def __init__(self,
in_channels=3,
classes=1000,
k=192,
l=224,
m=256,
n=384,
is_train=True):
super(Inceptionv4, self).__init__()
blocks = []
blocks.append(Stem(in_channels))
for _ in range(4):
blocks.append(InceptionA(384))
blocks.append(ReductionA(384, k, l, m, n))
for _ in range(7):
blocks.append(InceptionB(1024))
blocks.append(ReductionB(1024))
for _ in range(3):
blocks.append(InceptionC(1536))
self.features = nn.SequentialCell(blocks)
self.avgpool = P.ReduceMean(keep_dims=False)
self.softmax = nn.DenseBnAct(1536,
classes,
weight_init="XavierUniform",
has_bias=True,
has_bn=True,
activation="logsoftmax")
if is_train:
self.dropout = nn.Dropout(0.20)
else:
self.dropout = nn.Dropout(1)
self.bn0 = nn.BatchNorm1d(1536, eps=0.001, momentum=0.1)
def __init__(self):
super(BNReshapeDenseBNNet, self).__init__()
self.batch_norm = bn_with_initialize(2)
self.reshape = P.Reshape()
self.cast = P.Cast()
self.batch_norm2 = nn.BatchNorm1d(512, affine=False)
self.fc = fc_with_initialize(2 * 32 * 32, 512)
def __init__(self):
super(BNReshapeDenseBNNet, self).__init__()
self.batch_norm = bn_with_initialize(2)
self.reshape = P.Reshape()
self.batch_norm2 = nn.BatchNorm1d(512, affine=False)
self.fc = fc_with_initialize(2 * 32 * 32, 512)
self.loss = SemiAutoOneHotNet(args=Args(), strategy=StrategyBatch())
def test_bn1d():
"""ut of nn.BatchNorm1d"""
bn = nn.BatchNorm1d(3)
input_data = Tensor(np.random.randint(0, 1, [1, 3]).astype(np.float32))
output = bn(input_data)
output_np = output.asnumpy()
assert isinstance(output_np[0][0], (np.float32, np.float64))
def __init__(self, num_class=10): # 一共分十类，图片通道数是1
super(ForwardFashionRegularization, self).__init__()
self.num_class = num_class
self.conv1 = nn.Conv2d(1,
32,
kernel_size=3,
stride=1,
padding=0,
has_bias=False,
pad_mode="valid")
self.conv2 = nn.Conv2d(32,
64,
kernel_size=3,
stride=1,
padding=0,
has_bias=False,
pad_mode="valid")
self.conv3 = nn.Conv2d(64,
128,
kernel_size=3,
stride=1,
padding=0,
has_bias=False,
pad_mode="valid")
self.maxpool2d = nn.MaxPool2d(kernel_size=2, stride=2)
self.relu = nn.ReLU()
self.dropout = nn.Dropout()
self.flatten = nn.Flatten()
self.fc1 = nn.Dense(3200, 128)
self.bn = nn.BatchNorm1d(128)
self.fc2 = nn.Dense(128, self.num_class)
def _bn1_kaiming(channel):
return nn.BatchNorm1d(channel,
eps=1e-4,
momentum=0.9,
gamma_init=1,
beta_init=0,
moving_mean_init=0,
moving_var_init=1,
use_batch_statistics=None)
def __init__(self, flat_dim, fc_dim, attri_num_list):
super(AttriHead, self).__init__()
self.fc1 = fc_with_initialize(flat_dim, fc_dim)
self.fc1_relu = P.ReLU()
self.fc1_bn = nn.BatchNorm1d(fc_dim, affine=False)
self.attri_fc1 = fc_with_initialize(fc_dim, attri_num_list[0])
self.attri_fc1_relu = P.ReLU()
self.attri_bn1 = nn.BatchNorm1d(attri_num_list[0], affine=False)
self.softmax1 = P.Softmax()
self.fc2 = fc_with_initialize(flat_dim, fc_dim)
self.fc2_relu = P.ReLU()
self.fc2_bn = nn.BatchNorm1d(fc_dim, affine=False)
self.attri_fc2 = fc_with_initialize(fc_dim, attri_num_list[1])
self.attri_fc2_relu = P.ReLU()
self.attri_bn2 = nn.BatchNorm1d(attri_num_list[1], affine=False)
self.softmax2 = P.Softmax()
self.fc3 = fc_with_initialize(flat_dim, fc_dim)
self.fc3_relu = P.ReLU()
self.fc3_bn = nn.BatchNorm1d(fc_dim, affine=False)
self.attri_fc3 = fc_with_initialize(fc_dim, attri_num_list[2])
self.attri_fc3_relu = P.ReLU()
self.attri_bn3 = nn.BatchNorm1d(attri_num_list[2], affine=False)
self.softmax3 = P.Softmax()
def __init__(self, mul_size, test_size, strategy=None, strategy2=None):
super().__init__()
mul_np = np.full(mul_size, 0.5, dtype=np.float32)
floordiv_np = np.full(test_size, 0.1, dtype=np.float32)
self.mul_weight = Parameter(Tensor(mul_np), name="mul_weight")
self.floordiv_weight = Parameter(Tensor(floordiv_np),
name="floordiv_weight")
self.mul = TwoInputBpropOperator()
self.floor_div = P.FloorDiv()
self.bn = nn.BatchNorm1d(num_features=96)
if strategy is not None:
self.mul.op.shard(strategy2)
self.mul.bp.shard(strategy2)
self.floor_div.shard(strategy)
def __init__(self, num_layers=36, feature_dim=128, shape=(96, 64)):
super(SphereNet_float32, self).__init__()
assert num_layers in [12, 20, 36, 64], 'SphereNet num_layers should be 12, 20 or 64'
if num_layers == 12:
layers = [1, 1, 1, 1]
filter_list = [3, 16, 32, 64, 128]
fc_size = 128 * 6 * 4
elif num_layers == 20:
layers = [1, 2, 4, 1]
filter_list = [3, 64, 128, 256, 512]
fc_size = 512 * 6 * 4
elif num_layers == 36:
layers = [2, 4, 4, 2]
filter_list = [3, 32, 64, 128, 256]
fc_size = 256 * 6 * 4
elif num_layers == 64:
layers = [3, 7, 16, 3]
filter_list = [3, 64, 128, 256, 512]
fc_size = 512 * 6 * 4
else:
raise ValueError('sphere' + str(num_layers) + " IS NOT SUPPORTED! (sphere20 or sphere64)")
self.shape = P.Shape()
self.reshape = P.Reshape()
block = BaseBlock
self.layer1 = MakeLayer(block, filter_list[0], filter_list[1], layers[0], stride=2)
self.layer2 = MakeLayer(block, filter_list[1], filter_list[2], layers[1], stride=2)
self.layer3 = MakeLayer(block, filter_list[2], filter_list[3], layers[2], stride=2)
self.layer4 = MakeLayer(block, filter_list[3], filter_list[4], layers[3], stride=2)
self.fc = fc_with_initialize(fc_size, feature_dim)
self.last_bn = nn.BatchNorm1d(feature_dim, momentum=0.9).add_flags_recursive(fp32=True)
self.last_bn_sub = nn.BatchNorm2d(feature_dim, momentum=0.9).add_flags_recursive(fp32=True)
self.cast = P.Cast()
self.l2norm = P.L2Normalize(axis=1)
for _, cell in self.cells_and_names():
if isinstance(cell, (nn.Conv2d, nn.Dense)):
if cell.bias is not None:
cell.weight.set_data(initializer(me_init.ReidKaimingUniform(a=math.sqrt(5), mode='fan_out'),
cell.weight.shape))
cell.bias.set_data(initializer('zeros', cell.bias.shape))
else:
cell.weight.set_data(initializer(me_init.ReidXavierUniform(), cell.weight.shape))
self.device_target = context.get_context('device_target')
def __init__(self,
low_dim,
class_num=200,
drop=0.2,
part=0,
alpha=0.2,
nheads=4,
arch="resnet50"):
super(embed_net, self).__init__()
# print("class_num is :", class_num)
self.thermal_module = thermal_module(arch=arch)
self.visible_module = visible_module(arch=arch)
self.base_resnet = base_resnet(arch=arch)
pool_dim = 2048
self.dropout = drop
self.part = part
self.l2norm = Normalize(2)
self.bottleneck = nn.BatchNorm1d(num_features=pool_dim)
self.bottleneck.requires_grad = False # Maybe problematic? Original in PyTorch bottleneck.bias.requires_grad(False)
self.classifier = nn.Dense(pool_dim, class_num, has_bias=False)
# self.classifier1 = nn.Dense(pool_dim, class_num, has_bias=False)
# self.classifier2 = nn.Dense(pool_dim, class_num, has_bias=False)
# TODO:add weights initialization module
# self.bottleneck.apply(weights_init_kaiming)
# self.classifier.apply(weights_init_classifier)
# self.classifier1.apply(weights_init_classifier)
# self.classifier2.apply(weights_init_classifier)
self.classifier.weight.set_data(
weight_init.initializer(weight_init.Normal(sigma=0.001), \
self.classifier.weight.shape, self.classifier.weight.dtype))
self.avgpool = P.ReduceMean(keep_dims=True)
if self.part > 0:
self.wpa = IWPA(pool_dim, self.part)
else:
self.wpa = IWPA(pool_dim, 3)
def __init__(self, emb_size, args=None):
super(Head0, self).__init__()
if args.pre_bn == 1:
self.bn1 = bn_with_initialize(512, use_inference=args.inference)
else:
self.bn1 = Cut()
if args is not None:
if args.use_drop == 1:
self.drop = nn.Dropout(keep_prob=0.4)
else:
self.drop = Cut()
else:
self.drop = nn.Dropout(keep_prob=0.4)
self.fc1 = fc_with_initialize(512 * 7 * 7, emb_size)
if args.inference == 1:
self.bn2 = Cut()
else:
self.bn2 = nn.BatchNorm1d(
emb_size, affine=False,
momentum=0.9).add_flags_recursive(fp32=True)
self.reshape = P.Reshape()
self.shape = P.Shape()
def __init__(self):
super(BatchNormReshapeNet, self).__init__()
self.vd = P._VirtualDataset()
self.batch_norm = nn.BatchNorm1d(512, affine=False)
self.reshape = P.Reshape()
self.prelu = nn.PReLU(channel=256)
def __init__(self):
super(Blockcell, self).__init__()
self.bn = nn.BatchNorm1d(64, momentum=0.9)
def __init__(self,
batch_size,
input_size,
hidden_size,
num_layers,
bidirectional=False,
batch_norm=False,
rnn_type='LSTM'):
super(BatchRNN, self).__init__()
self.batch_size = batch_size
self.input_size = input_size
self.hidden_size = hidden_size
self.num_layers = num_layers
self.rnn_type = rnn_type
self.bidirectional = bidirectional
self.has_bias = True
self.is_batch_norm = batch_norm
self.num_directions = 2 if bidirectional else 1
self.reshape_op = P.Reshape()
self.shape_op = P.Shape()
self.sum_op = P.ReduceSum()
input_size_list = [input_size]
for i in range(num_layers - 1):
input_size_list.append(hidden_size)
layers = []
for i in range(num_layers):
layers.append(
nn.LSTMCell(input_size=input_size_list[i],
hidden_size=hidden_size,
bidirectional=bidirectional,
has_bias=self.has_bias))
weights = []
for i in range(num_layers):
weight_size = (input_size_list[i] +
hidden_size) * hidden_size * self.num_directions * 4
if self.has_bias:
bias_size = self.num_directions * hidden_size * 4 * 2
weight_size = weight_size + bias_size
stdv = 1 / math.sqrt(hidden_size)
w_np = np.random.uniform(-stdv, stdv,
(weight_size, 1, 1)).astype(np.float32)
weights.append(
Parameter(initializer(Tensor(w_np), w_np.shape),
name="weight" + str(i)))
self.h, self.c = self.stack_lstm_default_state(
batch_size,
hidden_size,
num_layers=num_layers,
bidirectional=bidirectional)
self.lstms = layers
self.weight = ParameterTuple(tuple(weights))
if batch_norm:
batch_norm_layer = []
for i in range(num_layers - 1):
batch_norm_layer.append(nn.BatchNorm1d(hidden_size))
self.batch_norm_list = batch_norm_layer
def bn_with_initialize(out_channels):
bn = nn.BatchNorm1d(out_channels, momentum=0.1, eps=1e-5)
return bn
def __init__(self,
in_channels,
part=3,
inter_channels=None,
out_channels=None):
super(IWPA, self).__init__()
self.in_channels = in_channels
self.inter_channels = inter_channels
self.out_channels = out_channels
self.part = part
self.l2norm = L2Normalize()
self.softmax = nn.Softmax(axis=-1)
if self.inter_channels is None:
self.inter_channels = in_channels
if self.out_channels is None:
self.out_channels = in_channels
self.fc1 = nn.Conv2d(in_channels=self.in_channels,
out_channels=self.inter_channels,
kernel_size=1,
stride=1,
padding=0)
self.fc2 = nn.Conv2d(in_channels=self.in_channels,
out_channels=self.inter_channels,
kernel_size=1,
stride=1,
padding=0)
self.fc3 = nn.Conv2d(in_channels=self.in_channels,
out_channels=self.inter_channels,
kernel_size=1,
stride=1,
padding=0)
self.W = nn.SequentialCell(
nn.Conv2d(in_channels=self.inter_channels,
out_channels=self.out_channels,
kernel_size=1,
stride=1,
padding=0),
nn.BatchNorm2d(self.out_channels),
)
# self.W[1].weight.set_data(Constant(0.0))
# self.w[2].bias.set_data(Constant(0.0))
self.bottleneck = nn.BatchNorm1d(in_channels)
self.bottleneck.requires_grad = False # no shift
# self.bottleneck.weight.set_data(Normal(sigma=0.01))
#In original PyTorch code:nn.init.normal_(self.bottleneck.weight.data, 1.0, 0.01)
# weighting vector of the part features
self.gate = ms.Parameter(ms.Tensor(np.ones(self.part),
dtype=ms.float32),
name="w",
requires_grad=True)
self.gate.set_data(
weight_init.initializer(Constant(1 / self.part), self.gate.shape,
self.gate.dtype))