def get_iters():
# check data
get_mnist_ubyte()
batch_size = 100
train_dataiter = mx.io.MNISTIter(
image="s3://dlami-dataset-test/MXNet/train-images.idx3-ubyte",
label="s3://dlami-dataset-test/MXNet/train-labels.idx1-ubyte",
data_shape=(1, 28, 28),
label_name='sm_label',
batch_size=batch_size,
shuffle=True,
flat=False,
silent=False,
seed=10)
val_dataiter = mx.io.MNISTIter(
image="s3://dlami-dataset-test/MXNet/t10k-images.idx3-ubyte",
label="s3://dlami-dataset-test/MXNet/t10k-labels.idx1-ubyte",
data_shape=(1, 28, 28),
label_name='sm_label',
batch_size=batch_size,
shuffle=True,
flat=False,
silent=False)
return train_dataiter, val_dataiter
def get_iterator_impl_mnist(args, kv):
"""return train and val iterators for mnist"""
# download data
get_mnist_ubyte()
flat = False if len(data_shape) != 1 else True
train = mx.io.MNISTIter(
image="data/train-images-idx3-ubyte",
label="data/train-labels-idx1-ubyte",
input_shape=data_shape,
batch_size=args.batch_size,
shuffle=True,
flat=flat,
num_parts=kv.num_workers,
part_index=kv.rank)
val = mx.io.MNISTIter(
image="data/t10k-images-idx3-ubyte",
label="data/t10k-labels-idx1-ubyte",
input_shape=data_shape,
batch_size=args.batch_size,
flat=flat,
num_parts=kv.num_workers,
part_index=kv.rank)
return (train, val)
def get_iters():
# check data
path = str(tmpdir)
get_mnist_ubyte(path)
batch_size = 100
train_dataiter = mx.io.MNISTIter(
image=os.path.join(path, 'train-images-idx3-ubyte'),
label=os.path.join(path, 'train-labels-idx1-ubyte'),
data_shape=(1, 28, 28),
label_name='sm_label',
batch_size=batch_size,
shuffle=True,
flat=False,
silent=False,
seed=10)
val_dataiter = mx.io.MNISTIter(
image=os.path.join(path, 't10k-images-idx3-ubyte'),
label=os.path.join(path, 't10k-labels-idx1-ubyte'),
data_shape=(1, 28, 28),
label_name='sm_label',
batch_size=batch_size,
shuffle=True,
flat=False,
silent=False)
return train_dataiter, val_dataiter
def get_iters():
# check data
get_mnist_ubyte()
batch_size = 100
train_dataiter = mx.io.MNISTIter(
image="data/train-images-idx3-ubyte",
label="data/train-labels-idx1-ubyte",
data_shape=(1, 28, 28),
label_name='sm_label',
batch_size=batch_size, shuffle=True, flat=False, silent=False, seed=10)
val_dataiter = mx.io.MNISTIter(
image="data/t10k-images-idx3-ubyte",
label="data/t10k-labels-idx1-ubyte",
data_shape=(1, 28, 28),
label_name='sm_label',
batch_size=batch_size, shuffle=True, flat=False, silent=False)
return train_dataiter, val_dataiter
def get_binarized_mnist(batch_size, input_shape):
"""
Returns training and validation iterators for binarized MNIST dataset
"""
get_mnist_ubyte()
flat = False if len(input_shape) == 3 else True
train_dataiter = BinarizedMNISTIter(
image=str(DATA_FOLDER_PATH / "train-images-idx3-ubyte"),
label=str(DATA_FOLDER_PATH / "train-labels-idx1-ubyte"),
input_shape=input_shape,
batch_size=batch_size,
flat=flat)
val_dataiter = BinarizedMNISTIter(
image=str(DATA_FOLDER_PATH / "t10k-images-idx3-ubyte"),
label=str(DATA_FOLDER_PATH / "t10k-labels-idx1-ubyte"),
input_shape=input_shape,
batch_size=batch_size,
flat=flat)
return train_dataiter, val_dataiter
def get_iterator_impl_mnist(args, kv):
"""return train and val iterators for mnist"""
# download data
get_mnist_ubyte()
flat = False if len(data_shape) != 1 else True
train = mx.io.MNISTIter(image="data/train-images-idx3-ubyte",
label="data/train-labels-idx1-ubyte",
input_shape=data_shape,
batch_size=args.batch_size,
shuffle=True,
flat=flat,
num_parts=kv.num_workers,
part_index=kv.rank)
val = mx.io.MNISTIter(image="data/t10k-images-idx3-ubyte",
label="data/t10k-labels-idx1-ubyte",
input_shape=data_shape,
batch_size=args.batch_size,
flat=flat,
num_parts=kv.num_workers,
part_index=kv.rank)
return (train, val)
def test_mlp(tmpdir):
# symbol net
batch_size = 100
data = mx.symbol.Variable('data')
fc1 = mx.symbol.FullyConnected(data, name='fc1', num_hidden=128)
act1 = mx.symbol.Activation(fc1, name='relu1', act_type="relu")
fc2 = mx.symbol.FullyConnected(act1, name='fc2', num_hidden=64)
act2 = mx.symbol.Activation(fc2, name='relu2', act_type="relu")
fc3 = mx.symbol.FullyConnected(act2, name='fc3', num_hidden=10)
softmax = mx.symbol.SoftmaxOutput(fc3, name='sm')
def accuracy(label, pred):
py = np.argmax(pred, axis=1)
return np.sum(py == label.astype(py)) / float(label.size)
# currently mxnet.numpy (which used in gluon.metric) did not support "==" between different types
num_epoch = 4
prefix = './mlp'
#check data
path = str(tmpdir)
get_mnist_ubyte(path)
train_dataiter = mx.io.MNISTIter(
image=os.path.join(path, 'train-images-idx3-ubyte'),
label=os.path.join(path, 'train-labels-idx1-ubyte'),
data_shape=(784, ),
label_name='sm_label',
batch_size=batch_size,
shuffle=True,
flat=True,
silent=False,
seed=10)
val_dataiter = mx.io.MNISTIter(
image=os.path.join(path, 't10k-images-idx3-ubyte'),
label=os.path.join(path, 't10k-labels-idx1-ubyte'),
data_shape=(784, ),
label_name='sm_label',
batch_size=batch_size,
shuffle=True,
flat=True,
silent=False)
# print logging by default
logging.basicConfig(level=logging.DEBUG)
model = mx.model.FeedForward.create(
softmax,
X=train_dataiter,
eval_data=val_dataiter,
eval_metric=mx.gluon.metric.np(accuracy),
epoch_end_callback=mx.callback.do_checkpoint(prefix),
ctx=[mx.cpu(i) for i in range(2)],
num_epoch=num_epoch,
learning_rate=0.1,
wd=0.0004,
momentum=0.9)
logging.info('Finish traning...')
prob = model.predict(val_dataiter)
logging.info('Finish predict...')
val_dataiter.reset()
y = np.concatenate([batch.label[0].asnumpy()
for batch in val_dataiter]).astype('int')
py = np.argmax(prob, axis=1)
acc1 = float(np.sum(py == y)) / len(y)
logging.info('final accuracy = %f', acc1)
assert (acc1 > 0.94)
# predict internal featuremaps
internals = softmax.get_internals()
fc2 = internals['fc2_output']
mfeat = mx.model.FeedForward(symbol=fc2,
arg_params=model.arg_params,
aux_params=model.aux_params,
allow_extra_params=True)
feat = mfeat.predict(val_dataiter)
assert feat.shape == (10000, 64)
# pickle the model
smodel = pickle.dumps(model)
model2 = pickle.loads(smodel)
prob2 = model2.predict(val_dataiter)
assert np.sum(np.abs(prob - prob2)) == 0
# load model from checkpoint
model3 = mx.model.FeedForward.load(prefix, num_epoch)
prob3 = model3.predict(val_dataiter)
assert np.sum(np.abs(prob - prob3)) == 0
# save model explicitly
model.save(prefix, 128)
model4 = mx.model.FeedForward.load(prefix, 128)
prob4 = model4.predict(val_dataiter)
assert np.sum(np.abs(prob - prob4)) == 0
for i in range(num_epoch):
os.remove('%s-%04d.params' % (prefix, i + 1))
os.remove('%s-symbol.json' % prefix)
os.remove('%s-0128.params' % prefix)
fc2 = mx.symbol.FullyConnected(act1, name='fc2', num_hidden=64)
act2 = mx.symbol.Activation(fc2, name='relu2', act_type="relu")
fc3 = mx.symbol.FullyConnected(act2, name='fc3', num_hidden=10)
softmax = mx.symbol.SoftmaxOutput(fc3, name='sm')
def accuracy(label, pred):
py = np.argmax(pred, axis=1)
return np.sum(py == label) / float(label.size)
num_epoch = 4
prefix = './mlp'
#check data
get_mnist_ubyte()
train_dataiter = mx.io.MNISTIter(image="data/train-images-idx3-ubyte",
label="data/train-labels-idx1-ubyte",
data_shape=(784, ),
label_name='sm_label',
batch_size=batch_size,
shuffle=True,
flat=True,
silent=False,
seed=10)
val_dataiter = mx.io.MNISTIter(image="data/t10k-images-idx3-ubyte",
label="data/t10k-labels-idx1-ubyte",
data_shape=(784, ),
label_name='sm_label',
batch_size=batch_size,
def test_autograd(tmpdir):
# define network
def get_net():
net = nn.Sequential()
net.add(nn.Dense(128, activation='relu', prefix='fc1_'))
net.add(nn.Dense(64, activation='relu', prefix='fc2_'))
net.add(nn.Dense(10, prefix='fc3_'))
return net
path = str(tmpdir)
get_mnist_ubyte(path)
batch_size = 100
train_data = mx.io.MNISTIter(image=os.path.join(path,
'train-images-idx3-ubyte'),
label=os.path.join(path,
'train-labels-idx1-ubyte'),
data_shape=(784, ),
label_name='sm_label',
batch_size=batch_size,
shuffle=True,
flat=True,
silent=False,
seed=10)
val_data = mx.io.MNISTIter(image=os.path.join(path,
't10k-images-idx3-ubyte'),
label=os.path.join(path,
't10k-labels-idx1-ubyte'),
data_shape=(784, ),
label_name='sm_label',
batch_size=batch_size,
shuffle=True,
flat=True,
silent=False)
def score(net, ctx_list):
metric = mx.metric.Accuracy()
val_data.reset()
for batch in val_data:
datas = gluon.utils.split_and_load(batch.data[0],
ctx_list,
batch_axis=0)
labels = gluon.utils.split_and_load(batch.label[0],
ctx_list,
batch_axis=0)
outputs = []
for x in datas:
outputs.append(net(x))
metric.update(labels, outputs)
return metric.get()[1]
def train(net, epoch, ctx_list):
net.collect_params().initialize(mx.init.Xavier(magnitude=2.24),
ctx=ctx_list)
trainer = gluon.Trainer(net.collect_params(), 'sgd',
{'learning_rate': 0.5})
metric = mx.metric.Accuracy()
loss = gluon.loss.SoftmaxCrossEntropyLoss()
for i in range(epoch):
train_data.reset()
for batch in train_data:
datas = gluon.utils.split_and_load(batch.data[0],
ctx_list,
batch_axis=0)
labels = gluon.utils.split_and_load(batch.label[0],
ctx_list,
batch_axis=0)
outputs = []
with autograd.record():
for x, y in zip(datas, labels):
z = net(x)
L = loss(z, y)
L.backward()
outputs.append(z)
trainer.step(batch.data[0].shape[0])
metric.update(labels, outputs)
name, acc = metric.get()
metric.reset()
print('training acc at epoch %d: %s=%f' % (i, name, acc))
net1 = get_net()
train(net1, 5, [mx.cpu(0), mx.cpu(1)])
acc1 = score(net1, [mx.cpu(0)])
acc2 = score(net1, [mx.cpu(0), mx.cpu(1)])
assert acc1 > 0.95
assert abs(acc1 - acc2) < 0.01
net1.collect_params().save('mnist.params')
net2 = get_net()
net2.collect_params().load('mnist.params', ctx=[mx.cpu(0)])
acc3 = score(net2, [mx.cpu(0)])
assert abs(acc3 - acc1) < 0.0001
fc1 = mx.symbol.FullyConnected(data, name='fc1', num_hidden=128)
act1 = mx.symbol.Activation(fc1, name='relu1', act_type="relu")
fc2 = mx.symbol.FullyConnected(act1, name = 'fc2', num_hidden = 64)
act2 = mx.symbol.Activation(fc2, name='relu2', act_type="relu")
fc3 = mx.symbol.FullyConnected(act2, name='fc3', num_hidden=10)
softmax = mx.symbol.SoftmaxOutput(fc3, name = 'sm')
def accuracy(label, pred):
py = np.argmax(pred, axis=1)
return np.sum(py == label) / float(label.size)
num_epoch = 4
prefix = './mlp'
#check data
get_mnist_ubyte()
train_dataiter = mx.io.MNISTIter(
image="data/train-images-idx3-ubyte",
label="data/train-labels-idx1-ubyte",
data_shape=(784,),
label_name='sm_label',
batch_size=batch_size, shuffle=True, flat=True, silent=False, seed=10)
val_dataiter = mx.io.MNISTIter(
image="data/t10k-images-idx3-ubyte",
label="data/t10k-labels-idx1-ubyte",
data_shape=(784,),
label_name='sm_label',
batch_size=batch_size, shuffle=True, flat=True, silent=False)
def test_mlp():