def train(network, x, y, epochs=3):
"""Train network for given number of epochs"""
for epoch in range(epochs):
print('epoch', epoch + 1)
losses = (
zip(x, y) >> nf.PrintProgress(x) >> vec2img >> augment >>
nf.Shuffle(1000) >> build_batch >> network.train() >> nf.Collect())
print('train loss:', np.mean(losses))
def view_misclassified_images(network, x, y):
"""Show misclassified images"""
make_label = nf.Map(lambda s: (s[0], 'true:%d pred:%d' % (s[1], s[2])))
filter_error = nf.Filter(lambda s: s[1] != s[2])
view_image = nm.ViewImageAnnotation(0, 1, pause=1)
preds = (zip(x, y) >> vec2img >> build_pred_batch >> network.predict() >>
nf.Map(np.argmax) >> nf.Collect())
(zip(x, y, preds) >> vec2img >> filter_error >> make_label >> view_image >>
nf.Consume())
def train(network, epochs=3):
"""Train network for given number of epochs"""
print('loading data...')
filepath = download_mnist()
x_train, y_train, x_test, y_test = load_mnist(filepath)
plot = nm.PlotLines(None, every_sec=0.2)
build_batch = (nm.BuildBatch(128, verbose=False).input(
0, 'vector', 'float32').output(1, 'number', 'int64'))
for epoch in range(epochs):
print('epoch', epoch + 1)
losses = (zip(x_train,
y_train) >> nf.PrintProgress(x_train) >> nf.Shuffle(1000)
>> build_batch >> network.train() >> plot >> nf.Collect())
acc_test = evaluate(network, x_test, y_test)
acc_train = evaluate(network, x_train, y_train)
print('train loss : {:.6f}'.format(np.mean(losses)))
print('train acc : {:.1f}'.format(acc_train))
print('test acc : {:.1f}'.format(acc_test))
def train(model, samples, epochs):
"""Train network for given number of epochs"""
model.optimizer = optim.Adam(model.parameters(),
amsgrad=False,
lr=0.00057,
weight_decay=1.55e-05)
# plot = nm.PlotLines(None)
for epoch in range(epochs):
losses = (samples >> AdjChannels() >> BuildBatch(BATCHSIZE) >>
nf.Shuffle(100) >> Train(model) >> nf.Collect())
loss = np.mean(losses)
print(epoch, 'train loss:', loss)
# plot(loss)
model.save_best(loss, True, 'best_weights.pt')
if loss > 1.1 * model.best_score: # reset weights
model.load_weights('best_weights.pt')
if loss < 0.001:
break
return model.best_score
def predict(network, x, y):
"""Compute network outputs and print accuracy"""
preds = (zip(x, y) >> nf.PrintProgress(x) >> vec2img >> build_pred_batch >>
network.predict() >> nf.Collect())
acc = accuracy(y, preds)
print('test acc', acc)
def validate(network, x, y):
"""Compute validation/test loss (= mean over batch losses)"""
losses = (zip(x, y) >> nf.PrintProgress(x) >> vec2img >> build_batch >>
network.validate() >> nf.Collect())
print('val loss:', np.mean(losses))