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 __init__(self, dataset, config, downsample_matrices, upsample_matrices,
adjacency_matrices, num_nodes):
super(Coma, self).__init__()
self.n_layers = config['n_layers']
self.filters = config['num_conv_filters']
self.filters.insert(
0, dataset.num_features) # To get initial features per node
self.K = config['polygon_order']
self.z = config['z']
self.downsample_matrices = downsample_matrices
self.upsample_matrices = upsample_matrices
self.adjacency_matrices = adjacency_matrices
self.A_edge_index, self.A_norm = zip(*[
ChebConv_Coma.norm(self.adjacency_matrices[i]._indices(),
num_nodes[i]) for i in range(len(num_nodes))
])
self.cheb = torch.nn.ModuleList([
ChebConv_Coma(self.filters[i], self.filters[i + 1], self.K[i])
for i in range(len(self.filters) - 2)
])
self.cheb_dec = torch.nn.ModuleList([
ChebConv_Coma(self.filters[-i - 1], self.filters[-i - 2],
self.K[i]) for i in range(len(self.filters) - 1)
])
self.cheb_dec[-1].bias = None # No bias for last convolution layer
self.pool = Pool()
self.enc_lin_mu = torch.nn.Linear(
self.downsample_matrices[-1].shape[0] * self.filters[-1], self.z)
self.enc_lin_logvar = torch.nn.Linear(
self.downsample_matrices[-1].shape[0] * self.filters[-1], self.z)
self.dec_lin = torch.nn.Linear(
self.z, self.filters[-1] * self.upsample_matrices[-1].shape[1])
self.reset_parameters()
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, dataset, config, downsample_matrices, upsample_matrices, adjacency_matrices, num_nodes, V_ref):
super(ComaAtt, self).__init__()
self.n_layers = config['n_layers']
self.filters_enc = config['filter_enc']
self.filters_dec = config['filter_dec']
self.K = config['polygon_order']
self.z = config['z']
self.downsample_att = config['downsample_att']
self.upsample_att = config['upsample_att']
self.downsample_matrices = downsample_matrices
self.upsample_matrices = upsample_matrices
self.adjacency_matrices = adjacency_matrices
self.A_edge_index, self.A_norm = zip(*[ChebConv_Coma.norm(self.adjacency_matrices[i]._indices(),
num_nodes[i]) for i in range(len(num_nodes))])
self.cheb_enc = torch.nn.ModuleList([ChebConv_Coma(self.filters_enc[i], self.filters_enc[i+1], self.K[i])
for i in range(len(self.filters_enc)-1)])
self.cheb_dec = torch.nn.ModuleList([ChebConv_Coma(self.filters_dec[i], self.filters_dec[i+1], self.K[i])
for i in range(len(self.filters_dec)-1)])
self.cheb_dec[-1].bias = None # No bias for last convolution layer
self.pool = Pool()
self.enc_lin = torch.nn.Linear(self.downsample_matrices[-1].shape[0]*self.filters_enc[-1], self.z)
self.dec_lin = torch.nn.Linear(self.z, self.filters_dec[0]*self.upsample_matrices[-1].shape[1])
if self.upsample_att is True or self.downsample_att is True:
self.attpoolenc, self.attpooldec = self.init_attpool(config, V_ref)
self.attpoolenc = torch.nn.ModuleList(self.attpoolenc)
self.attpooldec = torch.nn.ModuleList(self.attpooldec)
self.reset_parameters()
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, 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, 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]
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
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
# Expand dataset