def fit2():
print("fit2")
model = Model([
nn.Dense(128, inshape=1, activation='relu'),
nn.Dense(256, activation='relu'),
nn.Dropout(0.80),
nn.Dense(1)
])
model.assemble()
sess = Session(
model,
loss=losses.Mse(),
optimizer=optimizers.Fixed(),
)
history = sess.fit(ds,
200000,
val_data=val_ds,
val_epochs=1000,
listeners=[
FitListener('val_end',
callback=lambda h: on_val_end(sess, h))
])
fit_report(history, report_path + '02.png', 15)
def setUpClass(cls):
model = Model(
[Simplelayer(4, inshape=3),
Simplelayer(5),
Simplelayer(1)])
model.assemble()
cls.sess = Session(model,
loss=losses.Mse(),
optimizer=optimizers.Fixed())
def fit(name, optimizer):
inshape = ds_train.data.shape[1:]
#pdb.set_trace()
model = Model([
cnn.Conv2D(32, (3, 3), inshape=inshape),
cnn.MaxPool2D((2, 2), strides=(2, 2)),
cnn.Conv2D(64, (3, 3)),
cnn.MaxPool2D((2, 2), strides=(2, 2)),
cnn.Conv2D(64, (3, 3)),
nn.Flatten(),
nn.Dense(64),
nn.Dropout(0.5),
nn.Dense(10)
])
model.assemble()
sess = Session(model,
loss=losses.CategoricalCrossentropy(),
optimizer=optimizer)
stop_fit = session.condition_callback(lambda: sess.stop_fit(), 'val_loss',
30)
accuracy = lambda h: fit_tools.accuracy(sess, ds_test, h)
def save_and_report(history):
#pdb.set_trace()
fit_tools.fit_report(history, report_path + name + ".png")
model.save(model_path + name)
#pdb.set_trace()
history = sess.fit(ds_train,
200,
val_data=ds_test,
val_steps=100,
listeners=[
stop_fit,
session.FitListener('val_end', callback=accuracy),
session.FitListener('val_end',
callback=save_and_report)
])
save_and_report(history)
def fit(name, model):
sess = Session(model,
loss = losses.SparseCategoricalCrossentropy(),
optimizer = optimizers.Adam()
)
stop_fit = session.condition_callback(lambda :sess.stop_fit(), 'loss', 100)
def save_and_report(history):
#pdb.set_trace()
fit_tools.fit_report(history, report_path+name+".png")
model.save(model_path+name)
#pdb.set_trace()
history = sess.fit(ds_train, 1000,
listeners=[
stop_fit,
session.FitListener('epoch_end', callback=save_and_report)
]
)
save_and_report(history)
def fit(name, model):
sess = Session(model,
loss = losses.BinaryCrossentropy(),
optimizer = optimizers.Adam()
)
stop_fit = session.condition_callback(lambda :sess.stop_fit(), 'val_loss', 10)
def save_and_report(history):
#pdb.set_trace()
fit_tools.fit_report(history, report_path+name+".png")
model.save(model_path+name)
#pdb.set_trace()
history = sess.fit(ds_train, 100, val_data=ds_test, val_batches=20,
listeners=[
stop_fit,
session.FitListener('val_end', callback=fit_tools.binary_accuracy),
session.FitListener('epoch_end', callback=save_and_report)
]
)
save_and_report(history)
def fit():
inshape = ds_train.data.shape[1]
model = Model([nn.Dense(10, inshape=inshape, activation='relu')])
model.assemble()
sess = Session(model,
loss=losses.CategoricalCrossentropy(),
optimizer=optimizers.Fixed(0.001))
stop_fit = session.condition_callback(lambda: sess.stop_fit(), 'val_loss',
10)
#pdb.set_trace()
history = sess.fit(ds_train,
20000,
val_epochs=5,
val_data=ds_test,
listeners=[
stop_fit,
session.FitListener('val_end', callback=accuracy)
])
fit_report(history, report_path + "0.png")
def test_save_load(self):
fpath = "models/test_m1"
self.sess.save(fpath)
sess = Session.load(fpath)
def train(epochs, ds, model=None, batch_size=64, record_epochs=1):
#加载/构建session
sess = None
if model is None:
sess = Session.load(model_path)
else:
sess = Session(model, loss=losses.Mse(), optimizer=optimizers.Fixed())
train_x = ds['train_x']
train_y = ds['train_y']
test_x = ds['test_x']
test_y = ds['test_y']
batchs = int(train_x.shape[0] / batch_size)
print("epochs:%d, batchs=%d" % (epochs, batchs))
#记录训练历史
history = {
'loss': [],
'val_loss': [],
'epochs': [],
'val_x': test_x,
'val_y': test_y,
'val_pred': None
}
print("start training ")
t_start = time.time()
steps = epochs * batchs
epoch = 1
#循环训练
for step in range(steps):
start = (step % batchs) * batch_size
end = start + batch_size
batch_x = train_x[start:end]
batch_y = train_y[start:end]
loss = sess.batch_train(batch_x, batch_y)
cur_epoch = int(step / batchs) + 1
#每轮打印一次
if step > 0 and step % batchs == 0:
print((('epoch:%05d/%d loss=%f' % (cur_epoch, epochs, loss)) +
' ' * 50)[:50],
end='\r')
#记录
if step % batchs == 0 and (cur_epoch - epoch == record_epochs
or cur_epoch == epochs):
epoch = cur_epoch
y_pred = sess.model.predict(test_x)
val_loss = sess.loss(test_y, y_pred)
history['loss'].append(loss)
history['val_loss'].append(val_loss)
history['epochs'].append(epoch)
history['val_pred'] = y_pred
print((('epoch:%05d/%d loss=%f, val_loss=%f' %
(cur_epoch, epochs, loss, val_loss)) + ' ' * 50)[:50],
end='\r')
print("")
sess.save(model_path)
print("training finished cost:%f" % (time.time() - t_start))
return history