def squeezenetmxnetload():
net = vision.squeezenet1_1(pretrained=True)
net.hybridize()
return net
# print(data.shape)
# data = mx.nd.transpose(data, (1,2,0))
# plt.imshow(data.astype(np.uint8).asnumpy())
# plt.show()
# import ipdb; ipdb.set_trace()
# if label == 0:
# print("Forward")
# else:
# print("Backward")
# if i == 20:
# break
from mxnet.gluon.model_zoo import vision as models
# lets use a pretrained squeezenet, this a model known for being decently good accuracy at a low computational cost
squeezenet = models.squeezenet1_1(pretrained=True,
prefix="direction_",
ctx=ctx)
# create a new copy of squeezenet, this time though only have 2 output classes (hotdog or not hotdog)
dognet = models.squeezenet1_1(classes=2, prefix="direction_")
dognet.collect_params().initialize(ctx=ctx)
# use the the features chunk of squeezenet, only leave the output untouched
dognet.features = squeezenet.features
# in the trainer, specify that we only want to update the output chunk of dognet
trainer = gluon.Trainer(dognet.output.collect_params(), 'sgd',
{'learning_rate': .01})
loss = gluon.loss.SoftmaxCrossEntropyLoss()
##################
# Hyperparameter #
#----------------#
ctx = mx.cpu()
lr=0.05
epochs=10
momentum=0.9
batch_size=64
#----------------#
# Hyperparameter #
##################
################## model
from mxnet.gluon.model_zoo import vision
net = vision.squeezenet1_1(classes=10, pretrained=False, ctx=ctx)
# net = vision.squeezenet1_0(classes=10, pretrained=False, ctx=ctx)
################## 그래프
import gluoncv
inputShape = (1,3,224,224)
gluoncv.utils.viz.plot_network(net, shape=inputShape)
##### 전처리 ##############################################
def transformer(data, label):
data = mx.image.imresize(data, 224, 224)
data = mx.nd.transpose(data, (2, 0, 1))
data = data.astype(np.float32)
return data, label
rand_crop=True,
shuffle=True,
batch_size=batch_size,
max_random_scale=1.5,
min_random_scale=0.75,
rand_mirror=True)
val_iter = mx.io.ImageRecordIter(path_imgrec=validation_dataset,
min_img_size=256,
data_shape=(3, 224, 224),
batch_size=batch_size)
from mxnet.gluon import nn
from mxnet.gluon.model_zoo import vision as models
from mxnet.gluon import nn
# get pretrained squeezenet
net = models.squeezenet1_1(pretrained=True, prefix='mask_', ctx=contexts)
with net.name_scope():
net.output = nn.HybridSequential()
net.output.add(
nn.Conv2D(2, in_channels=512, kernel_size=(1, 1), strides=(1, 1)))
net.output.add(nn.Activation('relu'))
net.output.add(nn.AvgPool2D(pool_size=(13, 13), strides=(13, 13)))
net.output.add(nn.Flatten())
net.output.initialize(ctx=contexts)
print(net)
def metric_str(names, accs):
return ', '.join(['%s=%f' % (name, acc) for name, acc in zip(names, accs)])
data_shape=(3, 224, 224),
rand_crop=True,
shuffle=True,
batch_size=batch_size,
max_random_scale=1.5,
min_random_scale=0.75,
rand_mirror=True)
val_iter = mx.io.ImageRecordIter(path_imgrec=validation_dataset,
min_img_size=256,
data_shape=(3, 224, 224),
batch_size=batch_size)
from mxnet.gluon import nn
from mxnet.gluon.model_zoo import vision as models
net = models.squeezenet1_1(pretrained=True, prefix='rejector_', ctx=ctx)
rejector_net = models.squeezenet1_1(prefix='rejector_', classes=3)
rejector_net.collect_params().initialize(ctx=ctx)
rejector_net.features = net.features
# return metrics string representation
def metric_str(names, accs):
return ', '.join(['%s=%f' % (name, acc) for name, acc in zip(names, accs)])
metric = mx.metric.create(['acc'])
import mxnet.gluon as gluon
from mxnet.image import color_normalize