def __init__(self,
dim,
hidden_dim,
autoencoder,
nonlinearity='tanh',
dt=1e-3,
device=None):
super(PixelLagrangian, self).__init__()
self.autoencoder = autoencoder
S_net = MLP(dim, hidden_dim, dim**2, nonlinearity).to(device)
U_net = MLP(dim, hidden_dim, 1, nonlinearity).to(device)
self.lag = LagrangianFriction(dim, S_net, U_net, dt=dt, device=device)
def train():
data = get_dataset()
x = torch.tensor(data['x'], requires_grad=True,
dtype=torch.float32) # x 实际上是位置和速度
test_x = torch.tensor(data['test_x'],
requires_grad=True,
dtype=torch.float32)
_, acce = torch.Tensor(data['test_dx']).chunk(2, 1)
_, test_acce = torch.Tensor(data['test_dx']).chunk(2, 1)
N, freedom = x.shape
freedom /= 2
input_dim = int(freedom * 2)
output_dim = int(freedom)
model_nn = MLP(input_dim, 50, output_dim, 'tanh')
model = LNN(input_dim, differentiable_model=model_nn)
optim = torch.optim.Adam(model.parameters(), 5e-3, weight_decay=1e-4)
# vanilla train loop
stats = {'train_loss': [], 'test_loss': []}
torch.autograd.set_detect_anomaly(True)
for step in range(500): #500 epoch
# train step
loss = 0
for i in range(100):
acce_hat = model.forward_new(x[i])
loss = loss + L1_loss(acce[i], acce_hat)
loss.backward()
loss /= 100
optim.step()
optim.zero_grad()
print("step {}, train_loss {:.4e}, ".format(step, loss))
writer.add_scalar('LNN/spring_train_loss', loss, step)
def __init__(self, input_dim, hidden_dim, autoencoder,
field_type='solenoidal', nonlinearity='tanh', baseline=False):
super(PixelHNN, self).__init__()
self.autoencoder = autoencoder
self.baseline = baseline
output_dim = input_dim if baseline else 2
nn_model = MLP(input_dim, hidden_dim, output_dim, nonlinearity)
self.hnn = HNN(input_dim, differentiable_model=nn_model, field_type=field_type, baseline=baseline)
def train(args):
# set random seed
torch.manual_seed(args.seed)
np.random.seed(args.seed)
# init model and optimizer
if args.verbose:
print("Training baseline model:" if args.baseline else "Training HNN model:")
output_dim = args.input_dim if args.baseline else 2
nn_model = MLP(args.input_dim, args.hidden_dim, output_dim, args.nonlinearity)
model = HNN(args.input_dim, differentiable_model=nn_model,
field_type=args.field_type, baseline=args.baseline)
optim = torch.optim.Adam(model.parameters(), args.learn_rate, weight_decay=0)
# arrange data
data = get_dataset(args.name, args.save_dir, verbose=True)
x = torch.tensor( data['coords'], requires_grad=True, dtype=torch.float32)
test_x = torch.tensor( data['test_coords'], requires_grad=True, dtype=torch.float32)
dxdt = torch.Tensor(data['dcoords'])
test_dxdt = torch.Tensor(data['test_dcoords'])
# vanilla train loop
stats = {'train_loss': [], 'test_loss': []}
for step in range(args.total_steps+1):
# train step
ixs = torch.randperm(x.shape[0])[:args.batch_size]
dxdt_hat = model.time_derivative(x[ixs])
dxdt_hat += args.input_noise * torch.randn(*x[ixs].shape) # add noise, maybe
loss = L2_loss(dxdt[ixs], dxdt_hat)
loss.backward()
grad = torch.cat([p.grad.flatten() for p in model.parameters()]).clone()
optim.step() ; optim.zero_grad()
# run test data
test_ixs = torch.randperm(test_x.shape[0])[:args.batch_size]
test_dxdt_hat = model.time_derivative(test_x[test_ixs])
test_dxdt_hat += args.input_noise * torch.randn(*test_x[test_ixs].shape) # add noise, maybe
test_loss = L2_loss(test_dxdt[test_ixs], test_dxdt_hat)
# logging
stats['train_loss'].append(loss.item())
stats['test_loss'].append(test_loss.item())
if args.verbose and step % args.print_every == 0:
print("step {}, train_loss {:.4e}, test_loss {:.4e}, grad norm {:.4e}, grad std {:.4e}"
.format(step, loss.item(), test_loss.item(), grad@grad, grad.std()))
train_dxdt_hat = model.time_derivative(x)
train_dist = (dxdt - train_dxdt_hat)**2
test_dxdt_hat = model.time_derivative(test_x)
test_dist = (test_dxdt - test_dxdt_hat)**2
print('Final train loss {:.4e} +/- {:.4e}\nFinal test loss {:.4e} +/- {:.4e}'
.format(train_dist.mean().item(), train_dist.std().item()/np.sqrt(train_dist.shape[0]),
test_dist.mean().item(), test_dist.std().item()/np.sqrt(test_dist.shape[0])))
return model, stats
def get_hnn_model(args, baseline):
output_dim = args.input_dim if args.baseline else 2
nn_model = MLP(args.input_dim, 400, output_dim, args.nonlinearity)
model = HNN(args.input_dim,
differentiable_model=nn_model,
field_type=args.field_type,
baseline=args.baseline)
label = '-baseline' if args.baseline else '-hnn'
label = label + '-rad' if args.rad else label
path = '{}/{}{}.tar'.format(args.save_dir, args.name, label)
return model
def __init__(self,
dim,
autoencoder,
nonlinearity='tanh',
dt=1e-3,
M_hidden=300,
V_hidden=50,
g_hidden=200,
device=None):
super(PixelSymODEN_R, self).__init__()
self.autoencoder = autoencoder
M_net = PSD(dim, M_hidden, dim).to(device)
V_net = MLP(dim, V_hidden, 1).to(device)
g_net = MLP(dim, g_hidden, dim).to(device)
self.symoden = SymODEN_R(dim * 2,
M_net=M_net,
V_net=V_net,
g_net=g_net,
device=device,
baseline=False,
structure=True)
def get_model(args, baseline, structure, naive, num_points):
M_net = PSD(3, 400, 2).to(device)
g_net = MatrixNet(3, 300, 4, shape=(2,2)).to(device)
if structure == False:
if naive and baseline:
raise RuntimeError('argument *baseline* and *naive* cannot both be true')
elif naive:
input_dim = 6
output_dim = 5
nn_model = MLP(input_dim, 1000, output_dim, args.nonlinearity).to(device)
model = SymODEN_R1_T1(args.num_angle, H_net=nn_model, device=device, baseline=baseline, naive=naive, u_dim=2)
elif baseline:
input_dim = 6
output_dim = 4
nn_model = MLP(input_dim, 700, output_dim, args.nonlinearity).to(device)
model = SymODEN_R1_T1(args.num_angle, H_net=nn_model, M_net=M_net, device=device, baseline=baseline, naive=naive, u_dim=2)
else:
input_dim = 5
output_dim = 1
nn_model = MLP(input_dim, 500, output_dim, args.nonlinearity).to(device)
model = SymODEN_R1_T1(args.num_angle, H_net=nn_model, M_net=M_net, g_net=g_net, device=device, baseline=baseline, naive=naive, u_dim=2)
elif structure == True and baseline ==False and naive==False:
V_net = MLP(3, 300, 1).to(device)
model = SymODEN_R1_T1(args.num_angle, M_net=M_net, V_net=V_net, g_net=g_net, device=device, baseline=baseline, structure=True, u_dim=2).to(device)
else:
raise RuntimeError('argument *structure* is set to true, no *baseline* or *naive*!')
if naive:
label = '-naive_ode'
elif baseline:
label = '-baseline_ode'
else:
label = '-hnn_ode'
struct = '-struct' if structure else ''
path = '{}/{}{}{}-{}-p{}.tar'.format(args.save_dir, args.name, label, struct, args.solver, args.num_points)
model.load_state_dict(torch.load(path, map_location=device))
path = '{}/{}{}{}-{}-p{}-stats.pkl'.format(args.save_dir, args.name, label, struct, args.solver, args.num_points)
stats = from_pickle(path)
return model, stats
def get_model(args, baseline, structure, damping, num_points, gym=False):
if structure == False and baseline == True:
nn_model = MLP(args.input_dim, 600, args.input_dim,
args.nonlinearity).to(device)
model = SymODEN_R(args.input_dim,
H_net=nn_model,
device=device,
baseline=True)
elif structure == False and baseline == False:
H_net = MLP(args.input_dim, 400, 1, args.nonlinearity).to(device)
g_net = MLP(int(args.input_dim / 2), 200,
int(args.input_dim / 2)).to(device)
model = SymODEN_R(args.input_dim,
H_net=H_net,
g_net=g_net,
device=device,
baseline=False)
elif structure == True and baseline == False:
# M_net = MLP(1, args.hidden_dim, 1).to(device)
M_net = MLP(int(args.input_dim / 2), 300, int(args.input_dim / 2))
V_net = MLP(int(args.input_dim / 2), 50, 1).to(device)
g_net = MLP(int(args.input_dim / 2), 200,
int(args.input_dim / 2)).to(device)
model = SymODEN_R(args.input_dim,
M_net=M_net,
V_net=V_net,
g_net=g_net,
device=device,
baseline=False,
structure=True).to(device)
else:
raise RuntimeError(
'argument *baseline* and *structure* cannot both be true')
model_name = 'baseline_ode' if baseline else 'hnn_ode'
struct = '-struct' if structure else ''
rad = '-rad' if args.rad else ''
path = '{}pend-{}{}-{}-p{}{}.tar'.format(args.save_dir, model_name, struct,
args.solver, num_points, rad)
model.load_state_dict(torch.load(path, map_location=device))
path = '{}/pend-{}{}-{}-p{}-stats{}.pkl'.format(args.save_dir, model_name,
struct, args.solver,
num_points, rad)
stats = from_pickle(path)
return model, stats
def train(args):
# import ODENet
# from torchdiffeq import odeint
from torchdiffeq import odeint_adjoint as odeint
device = torch.device('cuda:' + str(args.gpu) if torch.cuda.is_available() else 'cpu')
# reproducibility: set random seed
torch.manual_seed(args.seed)
np.random.seed(args.seed)
torch.backends.cudnn.deterministic = True
torch.backends.cudnn.benchmark = False
# init model and optimizer
if args.verbose:
print("Start training with num of points = {} and solver {}.".format(args.num_points, args.solver))
if args.structure == False and args.baseline == True:
nn_model = MLP(args.input_dim, 600, args.input_dim, args.nonlinearity).to(device)
model = SymODEN_R(args.input_dim, H_net=nn_model, device=device, baseline=True)
elif args.structure == False and args.baseline == False:
H_net = MLP(args.input_dim, 400, 1, args.nonlinearity).to(device)
g_net = MLP(int(args.input_dim/2), 200, int(args.input_dim/2)).to(device)
model = SymODEN_R(args.input_dim, H_net=H_net, g_net=g_net, device=device, baseline=False)
elif args.structure == True and args.baseline ==False:
M_net = MLP(int(args.input_dim/2), 300, int(args.input_dim/2))
V_net = MLP(int(args.input_dim/2), 50, 1).to(device)
g_net = MLP(int(args.input_dim/2), 200, int(args.input_dim/2)).to(device)
model = SymODEN_R(args.input_dim, M_net=M_net, V_net=V_net, g_net=g_net, device=device, baseline=False, structure=True).to(device)
else:
raise RuntimeError('argument *baseline* and *structure* cannot both be true')
num_parm = get_model_parm_nums(model)
print('model contains {} parameters'.format(num_parm))
optim = torch.optim.Adam(model.parameters(), args.learn_rate, weight_decay=1e-4)
# arrange data
us = [-2.0, -1.0, 0.0, 1.0, 2.0]
# us = [0.0]
data = get_dataset(seed=args.seed, timesteps=45,
save_dir=args.save_dir, rad=args.rad, us=us, samples=50)
train_x, t_eval = arrange_data(data['x'], data['t'], num_points=args.num_points)
test_x, t_eval = arrange_data(data['test_x'], data['t'], num_points=args.num_points)
train_x = torch.tensor(train_x, requires_grad=True, dtype=torch.float32).to(device)
test_x = torch.tensor(test_x, requires_grad=True, dtype=torch.float32).to(device)
t_eval = torch.tensor(t_eval, requires_grad=True, dtype=torch.float32).to(device)
# training loop
stats = {'train_loss': [], 'test_loss': [], 'forward_time': [], 'backward_time': [],'nfe': []}
for step in range(args.total_steps+1):
train_loss = 0
test_loss = 0
for i in range(train_x.shape[0]):
t = time.time()
train_x_hat = odeint(model, train_x[i, 0, :, :], t_eval, method=args.solver)
forward_time = time.time() - t
train_loss_mini = L2_loss(train_x[i,:,:,:], train_x_hat)
train_loss = train_loss + train_loss_mini
t = time.time()
train_loss_mini.backward() ; optim.step() ; optim.zero_grad()
backward_time = time.time() - t
# run test data
test_x_hat = odeint(model, test_x[i, 0, :, :], t_eval, method=args.solver)
test_loss_mini = L2_loss(test_x[i,:,:,:], test_x_hat)
test_loss = test_loss + test_loss_mini
# logging
stats['train_loss'].append(train_loss.item())
stats['test_loss'].append(test_loss.item())
stats['forward_time'].append(forward_time)
stats['backward_time'].append(backward_time)
stats['nfe'].append(model.nfe)
if args.verbose and step % args.print_every == 0:
print("step {}, train_loss {:.4e}, test_loss {:.4e}".format(step, train_loss.item(), test_loss.item()))
# calculate loss mean and std for each traj.
train_x, t_eval = data['x'], data['t']
test_x, t_eval = data['test_x'], data['t']
train_x = torch.tensor(train_x, requires_grad=True, dtype=torch.float32).to(device)
test_x = torch.tensor(test_x, requires_grad=True, dtype=torch.float32).to(device)
t_eval = torch.tensor(t_eval, requires_grad=True, dtype=torch.float32).to(device)
train_loss = []
test_loss = []
for i in range(train_x.shape[0]):
train_x_hat = odeint(model, train_x[i, 0, :, :], t_eval, method=args.solver)
train_loss.append((train_x[i,:,:,:] - train_x_hat)**2)
# run test data
test_x_hat = odeint(model, test_x[i, 0, :, :], t_eval, method=args.solver)
test_loss.append((test_x[i,:,:,:] - test_x_hat)**2)
train_loss = torch.cat(train_loss, dim=1)
train_loss_per_traj = torch.sum(train_loss, dim=(0,2))
test_loss = torch.cat(test_loss, dim=1)
test_loss_per_traj = torch.sum(test_loss, dim=(0,2))
print('Final trajectory train loss {:.4e} +/- {:.4e}\nFinal trajectory test loss {:.4e} +/- {:.4e}'
.format(train_loss_per_traj.mean().item(), train_loss_per_traj.std().item(),
test_loss_per_traj.mean().item(), test_loss_per_traj.std().item()))
stats['traj_train_loss'] = train_loss_per_traj.detach().cpu().numpy()
stats['traj_test_loss'] = test_loss_per_traj.detach().cpu().numpy()
return model, stats
def train(args):
# set random seed
torch.manual_seed(args.seed)
np.random.seed(args.seed)
# init model and optimizer
if args.verbose:
print("Training baseline model:" if args.baseline else "Training HNN model:")
output_dim = args.input_dim if args.baseline else 2
nn_model = MLP(args.input_dim, 400, output_dim, args.nonlinearity)
model = HNN(args.input_dim, differentiable_model=nn_model,
field_type=args.field_type, baseline=args.baseline)
optim = torch.optim.Adam(model.parameters(), args.learn_rate, weight_decay=1e-4)
# the data API is different
# make sure it is a fair comparison
# generate the data the same way as in the SymODEN
# compute the time derivative based on the generated data
us = [0.0]
data = get_dataset(seed=args.seed, save_dir=args.save_dir,
rad=args.rad, us=us, samples=50, timesteps=45)
# arrange data
train_x, t_eval = data['x'][0,:,:,0:2], data['t']
test_x, t_eval = data['test_x'][0,:,:,0:2], data['t']
train_dxdt = (train_x[1:,:,:] - train_x[:-1,:,:]) / (t_eval[1] - t_eval[0])
test_dxdt = (test_x[1:,:,:] - test_x[:-1,:,:]) / (t_eval[1] - t_eval[0])
train_x = train_x[0:-1,:,:].reshape((-1,2))
test_x = test_x[0:-1,:,:].reshape((-1,2))
test_dxdt = test_dxdt.reshape((-1,2))
train_dxdt = train_dxdt.reshape((-1,2))
x = torch.tensor( train_x, requires_grad=True, dtype=torch.float32)
test_x = torch.tensor( test_x, requires_grad=True, dtype=torch.float32)
dxdt = torch.Tensor(train_dxdt)
test_dxdt = torch.Tensor(test_dxdt)
# vanilla train loop
stats = {'train_loss': [], 'test_loss': []}
for step in range(args.total_steps+1):
# train step
dxdt_hat = model.rk4_time_derivative(x) if args.use_rk4 else model.time_derivative(x)
loss = L2_loss(dxdt, dxdt_hat)
loss.backward() ; optim.step() ; optim.zero_grad()
# run test data
test_dxdt_hat = model.rk4_time_derivative(test_x) if args.use_rk4 else model.time_derivative(test_x)
test_loss = L2_loss(test_dxdt, test_dxdt_hat)
# logging
stats['train_loss'].append(loss.item())
stats['test_loss'].append(test_loss.item())
if args.verbose and step % args.print_every == 0:
print("step {}, train_loss {:.4e}, test_loss {:.4e}".format(step, loss.item(), test_loss.item()))
train_dxdt_hat = model.time_derivative(x)
train_dist = (dxdt - train_dxdt_hat)**2
test_dxdt_hat = model.time_derivative(test_x)
test_dist = (test_dxdt - test_dxdt_hat)**2
print('Final train loss {:.4e} +/- {:.4e}\nFinal test loss {:.4e} +/- {:.4e}'
.format(train_dist.mean().item(), train_dist.std().item()/np.sqrt(train_dist.shape[0]),
test_dist.mean().item(), test_dist.std().item()/np.sqrt(test_dist.shape[0])))
return model, stats
def train(args):
# set random seed
torch.manual_seed(args.seed)
np.random.seed(args.seed)
# init model and optimizer
if args.verbose:
print("Training baseline model:" if args.
baseline else "Training HNN model:")
output_dim = args.input_dim if args.baseline else 2
nn_model = MLP(args.input_dim, args.hidden_dim, output_dim,
args.nonlinearity)
model = HNN(args.input_dim,
differentiable_model=nn_model,
field_type=args.field_type,
baseline=args.baseline)
optim = torch.optim.Adam(model.parameters(),
args.learn_rate,
weight_decay=1e-4)
# arrange data
data = get_dataset(seed=args.seed)
x = torch.tensor(data['x'], requires_grad=True, dtype=torch.float32)
test_x = torch.tensor(data['test_x'],
requires_grad=True,
dtype=torch.float32)
dxdt = torch.Tensor(data['dx'])
test_dxdt = torch.Tensor(data['test_dx'])
# vanilla train loop
stats = {'train_loss': [], 'test_loss': []}
for step in range(args.total_steps + 1):
# train step
dxdt_hat = model.rk4_time_derivative(
x) if args.use_rk4 else model.time_derivative(x)
loss = L2_loss(dxdt, dxdt_hat)
loss.backward()
optim.step()
optim.zero_grad()
# run test data
test_dxdt_hat = model.rk4_time_derivative(
test_x) if args.use_rk4 else model.time_derivative(test_x)
test_loss = L2_loss(test_dxdt, test_dxdt_hat)
# logging
stats['train_loss'].append(loss.item())
stats['test_loss'].append(test_loss.item())
if args.verbose and step % args.print_every == 0:
print("step {}, train_loss {:.4e}, test_loss {:.4e}".format(
step, loss.item(), test_loss.item()))
train_dxdt_hat = model.time_derivative(x)
train_dist = (dxdt - train_dxdt_hat)**2
test_dxdt_hat = model.time_derivative(test_x)
test_dist = (test_dxdt - test_dxdt_hat)**2
print(
'Final train loss {:.4e} +/- {:.4e}\nFinal test loss {:.4e} +/- {:.4e}'
.format(train_dist.mean().item(),
train_dist.std().item() / np.sqrt(train_dist.shape[0]),
test_dist.mean().item(),
test_dist.std().item() / np.sqrt(test_dist.shape[0])))
return model, stats
def train(args):
if torch.cuda.is_available() and not args.cpu:
device = torch.device("cuda:0")
torch.set_default_tensor_type('torch.cuda.FloatTensor')
torch.cuda.empty_cache()
print("Running on the GPU")
else:
device = torch.device("cpu")
print("Running on the CPU")
# set random seed
torch.manual_seed(args.seed)
np.random.seed(args.seed)
print("{} {}".format(args.folder, args.speed))
print("Training scaled model:" if args.scaled else "Training noisy model:")
print('{} pairs of coords in latent space '.format(args.latent_dim))
#using universal autoencoder, pre-encode the training points
autoencoder = MLPAutoencoder(args.input_dim_ae,
args.hidden_dim,
args.latent_dim * 2,
nonlinearity='relu')
full_model = PixelHNN(args.latent_dim * 2,
args.hidden_dim,
autoencoder=autoencoder,
nonlinearity=args.nonlinearity,
baseline=args.baseline)
path = "{}/saved_models/{}.tar".format(args.save_dir, args.ae_path)
full_model.load_state_dict(torch.load(path))
full_model.eval()
autoencoder_model = full_model.autoencoder
# get dataset (no test data for now)
data = get_dataset(args.folder,
args.speed,
scaled=args.scaled,
split=args.split_data,
experiment_dir=args.experiment_dir,
tensor=True)
gcoords = autoencoder_model.encode(data).cpu().detach().numpy()
x = torch.tensor(gcoords, dtype=torch.float, requires_grad=True)
dx_np = full_model.time_derivative(
torch.tensor(gcoords, dtype=torch.float,
requires_grad=True)).cpu().detach().numpy()
dx = torch.tensor(dx_np, dtype=torch.float)
nnmodel = MLP(args.input_dim, args.hidden_dim, args.output_dim)
model = HNN(2, nnmodel)
model.to(device)
optim = torch.optim.Adam(model.parameters(),
args.learn_rate,
weight_decay=args.weight_decay)
# vanilla ae train loop
stats = {'train_loss': [], 'test_loss': []}
for step in range(args.total_steps + 1):
# train step
ixs = torch.randperm(x.shape[0])[:args.batch_size]
x_train, dxdt = x[ixs].to(device), dx[ixs].to(device)
dxdt_hat = model.time_derivative(x_train)
loss = L2_loss(dxdt, dxdt_hat)
loss.backward()
optim.step()
optim.zero_grad()
stats['train_loss'].append(loss.item())
if step % args.print_every == 0:
print("step {}, train_loss {:.4e}".format(step, loss.item()))
# train_dist = hnn_ae_loss(x, x_next, model, return_scalar=False)
# print('Final train loss {:.4e} +/- {:.4e}'
# .format(train_dist.mean().item(), train_dist.std().item() / np.sqrt(train_dist.shape[0])))
return model
def train(args):
# set random seed
torch.manual_seed(args.seed)
np.random.seed(args.seed)
# init model and optimizer
if args.verbose:
print("Training baseline model:" if args.baseline else "Training HNN model:")
S_net = MLP(int(args.input_dim/2), 140, int(args.input_dim/2)**2, args.nonlinearity)
U_net = MLP(int(args.input_dim/2), 140, 1, args.nonlinearity)
model = Lagrangian(int(args.input_dim/2), S_net, U_net, dt=1e-3)
num_parm = get_model_parm_nums(model)
print('model contains {} parameters'.format(num_parm))
optim = torch.optim.Adam(model.parameters(), args.learn_rate, weight_decay=1e-4)
# arrange data
data = get_lag_dataset(seed=args.seed)
x = torch.tensor( data['x'], requires_grad=False, dtype=torch.float32)
# append zero control
u = torch.zeros_like(x[:,0]).unsqueeze(-1)
x = torch.cat((x, u), -1)
test_x = torch.tensor( data['test_x'], requires_grad=False, dtype=torch.float32)
# append zero control
test_x = torch.cat((test_x, u), -1)
dxdt = torch.Tensor(data['dx'])
test_dxdt = torch.Tensor(data['test_dx'])
# vanilla train loop
stats = {'train_loss': [], 'test_loss': []}
for step in range(args.total_steps+1):
# train step
dq, dp, du = model.time_derivative(x).split(1,1)
dxdt_hat = torch.cat((dq, dp), -1)
loss = L2_loss(dxdt, dxdt_hat)
loss.backward() ; optim.step() ; optim.zero_grad()
# run test data
dq_test, dp_test, du_test = model.time_derivative(test_x).split(1,1)
test_dxdt_hat = torch.cat((dq_test, dp_test), -1)
test_loss = L2_loss(test_dxdt, test_dxdt_hat)
# logging
stats['train_loss'].append(loss.item())
stats['test_loss'].append(test_loss.item())
if args.verbose and step % args.print_every == 0:
print("step {}, train_loss {:.4e}, test_loss {:.4e}".format(step, loss.item(), test_loss.item()))
train_dq, train_dp, train_du = model.time_derivative(x).split(1,1)
train_dxdt_hat = torch.cat((train_dq, train_dp), -1)
train_dist = (dxdt - train_dxdt_hat)**2
test_dq, test_dp, test_du = model.time_derivative(test_x).split(1,1)
test_dxdt_hat = torch.cat((test_dq, test_dp), -1)
test_dist = (test_dxdt - test_dxdt_hat)**2
print('Final train loss {:.4e} +/- {:.4e}\nFinal test loss {:.4e} +/- {:.4e}'
.format(train_dist.mean().item(), train_dist.std().item()/np.sqrt(train_dist.shape[0]),
test_dist.mean().item(), test_dist.std().item()/np.sqrt(test_dist.shape[0])))
return model, stats
def train(args):
device = torch.device(
'cuda:' + str(args.gpu) if torch.cuda.is_available() else 'cpu')
# reproducibility: set random seed
torch.manual_seed(args.seed)
np.random.seed(args.seed)
torch.backends.cudnn.deterministic = True
torch.backends.cudnn.benchmark = False
# init model and optimizer
if args.verbose:
print("Start training with num of points = {} and solver {}.".format(
args.num_points, args.solver))
if args.structure == False and args.baseline == True:
nn_model = MLP(args.input_dim, 600, args.input_dim, args.nonlinearity)
model = SymODEN_R(args.input_dim,
H_net=nn_model,
device=device,
baseline=True)
elif args.structure == False and args.baseline == False:
H_net = MLP(args.input_dim, 400, 1, args.nonlinearity)
g_net = MLP(int(args.input_dim / 2), 200, int(args.input_dim / 2))
model = SymODEN_R(args.input_dim,
H_net=H_net,
g_net=g_net,
device=device,
baseline=False)
elif args.structure == True and args.baseline == False:
M_net = MLP(int(args.input_dim / 2), 300, int(args.input_dim / 2))
V_net = MLP(int(args.input_dim / 2), 50, 1)
g_net = MLP(int(args.input_dim / 2), 200, int(args.input_dim / 2))
model = SymODEN_R(args.input_dim,
M_net=M_net,
V_net=V_net,
g_net=g_net,
device=device,
baseline=False,
structure=True)
else:
raise RuntimeError(
'argument *baseline* and *structure* cannot both be true')
num_parm = get_model_parm_nums(model)
print('model contains {} parameters'.format(num_parm))
optim = torch.optim.Adam(model.parameters(),
args.learn_rate,
weight_decay=1e-4)
data = get_dataset(seed=args.seed)
# modified to use the hnn stuff
x = torch.tensor(data['x'], requires_grad=True,
dtype=torch.float32) # [1125, 2] Bx2
# append zero control
u = torch.zeros_like(x[:, 0]).unsqueeze(-1)
x = torch.cat((x, u), -1)
test_x = torch.tensor(data['test_x'],
requires_grad=True,
dtype=torch.float32)
# append zero control
test_x = torch.cat((test_x, u), -1)
dxdt = torch.Tensor(data['dx']) # [1125, 2] Bx2
test_dxdt = torch.Tensor(data['test_dx'])
# training loop
stats = {'train_loss': [], 'test_loss': []}
for step in range(args.total_steps + 1):
# modified to match hnn
dq, dp, du = model.time_derivative(x).split(1, 1)
dxdt_hat = torch.cat((dq, dp), -1)
loss = L2_loss(dxdt, dxdt_hat)
loss.backward()
optim.step()
optim.zero_grad()
# run test data
dq_test, dp_test, du_test = model.time_derivative(test_x).split(1, 1)
test_dxdt_hat = torch.cat((dq_test, dp_test), -1)
test_loss = L2_loss(test_dxdt, test_dxdt_hat)
# logging
stats['train_loss'].append(loss.item())
stats['test_loss'].append(test_loss.item())
if args.verbose and step % args.print_every == 0:
print("step {}, train_loss {:.4e}, test_loss {:.4e}".format(
step, loss.item(), test_loss.item()))
train_dq, train_dp, train_du = model.time_derivative(x).split(1, 1)
train_dxdt_hat = torch.cat((train_dq, train_dp), -1)
train_dist = (dxdt - train_dxdt_hat)**2
test_dq, test_dp, test_du = model.time_derivative(test_x).split(1, 1)
test_dxdt_hat = torch.cat((test_dq, test_dp), -1)
test_dist = (test_dxdt - test_dxdt_hat)**2
print(
'Final train loss {:.4e} +/- {:.4e}\nFinal test loss {:.4e} +/- {:.4e}'
.format(train_dist.mean().item(),
train_dist.std().item() / np.sqrt(train_dist.shape[0]),
test_dist.mean().item(),
test_dist.std().item() / np.sqrt(test_dist.shape[0])))
return model, stats
def train(args):
if torch.cuda.is_available() and not args.cpu:
device = torch.device("cuda:0")
torch.set_default_tensor_type('torch.cuda.FloatTensor')
torch.cuda.empty_cache()
print("Running on the GPU")
else:
device = torch.device("cpu")
print("Running on the CPU")
# set random seed
torch.manual_seed(args.seed)
np.random.seed(args.seed)
# get dataset (no test data for now)
angular_velo, acc_1, acc_2, sound = get_dataset_split(
args.folder,
args.speed,
scaled=args.scaled,
experiment_dir=args.experiment_dir,
tensor=True)
sub_col = {
0: [angular_velo, 1, 'v'],
1: [acc_1, 3, 'a1'],
2: [acc_2, 3, 'a2'],
3: [sound, 1, 's']
}
col2use = sub_col[args.sub_columns][0]
# using universal autoencoder, pre-encode the training points
autoencoder = MLPAutoencoder(sub_col[args.sub_columns][1],
args.hidden_dim,
args.latent_dim * 2,
dropout_rate=args.dropout_rate_ae)
full_model = PixelHNN(args.latent_dim * 2,
args.hidden_dim,
autoencoder=autoencoder,
nonlinearity=args.nonlinearity,
baseline=args.baseline,
dropout_rate=args.dropout_rate)
path = "{}/saved_models/{}-{}.tar".format(args.save_dir, args.ae_path,
sub_col[args.sub_columns][2])
full_model.load_state_dict(torch.load(path))
full_model.eval()
autoencoder_model = full_model.autoencoder
gcoords = autoencoder_model.encode(col2use).cpu().detach().numpy()
x = torch.tensor(gcoords, dtype=torch.float, requires_grad=True)
dx_np = full_model.time_derivative(
torch.tensor(gcoords, dtype=torch.float,
requires_grad=True)).cpu().detach().numpy()
dx = torch.tensor(dx_np, dtype=torch.float)
nnmodel = MLP(args.input_dim, args.hidden_dim, args.output_dim)
model = HNN(2, nnmodel)
model.to(device)
optim = torch.optim.Adam(model.parameters(),
args.learn_rate,
weight_decay=args.weight_decay)
print("Data from {} {}, column: {}".format(args.folder, args.speed,
sub_col[args.sub_columns][2]))
# x = torch.tensor(col2use[:-1], dtype=torch.float)
# x_next = torch.tensor(col2use[1:], dtype=torch.float)
#
# autoencoder = MLPAutoencoder(sub_col[args.sub_columns][1], args.hidden_dim, args.latent_dim * 2, dropout_rate=args.dropout_rate_ae)
# model = PixelHNN(args.latent_dim * 2, args.hidden_dim,
# autoencoder=autoencoder, nonlinearity=args.nonlinearity, baseline=args.baseline, dropout_rate=args.dropout_rate)
# model.to(device)
# optim = torch.optim.Adam(model.parameters(), args.learn_rate, weight_decay=args.weight_decay)
# vanilla ae train loop
stats = {'train_loss': []}
for step in range(args.total_steps + 1):
# train step
ixs = torch.randperm(x.shape[0])[:args.batch_size]
x_train, dxdt = x[ixs].to(device), dx[ixs].to(device)
dxdt_hat = model.time_derivative(x_train)
loss = L2_loss(dxdt, dxdt_hat)
loss.backward()
optim.step()
optim.zero_grad()
stats['train_loss'].append(loss.item())
if step % args.print_every == 0:
print("step {}, train_loss {:.4e}".format(step, loss.item()))
# train_dist = hnn_ae_loss(x, x_next, model, return_scalar=False)
# print('Final train loss {:.4e} +/- {:.4e}'
# .format(train_dist.mean().item(), train_dist.std().item() / np.sqrt(train_dist.shape[0])))
return model