return pp
if __name__ == "__main__":
BATCH_SIZE = 128
mnist = MNIST(batch_class=MyBatch)
config = dict(some=1, conv=dict(arg1=10))
print()
print("Start training...")
t = time()
train_tp = (Pipeline(config=config)
.init_variable('model', VGG16)
.init_variable('loss_history', init_on_each_run=list)
.init_variable('current_loss', init_on_each_run=0)
.init_variable('input_tensor_name', 'images')
.init_model('dynamic', V('model'), 'conv',
config={'session': {'config': tf.ConfigProto(allow_soft_placement=True)},
'loss': 'ce',
'optimizer': {'name':'Adam', 'use_locking': True},
'inputs': dict(images={'shape': (None, None, 1)}, #'shape': (28, 28, 1), 'transform': 'mip @ 1'},
#labels={'shape': 10, 'dtype': 'uint8',
labels={'classes': (10+np.arange(10)).astype('str'),
'transform': 'ohe', 'name': 'targets'}),
'input_block/inputs': 'images',
'output': dict(ops=['labels', 'accuracy'])})
.make_digits()
.train_model('conv', fetches='loss',
feed_dict={V('input_tensor_name'): B('images'),
'labels': B('digits')},
save_to=V('current_loss'))
#.print_variable('current_loss')
if __name__ == "__main__":
BATCH_SIZE = 64
mnist = MNIST()
train_template = (
Pipeline(config=dict(model=VGG7)).init_variable(
'model', ResNet18).init_variable(
'loss_history',
init_on_each_run=list).init_variable('current_loss',
init_on_each_run=0).
init_variable('pred_label', init_on_each_run=list).init_model(
'dynamic',
V('model'),
'conv',
config={
'inputs':
dict(images={'shape': B('image_shape')},
labels={
'classes': 10,
'transform': 'ohe',
'name': 'targets'
}),
'input_block/inputs':
'images',
'input_block/filters':
16,
#'body/block/bottleneck': 1,
#'head/units': [100, 100, 10],
#'input_block/filters': 32,
'input_block/inputs':
'images',
#'body/filters': [16,32,64,128],
})
#'output': dict(ops=['labels', 'accuracy'])})
.train_model(
'conv',
fetches='loss',
feed_dict={
'images': B('images'),
'masks': F(make_masks)
},
#feed_dict={'images': F(make3d_images), #B('images'),
# 'masks': F(make3d_masks)},
save_to=V('current_loss')).print(
V('current_loss')).update_variable('loss_history',
V('current_loss'),
mode='a'))
train_pp = (train_template << mnist.train)
print("Start training...")
t = time()
train_pp.run(BATCH_SIZE,
shuffle=True,
n_epochs=1,
drop_last=False,
prefetch=0)
print("End training", time() - t)
print()
pass
if __name__ == "__main__":
BATCH_SIZE = 64
#mnist = MNIST()
mnist = CIFAR10()
train_template = (
Pipeline(config=dict(model=MobileNet_v2)).init_variable(
'model', C('model')).init_variable('loss_history',
init_on_each_run=list).
init_variable('current_loss', init_on_each_run=0).init_model(
'dynamic',
V('model'),
'conv',
config={
'inputs':
dict(images={'shape': B('image_shape')},
labels={
'classes': 10,
'transform': 'ohe',
'name': 'targets'
}),
'input_block/inputs':
'images',
#'input_block/filters': 16,
#'body/block/bottleneck': 1,
#'head/units': [100, 100, 10],
#'nothing': F(lambda batch: batch.images.shape[1:]),
class MyModel(TFModel):
def _build(self, config=None):
tf.losses.add_loss(1.)
pass
if __name__ == "__main__":
BATCH_SIZE = 64
mnist = MNIST()
train_template = (Pipeline(config=dict(model=VGG7))
.init_variable('model', ResNetAttention56)
.init_variable('loss_history', init_on_each_run=list)
.init_variable('current_loss', init_on_each_run=0)
.init_variable('pred_label', init_on_each_run=list)
.init_model('dynamic', V('model'), 'conv',
config={'inputs': dict(images={'shape': B('image_shape')},
labels={'classes': 10, 'transform': 'ohe', 'name': 'targets'}),
'input_block/inputs': 'images',
'input_block/filters': 16,
#'body/block/bottleneck': 1,
#'head/units': [100, 100, 10],
#'nothing': F(lambda batch: batch.images.shape[1:]),
#'filters': 16, 'width_factor': 1,
#'body': dict(se_block=1, se_factor=4, resnext=1, resnext_factor=4, bottleneck=1),
'output': dict(ops=['accuracy'])})
#.resize(shape=(64, 64))
.train_model('conv', fetches='loss',
feed_dict={'images': B('images'),
'labels': B('labels')},
save_to=V('current_loss'), use_lock=True)