from base import base_mnist, base_fashion_mnist, base_cifar10 from keras.datasets import mnist, fashion_mnist, cifar10 from keras.callbacks import ModelCheckpoint # Parameters dataset = cifar10 base_model = base_cifar10 emb_d = 512 loss = triplet.contrastive_loss1 optimizer = 'adam' n_tr = 100 # Number of train samples per class trip = 3 # Number of triplets per each sample batch = 16 # Number of samples per class in each epoch epochs = 300 # Data mat_tr, mat_te = datatool.load(dataset, n_tr) gen_tr = tripletgen.tripletgen(mat_tr, trip, batch) gen_te = tripletgen.tripletgen(mat_te, trip, batch) # Model base_model = base_model(mat_tr.shape[2:], emb_d) trip_model, emb_model = triplet.create(base_model, mat_tr.shape[0]) triplet.compile(trip_model, loss, optimizer) # Train trip_model.fit_generator(gen_tr, epochs=epochs, verbose=2) # Obtain embedded spaces x_tr = np.reshape(mat_tr, (-1, ) + mat_tr.shape[2:]) e_tr = emb_model.predict(x_tr) y_tr = np.repeat(np.arange(mat_tr.shape[0]), mat_tr.shape[1]) x_te = np.reshape(mat_te, (-1, ) + mat_te.shape[2:]) e_te = emb_model.predict(x_te) y_te = np.repeat(np.arange(mat_te.shape[0]), mat_te.shape[1])
from keras.datasets import mnist, fashion_mnist, cifar10
import numpy as np
import gc
from keras import backend as K
import sys
# Parameters
base = base_cifar10
emb_d = 256
loss = siamese.contrastive_loss1
optimizer = 'adam'
batch = 16 # Number of samples per class in each epoch
epochs = 100
for n in range(39):
# Data
x_tr, y_tr, x_te, y_te = datatool.load(n)
gen_tr = pairgen.pairgen(x_tr, y_tr, batch)
# Model
base_model = base(x_tr.shape[1:], emb_d)
siam_model, emb_model = siamese.create(base_model, len(np.unique(y_tr)))
siamese.compile(siam_model, loss, optimizer)
# Train
siam_model.fit_generator(gen_tr, epochs=epochs, verbose=0)
# Obtain embedded spaces
e_tr = emb_model.predict(x_tr)
e_te = emb_model.predict(x_te)
# Evaluate
print("Imagen {:02}".format(n + 1))
predtool.hist2(x_tr, e_tr, y_tr, x_te, e_te, y_te, n)
sys.stdout.flush()
K.clear_session()