# transforms.RandomApply([transformx],p=0.15),
# transforms.RandomChoice([
# transforms.ColorJitter(brightness=(0.6,1.4)),
# transforms.RandomRotation(degrees=(-10,10),resample=3),
# transforms.RandomResizedCrop((112,112),scale=(0.9,1.0),ratio=(1.0,1.0),interpolation=3),
# ]),
transforms.ToTensor(),
transforms.Normalize(mean = RGB_MEAN,
std = RGB_STD),
])
dataset_train = datasets.ImageFolder(os.path.join(DATA_ROOT, train_name), train_transform)
# create a weighted random sampler to process imbalanced data
weights = make_weights_for_balanced_classes_old(dataset_train.imgs, len(dataset_train.classes))
weights = torch.DoubleTensor(weights)
sampler = torch.utils.data.sampler.WeightedRandomSampler(weights, len(weights))
train_loader = torch.utils.data.DataLoader(
dataset_train, batch_size = BATCH_SIZE, sampler = sampler, pin_memory = PIN_MEMORY,
num_workers = NUM_WORKERS, drop_last = DROP_LAST
)
# train_loader = torch.utils.data.DataLoader(
# dataset_train, batch_size = BATCH_SIZE, shuffle= True, pin_memory = PIN_MEMORY,
# num_workers = NUM_WORKERS, drop_last = DROP_LAST
# )
NUM_CLASS = len(train_loader.dataset.classes)
print("Number of Training Classes: {}".format(NUM_CLASS))
# lfw, cfp_ff, cfp_fp, agedb, calfw, cplfw, vgg2_fp, lfw_issame, cfp_ff_issame, cfp_fp_issame, agedb_issame, calfw_issame, cplfw_issame, vgg2_fp_issame = get_val_data(DATA_ROOT)
def test(model, dataloader, dataset, sub_goal_indexes, best_of_n=20):
model.eval()
total_dest_err = 0.
total_overall_err = 0.
for i, trajx in enumerate(dataloader):
x = trajx['src'][:, :, :2]
y = trajx['trg'][:, :, :2]
x = x - trajx['src'][:, -1:, :2]
y = y - trajx['src'][:, -1:, :2]
x *= args.data_scale
y *= args.data_scale
x = x.double().cuda()
y = y.double().cuda()
y = y.cpu().numpy()
x = x.view(-1, x.shape[1] * x.shape[2])
plan = y[:, sub_goal_indexes, :].reshape(y.shape[0], -1)
all_plan_errs = []
all_plans = []
for _ in range(best_of_n):
# dest_recon = model.forward(x, initial_pos, device=device)
# modes = torch.tensor(k % args.ny, device=device).long().repeat(batch_size)
plan_recon = model.forward(x, mask=None)
plan_recon = plan_recon.detach().cpu().numpy()
all_plans.append(plan_recon)
plan_err = np.linalg.norm(plan_recon - plan, axis=-1)
all_plan_errs.append(plan_err)
all_plan_errs = np.array(all_plan_errs)
all_plans = np.array(all_plans)
indices = np.argmin(all_plan_errs, axis=0)
best_plan = all_plans[indices, np.arange(x.shape[0]), :]
# FDE
best_dest_err = np.linalg.norm(best_plan[:, -2:] - plan[:, -2:],
axis=1).sum()
best_plan = torch.DoubleTensor(best_plan).cuda()
interpolated_future = model.predict(x, best_plan)
interpolated_future = interpolated_future.detach().cpu().numpy()
# ADE
predicted_future = np.reshape(interpolated_future,
(-1, args.future_length, 2))
overall_err = np.linalg.norm(y - predicted_future,
axis=-1).mean(axis=-1).sum()
overall_err /= args.data_scale
best_dest_err /= args.data_scale
total_overall_err += overall_err
total_dest_err += best_dest_err
total_overall_err /= len(dataset)
total_dest_err /= len(dataset)
return total_overall_err, total_dest_err
def drosenbrock(tensor):
x, y = tensor
return torch.DoubleTensor(
(-400 * x * (y - x**2) - 2 * (1 - x), 200 * (y - x**2)))
def main():
############## Hyperparameters ##############
update_timestep = 1 #TD use == 1 # update policy every n timesteps set for TD
K_epochs = 2 # update policy for K epochs lr太大会出现NAN?
eps_clip = 0.2
gamma = 0.9
episode = 0
training_stage = 65
sample_lr = [
0.0001, 0.00009, 0.00008, 0.00007, 0.00006, 0.00005, 0.00004, 0.00003,
0.00002, 0.00001, 0.000009, 0.000008, 0.000007, 0.000006, 0.000005,
0.000004, 0.000003, 0.000002, 0.000001
]
lr = 0.0001 #random_seed = None
state_dim = 3
action_dim = 1
#(self, state_dim, action_dim, lr, betas, gamma, K_epochs, eps_clip)
actor_path = os.getcwd()+'/PPO_Mixedinput_Navigation_Model/weight/ppo_TD3lstm_actor.pkl'
critic_path = os.getcwd()+'/PPO_Mixedinput_Navigation_Model/weight/ppo_TD3lstm_critic.pkl'
################ load ###################
if episode >training_stage : #50 100
try:
lr = sample_lr[int(episode //training_stage)]
except(IndexError):
lr = 0.000001#0.000001*(0.99 ** ((episode-1000) // 10))
ppo = PPO(state_dim, action_dim, lr, gamma, K_epochs, eps_clip)
if os.path.exists(actor_path):
ppo.actor.load_state_dict(torch.load(actor_path))
print('Actor Model loaded')
if os.path.exists(critic_path):
ppo.critic.load_state_dict(torch.load(critic_path))
print('Critic Model loaded')
print("Waiting for GAMA...")
################### initialization ########################
save_curve_pic = os.getcwd()+'/PPO_Mixedinput_Navigation_Model/result/PPO_3LSTM_loss_curve.png'
save_critic_loss = os.getcwd()+'/PPO_Mixedinput_Navigation_Model/training_data/PPO_TD3_critic_loss.csv'
save_reward = os.getcwd()+'/PPO_Mixedinput_Navigation_Model/training_data/PPO_TD3_reward.csv'
save_speed = os.path.abspath(os.curdir)+'/PPO_Mixedinput_Navigation_Model/training_data/AC_average_speed.csv'
save_NPC_speed = os.path.abspath(os.curdir)+'/PPO_Mixedinput_Navigation_Model/training_data/NPC_speed.csv'
reset()
memory = Memory()
advantages =0 #global value
loss = []
total_loss = []
rewards = []
total_rewards = []
average_speed = []
test = "GAMA"
state,reward,done,time_pass,over ,average_speed_NPC = GAMA_connect(test) #connect
#[real_speed/10, target_speed/10, elapsed_time_ratio, distance_left/100,distance_front_car/10,distance_behind_car/10,reward,done,over]
print("done:",done,"timepass:"******"state ",state)
#print("reward",reward)
average_speed.append(state[0])
rewards.append(reward)
memory.rewards.append(reward)
memory.is_terminals.append(done)
state = torch.DoubleTensor(state).reshape(1,state_dim).to(device)
state_img = generate_img()
tensor_cv = torch.from_numpy(np.transpose(state_img, (2, 0, 1))).double().to(device)
if len(memory.states_next) ==0:
for _ in range(3):
memory.states_next = memory.states
memory.states_next[2] = state
memory.states_img_next = memory.states_img
memory.states_img_next [2]= tensor_cv
else:
del memory.states_next[:1]
del memory.states_img_next[:1]
memory.states_next.append(state)
memory.states_img_next.append(tensor_cv)
loss_ = ppo.update(memory,lr,advantages,done)
loss.append(loss_)
del memory.logprobs[:]
del memory.rewards[:]
del memory.is_terminals[:]
#memory.clear_memory()
action = ppo.select_action(state,tensor_cv, memory)
send_to_GAMA([[1,float(action*10)]])
#print("acceleration ",float(action))
# 終わり
elif done == 1:
average_speed.append(state[0])
#先传后计算
print("state_last",state)
send_to_GAMA( [[1,0]] )
rewards.append(reward)
del memory.states_next[:1]
del memory.states_img_next[:1]
state = torch.DoubleTensor(state).reshape(1,state_dim).to(device) #转化成1行
memory.states_next.append(state)
state_img = generate_img()
tensor_cv = torch.from_numpy(np.transpose(state_img, (2, 0, 1))).double().to(device)
memory.states_img_next.append(tensor_cv)
memory.rewards.append(reward)
memory.is_terminals.append(done)
loss_ = ppo.update(memory,lr,advantages,done)
loss.append(loss_)
memory.clear_memory()
print("----------------------------------Net_Trained---------------------------------------")
print('--------------------------Iteration:',episode,'over--------------------------------')
episode += 1
loss_sum = sum(loss).cpu().detach().numpy()
total_loss.append(loss_sum)
total_reward = sum(rewards)
total_rewards.append(total_reward)
cross_loss_curve(loss_sum.squeeze(0),total_reward,save_curve_pic,save_critic_loss,save_reward,np.mean(average_speed)*10,save_speed,average_speed_NPC,save_NPC_speed)
rewards = []
loss = []
if episode >training_stage : #50 100
try:
lr = sample_lr[int(episode //training_stage)]
except(IndexError):
lr = 0.000003#0.000001*(0.99 ** ((episode-1000) // 10))
torch.save(ppo.actor.state_dict(),actor_path)
torch.save(ppo.critic.state_dict(),critic_path)
#最初の時
else:
print('Iteration:',episode,"lr",lr)
state = torch.DoubleTensor(state).reshape(1,state_dim).to(device)
state_img = generate_img() # numpy image: H x W x C (500, 500, 3) -> (3,500,500)
tensor_cv = torch.from_numpy(np.transpose(state_img, (2, 0, 1))).double().to(device) # np.transpose( xxx, (2, 0, 1)) torch image: C x H x W
action = ppo.select_action(state, tensor_cv,memory)
print("acceleration: ",action)
send_to_GAMA([[1,float(action*10)]])
state,reward,done,time_pass,over ,average_speed_NPC= GAMA_connect(test)
return None
def tensor(data, **kwargs):
"""Returns a torch.DoubleTensor, unlike torch.tensor which returns FloatTensor by default"""
return torch.DoubleTensor(data, **kwargs)
def greedy_matching():
global batch_size, test_data_t, test_data_f, model, optimizer, emb_indexer_inv, gt_mappings, all_metrics, direct_inputs, direct_targets
all_results = OrderedDict()
direct_inputs, direct_targets = [], []
with torch.no_grad():
all_pred = []
np.random.shuffle(test_data_t)
np.random.shuffle(test_data_f)
inputs_pos, targets_pos = generate_input(test_data_t, 1)
inputs_neg, targets_neg = generate_input(test_data_f, 0)
indices_pos = np.random.permutation(len(inputs_pos))
indices_neg = np.random.permutation(len(inputs_neg))
inputs_pos, targets_pos = inputs_pos[indices_pos], targets_pos[
indices_pos]
inputs_neg, targets_neg = inputs_neg[indices_neg], targets_neg[
indices_neg]
batch_size = min(batch_size, len(inputs_pos))
num_batches = int(ceil(len(inputs_pos) / batch_size))
batch_size_f = int(ceil(len(inputs_neg) / num_batches))
for batch_idx in range(num_batches):
batch_start = batch_idx * batch_size
batch_end = (batch_idx + 1) * batch_size
batch_start_f = batch_idx * batch_size_f
batch_end_f = (batch_idx + 1) * batch_size_f
inputs = np.concatenate((inputs_pos[batch_start:batch_end],
inputs_neg[batch_start_f:batch_end_f]))
targets = np.concatenate((targets_pos[batch_start:batch_end],
targets_neg[batch_start_f:batch_end_f]))
inp = inputs.transpose(1, 0, 2)
inp_elems = torch.LongTensor(inputs).to(device)
targ_elems = torch.DoubleTensor(targets)
outputs = model(inp_elems)
outputs = [el.item() for el in outputs]
targets = [True if el.item() else False for el in targets]
for idx, pred_elem in enumerate(outputs):
ent1 = emb_indexer_inv[inp[0][idx][0]]
ent2 = emb_indexer_inv[inp[1][idx][0]]
if (ent1, ent2) in all_results:
print("Error: ", ent1, ent2, "already present")
all_results[(ent1, ent2)] = (pred_elem, targets[idx])
direct_targets = [True if el else False for el in direct_targets]
print("Len (direct inputs): ", len(direct_inputs))
for idx, direct_input in enumerate(direct_inputs):
ent1 = emb_indexer_inv[direct_input[0]]
ent2 = emb_indexer_inv[direct_input[1]]
sim = cos_sim(emb_vals[direct_input[0]], emb_vals[direct_input[1]])
all_results[(ent1, ent2)] = (sim, direct_targets[idx])
optimum_metrics, opt_threshold = [-1000 for i in range(5)], -1000
low_threshold = np.min([el[0] for el in all_results.values()]) - 0.02
high_threshold = np.max([el[0] for el in all_results.values()]) + 0.02
threshold = low_threshold
step = 0.001
opt_fn, opt_fp = [], []
while threshold < high_threshold:
res = []
for i, key in enumerate(all_results):
if all_results[key][0] > threshold:
res.append(key)
fn_list = [(key, all_results[key][0]) for key in gt_mappings
if key not in set(res) and not is_valid(test_onto, key)]
fp_list = [(elem, all_results[elem][0]) for elem in res
if not all_results[elem][1]]
tp_list = [(elem, all_results[elem][0]) for elem in res
if all_results[elem][1]]
tp, fn, fp = len(tp_list), len(fn_list), len(fp_list)
exception = False
try:
precision = tp / (tp + fp)
recall = tp / (tp + fn)
f1score = 2 * precision * recall / (precision + recall)
f2score = 5 * precision * recall / (4 * precision + recall)
f0_5score = 1.25 * precision * recall / (0.25 * precision +
recall)
except Exception as e:
print(e)
exception = True
step = 0.001
threshold += step
continue
print("Threshold: ", threshold, precision, recall, f1score,
f2score, f0_5score)
if f1score > optimum_metrics[2]:
optimum_metrics = [
precision, recall, f1score, f2score, f0_5score
]
opt_threshold = threshold
opt_fn = fn_list
opt_fp = fp_list
if threshold > 0.98 and not exception:
step = 0.0001
else:
step = 0.001
print(step, threshold, exception)
threshold += step
print(
"Precision: {} Recall: {} F1-Score: {} F2-Score: {} F0.5-Score: {}"
.format(*optimum_metrics))
all_fn.extend(opt_fn)
all_fp.extend(opt_fp)
if optimum_metrics[2] != -1000:
all_metrics.append((opt_threshold, optimum_metrics))
return all_results
def __init__(self, weights, num_samples, replacement=True):
self.weights = torch.DoubleTensor(weights)
self.num_samples = num_samples
self.replacement = replacement
output_dim = 3
sigma_dim = 2
batch_size = 7
BBTT_steps = 10
input_dim = 5
nnLayerSizes = [20, 20, 20]
net_type = "LSTM"
# net_type = "FFNN"
input_ = torch.randn([batch_size, BBTT_steps + 1, input_dim])
net = NET(net_type, BBTT_steps, input_dim, nnLayerSizes, output_dim,
sigma_dim)
print(net)
output, sigma = net.forwardVector(input_)
input_ = torch.DoubleTensor(input_)
print("Input size KxTxD")
print(input_.size())
print("Output size KxN")
print(output.size())
print("Sigma size KxN")
print(sigma.size())
# if(PRINT_OUTPUT): print(dir(output))
# if(PRINT_OUTPUT): print(output.__dict__)
# if(PRINT_OUTPUT): print(output)
# if(PRINT_OUTPUT): print(output.detach().numpy())
def getDataloader(self):
trainTransform = A.Compose([
A.transforms.HorizontalFlip(p=0.5),
A.transforms.VerticalFlip(p=0.5),
A.transforms.Normalize(mean=(0.57, 0.94, 0.45),
std=(0.15, 0.17, 0.10)),
ToTensorV2()
])
testTransform = A.Compose([
A.transforms.Normalize(mean=(0.57, 0.94, 0.45),
std=(0.15, 0.17, 0.10)),
])
working_path = Path(os.getcwd())
working_path = working_path.parent.parent
dataPath = Path.joinpath(working_path, 'Data')
trainPath = Path.joinpath(dataPath, 'Train')
valPath = Path.joinpath(dataPath, 'Validation')
trainImgPath = Path.joinpath(trainPath, 'TrainImages')
valImgPath = Path.joinpath(valPath, 'ValidationImages')
testImgPath = Path.joinpath(dataPath, 'Test', 'TestImages')
trainLabels = str(trainPath) + '\\trainLbls.csv'
valLabels = str(valPath) + '\\valLbls.csv'
listTrainData = natsorted(
[os.path.join(trainImgPath, f) for f in os.listdir(trainImgPath)])
listTestData = natsorted(
[os.path.join(testImgPath, f) for f in os.listdir(testImgPath)])
listValData = natsorted(
[os.path.join(valImgPath, f) for f in os.listdir(valImgPath)])
albuTrainData = ImageDataset(imageList=listTrainData,
cvsFile=trainLabels,
transform=trainTransform)
albuTestData = ImageDataset(imageList=listTestData,
cvsFile=None,
transform=testTransform)
albuValData = ImageDataset(imageList=listValData,
cvsFile=valLabels,
transform=testTransform)
classes = pd.read_csv(trainLabels, header=None, names=['Labels'])
classes = classes['Labels'].unique()
balanced_weights = self.utils.create_weights_to_balance_classes(
albuTrainData.lbls.Labels - 1, len(classes))
balanced_weights = torch.DoubleTensor(balanced_weights)
balanced_sampler = torch.utils.data.sampler.WeightedRandomSampler(
balanced_weights, len(balanced_weights))
trainDataLoader = torch.utils.data.DataLoader(
albuTrainData,
batch_size=self.batch_size,
num_workers=self.num_workers,
sampler=balanced_sampler,
drop_last=True,
pin_memory=True)
testDataLoader = torch.utils.data.DataLoader(
albuTestData,
batch_size=self.batch_size,
num_workers=self.num_workers,
drop_last=True,
pin_memory=True)
valDataLoader = torch.utils.data.DataLoader(
albuValData,
batch_size=self.batch_size,
num_workers=self.num_workers,
drop_last=True,
pin_memory=True)
return classes, trainDataLoader, testDataLoader, valDataLoader
import odl
from odl.contrib.pytorch import TorchOperator
# --- Forward --- #
# Define ODL operator
matrix = np.array([[1, 2, 3], [4, 5, 6]], dtype=float)
odl_op = odl.MatrixOperator(matrix)
# Wrap as torch operator
torch_op = TorchOperator(odl_op)
# Define evaluation point and wrap into a variable. Mark as
# `requires_gradient`, otherwise `backward()` doesn't do anything.
# This is supported by the ODL wrapper.
x = torch.DoubleTensor([1, 1, 1])
x_var = torch.autograd.Variable(x, requires_grad=True)
# Evaluate torch operator
res_var = torch_op(x_var)
# ODL result
odl_res = odl_op(x.numpy())
print('pytorch result: ', res_var.data.numpy())
print('ODL result : ', odl_res.asarray())
# --- Gradient (backward) --- #
# Define ODL cost functional
odl_cost = odl.solvers.L2NormSquared(odl_op.range)
self.k = k
self.DOF = DOF
def get_f(self, x):
f = self.k * x**3
return f
""" 主程序"""
if __name__ == "__main__":
Nt = 2048
Nh = 5
Nf = 2
K = torch.DoubleTensor([[3.0e4, -1.5e4], [-1.5e4, 1.5e4]])
M = torch.DoubleTensor([[1, 0], [0, 1]])
C = torch.DoubleTensor([[2, -1], [-1, 1]])
model = Model(M, C, K, Nh)
force_nl = Force_gap(DOF=(0, ), g=0.0001, k=1e7)
# force_nl = Force_cubic(DOF=(0, ), k=5e9)
force_ex = Force_EX(DOF=(0, ), Nh=Nh, Nf=Nf)
aft = AFT.AFT(Nh=Nh, Nt=Nt, Nf=Nf, force=force_nl)
w1 = (50 + 0) * 2 * np.pi
w2 = (50 + 1) * 2 * np.pi
N_solver = Solver.Newton_Solver(model=model,
force_ex=force_ex,
def inner_objective(f, seed):
torch.manual_seed(seed)
batch_size = 500
x = torch.DoubleTensor(batch_size, 2).uniform_(-5, 5)
y = f(x)
return torch.nn.MSELoss()(f_target(x), y[:, 0])
beta2,
group["lr"],
group["weight_decay"],
group["eps"],
)
return loss
if __name__ == "__main__":
def rosenbrock(tensor):
x, y = tensor
return (1 - x)**2 + 100 * (y - x**2)**2
def drosenbrock(tensor):
x, y = tensor
return torch.DoubleTensor(
(-400 * x * (y - x**2) - 2 * (1 - x), 200 * (y - x**2)))
params = torch.DoubleTensor((1.5, 1.5))
gradclipadam = GradClipAdam(params, lr=1e-4)
gradclipadam.zero_grad()
y_hat = rosenbrock(params)
y_hat.backward()
gradclipadam.step()
# for i in range(100):
# algorithm(lambda x: (rosenbrock(x), drosenbrock(x)), params, config)
# print("{:.8f}\t{:.8f}\t".format(params[0], params[1]))
# This is a test among maxpooling, convolution and RRSVM
from torch.autograd import Variable
import torch
from torch.autograd import gradcheck
import numpy as np
import torch
import torch.nn.functional as F
if __name__ == '__main__':
kernel_size = 2
n_channel = 100
feature_size = 6
batch_size = 3
input = (Variable(torch.DoubleTensor(torch.randn(1, n_channel, feature_size, feature_size).double()), requires_grad=True), kernel_size)
f_max_pooling = F.max_pool2d
print gradcheck(f_max_pooling, input, eps=1e-3)
model.eval()
seq = test_input
future = []
with torch.no_grad():
model.hidden_cell = (torch.zeros(1, 1, model.hidden_layer_size).double().to(device),
torch.zeros(1, 1, model.hidden_layer_size).double().to(device))
model.hidden_cell_2 = (torch.zeros(1, 1, model.input_size).double().to(device),
torch.zeros(1, 1, model.input_size).double().to(device))
for i in range(nfuture):
y = model(seq.double().to(device)).detach().cpu().numpy()
future.append(y[-1])
seq = seq.detach().cpu().tolist()
seq.append(y[-1])
seq.pop(0)
seq = torch.DoubleTensor(seq)
y = np.array(future)
x = [test_input_df[t_].max() + datetime.timedelta(hours=i) for i in range(len(future))]
assert len(x)==len(y)
"""
Plotting
"""
import plotly.graph_objects as go
fig = go.Figure()
fig.add_trace(go.Scatter(
y=df[T_].to_numpy(),
x=timestamp,
mode='lines+markers',
def count_parameters(m, x, y):
total_params = 0
for p in m.parameters():
total_params += torch.DoubleTensor([p.numel()])
m.total_params[0] = total_params
num_batches = int(ceil(len(inputs_pos) / batch_size))
batch_size_f = int(ceil(len(inputs_neg) / num_batches))
for batch_idx in range(num_batches):
batch_start = batch_idx * batch_size
batch_end = (batch_idx + 1) * batch_size
batch_start_f = batch_idx * batch_size_f
batch_end_f = (batch_idx + 1) * batch_size_f
inputs = np.concatenate((inputs_pos[batch_start:batch_end],
inputs_neg[batch_start_f:batch_end_f]))
targets = np.concatenate((targets_pos[batch_start:batch_end],
targets_neg[batch_start_f:batch_end_f]))
inp_elems = torch.LongTensor(inputs).to(device)
targ_elems = torch.DoubleTensor(targets).to(device)
optimizer.zero_grad()
outputs = model(inp_elems)
loss = F.mse_loss(outputs, targ_elems)
loss.backward()
optimizer.step()
if batch_idx % 500 == 0:
print("Epoch: {} Idx: {} Loss: {}".format(
epoch, batch_idx, loss.item()))
model.eval()
print("Trained value of W: ", model.output.weight)
torch.save(model.state_dict(), "/u/vivek98/attention.pt")
test_data_t = [key for key in test_data if test_data[key]]
from models.value_network import Value
print(gym.__version__)
use_gpu = torch.cuda.is_available()
Tensor = torch.cuda.DoubleTensor if use_gpu else torch.DoubleTensor
LongTensor = torch.cuda.LongTensor if use_gpu else torch.LongTensor
if use_gpu:
torch.set_default_tensor_type('torch.cuda.DoubleTensor')
PI = torch.cuda.DoubleTensor([3.1415926])
else:
torch.set_default_tensor_type('torch.DoubleTensor')
PI = torch.DoubleTensor([3.1415926])
env_name = 'Humanoid-v2'
render = False
log_std = 0
gamma = 0.99
tau = 0.95
l2_reg = 1e-3
learning_rate = 0.0003
clip_epsilon = 0.2 # for PPO
seed = 1
min_batch_size = 10000
total_iterations = 1000
log_interval = 1
optim_epochs = 5
optim_batch_size = 64
def conjugate_gradiant(A_grad, b, model, is_cuda, hessian=None):
x = torch.DoubleTensor(np.zeros((b.size(0), b.size(1))) + 1)
b_norm = max(b.norm(), 1e-20)
k = 0
if hessian is not None:
I = torch.DoubleTensor(np.identity(hessian.data.size(0)) * 1e-4)
if is_cuda:
I = I.cuda()
hessian += I
if is_cuda:
x = x.cuda()
if hessian is not None:
r_i = b - hessian @ x
else:
r_i = b - hessian_vector_product(A_grad, x, model, k)
d_i = r_i.clone()
tolerance = 1e-3
i = 0
no_update_i = 0
max_iter = 5000
print('i:{}, r norm:{}'.format(i, r_i.norm()))
min_r = (b - hessian_vector_product(A_grad, x, model, 0)).norm()
best_x = x.clone()
while min_r / b_norm > tolerance:
rr_i = r_i.permute(1, 0) @ r_i
if hessian is not None:
A_d = hessian @ d_i
else:
A_d = hessian_vector_product(A_grad, d_i, model, k)
alpha = rr_i / (d_i.permute(1, 0) @ A_d)
x += alpha * d_i
r_i -= alpha * A_d
rr_i_new = r_i.permute(1, 0) @ r_i
d_i = r_i + (rr_i_new / rr_i) * d_i
i += 1
no_update_i += 1
if rr_i_new.sqrt() < min_r:
r = b - hessian_vector_product(A_grad, x, model, 0)
if r.norm() < min_r:
min_r = r.norm()
best_x = x.clone()
no_update_i = 0
if i % 100 == 0:
print(
'i:{}, r norm:{}, min r norm:{}, releative: {} same best: {}'.
format(i, rr_i_new.sqrt(), min_r, min_r / b_norm, no_update_i))
if i == max_iter or no_update_i == max_iter / 10:
if min_r / b_norm > 1:
if max_iter > 10000:
break
max_iter += 1000
k = b_norm * 4 * 1e-3
x = best_x.clone()
r_i = b - hessian_vector_product(A_grad, x, model, k)
d_i = r_i.clone()
else:
break
r = b - hessian_vector_product(A_grad, best_x, model, 0)
print(r.norm())
print(b_norm)
best_x = best_x.cpu()
return best_x, i
def _number_format(tensor, min_sz=-1):
min_sz = max(min_sz, 2)
tensor = torch.DoubleTensor(tensor.nelement()).copy_(tensor).abs_()
pos_inf_mask = tensor.eq(float('inf'))
neg_inf_mask = tensor.eq(float('-inf'))
nan_mask = tensor.ne(tensor)
invalid_value_mask = pos_inf_mask + neg_inf_mask + nan_mask
if invalid_value_mask.all():
example_value = 0
else:
example_value = tensor[invalid_value_mask.eq(0)][0]
tensor[invalid_value_mask] = example_value
if invalid_value_mask.any():
min_sz = max(min_sz, 3)
int_mode = True
# TODO: use fmod?
for value in tensor:
if value != math.ceil(value):
int_mode = False
break
exp_min = tensor.min()
if exp_min != 0:
exp_min = math.floor(math.log10(exp_min)) + 1
else:
exp_min = 1
exp_max = tensor.max()
if exp_max != 0:
exp_max = math.floor(math.log10(exp_max)) + 1
else:
exp_max = 1
scale = 1
exp_max = int(exp_max)
prec = PRINT_OPTS.precision
if int_mode:
if exp_max > prec + 1:
format = '{{:11.{}e}}'.format(prec)
sz = max(min_sz, 7 + prec)
else:
sz = max(min_sz, exp_max + 1)
format = '{:' + str(sz) + '.0f}'
else:
if exp_max - exp_min > prec:
sz = 7 + prec
if abs(exp_max) > 99 or abs(exp_min) > 99:
sz = sz + 1
sz = max(min_sz, sz)
format = '{{:{}.{}e}}'.format(sz, prec)
else:
if exp_max > prec + 1 or exp_max < 0:
sz = max(min_sz, 7)
scale = math.pow(10, exp_max - 1)
else:
if exp_max == 0:
sz = 7
else:
sz = exp_max + 6
sz = max(min_sz, sz)
format = '{{:{}.{}f}}'.format(sz, prec)
return format, scale, sz
def reset(self):
self.scores = torch.DoubleTensor(torch.DoubleStorage()).numpy()
self.targets = torch.LongTensor(torch.LongStorage()).numpy()
def testGrad(self):
def make_world(learned_force):
bodies = []
joints = []
target = Circle([500, 300], 30)
bodies.append(target)
c1 = Circle([250, 210], 30)
bodies.append(c1)
c1.add_force(ExternalForce(learned_force))
c1.add_no_collision(target)
c2 = Circle([400, 250], 30)
bodies.append(c2)
c2.add_no_collision(target)
world = World(bodies, joints, dt=DT)
return world, c2, target
initial_force = torch.DoubleTensor([0, 3, 0])
initial_force[2] = 0
initial_force = Variable(initial_force, requires_grad=True)
# Initial demo
learned_force = lambda t: initial_force if t < 0.1 else ExternalForce.ZEROS
# learned_force = gravity
world, c, target = make_world(learned_force)
# initial_state = world.save_state()
# next_fric_coeff = Variable(torch.DoubleTensor([1e-7]), requires_grad=True)
# c.fric_coeff = next_fric_coeff
# initial_state = world.save_state()
run_world(world, run_time=TIME, screen=None)
learning_rate = 0.001
max_iter = 100
dist_hist = []
last_dist = 1e10
for i in range(max_iter):
learned_force = lambda t: initial_force if t < 0.1 else ExternalForce.ZEROS
world, c, target = make_world(learned_force)
# world.load_state(initial_state)
# world.reset_engine()
# c = world.bodies[0]
# c.fric_coeff = next_fric_coeff
run_world(world, run_time=TIME, screen=None)
dist = (target.pos - c.pos).norm()
dist.backward()
grad = initial_force.grad.data
# grad.clamp_(-10, 10)
initial_force = Variable(initial_force.data - learning_rate * grad,
requires_grad=True)
# grad = c.fric_coeff.grad.data
# grad.clamp_(-10, 10)
# temp = c.fric_coeff.data - learning_rate * grad
# temp.clamp_(1e-7, 1)
learning_rate /= 1.1
# next_fric_coeff = Variable(temp, requires_grad=True)
# print(next_fric_coeff)
if abs((last_dist - dist).data[0]) < 1e-5:
break
last_dist = dist
dist_hist.append(dist)
world = make_world(learned_force)[0]
# c.fric_coeff = next_fric_coeff
# world.load_state(initial_state)
# world.reset_engine()
run_world(world, run_time=TIME, screen=None, recorder=None)
dist = (target.pos - c.pos).norm()
def as_tensor(*args, **kwargs):
r = tuple(
torch.DoubleTensor(arg, **kwargs) if arg is not None else None
for arg in args)
return r if len(r) > 1 else r[0]
def testInference(self):
def make_world(forces, mass):
bodies = []
joints = []
# make chain of rectangles
r = Rect([300, 50], [20, 60])
bodies.append(r)
joints.append(Joint(r, None, [300, 30]))
for i in range(1, 10):
if i < 9:
r = Rect([300, 50 + 50 * i], [20, 60])
else:
r = Rect([300, 50 + 50 * i], [20, 60], mass=mass)
bodies.append(r)
joints.append(Joint(bodies[-1], bodies[-2],
[300, 25 + 50 * i]))
bodies[-1].add_no_collision(bodies[-2])
bodies[-1].add_force(ExternalForce(gravity, multiplier=100))
# make projectile
m = 13
c1 = Circle([50, 500], 20)
bodies.append(c1)
for f in forces:
c1.add_force(ExternalForce(f, multiplier=100 * m))
world = World(bodies, joints, dt=DT)
return world, r
def positions_run_world(world,
dt=Params.DEFAULT_DT,
run_time=10,
screen=None,
recorder=None):
positions = [torch.cat([b.p for b in world.bodies])]
while world.t < run_time:
world.step()
positions.append(torch.cat([b.p for b in world.bodies]))
return positions
MASS_EPS = 1e-7
forces = [hor_impulse]
ground_truth_mass = Variable(torch.DoubleTensor([7]))
world, c = make_world(forces, ground_truth_mass)
ground_truth_pos = positions_run_world(world,
run_time=10,
screen=None,
recorder=None)
ground_truth_pos = [p.data for p in ground_truth_pos]
ground_truth_pos = Variable(torch.cat(ground_truth_pos))
learning_rate = 0.01
max_iter = 100
next_mass = Variable(torch.DoubleTensor([1.3]), requires_grad=True)
loss_hist = []
mass_hist = [next_mass]
last_dist = 1e10
for i in range(max_iter):
print(i, end='\r')
world, c = make_world(forces, next_mass)
# world.load_state(initial_state)
# world.reset_engine()
positions = positions_run_world(world, run_time=10, screen=None)
positions = torch.cat(positions)
positions = positions[:len(ground_truth_pos)]
# temp_ground_truth_pos = ground_truth_pos[:len(positions)]
loss = torch.nn.MSELoss()(positions, ground_truth_pos)
loss.backward()
grad = c.mass.grad.data
# clip gradient
grad = torch.max(torch.min(grad, torch.DoubleTensor([100])),
torch.DoubleTensor([-100]))
temp = c.mass.data - learning_rate * grad
temp = max(MASS_EPS, temp[0])
next_mass = Variable(torch.DoubleTensor([temp]),
requires_grad=True)
# learning_rate /= 1.1
if abs((last_dist - loss).data[0]) < 1e-3:
break
last_dist = loss
loss_hist.append(loss)
mass_hist.append(next_mass)
assert abs(next_mass.data[0] - ground_truth_mass.data[0]) <= 1e-1
def forward(self, x, dest=None, mask=None, iteration=1, y=None):
ftraj = self.encoder_past(x)
if mask:
for _ in range(self.nonlocal_pools):
ftraj = self.non_local_social_pooling(ftraj, mask)
if self.training:
pcd = True if len(
self.replay_memory) == args.memory_size else False
if pcd:
z_e_0 = self.replay_memory.sample(
n=ftraj.size(0)).clone().detach().cuda()
else:
z_e_0 = sample_p_0(n=ftraj.size(0), nz=self.zdim)
z_e_k, _ = self.sample_langevin_prior_z(Variable(z_e_0),
ftraj,
pcd=pcd,
verbose=(iteration %
1000 == 0))
for _z_e_k in z_e_k.clone().detach().cpu().split(1):
self.replay_memory.push(_z_e_k)
else:
z_e_0 = sample_p_0(n=ftraj.size(0), nz=self.zdim)
z_e_k, _ = self.sample_langevin_prior_z(Variable(z_e_0),
ftraj,
pcd=False,
verbose=(iteration %
1000 == 0),
y=y)
z_e_k = z_e_k.double().cuda()
if self.training:
dest_features = self.encoder_dest(dest)
features = torch.cat((ftraj, dest_features), dim=1)
latent = self.encoder_latent(features)
mu = latent[:, 0:self.zdim]
logvar = latent[:, self.zdim:]
var = logvar.mul(0.5).exp_()
eps = torch.DoubleTensor(var.size()).normal_().cuda()
z_g_k = eps.mul(var).add_(mu)
z_g_k = z_g_k.double().cuda()
if self.training:
decoder_input = torch.cat((ftraj, z_g_k), dim=1)
else:
decoder_input = torch.cat((ftraj, z_e_k), dim=1)
generated_dest = self.decoder(decoder_input)
if self.training:
generated_dest_features = self.encoder_dest(generated_dest)
prediction_features = torch.cat(
(ftraj, generated_dest_features), dim=1)
pred_future = self.predictor(prediction_features)
en_pos = self.ebm(z_g_k, ftraj).mean()
en_neg = self.ebm(z_e_k.detach().clone(), ftraj).mean()
cd = en_pos - en_neg
return generated_dest, mu, logvar, pred_future, cd, en_pos, en_neg, pcd
return generated_dest
"""
import torch
a = torch.tensor([2,2,4])
b = torch.tensor([[1,2,3],[2,3,4],[3,4,5],[4,5,6]])
print(a)
print(b)
print(a.shape)
print(b.shape)
print(a.size())
print(b.size())
c = torch.FloatTensor([[1,2,3],[2,3,4],[3,4,5],[4,5,6]])
d = torch.DoubleTensor([[1,2,3],[2,3,4],[3,4,5],[4,5,6]])
print(c)
print(c.type)
print(d)
print(d.type)
print(c.mean())
print(d.mean())
print(c.std())
print(d.std())
print("Reshape")
print(b.view(-1,1))
print(b.view(12))
print(b.view(-1,4))
image_datasets = {
x: datasets.ImageFolder(
os.path.join(tar_extract_path, tar_name, fold_lst[fold_num], x),
data_transforms[x])
for x in ['train', 'val']
}
weights_dict = {
x: make_weights_for_balanced_classes(image_datasets[x].imgs,
len(image_datasets[x].classes))
for x in ['train', 'val']
}
sampler_dict = {
x: torch.utils.data.sampler.WeightedRandomSampler(
torch.DoubleTensor(weights_dict[x]),
len(torch.DoubleTensor(weights_dict[x])))
for x in ['train', 'val']
}
dataloaders_dict_sampler = {
x: torch.utils.data.DataLoader(image_datasets[x],
batch_size=batch_size,
shuffle=False,
sampler=sampler_dict[x],
num_workers=8)
for x in ['train', 'val']
}
dataloaders_dict = {
x: torch.utils.data.DataLoader(image_datasets[x],
offset = 0
n = 40336
n=20000
split = 0.8
ind = list(range(n))
div = int(n * split)
partition = dict([])
partition['train'] = list(ind[:div])
partition['validation'] = list(ind[div:n])
'''TCN'''
model = TCN(40, 1, args.layers, fcl=args.fcl, kernel_size=2, dropout=args.drop).to(args.device)
criterion = nn.BCEWithLogitsLoss(pos_weight=torch.DoubleTensor([1.8224]).to(args.device), reduction='none')
optimizer = optim.SGD(model.parameters(), lr=args.lr, momentum=args.momentum)
scheduler = optim.lr_scheduler.ReduceLROnPlateau(optimizer, verbose=True, factor=0.1, patience=5)
if load_model:
model.load_state_dict(torch.load(load_model))
train_losses = np.concatenate((np.load(load_folder +'/train_losses.npy'), np.zeros(epochs)))
train_pos_acc = np.concatenate((np.load(load_folder +'/train_pos_acc.npy'), np.zeros(epochs)))
train_neg_acc = np.concatenate((np.load(load_folder +'/train_neg_acc.npy'), np.zeros(epochs)))
val_losses = np.concatenate((np.load(load_folder +'/val_losses.npy'), np.zeros(epochs)))
val_pos_acc = np.concatenate((np.load(load_folder +'/val_pos_acc.npy'), np.zeros(epochs)))
val_neg_acc = np.concatenate((np.load(load_folder +'/val_neg_acc.npy'), np.zeros(epochs)))
utility = np.concatenate((np.load(load_folder +'/utility.npy'), np.zeros(epochs)))
plotter(model_name, utility, train_losses, train_pos_acc, train_neg_acc,
val_losses, val_pos_acc, val_neg_acc)
else:
def count_matmul(m, x, y):
num_mul = x[0].numel() * x[1].size(-1)
# m.total_ops += torch.DoubleTensor([int(num_mul)])
m.total_ops += torch.DoubleTensor([int(0)])
n_steps = 40
tdim = 5
sdim = 10
vol_floor = torch.zeros((sdim, tdim)).double()
vol = torch.tensor([[0.2] * tdim] * sdim).double()
m = tdim * sdim
n = 11
r = 0.02
###------------------------- Calculating Data -------------------------###
iv = torch.zeros(n)
spot = float(35.)
expiries = np.array([0.25] * n)
strike = np.linspace(30., 40., n).astype(float)
vals = torch.DoubleTensor(f.Bachelier(spot, 7., expiries, strike, r))
for i in range(0, vals.size()[0]):
iv[i] = f.impliedVol(spot, vals[i], expiries[i], strike[i], r)
###------------------------- ADAM Parameters -------------------------###
expiries = torch.tensor(expiries)
beta1 = 0.9
beta2 = 0.999
mt = 0.0
vt = 0.0
vhat = torch.zeros((sdim, tdim)).double()
LearningRate = 0.005
#