+    model=models.Model(inputs=entree, outputs=embeddings)
+    return model

Tensorflow/tutoriel38/mnist.py

@@ -0,0 +1,29 @@
+import numpy as np
+import tensorflow as tf
+import tensorflow_addons as tfa
+
+import common
+
+batch_size=64
+nbr_cc=64 # Nombre de cartes de caracteristique du modele
+embeddings_size=256 # Taille du vecteur embedding (sortie du réseau de neurone)
+
+(x_train, y_train),(x_test, y_test)=tf.keras.datasets.mnist.load_data()
+
+train_dataset=(x_train.reshape(-1, 28, 28, 1)/255).astype(np.float32)
+test_dataset=(x_test.reshape(-1, 28, 28, 1)/255).astype(np.float32)
+
+train_ds=tf.data.Dataset.from_tensor_slices((train_dataset, y_train)).batch(batch_size)
+test_ds=tf.data.Dataset.from_tensor_slices((test_dataset, y_test)).batch(batch_size)
+
+model=common.model_embedding(nbr_cc, embeddings_size)
+
+common.write_labels_embs(model, test_ds, 'embeddings1.tsv', 'labels.tsv')
+
+model.compile(
+    optimizer=tf.keras.optimizers.Adam(1E-3),
+    loss=tfa.losses.TripletSemiHardLoss())
+
+model.fit(train_ds, epochs=5)
+
+common.write_labels_embs(model, test_ds, 'embeddings2.tsv', None)

Tensorflow/tutoriel38/mnist_avec_predictions.py

@@ -0,0 +1,51 @@
+import numpy as np
+import tensorflow as tf
+import tensorflow_addons as tfa
+from tensorflow.keras import layers, models
+from sklearn import cluster
+import sklearn
+
+import common
+
+batch_size=16
+nbr_cc=64 # Nombre de cartes de caracteristique du modele
+embeddings_size=256 # Taille du vecteur embedding (sortie du réseau de neurone)
+
+(x_train, y_train), (x_test, y_test)=tf.keras.datasets.mnist.load_data()
+
+train_dataset=(x_train.reshape(-1, 28, 28, 1)/255).astype(np.float32)
+test_dataset=(x_test.reshape(-1, 28, 28, 1)/255).astype(np.float32)
+
+train_ds=tf.data.Dataset.from_tensor_slices((train_dataset, y_train)).batch(batch_size)
+test_ds=tf.data.Dataset.from_tensor_slices((test_dataset, y_test)).batch(batch_size)
+
+model=common.model_embedding(nbr_cc, embeddings_size)
+
+model.compile(
+    optimizer=tf.keras.optimizers.Adam(1E-3),
+    loss=tfa.losses.TripletSemiHardLoss())
+
+model.fit(train_ds, epochs=5)
+
+# Recherche des centroids des clusters
+embeddings=model.predict(train_dataset)
+kmeans=cluster.KMeans(n_clusters=len(set(y_train)))
+kmeans.fit(embeddings)
+centroids=kmeans.cluster_centers_
+
+# Recherche du label des centroids
+distances=sklearn.metrics.pairwise_distances(embeddings, centroids)
+# shape distances: 6000, 10
+# lmin contient le vecteur embeddings le plus proche de chacun des centroids
+lmin=np.argmin(distances, axis=0)
+labels_centroids=y_train[lmin]
+
+# Calcul de précision de la base d'entrainement
+result=np.equal(y_train, labels_centroids[np.argmin(distances, axis=-1)]).astype(np.float32)
+print("Train: précision {:4.2%}".format(np.mean(result)))
+
+# Calcul de précision de la base de test
+embeddings=model.predict(test_dataset)
+distances=sklearn.metrics.pairwise_distances(embeddings, centroids)
+result=np.equal(y_test.astype(np.int32), labels_centroids[np.argmin(distances, axis=-1)]).astype(np.float32)
+print("Test : précision {:4.2%}".format(np.mean(result)))