def Inference(in_op,num_labels):
conv1 = Conv(in_op,'conv1',3,1,80)
conv2 = Conv(conv1,'conv2',3,1,64)
pool1 = Maxpool(conv2,'pool1',2,2)
conv3 = Conv(pool1,'conv3',3,1,64)
# conv4 = Conv(conv3,'conv4',3,1,64)
pool2 = Maxpool(conv3,'pool2',2,2)
# conv5 = Conv(pool2,'conv5',3,1,32)
# conv6 = Conv(conv5,'conv6',3,1,32)
# pool3 = Maxpool(conv6,'pool3',2,2)
##
# conv7 = Conv(pool3,'conv7',3,1,246)
# conv8 = Conv(conv7,'conv8',3,1,256)
# pool4 = Maxpool(conv8,'pool4',2,2)
#
# conv9 = Conv(pool4,'conv9',3,1,128)
# conv10= Conv(conv9,'conv10',3,1,128)
# pool5 = Maxpool(conv10,'pool5',2,2)
flat = Flatten(pool2)
fc1 = Fc(flat, 'fc1', 128, activation = 'relu')
# drop1 = tf.nn.dropout(fc1,0.75)
# fc2 = Fc(drop1,'fc2',32)
# drop2 = tf.nn.dropout(fc2,0.5)
logit = Fc(fc1, 'logit', num_labels)
return logit
def __init__(self, nstack, inp_dim, oup_dim, bn=False, increase=0):
super(PoseNet, self).__init__()
self.nstack = nstack
self.pre = nn.Sequential(
Conv(3, 64, 7, 2, bn=True, relu=True),
Residual(64, 128),
Pool(2, 2),
Residual(128, 128),
Residual(128, inp_dim)
)
self.hgs = nn.ModuleList( [
nn.Sequential(
Hourglass(4, inp_dim, bn, increase),
) for i in range(nstack)] )
self.features = nn.ModuleList( [
nn.Sequential(
Residual(inp_dim, inp_dim),
Conv(inp_dim, inp_dim, 1, bn=True, relu=True)
) for i in range(nstack)] )
self.outs = nn.ModuleList( [Conv(inp_dim, oup_dim, 1, relu=False, bn=False) for i in range(nstack)] )
self.merge_features = nn.ModuleList( [Merge(inp_dim, inp_dim) for i in range(nstack-1)] )
self.merge_preds = nn.ModuleList( [Merge(oup_dim, inp_dim) for i in range(nstack-1)] )
self.nstack = nstack
def __init__(self, nstack=1, inp_dim=256, oup_dim=6, bn=False, increase=0):
super(DiscConfNet, self).__init__()
self.nstack = nstack
self.pre = nn.Sequential(
Conv(6, 64, 7, 2, bn=True,
relu=True), ## in place of 15 3 was there
Residual(64, 128),
Pool(2, 2),
Residual(128, 128),
Residual(128, inp_dim))
self.hgs = nn.ModuleList([
nn.Sequential(Hourglass(4, inp_dim, bn, increase), )
for i in range(nstack)
])
self.features = nn.ModuleList([
nn.Sequential(Residual(inp_dim, inp_dim),
Conv(inp_dim, inp_dim, 1, bn=True, relu=True))
for i in range(nstack)
])
self.outs = nn.ModuleList([
Conv(inp_dim, oup_dim, 1, relu=False, bn=False)
for i in range(nstack)
])
self.merge_features = nn.ModuleList(
[Merge(inp_dim, inp_dim) for i in range(nstack - 1)])
self.merge_preds = nn.ModuleList(
[Merge(oup_dim, inp_dim) for i in range(nstack - 1)])
self.nstack = nstack
self.fc1 = nn.Linear(64 * 64 * 6, 6)
def __init__(self,
nstack=8,
layer=4,
in_channel=256,
out_channel=16,
increase=0):
super(PoseNet, self).__init__()
self.nstack = nstack
self.pre = nn.Sequential(Conv(3, 64, 7, 2, bn=True, relu=True),
Residual(64, 128), nn.MaxPool2d(2, 2),
Residual(128, 128), Residual(128, in_channel))
self.hourglass = nn.ModuleList([
nn.Sequential(Hourglass(layer, in_channel, inc=increase))
for _ in range(nstack)
])
self.feature = nn.ModuleList([
nn.Sequential(Residual(in_channel, in_channel),
Conv(in_channel, in_channel, 1, bn=True, relu=True))
for _ in range(nstack)
])
self.outs = nn.ModuleList([
Conv(in_channel, out_channel, 1, bn=False, relu=False)
for _ in range(nstack)
])
self.merge_feature = nn.ModuleList(
[Convert(in_channel, in_channel) for _ in range(nstack - 1)])
self.merge_pred = nn.ModuleList(
[Convert(out_channel, in_channel) for _ in range(nstack - 1)])
def __init__(self, device=torch.device('cuda:0')):
super().__init__(device)
# self.fc1 = Linear(28*28, 25, noise_std=1e-0, device=self.device)
self.fc1 = Conv(1, 25, kernel_size=25, noise_std=1e-0, act='TanH', device=self.device)
self.act1 = Activation('TanH')
self.fc2 = Linear(16*25, 10, noise_std=1e-0, device=self.device)
self.softmax = Activation('Softmax')
self.layers = [self.fc1, self.fc2]
def __init__(self, in_channels, growth_rate, args):
super(_DenseLayer, self).__init__()
self.group_1x1 = args.group_1x1
self.group_3x3 = args.group_3x3
### 1x1 conv i --> b*k
self.conv_1 = Conv(in_channels, args.bottleneck * growth_rate,
kernel_size=1, groups=self.group_1x1)
### 3x3 conv b*k --> k
self.conv_2 = Conv(args.bottleneck * growth_rate, growth_rate,
kernel_size=3, padding=1, groups=self.group_3x3)
def __init__(self, device=torch.device('cuda:0')):
super().__init__(device)
self.conv1 = Conv(1, 6, kernel_size=5, noise_std=1e-0, act='TanH', device=self.device)
self.act1 = Activation('TanH')
self.pool1 = Pool(2, device=self.device)
self.fc1 = Linear(6*12*12, 100, noise_std=1e-0, act='TanH', device=self.device)
self.act2 = Activation('TanH')
self.fc2 = Linear(100, 10, noise_std=1e-0, act='TanH', device=self.device)
self.softmax = Activation('Softmax')
self.layers = [self.conv1, self.fc1, self.fc2]
class ShallowConvNet(Net):
def __init__(self, device=torch.device('cuda:0')):
super().__init__(device)
self.conv1 = Conv(1, 6, kernel_size=5, noise_std=1e-0, act='TanH', device=self.device)
self.act1 = Activation('TanH')
self.pool1 = Pool(2, device=self.device)
self.fc1 = Linear(6*12*12, 100, noise_std=1e-0, act='TanH', device=self.device)
self.act2 = Activation('TanH')
self.fc2 = Linear(100, 10, noise_std=1e-0, act='TanH', device=self.device)
self.softmax = Activation('Softmax')
self.layers = [self.conv1, self.fc1, self.fc2]
def forward(self, input):
conv_out_1 = self.conv1.forward(input)
act_out_1 = self.act1.forward(conv_out_1)
pool_out_1 = self.pool1.forward(act_out_1)
pool_out_1 = pool_out_1.reshape(len(pool_out_1), -1)
fc_out_1 = self.fc1.forward(pool_out_1)
act_out_2 = self.act2.forward(fc_out_1)
fc_out_2 = self.fc2.forward(act_out_2)
output = self.softmax.forward(fc_out_2)
return output
def __init__(self, in_channels, out_channels, args):
super(_Transition, self).__init__()
self.conv = Conv(in_channels,
out_channels,
kernel_size=1,
groups=args.group_1x1)
self.pool = nn.AvgPool2d(kernel_size=2, stride=2)
def __init__(self, in_channels, growth_rate, args):
super(_DenseLayer, self).__init__()
self.group_1x1 = args.group_1x1
self.group_3x3 = args.group_3x3
### 1x1 conv i --> b*k
#self.conv_1 = CSDN_Tem(in_channels, growth_rate, dropout=args.dropout_rate)
### 3x3 conv b*k --> k
self.conv_2 = Conv(in_channels, growth_rate, dropout=args.dropout_rate)
def __init__(self, device=torch.device('cuda:0')):
super().__init__(device)
self.conv1 = Conv(1, 6, kernel_size=5, noise_std=1e-0, act='ReLU', device=self.device)
self.act1 = Activation('ReLU')
self.pool1 = Pool(2, device=self.device)
self.conv2 = Conv(6, 16, kernel_size=5, noise_std=1e-0, act='ReLU', device=self.device)
self.act2 = Activation('ReLU')
self.pool2 = Pool(2, device=self.device)
self.fc1 = Linear(256, 120, noise_std=1e-0, act='ReLU', device=self.device)
self.act3 = Activation('ReLU')
self.fc2 = Linear(120, 84, noise_std=1e-0, act='ReLU', device=self.device)
self.act4 = Activation('ReLU')
self.fc3 = Linear(84, 10, noise_std=1e-0, act='ReLU', device=self.device)
self.softmax = Activation('Softmax')
self.layers = [self.conv1, self.conv2, self.fc1, self.fc2, self.fc3]
def main():
c = color_codes()
mnist = input_data.read_data_sets("MNIST_data/", one_hot=True)
try:
net = load_model('/home/mariano/Desktop/test.tf')
except IOError:
x = Input([784])
x_image = Reshape([28, 28, 1])(x)
x_conv1 = Conv(filters=32,
kernel_size=(5, 5),
activation='relu',
padding='same')(x_image)
h_pool1 = MaxPool((2, 2), padding='same')(x_conv1)
h_conv2 = Conv(filters=64,
kernel_size=(5, 5),
activation='relu',
padding='same')(h_pool1)
h_pool2 = MaxPool((2, 2), padding='same')(h_conv2)
h_fc1 = Dense(1024, activation='relu')(h_pool2)
h_drop = Dropout(0.5)(h_fc1)
y_conv = Dense(10)(h_drop)
net = Model(x,
y_conv,
optimizer='adam',
loss='categorical_cross_entropy',
metrics='accuracy')
print(c['c'] + '[' + strftime("%H:%M:%S") + '] ' + c['g'] + c['b'] +
'Original (MNIST)' + c['nc'] + c['g'] + ' net ' + c['nc'] + c['b'] +
'(%d parameters)' % net.count_trainable_parameters() + c['nc'])
net.fit(mnist.train.images,
mnist.train.labels,
val_data=mnist.test.images,
val_labels=mnist.test.labels,
patience=10,
epochs=200,
batch_size=1024)
save_model(net, '/home/mariano/Desktop/test.tf')
class LeNet5(Net):
def __init__(self, device=torch.device('cuda:0')):
super().__init__(device)
self.conv1 = Conv(1, 6, kernel_size=5, noise_std=1e-0, act='ReLU', device=self.device)
self.act1 = Activation('ReLU')
self.pool1 = Pool(2, device=self.device)
self.conv2 = Conv(6, 16, kernel_size=5, noise_std=1e-0, act='ReLU', device=self.device)
self.act2 = Activation('ReLU')
self.pool2 = Pool(2, device=self.device)
self.fc1 = Linear(256, 120, noise_std=1e-0, act='ReLU', device=self.device)
self.act3 = Activation('ReLU')
self.fc2 = Linear(120, 84, noise_std=1e-0, act='ReLU', device=self.device)
self.act4 = Activation('ReLU')
self.fc3 = Linear(84, 10, noise_std=1e-0, act='ReLU', device=self.device)
self.softmax = Activation('Softmax')
self.layers = [self.conv1, self.conv2, self.fc1, self.fc2, self.fc3]
def forward(self, input):
conv_out_1 = self.conv1.forward(input)
act_out_1 = self.act1.forward(conv_out_1)
pool_out_1 = self.pool1.forward(act_out_1)
conv_out_2 = self.conv2.forward(pool_out_1)
act_out_2 = self.act2.forward(conv_out_2)
pool_out_2 = self.pool2.forward(act_out_2)
pool_out_2 = pool_out_2.reshape(len(pool_out_2), -1)
fc_out_1 = self.fc1.forward(pool_out_2)
act_out_3 = self.act3.forward(fc_out_1)
fc_out_2 = self.fc2.forward(act_out_3)
act_out_4 = self.act4.forward(fc_out_2)
fc_out_3 = self.fc3.forward(act_out_4)
output = self.softmax.forward(fc_out_3)
return output
def make_cnn(X_dim, num_class):
conv = Conv(X_dim,
n_filter=16,
h_filter=5,
w_filter=5,
stride=1,
padding=2)
relu = ReLU()
maxpool = Maxpool(conv.out_dim, size=2, stride=2)
conv2 = Conv(maxpool.out_dim,
n_filter=20,
h_filter=5,
w_filter=5,
stride=1,
padding=2)
relu2 = ReLU()
maxpool2 = Maxpool(conv2.out_dim, size=2, stride=2)
flat = Flatten()
fc = FullyConnected(np.prod(maxpool2.out_dim), num_class)
return [conv, relu, maxpool, conv2, relu2, maxpool2, flat, fc]
def __init__(self, in_channels, growth_rate, args):
super(_DenseLayer, self).__init__()
self.group_1x1 = args.group_1x1
self.group_3x3 = args.group_3x3
### 1x1 conv i --> b*k
self.conv_1 = LearnedGroupConv(in_channels, args.bottleneck * growth_rate,
kernel_size=1, groups=self.group_1x1,
condense_factor=args.condense_factor,
dropout_rate=args.dropout_rate)
### 3x3 conv b*k --> k
self.conv_2 = Conv(args.bottleneck * growth_rate, in_channels,
kernel_size=3, padding=1, groups=self.group_3x3)
def build_model():
tf.compat.v1.reset_default_graph()
x = tf.compat.v1.placeholder(tf.float32, [None, 32, 32, 3])
t = tf.compat.v1.placeholder(tf.float32, [None, 10])
is_training = tf.compat.v1.placeholder(tf.bool)
layers = [
Conv((3, 3, 3, 64), tf.nn.relu),
Conv((3, 3, 64, 64), tf.nn.relu),
Pooling((1, 2, 2, 1)),
Conv((3, 3, 64, 128), tf.nn.relu),
Conv((3, 3, 128, 128), tf.nn.relu),
Pooling((1, 2, 2, 1)),
Flatten(),
Dense(3200, 256, tf.nn.relu),
Dense(256, 256, tf.nn.relu),
Dense(256, 10, tf.nn.softmax)
]
y = f_props(layers, x)
params = get_params(layers)
cost = - tf.reduce_mean(tf.reduce_sum(t * tf_log(y), axis=1))
return x, t, is_training, y, cost, params
def inner_model(trainable, x):
layers_list = [
Reshape([-1, 28, 28, 1]),
Conv(32),
BatchNormalization(),
Relu(),
MaxPool(),
Conv(64),
BatchNormalization(),
Relu(),
MaxPool(),
Reshape([-1, 7 * 7 * 64]),
FullyConnected(1024),
Relu(),
FullyConnected(10)
]
variable_saver = VariableSaver()
signal = x
print('shape', signal.get_shape())
for idx, layer in enumerate(layers_list):
signal = layer.contribute(signal, idx, trainable,
variable_saver.save_variable)
print('shape', signal.get_shape())
return signal, variable_saver.var_list
def __init__(self, in_channels, growth_rate, args):
super(_DenseLayer, self).__init__()
### 1x1 conv: i --> bottleneck * k
self.conv_1 = DynamicMultiHeadConv(in_channels,
args.bottleneck * growth_rate,
kernel_size=1,
heads=args.heads,
squeeze_rate=args.squeeze_rate,
gate_factor=args.gate_factor)
### 3x3 conv: bottleneck * k --> k
self.conv_2 = Conv(args.bottleneck * growth_rate,
growth_rate,
kernel_size=3,
padding=1,
groups=args.group_3x3)
def __init__(self, in_channels, out_channels, args, width, height):
super(_Transition, self).__init__()
self.conv = Conv(in_channels,
out_channels,
kernel_size=1,
groups=args.group_1x1)
padding = 0
if args.kernel_size == 3:
padding = 1
self.pool = cmaxgb(out_channels,
out_channels,
args=args,
kernel_size=args.kernel_size,
stride=2,
padding=padding,
width=width,
height=height)
class DenseNet_CNN(Net):
def __init__(self, device=torch.device('cuda:0')):
super().__init__(device)
# self.fc1 = Linear(28*28, 25, noise_std=1e-0, device=self.device)
self.fc1 = Conv(1, 25, kernel_size=25, noise_std=1e-0, act='TanH', device=self.device)
self.act1 = Activation('TanH')
self.fc2 = Linear(16*25, 10, noise_std=1e-0, device=self.device)
self.softmax = Activation('Softmax')
self.layers = [self.fc1, self.fc2]
def forward(self, input):
#input = input.reshape(len(input), -1)
fc_out_1 = self.fc1.forward(input)
act_out_1 = self.act1.forward(fc_out_1)
act_out_1 = act_out_1.reshape(len(act_out_1), -1)
fc_out_2 = self.fc2.forward(act_out_1)
output = self.softmax.forward(fc_out_2)
return output
def __init__(self, in_channels, out_channels, args):
super(_Transition, self).__init__()
padding = 0
if args.kernel_size == 3:
padding = 1
if args.convs:
if args.no1x1:
# purely 3x3 conv stride 2
self.conv = Conv(in_channels, out_channels,
kernel_size=args.kernel_size, padding=padding, stride=2)
self.pool = nn.Sequential()
elif args.dw:
# depthwise separable convolutions
self.conv = Conv(in_channels, in_channels, kernel_size=args.kernel_size,
padding=padding, groups=in_channels, stride=2)
#self.pool = Conv(in_channels, out_channels, kernel_size=1)
self.pool = nn.Conv2d(in_channels, out_channels,
kernel_size=1, bias=True) #adding bias because no BN before
else:
# 1x1 into 3x3 conv stride 2
self.conv = Conv(in_channels, out_channels,
kernel_size=1, padding=padding, groups=args.group_1x1)
#self.pool = Conv(out_channels, out_channels, kernel_size=args.kernel_size,
# padding=padding, stride=2)
self.pool = nn.Conv2d(out_channels, out_channels, kernel_size=args.kernel_size,
padding=padding, stride=2, bias=True) #adding bias because no BN before
elif args.dw:
# 1x1 into 3x3 conv stride 2 dw
self.conv = Conv(in_channels, out_channels, kernel_size=1)
# self.pool = Conv(out_channels, out_channels, kernel_size=args.kernel_size,
# padding=padding, stride=2, groups=out_channels)
self.pool = nn.Conv2d(out_channels, out_channels, kernel_size=args.kernel_size,
padding=padding, stride=2, bias=True, groups=out_channels) #adding bias because no BN before
elif args.noavg and args.nomax:
# 1x1 stride 2
self.conv = Conv(in_channels, out_channels, kernel_size=1, stride=2)
self.pool = nn.Sequential()
else:
self.conv = Conv(in_channels, out_channels,
kernel_size=1, groups=args.group_1x1)
self.pool = mixgb(args, padding=padding)
def get_brats_nets(input_shape, filters_list, kernel_size_list, dense_size,
nlabels):
inputs = Input(shape=input_shape)
conv = inputs
for filters, kernel_size in zip(filters_list, kernel_size_list):
conv = Conv(filters,
kernel_size=(kernel_size, ) * 3,
activation='relu',
data_format='channels_first')(conv)
full = Conv(dense_size,
kernel_size=(1, 1, 1),
data_format='channels_first',
name='fc_dense',
activation='relu')(conv)
full_roi = Conv(nlabels[0],
kernel_size=(1, 1, 1),
data_format='channels_first',
name='fc_roi')(full)
full_sub = Conv(nlabels[1],
kernel_size=(1, 1, 1),
data_format='channels_first',
name='fc_sub')(full)
rf_roi = Concatenate(axis=1)([conv, full_roi])
rf_sub = Concatenate(axis=1)([conv, full_sub])
rf_num = 1
while np.product(rf_roi.shape[2:]) > 1:
rf_roi = Conv(dense_size,
kernel_size=(3, 3, 3),
data_format='channels_first',
name='rf_roi%d' % rf_num)(rf_roi)
rf_sub = Conv(dense_size,
kernel_size=(3, 3, 3),
data_format='channels_first',
name='rf_sub%d' % rf_num)(rf_sub)
rf_num += 1
full_roi = Reshape((nlabels[0], -1))(full_roi)
full_sub = Reshape((nlabels[1], -1))(full_sub)
full_roi = Permute((2, 1))(full_roi)
full_sub = Permute((2, 1))(full_sub)
full_roi_out = Activation('softmax', name='fc_roi_out')(full_roi)
full_sub_out = Activation('softmax', name='fc_sub_out')(full_sub)
combo_roi = Concatenate(axis=1)([Flatten()(conv), Flatten()(rf_roi)])
combo_sub = Concatenate(axis=1)([Flatten()(conv), Flatten()(rf_sub)])
tumor_roi = Dense(nlabels[0], activation='softmax',
name='tumor_roi')(combo_roi)
tumor_sub = Dense(nlabels[1], activation='softmax',
name='tumor_sub')(combo_sub)
outputs_roi = [tumor_roi, full_roi_out]
net_roi = Model(inputs=inputs,
outputs=outputs_roi,
optimizer='adadelta',
loss='categorical_cross_entropy',
metrics='accuracy')
outputs_sub = [tumor_sub, full_sub_out]
net_sub = Model(inputs=inputs,
outputs=outputs_sub,
optimizer='adadelta',
loss='categorical_cross_entropy',
metrics='accuracy')
return net_roi, net_sub
def main():
train = SurfaceNormalsDataset(args.dataPath + '/actual_training', test = False, normalize=True)
trainloader = torch.utils.data.DataLoader(train, batch_size=args.batchSize, shuffle=True)
valid = SurfaceNormalsDataset(args.dataPath + '/validation', test = False, normalize=True)
# model = torch.nn.Sequential(
# Conv(3,10),
# Hourglass(4,10,bn=None,increase=64),
# Conv(10,10),
# Conv(10,3),
# )
if args.model == "":
model = torch.nn.Sequential(
Conv(3,args.numFilters),
Hourglass(args.hourglassSteps,args.numFilters,bn=None,increase=args.increase),
Conv(args.numFilters,args.numFilters),
Conv(args.numFilters,3, relu = True),
)
else:
model = torch.load("../models/" + args.model)
if args.gpu:
torch.cuda.set_device(0)
model = model.cuda()
# criterion = torch.nn.MSELoss(size_average=False)
criterion = MAELoss()
# criterion = cosDist()
optimizer = torch.optim.Adam(model.parameters(), lr=args.learn)
start = time.time()
for epoch in range(args.numEpochs): # loop over the dataset multiple times
running_loss = 0.0
for i, data in enumerate(trainloader, 0):
# get the inputs
inputs = data['image']
labels = data['normal']
# wrap them in Variable
if args.gpu:
inputs, labels = Variable(inputs).cuda(), Variable(labels).cuda()
else:
inputs, labels = Variable(inputs), Variable(labels)
# zero the parameter gradients
optimizer.zero_grad()
# forward + backward + optimize
outputs = model(inputs)
#print(outputs)
loss = criterion(outputs, labels)
loss.backward()
optimizer.step()
# print statistics
running_loss += loss.data[0]
if i % 10 == 9: # print every 10 mini-batches
divisor = (10 * 3 * 128 * 128 * args.batchSize)
divisor = 10
print('[%d, %5d] loss: %.8f' %
(epoch + 1, i + 1, running_loss / (divisor)))
running_loss = 0.0
print("Time taken: " + timeStr(time.time() - start))
print("Time remaining: " + timeRemaining(start, time.time(), epoch, args.numEpochs, i, len(trainloader)))
torch.save(model, '../models/' + args.outputName)
valid_loss = get_validation_loss(model, valid, criterion, args)
print('validation loss:', valid_loss.data[0], '\n')
sys.stdout.flush()
torch.save(model, '../models/' + args.outputName)
print('Finished Training')
def __init__(self,resgroups=1,expansion=6, num_levels=4,
down_depth = [1,2,2,2], up_depth = [1,1,1],
filters=[16,24,32,48],endchannels=[16,1],groupings=(1,1),
upkern=3,use_JPU=False,dilate_channels=32,bias_ll=False):
super().__init__()
self.useJPU = False #use_JPU
self.fused = False
assert num_levels >= 3 and num_levels <= 6
assert len(filters) == num_levels
assert len(down_depth) == num_levels
assert len(up_depth) == num_levels-1
self.num_levels = num_levels
# drop to 1/2
self.encoder0 = Sequential(Conv(3,filters[0],DWS=False,stride=2))
for j in range(down_depth[0]):
name = "DownIR_{}_{}".format(0,j)
self.encoder0.add_module(name,InvertedResidual(filters[0],
filters[0],
expansion))
self.decoder0 = Sequential(InvertedResidual(2*filters[0],filters[0],
expansion))
for j in range(up_depth[0]):
name = "UpIR_{}_{}".format(0,j)
self.decoder0.add_module(name,InvertedResidual(filters[0],
filters[0],
expansion))
if upkern==3:
self.upconv0 = UpConvUS(filters[0],endchannels[0],upsample=2,DWS=True)
else:
self.upconv0 = UpConv(filters[0],endchannels[0],upsample=2,DWS=True)
# drop to 1/4
i = 1
self.encoder1 = Sequential(InvertedResidual(filters[i-1],
filters[i],
expansion,
stride=2))
for j in range(down_depth[i]):
name = "DownIR_{}_{}".format(i,j)
self.encoder1.add_module(name,InvertedResidual(filters[i],
filters[i],
expansion))
self.decoder1 = Sequential(InvertedResidual(2*filters[i],
filters[i],
expansion))
for j in range(up_depth[i]):
name = "UpIR_{}_{}".format(i,j)
self.decoder1.add_module(name,InvertedResidual(filters[i],
filters[i],
expansion))
if upkern==3:
self.upconv1 = UpConvUS(filters[i],filters[i-1],upsample=2,DWS=True)
else:
self.upconv1 = UpConv(filters[i],filters[i-1],upsample=2,DWS=True)
# drop to 1/8
i = 2
self.encoder2 = Sequential(InvertedResidual(filters[i-1],
filters[i],
expansion,
stride=2))
for j in range(down_depth[i]):
name = "DownIR_{}_{}".format(i,j)
self.encoder2.add_module(name,InvertedResidual(filters[i],
filters[i],
expansion))
if upkern==3:
self.upconv2 = UpConvUS(filters[i],filters[i-1],upsample=2,DWS=True)
else:
self.upconv2 = UpConv(filters[i],filters[i-1],upsample=2,DWS=True)
if num_levels > 3:
# note: decoders only need one fewer level
self.decoder2 = Sequential(InvertedResidual(2*filters[i],
filters[i],
expansion))
for j in range(up_depth[i]):
name = "UpIR_{}_{}".format(i,j)
self.decoder2.add_module(name,InvertedResidual(filters[i],
filters[i],
expansion))
# drop to 1/16
i = 3
self.encoder3 = Sequential(InvertedResidual(filters[i-1],
filters[i],
expansion,
stride=2))
for j in range(down_depth[i]):
name = "DownIR_{}_{}".format(i,j)
self.encoder3.add_module(name,InvertedResidual(filters[i],
filters[i],
expansion))
if upkern==3:
self.upconv3 = UpConvUS(filters[i],filters[i-1],upsample=2,DWS=True)
else:
self.upconv3 = UpConv(filters[i],filters[i-1],upsample=2,DWS=True)
if num_levels > 4:
self.decoder3 = Sequential(InvertedResidual(2*filters[i],
filters[i],
expansion))
for j in range(up_depth[i]):
name = "UpIR_{}_{}".format(i,j)
self.decoder3.add_module(name,InvertedResidual(filters[i],
filters[i],
expansion))
# drop to 1/32
i = 4
self.encoder4 = Sequential(InvertedResidual(filters[i-1],
filters[i],
expansion,
stride=2))
for j in range(down_depth[i]):
name = "DownIR_{}_{}".format(i,j)
self.encoder4.add_module(name,InvertedResidual(filters[i],
filters[i],
expansion))
if upkern==3:
self.upconv4 = UpConvUS(filters[i],filters[i-1],upsample=2,DWS=True)
else:
self.upconv4 = UpConv(filters[i],filters[i-1],upsample=2,DWS=True)
if num_levels > 5:
self.decoder4 = Sequential(InvertedResidual(2*filters[i],
filters[i],
expansion))
for j in range(up_depth[i]):
name = "UpIR_{}_{}".format(i,j)
self.decoder4.add_module(name,InvertedResidual(filters[i],
filters[i],
expansion))
# drop to 1/64
i = 5
self.encoder5 = Sequential(InvertedResidual(filters[i-1],
filters[i],
expansion,
stride=2))
for j in range(down_depth[i]):
name = "DownIR_{}_{}".format(i,j)
self.encoder5.add_module(name,InvertedResidual(filters[i],
filters[i],
expansion))
if upkern==3:
self.upconv5 = UpConvUS(filters[i],filters[i-1],upsample=2,DWS=True)
else:
self.upconv5 = UpConv(filters[i],filters[i-1],upsample=2,DWS=True)
self.pred = Conv(endchannels[0],endchannels[1],DWS=False,bias=bias_ll)
self.edge = Conv(endchannels[0],endchannels[1],DWS=False,bias=bias_ll)
self.quant = QuantStub()
self.dequant = DeQuantStub()
#%% LSTM: WORKS
NN = FF([LSTMFullCon(Tanh, Sigmoid, Softmax, (4, 50, 4))], RNNCrossEntropy)
NN.SGD(train, eta=1e-3)
#%% RNN: WORKS
NN = FF([RecurrentFullCon(Tanh, Softmax, (4, 50, 4))], RNNCrossEntropy)
NN.SGD(train, eta=1e-4)
#%% Networks: WORKS
NNS = [
FF([FullCon(Sigmoid, (28 * 28, 30)),
FullCon(Sigmoid, (30, 10))], CrossEntropy) for i in range(3)
]
NN = FF([Networks(Sigmoid, NNS)], CrossEntropy)
NN.SGD(part_train)
#%% Conv+Pool+Dropout: WORKS
NN = FF([
Conv(Identity, (25, 3)),
Pool('mean', (3, -1), (2, 2)),
Dropout('binomial', 0.9),
FullCon(Sigmoid, (3 * 13 * 13, 10))
], CrossEntropy)
NN.SGD(part_train)
#%% FullCon Swish: WORKS
b = 10.
NN = FF([FullCon(Swish(b), (28 * 28, 30)),
FullCon(Swish(b), (30, 10))], CrossEntropy)
NN.SGD(part_train)
#%% FullCon: WORKS
NN = FF([FullCon(Sigmoid, (28 * 28, 30)),
FullCon(Sigmoid, (30, 10))], CrossEntropy)
NN.SGD(part_train)
#%% TO DO
def __init__(self, x_dim, y_dim):
super(Merge, self).__init__()
self.conv = Conv(x_dim, y_dim, 1, relu=False, bn=False)
def __init__(self,leak=0,
norm_type='batch',
DWS=True,DWSFL=False,
outerK=3,resgroups=1,
filters=[8,16,16],
shuffle=False,
blocks=[2,2,2,1,1],
endgroups=(1,1),
upkern=3,
bias_ll=True,
quant=False):
super().__init__()
self.fused = False
self.leak = leak
self.norm_type = norm_type
# downward conv block (shrink to 1/4x1/4 image)
self.down_conv = torch.nn.Sequential(
Conv1stLayer(3, filters[0], outerK, 1, DWS=DWSFL,
norm_type=norm_type, leak=leak),
Conv(filters[0],filters[1], 3, 2, DWS=DWS,groups=endgroups[0],
norm_type=norm_type, leak=leak),
Conv(filters[1], filters[2], 3, 2, DWS=DWS,groups=endgroups[1],
norm_type=norm_type, leak=leak)
)
# resblock - most effort is here
if shuffle:
self.res_block = torch.nn.Sequential()
i=0
for block in blocks:
self.res_block.add_module(str(i),ResShuffleLayer(filters[2],
leak=leak,
norm_type=norm_type,
DWS=DWS,
groups=resgroups,
dilation=block))
i += 1
else:
self.res_block = torch.nn.Sequential()
i=0
for block in blocks:
self.res_block.add_module(str(i),ResLayer(filters[2],
leak=leak,
norm_type=norm_type,
DWS=DWS,
dilation=block,
groups=resgroups))
i += 1
# up conv block (grow to original size)
if upkern == 4:
self.up_conv = torch.nn.Sequential(
UpConv(filters[2], filters[1], 4, 2, DWS=DWS,groups=endgroups[1],
norm_type=norm_type, leak=leak),
UpConv(filters[1], filters[0], 4, 2, DWS=DWS,groups=endgroups[0],
norm_type=norm_type, leak=leak),
Conv(filters[0], 3, outerK, 1, DWS=DWSFL, bias=bias_ll)
)
if upkern == 3:
self.up_conv = torch.nn.Sequential(
UpConvUS(filters[2], filters[1], 3, 2, DWS=DWS,groups=endgroups[1],
norm_type=norm_type, leak=leak),
UpConvUS(filters[1], filters[0], 3, 2, DWS=DWS,groups=endgroups[0],
norm_type=norm_type, leak=leak),
Conv(filters[0], 3, outerK, 1, DWS=DWSFL, bias=bias_ll)
)
if quant:
self.quant = QuantStub()
self.dequant = DeQuantStub()
self.transformer = torch.nn.Sequential(self.quant,
self.down_conv,
self.res_block,
self.up_conv,
self.dequant)
else:
self.transformer = torch.nn.Sequential(self.down_conv,
self.res_block,
self.up_conv)