def train():
from keras.metrics import categorical_accuracy
transform = (TransformImage(0).by('rerange', 0, 255, 0, 1, 'float32'))
build_batch = (BuildBatch(BATCH_SIZE).by(0, 'image', 'float32').by(
1, 'one_hot', 'uint8', NUM_CLASSES))
plot = PlotLines((0, 1), layout=(2, 1), every_sec=1)
print('loading data...')
train_samples, test_samples = load_samples()
print('creating network ...')
network = create_network()
print('training...', NUM_EPOCHS)
for epoch in range(NUM_EPOCHS):
print('EPOCH:', epoch)
t_loss, t_acc = (train_samples >> PrintProgress(train_samples) >>
Pick(PICK) >> transform >> Shuffle(100) >> build_batch
>> network.train() >> plot >> Unzip())
print('train loss : {:.6f}'.format(t_loss >> Mean()))
print('train acc : {:.1f}'.format(100 * (t_acc >> Mean())))
e_acc = (test_samples >> transform >> build_batch >> network.evaluate(
[categorical_accuracy]))
print('test acc : {:.1f}'.format(100 * e_acc))
network.save_best(e_acc, isloss=False)
def train(cfg: CFG, net: Network, traindata: Samples, valdata: Samples):
plotlines = PlotLines((0, 1, 2),
layout=(3, 1),
figsize=(8, 12),
titles=('loss', 'train-acc', 'val-acc'))
train_cache = Cache(cfg.cachedir + '/train', cfg.cacheclear)
val_cache = Cache(cfg.cachedir + '/val', cfg.cacheclear)
optimizer = to.optim.Adam(net.parameters(), lr=cfg.lr)
max_val = 0
for epoch in range(cfg.n_epochs):
start = time.time()
net.train()
losses, accs = (
traindata >> LoadGraph() >> train_cache >> Shuffle(100) >>
MakeBatch(cfg.batchsize) >> TrainBatch(net, optimizer) >> Unzip())
loss, acc = np.mean(losses), np.mean(accs)
if cfg.verbose:
msg = "{:d}..{:d} {:s} : {:.4f} ({:.1f}% {:.1f}%)"
elapsed = time.strftime("%M:%S", time.gmtime(time.time() - start))
print(msg.format(epoch, cfg.n_epochs, elapsed, loss, acc, max_val))
val_acc = validate(cfg, net, valdata, val_cache)
if val_acc >= max_val:
max_val = val_acc
save_wgts(net)
if cfg.verbose > 1:
plotlines((loss, acc, max_val))
def train():
from keras.metrics import categorical_accuracy
rerange = TransformImage(0).by('rerange', 0, 255, 0, 1, 'float32')
build_batch = (BuildBatch(BATCH_SIZE).by(0, 'image', 'float32').by(
1, 'one_hot', 'uint8', NUM_CLASSES))
p = 0.1
augment = (AugmentImage(0).by('identical', 1.0).by(
'elastic', p, [5, 5], [100, 100],
[0, 100]).by('brightness', p,
[0.7, 1.3]).by('color', p, [0.7, 1.3]).by(
'shear', p,
[0, 0.1]).by('fliplr', p).by('rotate', p, [-10, 10]))
plot_eval = PlotLines((0, 1), layout=(2, 1))
print('creating network...')
network = create_network()
print('loading data...')
train_samples, test_samples = load_samples()
train_samples, val_samples = train_samples >> SplitRandom(0.8)
print('training...', len(train_samples), len(val_samples))
for epoch in range(NUM_EPOCHS):
print('EPOCH:', epoch)
t_loss, t_acc = (train_samples >> PrintProgress(train_samples) >>
Pick(PICK) >> augment >> rerange >> Shuffle(100) >>
build_batch >> network.train() >> Unzip())
t_loss, t_acc = t_loss >> Mean(), t_acc >> Mean()
print("train loss : {:.6f}".format(t_loss))
print("train acc : {:.1f}".format(100 * t_acc))
v_loss, v_acc = (val_samples >> rerange >> build_batch >>
network.validate() >> Unzip())
v_loss, v_acc = v_acc >> Mean(), v_acc >> Mean()
print('val loss : {:.6f}'.format(v_loss))
print('val acc : {:.1f}'.format(100 * v_acc))
network.save_best(v_acc, isloss=False)
plot_eval((t_acc, v_acc))
print('testing...', len(test_samples))
e_acc = (test_samples >> rerange >> build_batch >> network.evaluate(
[categorical_accuracy]))
print('test acc : {:.1f}'.format(100 * e_acc))