def __init__(self, state_dim, nf_particle, nf_effect, g_dim, u_dim, n_timesteps, deriv_in_state=False, num_sys=0, alpha=1, init_scale=1): super(KoopmanOperators, self).__init__() self.n_timesteps = n_timesteps self.u_dim = u_dim if deriv_in_state and n_timesteps > 2: first_deriv_dim = n_timesteps - 1 sec_deriv_dim = n_timesteps - 2 else: first_deriv_dim = 0 sec_deriv_dim = 0 ''' state ''' input_particle_dim = state_dim * ( n_timesteps + first_deriv_dim + sec_deriv_dim ) #+ g_dim # g_dim added for recursive sampling self.mapping = encoderNet(input_particle_dim, g_dim, ALPHA=alpha) self.composed_mapping = encoderNet(input_particle_dim, g_dim * 2, ALPHA=alpha) self.inv_mapping = decoderNet(state_dim, g_dim, ALPHA=alpha, SN=False) self.dynamics = dynamics(g_dim, init_scale) self.backdynamics = dynamics_back(g_dim, self.dynamics) self.u_dynamics = u_dynamics(g_dim, u_dim) # self.gru_u_mapping = nn.GRU(input_particle_dim, u_dim, num_layers = 1, batch_first=True) # self.linear_u_mapping = nn.Linear(u_dim, u_dim * 2) # self.nlinear_u_mapping = nn.Sequential( nn.Linear(input_particle_dim, state_dim), nn.ReLU(), nn.Linear(state_dim, u_dim * 2)) self.softmax = nn.Softmax(dim=-1) self.st_gumbel_softmax = tut.STGumbelSoftmax(-1) self.round = tut.Round() self.st_gumbel_sigmoid = tut.STGumbelSigmoid() self.system_identify = self.fit_block_diagonal # self.system_identify_with_A = self.fit_with_A # self.system_identify_with_compositional_A = self.fit_with_compositional_A # self.system_identify = self.fit_across_objects self.simulate = self.rollout self.simulate_no_input = self.rollout_no_input
def __init__(self, s, c, hidden_size, output_size, collision_margin=None, SN=False): super(Rel_Embedding, self).__init__() self.collision_margin = collision_margin self.output_size = output_size self.c = c self.g = output_size self.sigmoid = nn.Sigmoid() self.gumbel_sigmoid = tut.STGumbelSigmoid() # Input is s_o1 - s_o2 # if collision_margin is None: self.model_rel = nn.Sequential( nn.Linear(3 * s, hidden_size // 2), nn.ReLU(), nn.Linear(hidden_size // 2, hidden_size // 2), nn.ReLU(), nn.Linear(hidden_size // 2, 1)) self.model = nn.Sequential( nn.Linear(3 * s, hidden_size), nn.ReLU(), nn.Linear(hidden_size, hidden_size), nn.ReLU(), nn.Linear(hidden_size, output_size * c), ) for m in self.modules(): if isinstance(m, nn.Linear): nn.init.xavier_normal_(m.weight) if m.bias is not None: nn.init.constant_(m.bias, 0.0) # if SN: # self.model = SpectralNorm(self.model) # self.model_rel = SpectralNorm(self.model_u) self.round = tut.Round() self.sigmoid = nn.Sigmoid() self.relu = nn.ReLU()
def __init__(self, n_objects, dims): super(NTM, self).__init__() self.n_objects = n_objects conf_dim, layer_dim, pose_dim, shape_dim, app_dim = dims[ 'confidence'], dims['layer'], dims['pose'], dims['shape'], dims[ 'appearance'] dim_y = conf_dim + layer_dim + pose_dim + shape_dim + app_dim self.dims = dims self.dim_h_o = dims['tracker_state'] self.fcn = nn.Sequential(nn.Linear(self.dim_h_o, self.dim_h_o), nn.CELU(), nn.Linear(self.dim_h_o, self.dim_h_o), nn.CELU(), nn.Linear(self.dim_h_o, dim_y)) self.fcn = nn.Linear(self.dim_h_o, dim_y) self.softmax = nn.Softmax(dim=1) self.st_gumbel_sigmoid = ut.STGumbelSigmoid() self.st_gumbel_softmax = ut.STGumbelSoftmax(1) self.permutation_matrix_calculator = ut.PermutationMatrixCalculator() # dim, feat_dim, n_objects, H, W self.ntm_cell = NTMCell(self.dim_h_o, dims['features'], n_objects) self.t = 0 self.att = None self.mem = None
def __init__(self, state_dim, nf_particle, nf_effect, g_dim, u_dim, n_timesteps, n_blocks=1, residual=False, deriv_in_state=False, fixed_A=False, fixed_B=False, num_sys=0): super(KoopmanOperators, self).__init__() self.residual = residual self.n_timesteps = n_timesteps self.u_dim = u_dim if deriv_in_state and n_timesteps > 2: first_deriv_dim = n_timesteps - 1 sec_deriv_dim = n_timesteps - 2 else: first_deriv_dim = 0 sec_deriv_dim = 0 ''' state ''' # TODO: state_dim * n_timesteps if not hankel. Pass hankel as parameter. input_particle_dim = state_dim * (n_timesteps + first_deriv_dim + sec_deriv_dim) #+ g_dim #TODO: g_dim added for recursive sampling self.mapping = SimplePropagationNetwork( input_particle_dim=input_particle_dim, nf_particle=nf_particle, nf_effect=nf_effect, output_dim=g_dim, output_action_dim = u_dim, tanh=False, # use tanh to enforce the shape of the code space residual=residual) # g * 2 self.composed_mapping = SimplePropagationNetwork( input_particle_dim=input_particle_dim, nf_particle=nf_particle, nf_effect=nf_effect, output_dim=2 * g_dim, output_action_dim = u_dim, tanh=False, # use tanh to enforce the shape of the code space residual=residual) # g * 2 self.gru_u_mapping = nn.GRU(input_particle_dim, u_dim, num_layers = 1, batch_first=True) self.linear_u_mapping = nn.Linear(u_dim, u_dim * 2) # self.nlinear_u_mapping = nn.Sequential(nn.Linear(input_particle_dim, state_dim), nn.ReLU(), nn.Linear(state_dim, u_dim * 2)) # self.nlinear_u_mapping = nn.Sequential(nn.Linear(g_dim, state_dim), # nn.ReLU(), # nn.Linear(state_dim, u_dim * 2)) # the state for decoding phase is replaced with code of g_dim input_particle_dim = g_dim # print('state_decoder', 'node', input_particle_dim, 'edge', input_relation_dim) self.inv_mapping = SimplePropagationNetwork( input_particle_dim=input_particle_dim, nf_particle=nf_particle, nf_effect=nf_effect, output_dim=state_dim, tanh=False, residual=residual, spectral_norm=True) # TRUE ''' dynamical system coefficient: A''' # self.system_identify = self.fit # self.simulate = self.rollout # self.step = self.linear_forward self.A_reg = torch.eye(g_dim // n_blocks).unsqueeze(0) if fixed_A: # self.A = nn.Parameter( torch.randn((1, n_blocks, g_dim // n_blocks, g_dim // n_blocks), requires_grad=True) * .1 + (1 / (g_dim // n_blocks))) self.A = nn.Parameter(torch.zeros(1, n_blocks, g_dim // n_blocks, g_dim // n_blocks, requires_grad=True) + torch.eye(g_dim // n_blocks)[None, None]) if fixed_B: self.B = nn.Parameter( torch.randn((1, n_blocks, u_dim // n_blocks, g_dim // n_blocks), requires_grad=True) * .1 + (1 / (g_dim // n_blocks))) if num_sys > 0: # self.A = nn.Parameter( torch.randn((1, num_sys, g_dim // n_blocks, g_dim // n_blocks), requires_grad=True) * .1 + (1 / (g_dim // n_blocks))) self.A = nn.Parameter(torch.zeros(1, num_sys, g_dim // n_blocks, g_dim // n_blocks, requires_grad=True) + torch.eye(g_dim // n_blocks)[None, None]) # ids = torch.arange(0,self.A.shape[-1]) # self.A[..., ids, ids] = 1 # self.A = nn.Parameter(self.A) self.selector_fc = nn.Sequential(nn.Linear(g_dim, g_dim), nn.ReLU(), nn.Linear(g_dim, num_sys), nn.ReLU()) self.softmax = nn.Softmax(dim=-1) self.st_gumbel_softmax = tut.STGumbelSoftmax(-1) self.round = tut.Round() self.st_gumbel_sigmoid = tut.STGumbelSigmoid() self.system_identify = self.fit_block_diagonal self.system_identify_with_A = self.fit_with_A self.system_identify_with_compositional_A = self.fit_with_compositional_A # self.system_identify = self.fit_across_objects self.simulate = self.rollout_block_diagonal self.step = self.linear_forward_block_diagonal self.num_blocks = n_blocks
def __init__(self, state_dim, nf_particle, nf_effect, g_dim, u_dim, n_timesteps, deriv_in_state=False, num_sys=0, alpha=1, init_scale=1): super(KoopmanOperators, self).__init__() self.n_timesteps = n_timesteps self.u_dim = u_dim if deriv_in_state and n_timesteps > 2: first_deriv_dim = n_timesteps - 1 sec_deriv_dim = n_timesteps - 2 else: first_deriv_dim = 0 sec_deriv_dim = 0 ''' state ''' input_particle_dim = state_dim * ( n_timesteps + first_deriv_dim + sec_deriv_dim ) #+ g_dim # g_dim added for recursive sampling self.mapping = encoderNet(input_particle_dim, g_dim, ALPHA=alpha) self.composed_mapping = encoderNet(input_particle_dim, g_dim * 2, ALPHA=alpha) self.inv_mapping = decoderNet(state_dim, g_dim, ALPHA=alpha) self.dynamics = dynamics(g_dim, init_scale) self.backdynamics = dynamics_back(g_dim, self.dynamics) # self.gru_u_mapping = nn.GRU(input_particle_dim, u_dim, num_layers = 1, batch_first=True) # self.linear_u_mapping = nn.Linear(u_dim, u_dim * 2) # self.nlinear_u_mapping = nn.Sequential( nn.Linear(input_particle_dim, state_dim), nn.ReLU(), nn.Linear(state_dim, u_dim * 2)) # if num_sys > 0: # # self.A = nn.Parameter( torch.randn((1, num_sys, g_dim // n_blocks, g_dim // n_blocks), requires_grad=True) * .1 + (1 / (g_dim // n_blocks) **2)) # self.A = torch.zeros(1, num_sys, g_dim, g_dim) # ids = torch.arange(0,self.A.shape[-1]) # self.A[..., ids, ids] = 1 # self.A = nn.Parameter(self.A, requires_grad=True) # self.selector_fc = nn.Sequential(nn.Linear(g_dim, g_dim), # nn.ReLU(), # nn.Linear(g_dim, num_sys), # nn.ReLU()) self.softmax = nn.Softmax(dim=-1) self.st_gumbel_softmax = tut.STGumbelSoftmax(-1) self.round = tut.Round() self.st_gumbel_sigmoid = tut.STGumbelSigmoid() self.system_identify = self.fit_block_diagonal self.system_identify_with_A = self.fit_with_A self.system_identify_with_compositional_A = self.fit_with_compositional_A # self.system_identify = self.fit_across_objects self.simulate = self.rollout_block_diagonal self.step = self.linear_forward_block_diagonal
def __init__(self, input_size, out_ch, dyn_dim): super(ImageDecoder, self).__init__() self.dyn_dim = dyn_dim init_ch = 256 self.init_dec_cnn = nn.Sequential( nn.Conv2d(input_size, init_ch, 1), nn.CELU(), nn.BatchNorm2d(init_ch), ) # TO X CNN Maps appearance to object. 1 x 1 self.to_x = nn.Sequential( nn.Conv2d(init_ch, 128 * 2 * 2, 1), # 2, 2 nn.PixelShuffle(2), nn.CELU(), nn.BatchNorm2d(128), nn.Conv2d(128, 128, 3, 1, 1), nn.CELU(), nn.BatchNorm2d(128), nn.Conv2d(128, 128 * 2 * 2, 1), # 4, 4 nn.PixelShuffle(2), nn.CELU(), nn.BatchNorm2d(128), nn.Conv2d(128, 128, 3, 1, 1), nn.CELU(), nn.BatchNorm2d(128), # -- nn.Conv2d(128, 64 * 2 * 2, 1), # 8, 8 nn.PixelShuffle(2), nn.CELU(), nn.BatchNorm2d(64), nn.Conv2d(64, 64, 3, 1, 1), nn.CELU(), nn.BatchNorm2d(64), # -- nn.Conv2d(64, 32 * 2 * 2, 1), # 16, 16 nn.PixelShuffle(2), nn.CELU(), nn.BatchNorm2d(32), nn.Conv2d(32, 32, 3, 1, 1), nn.CELU(), nn.BatchNorm2d(32), # -- nn.Conv2d(32, 16 * 2 * 2, 1), # 32, 32 nn.PixelShuffle(2), nn.CELU(), nn.BatchNorm2d(16), nn.Conv2d(16, 16, 3, 1, 1), nn.CELU(), nn.BatchNorm2d(16), # -- nn.Conv2d(16, out_ch, 3, 1, 1) # output channels. If your input is RGB, out_ch=3 ) self.to_x = nn.Sequential( nn.Conv2d(init_ch, 128, 3, 1, 1), # 2, 2 nn.UpsamplingBilinear2d(scale_factor=2), nn.CELU(), nn.BatchNorm2d(128), nn.Conv2d(128, 128, 3, 1, 1), nn.CELU(), nn.BatchNorm2d(128), nn.Conv2d(128, 128, 3, 1, 1), # 4, 4 nn.UpsamplingBilinear2d(scale_factor=2), nn.CELU(), nn.BatchNorm2d(128), nn.Conv2d(128, 128, 3, 1, 1), nn.CELU(), nn.BatchNorm2d(128), # -- nn.Conv2d(128, 64, 3, 1, 1), # 8, 8 nn.UpsamplingBilinear2d(scale_factor=2), nn.CELU(), nn.BatchNorm2d(64), nn.Conv2d(64, 64, 3, 1, 1), nn.CELU(), nn.BatchNorm2d(64), # -- nn.Conv2d(64, 32, 3, 1, 1), # 16, 16 nn.UpsamplingBilinear2d(scale_factor=2), nn.CELU(), nn.BatchNorm2d(32), nn.Conv2d(32, 32, 3, 1, 1), nn.CELU(), nn.BatchNorm2d(32), # -- nn.Conv2d(32, 16, 3, 1, 1), # 32, 32 nn.UpsamplingBilinear2d(scale_factor=2), nn.CELU(), nn.BatchNorm2d(16), nn.Conv2d(16, 16, 3, 1, 1), nn.CELU(), nn.BatchNorm2d(16), # -- nn.Conv2d(16, out_ch, 3, 1, 1) # output channels. If your input is RGB, out_ch=3 ) self.st_gumbel_sigmoid = ut.STGumbelSigmoid()