def _get_random_data(n, **tkwargs):
train_x1 = torch.linspace(0, 0.95, n + 1, **tkwargs) + 0.05 * torch.rand(
n + 1, **tkwargs
)
train_x2 = torch.linspace(0, 0.95, n, **tkwargs) + 0.05 * torch.rand(n, **tkwargs)
train_y1 = torch.sin(train_x1 * (2 * math.pi)) + 0.2 * torch.randn_like(train_x1)
train_y2 = torch.cos(train_x2 * (2 * math.pi)) + 0.2 * torch.randn_like(train_x2)
return train_x1.unsqueeze(-1), train_x2.unsqueeze(-1), train_y1, train_y2
def forward(self, variableInput, variableFlow): if hasattr(self, 'tensorGrid') == False or self.tensorGrid.size(0) != variableInput.size(0) or self.tensorGrid.size(2) != variableInput.size(2) or self.tensorGrid.size(3) != variableInput.size(3): torchHorizontal = torch.linspace(-1.0, 1.0, variableInput.size(3)).view(1, 1, 1, variableInput.size(3)).expand(variableInput.size(0), 1, variableInput.size(2), variableInput.size(3)) torchVertical = torch.linspace(-1.0, 1.0, variableInput.size(2)).view(1, 1, variableInput.size(2), 1).expand(variableInput.size(0), 1, variableInput.size(2), variableInput.size(3)) self.tensorGrid = torch.cat([ torchHorizontal, torchVertical ], 1).cuda() # end variableFlow = torch.cat([ variableFlow[:, 0:1, :, :] / ((variableInput.size(3) - 1.0) / 2.0), variableFlow[:, 1:2, :, :] / ((variableInput.size(2) - 1.0) / 2.0) ], 1) variableGrid = torch.autograd.Variable(data=self.tensorGrid, volatile=not self.training) + variableFlow return torch.nn.functional.grid_sample(input=variableInput, grid=variableGrid.clamp(-1.0, 1.0).permute(0, 2, 3, 1), mode='bilinear')
def _setUp(self, double=False, cuda=False):
device = torch.device("cuda") if cuda else torch.device("cpu")
dtype = torch.double if double else torch.float
train_x = torch.linspace(0, 1, 10, device=device, dtype=dtype).unsqueeze(-1)
train_y = torch.sin(train_x * (2 * math.pi)).squeeze(-1)
train_yvar = torch.tensor(0.1 ** 2, device=device)
noise = torch.tensor(NOISE, device=device, dtype=dtype)
self.train_x = train_x
self.train_y = train_y + noise
self.train_yvar = train_yvar
self.bounds = torch.tensor([[0.0], [1.0]], device=device, dtype=dtype)
model_st = SingleTaskGP(self.train_x, self.train_y)
self.model_st = model_st.to(device=device, dtype=dtype)
self.mll_st = ExactMarginalLogLikelihood(
self.model_st.likelihood, self.model_st
)
self.mll_st = fit_gpytorch_model(self.mll_st, options={"maxiter": 5})
model_fn = FixedNoiseGP(
self.train_x, self.train_y, self.train_yvar.expand_as(self.train_y)
)
self.model_fn = model_fn.to(device=device, dtype=dtype)
self.mll_fn = ExactMarginalLogLikelihood(
self.model_fn.likelihood, self.model_fn
)
self.mll_fn = fit_gpytorch_model(self.mll_fn, options={"maxiter": 5})
def __call__(self, spec_f):
spec_f, is_variable = _check_is_variable(spec_f)
n_fft = spec_f.size(2)
m_min = 0. if self.f_min == 0 else 2595 * np.log10(1. + (self.f_min / 700))
m_max = 2595 * np.log10(1. + (self.f_max / 700))
m_pts = torch.linspace(m_min, m_max, self.n_mels + 2)
f_pts = (700 * (10**(m_pts / 2595) - 1))
bins = torch.floor(((n_fft - 1) * 2) * f_pts / self.sr).long()
fb = torch.zeros(n_fft, self.n_mels)
for m in range(1, self.n_mels + 1):
f_m_minus = bins[m - 1].item()
f_m = bins[m].item()
f_m_plus = bins[m + 1].item()
if f_m_minus != f_m:
fb[f_m_minus:f_m, m - 1] = (torch.arange(f_m_minus, f_m) - f_m_minus) / (f_m - f_m_minus)
if f_m != f_m_plus:
fb[f_m:f_m_plus, m - 1] = (f_m_plus - torch.arange(f_m, f_m_plus)) / (f_m_plus - f_m)
fb = Variable(fb)
spec_m = torch.matmul(spec_f, fb) # (c, l, n_fft) dot (n_fft, n_mels) -> (c, l, n_mels)
return spec_m if is_variable else spec_m.data
def main():
x = torch.unsqueeze(torch.linspace(-1, 1, 100), dim=1)
y = x.pow(2) + 0.2 * torch.rand(x.size())
x = Variable(x)
y = Variable(y)
net = RegreNN(1,1)
optm = torch.optim.SGD(net.parameters(),lr=0.5e-1)
loss_func = torch.nn.MSELoss()
plt.ion()
for i in range(600):
v = net(x)
loss = loss_func(v,y)
optm.zero_grad()
loss.backward()
optm.step()
if i % 100 == 0:
print(loss)
plt.cla()
plt.scatter(x.data.numpy(), y.data.numpy())
plt.plot(x.data.numpy(), v.data.numpy(), 'r-', lw=5)
plt.text(0.5, 0, 'Loss=%.4f' % loss.data[0], fontdict={'size': 20, 'color': 'red'})
plt.pause(0.1)
plt.ioff()
plt.show()
def __init__(self, n_bins=15):
"""
n_bins (int): number of confidence interval bins
"""
super(_ECELoss, self).__init__()
bin_boundaries = torch.linspace(0, 1, n_bins + 1)
self.bin_lowers = bin_boundaries[:-1]
self.bin_uppers = bin_boundaries[1:]
def forward(ctx, theta, size):
assert type(size) == torch.Size
N, C, H, W = size
ctx.size = size
if theta.is_cuda:
AffineGridGenerator._enforce_cudnn(theta)
assert False
ctx.is_cuda = False
base_grid = theta.new(N, H, W, 3)
linear_points = torch.linspace(-1, 1, W) if W > 1 else torch.Tensor([-1])
base_grid[:, :, :, 0] = torch.ger(torch.ones(H), linear_points).expand_as(base_grid[:, :, :, 0])
linear_points = torch.linspace(-1, 1, H) if H > 1 else torch.Tensor([-1])
base_grid[:, :, :, 1] = torch.ger(linear_points, torch.ones(W)).expand_as(base_grid[:, :, :, 1])
base_grid[:, :, :, 2] = 1
ctx.base_grid = base_grid
grid = torch.bmm(base_grid.view(N, H * W, 3), theta.transpose(1, 2))
grid = grid.view(N, H, W, 2)
return grid
def __init__(self, policy, cmdl):
"""Assumes policy returns an autograd.Variable"""
self.name = "CP"
self.cmdl = cmdl
self.policy = policy
self.dtype = dtype = TorchTypes(cmdl.cuda)
self.support = torch.linspace(cmdl.v_min, cmdl.v_max, cmdl.atoms_no)
self.support = self.support.type(dtype.FloatTensor)
def _getModel(self, double=False, cuda=False):
device = torch.device("cuda") if cuda else torch.device("cpu")
dtype = torch.double if double else torch.float
train_x = torch.linspace(0, 1, 10, device=device, dtype=dtype).unsqueeze(-1)
noise = torch.tensor(NOISE, device=device, dtype=dtype)
train_y = torch.sin(train_x.view(-1) * (2 * math.pi)) + noise
model = SingleTaskGP(train_x, train_y)
mll = ExactMarginalLogLikelihood(model.likelihood, model)
return mll.to(device=device, dtype=dtype)
def act(self, state, epsilon, exp_model, evaluation=False):
# self.T += 1
self.dqn.eval()
orig_state = state[:, :, -1:]
state = torch.from_numpy(state).float().transpose_(0, 2).unsqueeze(0)
q_values_distributions = self.dqn(Variable(state, volatile=True)).cpu().data[0]
# TODO: Log Q-Value distributions
# print(q_values_distributions)
values = torch.linspace(self.args.v_min, self.args.v_max, steps=self.args.atoms)
values = values.view(1, self.args.atoms)
values = values.expand(self.args.actions, self.args.atoms)
# print(values, q_values_distributions, torch.sum(q_values_distributions * values, dim=1))
q_value_expectations = torch.sum(q_values_distributions * values, dim=1)
q_values_numpy = q_value_expectations.numpy()
extra_info = {}
if self.args.optimistic_init and not evaluation:
raise NotImplementedError
q_values_pre_bonus = np.copy(q_values_numpy)
if not self.args.ucb:
for a in range(self.args.actions):
_, info = exp_model.bonus(orig_state, a, dont_remember=True)
action_pseudo_count = info["Pseudo_Count"]
# TODO: Log the optimism bonuses
optimism_bonus = self.args.optimistic_scaler / np.sqrt(action_pseudo_count + 0.01)
self.log("Bandit/Action_{}".format(a), optimism_bonus, step=self.T)
q_values[a] += optimism_bonus
else:
action_counts = []
for a in range(self.args.actions):
_, info = exp_model.bonus(orig_state, a, dont_remember=True)
action_pseudo_count = info["Pseudo_Count"]
action_counts.append(action_pseudo_count)
total_count = sum(action_counts)
for ai, a in enumerate(action_counts):
# TODO: Log the optimism bonuses
optimisim_bonus = self.args.optimistic_scaler * np.sqrt(2 * np.log(max(1, total_count)) / (a + 0.01))
self.log("Bandit/UCB/Action_{}".format(ai), optimisim_bonus, step=self.T)
q_values[ai] += optimisim_bonus
extra_info["Action_Bonus"] = q_values_numpy - q_values_pre_bonus
extra_info["Q_Values"] = q_values_numpy
if np.random.random() < epsilon:
action = np.random.randint(low=0, high=self.args.actions)
else:
action = int(np.argmax(q_values_numpy))
# action = q_values.max(0)[1][0] # Torch...
extra_info["Action"] = action
return action, extra_info
def gaussian_robustness(self, loader, device, sigmas_range=(0, 0.5, 30)):
'''Compute the robustness in the presence of Gaussian noise.
We first define the accuracy at a give input noise level sigma
as a function rho(sigma). Then, we define the robustness to be the
area under the curve of rho when sigma is in the given range.
Note: rho(sigma) is independent of the ground truth labels
and the noise is generated using `utils.rand.ring()`.
Args:
loader: A torch.utils.data.DataLoader without shuffling.
device: The device to do the forward passes on.
sigmas_range: (min_sigma, max_sigma, num_sigmas)
The noise levels are `linspace(*sigmas_range)`.
The generated noise will be multiplied by `self.input_range()`.
Returns:
(robustness: The robustness of the classifier under Gaussian noise,
(sigmas: The noise levels, accuracies: the accuracies))
'''
self.to(device).eval()
# compute the output predictions for the clean images
count = 0
clean_labels = []
with torch.no_grad():
for images, _ in loader:
labels = self.forward(images.to(device)).argmax(dim=-1)
clean_labels.append(labels)
count += images.size(0)
# get the noise levels (sigmas)
input_range = self.input_range()
sigmas = torch.linspace(*sigmas_range, device=device)
accuracies = torch.zeros(sigmas.numel(), device=device)
kwargs = {
'device': device,
'dtype': images.dtype,
'size': images[0, ...].size(),
'tolerance': float((sigmas[1] - sigmas[0]) / 2),
}
# seeding the device to get deterministic output
with RNG(seed=0, devices=[device]):
for i, sigma in enumerate(verbosify(sigmas)):
sigma = float(sigma * input_range)
for (images, _), labels in zip(loader, clean_labels):
noise = ring(batch=images.size(0), sigma=sigma, **kwargs)
out = self.forward(noise.add_(images.to(device)))
accuracies[i] += int(out.argmax(dim=-1).eq(labels).sum())
accuracies /= count
robustness = trapz(accuracies, x=sigmas) / (sigmas[-1] - sigmas[0])
return robustness, (sigmas, accuracies)
def _get_random_data(batch_shape, num_outputs, n=10, **tkwargs):
train_x = torch.linspace(0, 0.95, n, **tkwargs).unsqueeze(-1) + 0.05 * torch.rand(
n, 1, **tkwargs
).repeat(batch_shape + torch.Size([1, 1]))
train_y = torch.sin(train_x * (2 * math.pi)) + 0.2 * torch.randn(
n, num_outputs, **tkwargs
).repeat(batch_shape + torch.Size([1, 1]))
if num_outputs == 1:
train_y = train_y.squeeze(-1)
return train_x, train_y
def _get_random_mt_data(**tkwargs):
train_x = torch.linspace(0, 0.95, 10, **tkwargs) + 0.05 * torch.rand(10, **tkwargs)
train_y1 = torch.sin(train_x * (2 * math.pi)) + torch.randn_like(train_x) * 0.2
train_y2 = torch.cos(train_x * (2 * math.pi)) + torch.randn_like(train_x) * 0.2
train_i_task1 = torch.full_like(train_x, dtype=torch.long, fill_value=0)
train_i_task2 = torch.full_like(train_x, dtype=torch.long, fill_value=1)
full_train_x = torch.cat([train_x, train_x])
full_train_i = torch.cat([train_i_task1, train_i_task2])
full_train_y = torch.cat([train_y1, train_y2])
train_X = torch.stack([full_train_x, full_train_i.type_as(full_train_x)], dim=-1)
train_Y = full_train_y
return train_X, train_Y
def main():
LR = 1e-2
BATCH_SIZE=4
x = torch.unsqueeze(torch.linspace(-1, 1, 100), dim=1)
y = x.pow(2) + 0.2 * torch.rand(x.size())
torch_dataset = Data.TensorDataset(data_tensor=x, target_tensor=y)
loader = Data.DataLoader(dataset=torch_dataset, batch_size=BATCH_SIZE, shuffle=True, num_workers=2, )
net_SGD = Net()
net_Momentum = Net()
net_RMSprop = Net()
net_Adam = Net()
nets = [net_SGD, net_Momentum, net_RMSprop, net_Adam]
loss_func = torch.nn.MSELoss()
optimizer_list = [
torch.optim.SGD(net_SGD.parameters(),lr=LR),
torch.optim.SGD(net_Momentum.parameters(),lr=LR, momentum=0.8),
torch.optim.RMSprop(net_RMSprop.parameters(),lr=LR, alpha=0.9),
torch.optim.Adam(net_Adam.parameters(),lr=LR, betas=(0.9, 0.99)),
]
losses_his = [[], [], [], []] # record loss
for i in range(BATCH_SIZE):
for step, (batch_x, batch_y) in enumerate(loader):
b_x = Variable(x)
b_y = Variable(y)
for net, opt, l_his in zip(nets, optimizer_list, losses_his):
output = net(b_x) # get output for every net
loss = loss_func(output, b_y) # compute loss for every net
opt.zero_grad() # clear gradients for next train
loss.backward() # backpropagation, compute gradients
opt.step() # apply gradients
l_his.append(loss.data[0]) # loss recoder
labels = ['SGD', 'Momentum', 'RMSprop', 'Adam']
for i, l_his in enumerate(losses_his):
plt.plot(l_his, label=labels[i])
plt.legend(loc='best')
plt.xlabel('Steps')
plt.ylabel('Loss')
plt.ylim((0, 0.2))
plt.show()
def _get_categorical(self, next_states, rewards, mask):
batch_sz = next_states.size(0)
gamma = self.gamma
rewards = rewards.data
# Compute probabilities p(x, a)
probs = self.target_policy(next_states).data
argmax_a = torch.mul(
probs, self.support.expand_as(probs)).sum(2).max(1)[1].squeeze(1)
action_mask = argmax_a.unsqueeze(2).expand(batch_sz, 1, self.atoms_no)
qa_probs = probs.gather(1, action_mask).squeeze()
# Mask gamma and reshape it torgether with rewards to fit p(x,a).
rewards = rewards.unsqueeze(1).expand_as(qa_probs)
gamma = (mask.float() * gamma).unsqueeze(1).expand_as(qa_probs)
# Compute projection of the application of the Bellman operator.
bellman_op = rewards + gamma * self.support.unsqueeze(0).expand_as(rewards)
bellman_op = torch.clamp(bellman_op, self.v_min, self.v_max)
# Compute categorical indices for distributing the probability
m = torch.zeros(batch_sz, self.atoms_no).type(self.dtype.FloatTensor)
b = (bellman_op - self.v_min) / self.delta_z
l = b.floor().type(self.dtype.LongTensor)
u = b.ceil().type(self.dtype.LongTensor)
# Distribute probability
"""
for i in range(batch_sz):
for j in range(self.atoms_no):
uidx = u[i][j]
lidx = l[i][j]
m[i][lidx] = m[i][lidx] + qa_probs[i][j] * (uidx - b[i][j])
m[i][uidx] = m[i][uidx] + qa_probs[i][j] * (b[i][j] - lidx)
for i in range(batch_sz):
m[i].index_add_(0, l[i], qa_probs[i] * (u[i].float() - b[i]))
m[i].index_add_(0, u[i], qa_probs[i] * (b[i] - l[i].float()))
"""
# Optimized by https://github.com/tudor-berariu
offset = torch.linspace(0, ((batch_sz - 1) * self.atoms_no), batch_sz)\
.type(self.dtype.LongTensor)\
.unsqueeze(1).expand(batch_sz, self.atoms_no)
m.view(-1).index_add_(0, (l + offset).view(-1),
(qa_probs * (u.float() - b)).view(-1))
m.view(-1).index_add_(0, (u + offset).view(-1),
(qa_probs * (b - l.float())).view(-1))
return Variable(m.type(self.dtype.FloatTensor))
def get_precision_recall(args, score, label, num_samples, beta=1.0, sampling='log', predicted_score=None):
'''
:param args:
:param score: anomaly scores
:param label: anomaly labels
:param num_samples: the number of threshold samples
:param beta:
:param scale:
:return:
'''
if predicted_score is not None:
score = score - torch.FloatTensor(predicted_score).squeeze().to(args.device)
maximum = score.max()
if sampling=='log':
# Sample thresholds logarithmically
# The sampled thresholds are logarithmically spaced between: math:`10 ^ {start}` and: math:`10 ^ {end}`.
th = torch.logspace(0, torch.log10(torch.tensor(maximum)), num_samples).to(args.device)
else:
# Sample thresholds equally
# The sampled thresholds are equally spaced points between: attr:`start` and: attr:`end`
th = torch.linspace(0, maximum, num_samples).to(args.device)
precision = []
recall = []
for i in range(len(th)):
anomaly = (score > th[i]).float()
idx = anomaly * 2 + label
tn = (idx == 0.0).sum().item() # tn
fn = (idx == 1.0).sum().item() # fn
fp = (idx == 2.0).sum().item() # fp
tp = (idx == 3.0).sum().item() # tp
p = tp / (tp + fp + 1e-7)
r = tp / (tp + fn + 1e-7)
if p != 0 and r != 0:
precision.append(p)
recall.append(r)
precision = torch.FloatTensor(precision)
recall = torch.FloatTensor(recall)
f1 = (1 + beta ** 2) * (precision * recall).div(beta ** 2 * precision + recall + 1e-7)
return precision, recall, f1
def loss_categorical(self, transitions):
num_atoms = self.config.num_atoms
batch_size = len(transitions)
batch = Transition(*zip(*transitions))
non_final_mask = torch.tensor(tuple(map(lambda s: s is not None, batch.next_state)), dtype=torch.uint8, device=self.config.device)
state_batch = torch.cat(batch.state).to(torch.float32)
action_batch = torch.cat(batch.action)
reward_batch = torch.cat(batch.reward).to(torch.float32)
next_states = [s for s in batch.next_state if s is not None]
with torch.no_grad():
non_final_next_state = torch.cat(next_states).to(torch.float32)
best_actions = self._get_Q(self.model, non_final_next_state).argmax(dim=1)
self.target_model.reset_noise()
p_next = self.target_model(non_final_next_state, ApplySoftmax.NORMAL)
p_next_best = torch.zeros(0).to(self.config.device, dtype=torch.float32).new_full((batch_size, num_atoms), 1.0 / num_atoms)
p_next_best[non_final_mask] = p_next[range(len(non_final_next_state)), best_actions]
gamma = torch.zeros(batch_size, num_atoms).to(self.config.device)
gamma[non_final_mask] = self.config.gamma ** self.config.num_multi_step_reward
Tz = (reward_batch.unsqueeze(1) + gamma * self.support.unsqueeze(0)).clamp(self.Vmin, self.Vmax)
b = (Tz - self.Vmin) / self.delta_z
l = b.floor().long()
u = b.ceil().long()
l[(l == u) * (0 < l)] -= 1
# the values in l has already changed, so same index is not processed for u
u[(l == u) * (u < num_atoms - 1)] += 1
m = torch.zeros(batch_size, num_atoms).to(self.config.device, dtype=torch.float32)
offset = torch.linspace(0, ((batch_size-1) * num_atoms), batch_size).unsqueeze(1).expand(batch_size, num_atoms).to(l)
m.view(-1).index_add_(0, (l + offset).view(-1), (p_next_best * (u.float() - b)).view(-1))
m.view(-1).index_add_(0, (u + offset).view(-1), (p_next_best * (b - l.float())).view(-1))
self.model.reset_noise()
log_p = self.model(state_batch, ApplySoftmax.LOG)
log_p_a = log_p[range(batch_size), action_batch.squeeze()]
return -torch.sum(m * log_p_a, dim=1)
def plot_cont_traversal(model: infogan.InfoGAN, c, nrow, nstep=9):
values = torch.linspace(-2, 2, nstep).to(model.device)
latent = model.sample_latent(nrow).repeat(nstep, 1)
for r in range(nrow):
latent[r::nrow, model.cont_idx[c]] = values
samples = model.gen(latent).detach()
fig, axs = plot_grid(samples, nrow=nrow, figsize=(nstep, nrow),
gridspec_kw=dict(wspace=0, hspace=0))
# plt.suptitle(f"$c_{{{c + 2}}}$: Continuous (-2 to 2)")
for i in [0, -1]:
_prep_ax(axs[i, 0])
axs[0, 0].set_xlabel(f'${values[ 0]:+g}$', ha='center', va='bottom', size=_TICK_LABEL_SIZE)
axs[-1, 0].set_xlabel(f'${values[-1]:+g}$', ha='center', va='bottom', size=_TICK_LABEL_SIZE)
ypos = axs[0, 0].get_position().y1
fig.text(.5, ypos, f'$c_{{{c + 2}}}$', ha='center', va='bottom', size=_VAR_LABEL_SIZE)
def __init__(self, policy, target_policy, cmdl):
self.name = "Categorical-PI"
self.policy = policy
self.target_policy = target_policy
self.lr = cmdl.lr
self.gamma = cmdl.gamma
self.cmdl = cmdl
self.optimizer = optim.Adam(self.policy.parameters(), lr=self.lr)
self.optimizer.zero_grad()
self.grads_decoupled = False
self.dtype = dtype = TorchTypes(cmdl.cuda)
self.v_min, self.v_max = v_min, v_max = cmdl.v_min, cmdl.v_max
self.atoms_no = atoms_no = cmdl.atoms_no
self.support = torch.linspace(v_min, v_max, atoms_no)
self.support = self.support.type(dtype.FloatTensor)
self.delta_z = (cmdl.v_max - cmdl.v_min) / (cmdl.atoms_no - 1)
def _setUp(self, double=False, cuda=False, expand=False):
device = torch.device("cuda") if cuda else torch.device("cpu")
dtype = torch.double if double else torch.float
train_x = torch.linspace(0, 1, 10, device=device, dtype=dtype).unsqueeze(-1)
train_y = torch.sin(train_x * (2 * math.pi)).squeeze(-1)
noise = torch.tensor(NOISE, device=device, dtype=dtype)
self.train_x = train_x
self.train_y = train_y + noise
if expand:
self.train_x = self.train_x.expand(-1, 2)
ics = torch.tensor([[0.5, 1.0]], device=device, dtype=dtype)
else:
ics = torch.tensor([[0.5]], device=device, dtype=dtype)
self.initial_conditions = ics
self.f_best = self.train_y.max().item()
model = SingleTaskGP(self.train_x, self.train_y)
self.model = model.to(device=device, dtype=dtype)
self.mll = ExactMarginalLogLikelihood(self.model.likelihood, self.model)
self.mll = fit_gpytorch_model(self.mll, options={"maxiter": 1})
def __init__(self, scale, taps, samplerate):
super(SincLayer, self).__init__()
self.samplerate = int(samplerate)
self.taps = taps
self.scale = scale
# each filter requires two parameters to define the filter bandwidth
filter_parameters = torch.FloatTensor(len(scale), 2)
self.linear = nn.Parameter(
torch.linspace(-math.pi, math.pi, steps=taps), requires_grad=False)
self.window = nn.Parameter(
torch.hamming_window(self.taps), requires_grad=False)
for i, band in enumerate(scale):
start = self.samplerate / band.start_hz
stop = self.samplerate / band.stop_hz
filter_parameters[i, 0] = start
filter_parameters[i, 1] = stop
self.filter_parameters = nn.Parameter(filter_parameters)
def plot_cont_cont_traversal(model: infogan.InfoGAN, c1, c2, nstep=9):
values = torch.linspace(-1.5, 1.5, nstep).to(model.device)
latent = model.sample_latent(1).repeat(nstep ** 2, 1)
for s in range(nstep):
latent[s::nstep, model.cont_idx[c2]] = values
latent[s * nstep:(s + 1) * nstep, model.cont_idx[c1]] = values
samples = model.gen(latent).detach()
fig, axs = plot_grid(samples, nrow=nstep, figsize=(nstep, nstep),
gridspec_kw=dict(wspace=0, hspace=0))
# plt.suptitle(rf"$c_{{{c1 + 2}}} \times c_{{{c2 + 2}}}$: Continuous (-2 to 2)")
for i in [(0, 0), (0, -1), (-1, 0)]:
_prep_ax(axs[i])
axs[ 0, 0].set_xlabel(f'${values[ 0]:+g}$', ha='center', va='bottom', size=_TICK_LABEL_SIZE)
axs[-1, 0].set_xlabel(f'${values[-1]:+g}$', ha='center', va='bottom', size=_TICK_LABEL_SIZE)
axs[ 0, 0].set_ylabel(f'${values[ 0]:+g}$', ha='right', va='center', rotation=0, size=_TICK_LABEL_SIZE)
axs[ 0,-1].set_ylabel(f'${values[-1]:+g}$', ha='right', va='center', rotation=0, size=_TICK_LABEL_SIZE)
xpos = axs[ 0, 0].get_position().x0
ypos = axs[ 0, 0].get_position().y1
fig.text(.5, ypos, f'$c_{{{c1 + 2}}}$', ha='center', va='bottom', size=_VAR_LABEL_SIZE)
fig.text(xpos, .5, f'$c_{{{c2 + 2}}}$', ha='right', va='center', size=_VAR_LABEL_SIZE)
def _get_model(self, cuda=False, dtype=torch.float):
device = torch.device("cuda") if cuda else torch.device("cpu")
state_dict = {
"mean_module.constant": torch.tensor([-0.0066]),
"covar_module.raw_outputscale": torch.tensor(1.0143),
"covar_module.base_kernel.raw_lengthscale": torch.tensor([[-0.99]]),
"covar_module.base_kernel.lengthscale_prior.concentration": torch.tensor(
3.0
),
"covar_module.base_kernel.lengthscale_prior.rate": torch.tensor(6.0),
"covar_module.outputscale_prior.concentration": torch.tensor(2.0),
"covar_module.outputscale_prior.rate": torch.tensor(0.1500),
}
train_x = torch.linspace(0, 1, 10, device=device, dtype=dtype)
train_y = torch.sin(train_x * (2 * math.pi))
noise = torch.tensor(NEI_NOISE, device=device, dtype=dtype)
train_y += noise
train_yvar = torch.full_like(train_y, 0.25 ** 2)
train_x = train_x.view(-1, 1)
model = FixedNoiseGP(train_X=train_x, train_Y=train_y, train_Yvar=train_yvar)
model.load_state_dict(state_dict)
model.to(train_x)
model.eval()
return model
parser.add_argument('--version', type=str, choices=['standard','steer','normal'], default='steer')
args = parser.parse_args()
torch.manual_seed(6)
if args.adjoint:
from torchdiffeq import odeint_adjoint as odeint
from torchdiffeq import odeint_adjoint_stochastic_end_v3 as odeint_stochastic_end_v3
from torchdiffeq import odeint_adjoint_stochastic_end_normal as odeint_stochastic_end_normal
else:
from torchdiffeq import odeint_stochastic_end_v3
from torchdiffeq import odeint
device = torch.device('cuda:' + str(args.gpu) if torch.cuda.is_available() else 'cpu')
true_y0 = torch.tensor([0.])
#true_y0 = torch.tensor([1.])
t = torch.linspace(0., 15., args.data_size)
test_t = torch.linspace(0., 25., args.data_size)
true_A = torch.tensor([[-0.1, 2.0], [-2.0, -0.1]])
class Lambda(nn.Module):
def forward(self, t, y):
t = t.unsqueeze(0)
equation = -1000*y + 3000 - 2000 * torch.exp(-t) + 1000 * torch.sin(t)
#equation = -1000*y + 3000 - 2000 * torch.exp(-t)
#equation = -100*y + 300 - 200 * torch.exp(-1*t)
#equation = -1*y*args.stiffness_ratio + 3*args.stiffness_ratio - 2*args.stiffness_ratio * torch.exp(-1*t)# - 2*args.stiffness_ratio * torch.exp(-10000*t)
#equation = 0.12*y #-2000*torch.exp(-t)#+ 3000 - 2000 * torch.exp(-t)
#equation = 10 * torch.sin(t)
def get_region_boxes(output,
conf_thresh,
num_classes,
anchors,
num_anchors,
only_objectness=1,
validation=False):
anchor_step = len(anchors) / num_anchors
if output.dim() == 3:
output = output.unsqueeze(0)
batch = output.size(0)
assert (output.size(1) == (5 + num_classes) * num_anchors)
h = output.size(2)
w = output.size(3)
t0 = time.time()
all_boxes = []
output = output.view(batch * num_anchors, 5 + num_classes,
h * w).transpose(0, 1).contiguous().view(
5 + num_classes, batch * num_anchors * h * w)
grid_x = torch.linspace(0, w - 1, w).repeat(h, 1).repeat(
batch * num_anchors, 1, 1).view(batch * num_anchors * h * w).cuda()
grid_y = torch.linspace(0, h - 1, h).repeat(w, 1).t().repeat(
batch * num_anchors, 1, 1).view(batch * num_anchors * h * w).cuda()
xs = torch.sigmoid(output[0]) + grid_x
ys = torch.sigmoid(output[1]) + grid_y
anchor_w = torch.Tensor(anchors).view(num_anchors,
anchor_step).index_select(
1, torch.LongTensor([0]))
anchor_h = torch.Tensor(anchors).view(num_anchors,
anchor_step).index_select(
1, torch.LongTensor([1]))
anchor_w = anchor_w.repeat(batch, 1).repeat(1, 1, h * w).view(
batch * num_anchors * h * w).cuda()
anchor_h = anchor_h.repeat(batch, 1).repeat(1, 1, h * w).view(
batch * num_anchors * h * w).cuda()
ws = torch.exp(output[2]) * anchor_w
hs = torch.exp(output[3]) * anchor_h
det_confs = torch.sigmoid(output[4])
cls_confs = torch.nn.Softmax()(Variable(
output[5:5 + num_classes].transpose(0, 1))).data
cls_max_confs, cls_max_ids = torch.max(cls_confs, 1)
cls_max_confs = cls_max_confs.view(-1)
cls_max_ids = cls_max_ids.view(-1)
t1 = time.time()
sz_hw = h * w
sz_hwa = sz_hw * num_anchors
det_confs = convert2cpu(det_confs)
cls_max_confs = convert2cpu(cls_max_confs)
cls_max_ids = convert2cpu_long(cls_max_ids)
xs = convert2cpu(xs)
ys = convert2cpu(ys)
ws = convert2cpu(ws)
hs = convert2cpu(hs)
if validation:
cls_confs = convert2cpu(cls_confs.view(-1, num_classes))
t2 = time.time()
for b in range(batch):
boxes = []
for cy in range(h):
for cx in range(w):
for i in range(num_anchors):
ind = b * sz_hwa + i * sz_hw + cy * w + cx
det_conf = det_confs[ind]
if only_objectness:
conf = det_confs[ind]
else:
conf = det_confs[ind] * cls_max_confs[ind]
if conf > conf_thresh:
bcx = xs[ind]
bcy = ys[ind]
bw = ws[ind]
bh = hs[ind]
cls_max_conf = cls_max_confs[ind]
cls_max_id = cls_max_ids[ind]
box = [
bcx / w, bcy / h, bw / w, bh / h, det_conf,
cls_max_conf, cls_max_id
]
if (not only_objectness) and validation:
for c in range(num_classes):
tmp_conf = cls_confs[ind][c]
if c != cls_max_id and det_confs[
ind] * tmp_conf > conf_thresh:
box.append(tmp_conf)
box.append(c)
boxes.append(box)
all_boxes.append(boxes)
t3 = time.time()
if False:
print('---------------------------------')
print('matrix computation : %f' % (t1 - t0))
print(' gpu to cpu : %f' % (t2 - t1))
print(' boxes filter : %f' % (t3 - t2))
print('---------------------------------')
return all_boxes
def forward(self,
feed_dict_q,
feed_dict_k=None,
metadata=None,
is_eval=False,
cluster_result=None,
index=None,
is_viewpoint_eval=False,
feed_dicts_N=None,
forward_type=None):
"""
Input:
feed_dict_q: a batch of query images and bounding boxes
feed_dict_k: a batch of key images and bounding boxes
is_eval: return momentum embeddings (used for clustering)
cluster_result: cluster assignments, centroids, and density
index: indices for training samples
Output:
logits, targets, proto_logits, proto_targets
"""
mode = self.mode
hyp_N = feed_dict_q["objects"][0].item()
rel_viewpoint = metadata["rel_viewpoint"]
if mode == "node":
rel_viewpoint = None
if is_viewpoint_eval and mode == "spatial":
with torch.no_grad():
k = self.encoder_q(feed_dict_q)
k = self.merge_pose_with_scene_embeddings(
k, rel_viewpoint) #merge
for batch_ind in range(len(k)):
k[batch_ind][1] = self.spatial_viewpoint_transformation(
k[batch_ind][1]
) # Do viewpoint transformation on spatial embeddings
k = stack_features_across_batch(k, mode)
k = nn.functional.normalize(k, dim=1)
return k
if is_eval:
with torch.no_grad():
# the output from encoder is a list of features from the batch where each batch element (image)
# might contain different number of objects
k = self.encoder_q(feed_dict_q)
# encoder output features in the list are stacked to form a tensor of features across the batch
k = stack_features_across_batch(k, mode)
# normalize feature across the batch
k = nn.functional.normalize(k, dim=1)
return k
# k_o : spatial embeddings before viewpoint transformation
# k_t : spatial embeds after viewpoint transformation
# update the key encoder
k_o = self.encoder_q(feed_dict_k) # callculate the embeddings
# Do viewpoint transformation on embeddings if the pose is fed as input
if mode == "spatial" and rel_viewpoint is not None:
k_t = self.merge_pose_with_scene_embeddings(
k_o, rel_viewpoint) #merge pose with spatial embeddings
for batch_ind in range(len(k_t)):
k_t[batch_ind][1] = self.spatial_viewpoint_transformation(
k_t[batch_ind]
[1]) # Do viewpoint transformation on spatial embeddings
k_o = stack_features_across_batch(k_o, mode)
k_o = nn.functional.normalize(k_o, dim=1)
k_t = stack_features_across_batch(k_t, mode)
k_t = nn.functional.normalize(k_t, dim=1)
q = self.encoder_q(feed_dict_q) # queries: NxC
q = stack_features_across_batch(q, mode)
q = nn.functional.normalize(q, dim=1)
if forward_type == "scene":
# compute logits
# Einstein sum is more intuitive
# positive logits: Nx1
l_pos = torch.einsum('nc,nc->n', [q, k_t]).unsqueeze(-1)
# negative logits: Nxr
l_neg = torch.einsum('nc,ck->nk',
[q, self.queue_scene.clone().detach()])
# logits: Nx(1+r)
logits = torch.cat([l_pos, l_neg], dim=1)
# apply temperature
logits /= self.T
# labels: positive key indicators
labels = torch.zeros(logits.shape[0], dtype=torch.long).cuda()
# dequeue and enqueue
# self._dequeue_and_enqueue_scene(k_t.clone().detach())
self._dequeue_and_enqueue_scene(k_o.clone().detach())
return logits, labels, None, None
index = convert_indices(index, hyp_N, mode)
# prototypical contrast
if cluster_result is not None:
proto_labels = []
proto_logits = []
for n, (im2cluster, prototypes, density) in enumerate(
zip(cluster_result['im2cluster'],
cluster_result['centroids'],
cluster_result['density'])):
# get positive prototypes
pos_proto_id = im2cluster[index]
pos_prototypes = prototypes[pos_proto_id]
# sample negative prototypes
all_proto_id = [i for i in range(im2cluster.max())]
#print(len(pos_prototypes), len(all_proto_id))
neg_proto_id = set(all_proto_id) - set(
pos_proto_id.tolist())
neg_proto_id = sample(
neg_proto_id,
self.scene_r) #sample r negative prototypes
neg_prototypes = prototypes[neg_proto_id]
proto_selected = torch.cat(
[pos_prototypes, neg_prototypes], dim=0)
# compute prototypical logits
logits_proto = torch.mm(q, proto_selected.t())
# targets for prototype assignment
labels_proto = torch.linspace(
0, q.size(0) - 1, steps=q.size(0)).long().cuda()
# scaling temperatures for the selected prototypes
temp_proto = density[torch.cat(
[pos_proto_id,
torch.LongTensor(neg_proto_id).cuda()],
dim=0)]
logits_proto /= temp_proto
proto_labels.append(labels_proto)
proto_logits.append(logits_proto)
return logits, labels, proto_logits, proto_labels
elif forward_type == "view":
self.queue_view = torch.randn(self.dim, self.view_r).cuda()
self.queue_view_ptr[0] = 0
negative_view_index = [feed_n[1] for feed_n in feed_dicts_N]
# getting encoding for scene_negatives
for feed_dict_ in feed_dicts_N:
# with torch.no_grad():
k_n = self.encoder_q(feed_dict_[0])
# encoder output features in the list are stacked to form a tensor of features across the batch
k_n = stack_features_across_batch(k_n, mode)
# normalize feature across the batch
scene_negatives = nn.functional.normalize(k_n, dim=1)
# append negagives to queue_view
self._dequeue_and_enqueue_view(scene_negatives)
self._dequeue_and_enqueue_view(k_o)
# positive logits: Nx1
l_pos = torch.einsum('nc,nc->n', [q, k_t]).unsqueeze(-1)
# negative logits: Nxr
l_neg = torch.einsum('nc,ck->nk', [q, self.queue_view])
# logits: Nx(1+r)
logits = torch.cat([l_pos, l_neg], dim=1)
# apply temperature
logits /= self.T
# labels: positive key indicators
labels = torch.zeros(logits.shape[0], dtype=torch.long).cuda()
return logits, labels, None, None
else:
raise ValueError("Forward type of the mode must be defined")
def loss(input, target, border=10):
return torch.sum(
torch.square(input[:, :, border:-border, border:-border] -
target[:, :, border:-border, border:-border]))
cap = cv2.VideoCapture("IMG_0004.mp4")
# Stores the last k frames of the video
k = 25
prev_k_frames = []
# As I understand, the grid sample method uses the x, y values of the outputs of the pixel. It may be easier for the model to not need to learn this
# so I am adding it in back at the end
shape = 223
single_vector_x = torch.linspace(-1, 1, steps=shape).reshape(1, -1)
single_vector_y = torch.linspace(-1, 1, steps=shape).reshape(-1, 1)
baseline = torch.zeros(1, shape, shape, 2)
baseline[:, :, :, 0] = single_vector_x
baseline[:, :, :, 1] = single_vector_y
baseline = baseline.to(torch.device("cuda:0"))
frame_count = 0
# Opens video and trains model
while (cap.isOpened()):
ret, frame = cap.read()
if frame is None:
break
frame = Image.fromarray(frame)
#frame = frame.convert('1')
frame = frame.resize((shape, shape))
def train(self):
if self.T - self.target_sync_T > self.args.target:
self.sync_target_network()
self.target_sync_T = self.T
info = {}
for _ in range(self.args.iters):
self.dqn.eval()
batch, indices, is_weights = self.replay.Sample_N(self.args.batch_size, self.args.n_step, self.args.gamma)
columns = list(zip(*batch))
states = Variable(torch.from_numpy(np.array(columns[0])).float().transpose_(1, 3))
actions = Variable(torch.LongTensor(columns[1]))
terminal_states = Variable(torch.FloatTensor(columns[5]))
rewards = Variable(torch.FloatTensor(columns[2]))
# Have to clip rewards for DQN
rewards = torch.clamp(rewards, -1, 1)
steps = Variable(torch.FloatTensor(columns[4]))
new_states = Variable(torch.from_numpy(np.array(columns[3])).float().transpose_(1, 3))
target_dqn_qvals = self.target_dqn(new_states).cpu()
# Make a new variable with those values so that these are treated as constants
target_dqn_qvals_data = Variable(target_dqn_qvals.data)
q_value_gammas = (Variable(torch.ones(terminal_states.size()[0])) - terminal_states)
inter = Variable(torch.ones(terminal_states.size()[0]) * self.args.gamma)
# print(steps)
q_value_gammas = q_value_gammas * torch.pow(inter, steps)
values = torch.linspace(self.args.v_min, self.args.v_max, steps=self.args.atoms)
values = Variable(values)
values = values.view(1, 1, self.args.atoms)
values = values.expand(self.args.batch_size, self.args.actions, self.args.atoms)
# print(values)
q_value_gammas = q_value_gammas.view(self.args.batch_size, 1, 1)
q_value_gammas = q_value_gammas.expand(self.args.batch_size, self.args.actions, self.args.atoms)
# print(q_value_gammas)
gamma_values = q_value_gammas * values
# print(gamma_values)
rewards = rewards.view(self.args.batch_size, 1, 1)
rewards = rewards.expand(self.args.batch_size, self.args.actions, self.args.atoms)
# print(rewards)
operator_q_values = rewards + gamma_values
# print(operator_q_values)
clipped_operator_q_values = torch.clamp(operator_q_values, self.args.v_min, self.args.v_max)
delta_z = (self.args.v_max - self.args.v_min) / (self.args.atoms - 1)
# Using the notation from the categorical paper
b_j = (clipped_operator_q_values - self.args.v_min) / delta_z
# print(b_j)
lower_bounds = torch.floor(b_j)
upper_bounds = torch.ceil(b_j)
# Work out the max action
atom_values = Variable(torch.linspace(self.args.v_min, self.args.v_max, steps=self.args.atoms))
atom_values = atom_values.view(1, 1, self.args.atoms)
atom_values = atom_values.expand(self.args.batch_size, self.args.actions, self.args.atoms)
# Sum over the atoms dimension
target_expected_qvalues = torch.sum(target_dqn_qvals_data * atom_values, dim=2)
# Get the maximum actions index across the batch size
max_actions = target_expected_qvalues.max(dim=1)[1].view(-1)
# Project back onto the original support for the max actions
q_value_distribution_targets = torch.zeros(self.args.batch_size, self.args.atoms)
# Distributions for the max actions
# print(target_dqn_qvals_data, max_actions)
q_value_max_actions_distribs = target_dqn_qvals_data.index_select(dim=1, index=max_actions)[:,0,:]
# print(q_value_max_actions_distribs)
# Lower_bounds_actions
lower_bounds_actions = lower_bounds.index_select(dim=1, index=max_actions)[:,0,:]
upper_bounds_actions = upper_bounds.index_select(dim=1, index=max_actions)[:,0,:]
b_j_actions = b_j.index_select(dim=1, index=max_actions)[:,0,:]
lower_bound_values_to_add = q_value_max_actions_distribs * (upper_bounds_actions - b_j_actions)
upper_bound_values_to_add = q_value_max_actions_distribs * (b_j_actions - lower_bounds_actions)
# print(lower_bounds_actions)
# print(lower_bound_values_to_add)
# Naive looping
for b in range(self.args.batch_size):
for l, pj in zip(lower_bounds_actions.data.type(torch.LongTensor)[b], lower_bound_values_to_add[b].data):
q_value_distribution_targets[b][l] += pj
for u, pj in zip(upper_bounds_actions.data.type(torch.LongTensor)[b], upper_bound_values_to_add[b].data):
q_value_distribution_targets[b][u] += pj
self.dqn.train()
if self.args.gpu:
actions = actions.cuda()
# q_value_targets = q_value_targets.cuda()
q_value_distribution_targets = q_value_distribution_targets.cuda()
model_predictions = self.dqn(states).index_select(1, actions.view(-1))[:,0,:]
q_value_distribution_targets = Variable(q_value_distribution_targets)
# print(q_value_distribution_targets)
# print(model_predictions)
# Cross entropy loss
ce_loss = -torch.sum(q_value_distribution_targets * torch.log(model_predictions), dim=1)
ce_batch_loss = ce_loss.mean()
info = {}
self.log("DQN/X_Entropy_Loss", ce_batch_loss.data[0], step=self.T)
# Update
self.optimizer.zero_grad()
ce_batch_loss.backward()
# Taken from pytorch clip_grad_norm
# Remove once the pip version it up to date with source
gradient_norm = clip_grad_norm(self.dqn.parameters(), self.args.clip_value)
if gradient_norm is not None:
info["Norm"] = gradient_norm
self.optimizer.step()
if "States" in info:
states_trained = info["States"]
info["States"] = states_trained + columns[0]
else:
info["States"] = columns[0]
# Pad out the states to be of size batch_size
if len(info["States"]) < self.args.batch_size:
old_states = info["States"]
new_states = old_states[0] * (self.args.batch_size - len(old_states))
info["States"] = new_states
return info
def learn(self, mem):
# Sample transitions
idxs, states, actions, returns, next_states, nonterminals, weights = mem.sample(
self.batch_size)
# Calculate current state probabilities (online network noise already sampled)
log_ps = self.online_net(
states, log=True) # Log probabilities log p(s_t, ·; θonline)
log_ps_a = log_ps[range(self.batch_size),
actions] # log p(s_t, a_t; θonline)
with torch.no_grad():
# Calculate nth next state probabilities
pns = self.online_net(
next_states) # Probabilities p(s_t+n, ·; θonline)
dns = (self.support.expand_as(pns) * pns
) # Distribution d_t+n = (z, p(s_t+n, ·; θonline))
argmax_indices_ns = dns.sum(2).argmax(
1
) # Perform argmax action selection using online network: argmax_a[(z, p(s_t+n, a; θonline))]
self.target_net.reset_noise() # Sample new target net noise
pns = self.target_net(
next_states) # Probabilities p(s_t+n, ·; θtarget)
pns_a = pns[range(
self.batch_size
), argmax_indices_ns] # Double-Q probabilities p(s_t+n, argmax_a[(z, p(s_t+n, a; θonline))]; θtarget)
# Compute Tz (Bellman operator T applied to z)
Tz = returns.unsqueeze(1) + nonterminals * (
self.discount**self.n) * self.support.unsqueeze(
0) # Tz = R^n + (γ^n)z (accounting for terminal states)
Tz = Tz.clamp(min=self.Vmin,
max=self.Vmax) # Clamp between supported values
# Compute L2 projection of Tz onto fixed support z
b = (Tz - self.Vmin) / self.delta_z # b = (Tz - Vmin) / Δz
l, u = b.floor().to(torch.int64), b.ceil().to(torch.int64)
# Fix disappearing probability mass when l = b = u (b is int)
l[(u > 0) * (l == u)] -= 1
u[(l < (self.atoms - 1)) * (l == u)] += 1
# Distribute probability of Tz
m = states.new_zeros(self.batch_size, self.atoms)
offset = (torch.linspace(0, ((self.batch_size - 1) * self.atoms),
self.batch_size).unsqueeze(1).expand(
self.batch_size,
self.atoms).to(actions))
m.view(-1).index_add_(
0, (l + offset).view(-1),
(pns_a *
(u.float() - b)).view(-1)) # m_l = m_l + p(s_t+n, a*)(u - b)
m.view(-1).index_add_(
0, (u + offset).view(-1),
(pns_a *
(b - l.float())).view(-1)) # m_u = m_u + p(s_t+n, a*)(b - l)
loss = -torch.sum(
m * log_ps_a,
1) # Cross-entropy loss (minimises DKL(m||p(s_t, a_t)))
self.online_net.zero_grad()
(weights * loss).mean().backward(
) # Backpropagate importance-weighted minibatch loss
self.optimiser.step()
mem.update_priorities(idxs,
loss.detach().cpu().numpy()
) # Update priorities of sampled transitions
def forward(self, gt_depths):
"""
In the forward function we accept a Tensor of input data and we must return
a Tensor of output data. We can use Modules defined in the constructor as
well as arbitrary operators on Tensors.
"""
#### Calculate current coding functions based on learned parameters
ModFs_func = torch.zeros(self.N,
1,
device=device,
dtype=dtype,
requires_grad=False)
DemodFs_func = torch.zeros(self.N,
self.K,
device=device,
dtype=dtype,
requires_grad=False)
ModFs = torch.zeros(self.N,
1,
device=device,
dtype=dtype,
requires_grad=False)
DemodFs = torch.zeros(self.N,
self.K,
device=device,
dtype=dtype,
requires_grad=False)
p = torch.linspace(0, 2 * math.pi, self.N, device=device)
for order in range(0, self.order):
ModFs_func[:, 0] += self.alpha_mod[0, order] * torch.cos(
(p + self.phi_mod[0, order]) * (order + 1))
for k in range(0, self.K):
DemodFs_func[:, k] += self.alpha_demod[k, order] * torch.cos(
(p + self.phi_demod[k, order]) * (order + 1))
# Normalize ModFs and DemodFs
min_ModFs_func, _ = torch.min(ModFs_func, dim=0)
ModFs = ModFs_func - min_ModFs_func # ModFs can't be lower than zero (negative light)
min_DemodFs_func, _ = torch.min(DemodFs_func, dim=0)
max_DemodFs_func, _ = torch.max(DemodFs_func, dim=0)
DemodFs = (DemodFs_func - min_DemodFs_func) / (
max_DemodFs_func - min_DemodFs_func) # DemodFs can only be 0->1
#################### Simulation
## Set area under the curve of outgoing ModF to the totalEnergy
ModFs = ModFs.repeat(1, self.K)
ModFs_scaled = Utils.ScaleMod(ModFs,
device,
tau=self.tauMin,
pAveSource=self.pAveSourcePerPixel)
# Calculate correlation functions (NxK matrix) and normalize it (zero mean, unit variance)
CorrFs = Utils.GetCorrelationFunctions(ModFs_scaled,
DemodFs,
device,
dt=self.dt)
NormCorrFs = (CorrFs.t() - torch.mean(CorrFs, 1)) / torch.std(
CorrFs, 1)
NormCorrFs = NormCorrFs.t()
# Compute brightness values
BVals = Utils.ComputeBrightnessVals(ModFs=ModFs_scaled, DemodFs=DemodFs, CorrFs=CorrFs, depths=gt_depths, \
pAmbient=self.pAveAmbientPerPixel, beta=self.meanBeta, T=self.T, tau=self.tau, dt=self.dt, gamma=self.gamma)
#### Add noise
# Calculate variance
noiseVar = BVals * self.gamma + math.pow(self.readNoise * self.gamma,
2)
# Add noise to all brightness values
BVals = Utils.GetClippedBSamples(nSamples=1,
BMean=BVals,
BVar=noiseVar,
device=device)
BVals = BVals.permute(0, 3, 1,
2) # Put channel dimension at right position
# Normalize BVals
BVals_mean = torch.mean(BVals)
BVals_std = torch.std(BVals)
BVals = (BVals - BVals_mean) / BVals_std
#### CNN Network
out = CNN.network(self, self.architecture, BVals)
decodedDepths = torch.squeeze(out, 1) # Remove channel dimension
return decodedDepths
def load(self, file_path):
self.network_training.load_state_dict(torch.load(file_path))
self.network_release.load_state_dict(
self.network_training.state_dict())
if __name__ == "__main__":
input_dimension = 16
critic = Critic(input_dimension,
lr=1e-3,
lr_step_size=100,
lr_gamma=0.5,
weight_decay=1e-7,
lag=10)
x_data = torch.linspace(0., 10.,
input_dimension * 11).view(-1, input_dimension)
y_data = torch.linspace(0., 10., 11).view(-1, 1)
epochs = 100
for epoch in range(epochs):
critic.fit(x_data, y_data)
preds = critic(x_data)
SSQ = ((preds - y_data) * (preds - y_data)).sum()
preds_trainer = critic.network_training(x_data)
SSQ_trainer = ((preds_trainer - y_data) *
(preds_trainer - y_data)).sum()
print("epoch %d, error %f, (%f)" % (epoch, SSQ, SSQ_trainer))
def eval_conditional_density(
density: Any,
condition: Tensor,
limits: Tensor,
dim1: int,
dim2: int,
resolution: int = 50,
eps_margins1: Union[Tensor, float] = 1e-32,
eps_margins2: Union[Tensor, float] = 1e-32,
) -> Tensor:
r"""
Return the unnormalized conditional along `dim1, dim2` given parameters `condition`.
We compute the unnormalized conditional by evaluating the joint distribution:
$p(x1 | x2) = p(x1, x2) / p(x2) \propto p(x1, x2)$
Args:
density: Probability density function with `.log_prob()` method.
condition: Parameter set that all dimensions other than dim1 and dim2 will be
fixed to. Should be of shape (1, dim_theta), i.e. it could e.g. be
a sample from the posterior distribution. The entries at `dim1` and `dim2`
will be ignored.
limits: Bounds within which to evaluate the density. Shape (dim_theta, 2).
dim1: First dimension along which to evaluate the conditional.
dim2: Second dimension along which to evaluate the conditional.
resolution: Resolution of the grid along which the conditional density is
evaluated.
eps_margins1: We will evaluate the posterior along `dim1` at
`limits[0]+eps_margins` until `limits[1]-eps_margins`. This avoids
evaluations potentially exactly at the prior bounds.
eps_margins2: We will evaluate the posterior along `dim2` at
`limits[0]+eps_margins` until `limits[1]-eps_margins`. This avoids
evaluations potentially exactly at the prior bounds.
Returns: Conditional probabilities. If `dim1 == dim2`, this will have shape
(resolution). If `dim1 != dim2`, it will have shape (resolution, resolution).
"""
condition = ensure_theta_batched(condition)
theta_grid_dim1 = torch.linspace(
float(limits[dim1, 0] + eps_margins1),
float(limits[dim1, 1] - eps_margins1),
resolution,
)
theta_grid_dim2 = torch.linspace(
float(limits[dim2, 0] + eps_margins2),
float(limits[dim2, 1] - eps_margins2),
resolution,
)
if dim1 == dim2:
repeated_condition = condition.repeat(resolution, 1)
repeated_condition[:, dim1] = theta_grid_dim1
log_probs_on_grid = density.log_prob(repeated_condition)
else:
repeated_condition = condition.repeat(resolution**2, 1)
repeated_condition[:, dim1] = theta_grid_dim1.repeat(resolution)
repeated_condition[:, dim2] = torch.repeat_interleave(
theta_grid_dim2, resolution)
log_probs_on_grid = density.log_prob(repeated_condition)
log_probs_on_grid = torch.reshape(log_probs_on_grid,
(resolution, resolution))
# Subtract maximum for numerical stability.
return torch.exp(log_probs_on_grid - torch.max(log_probs_on_grid))
def plot(self,
axes=None,
block=False,
title=None,
plotting=False,
Ndiv=41,
showtickslabels=True,
showticks=True,
xlabel=None,
ylabel=None,
clear_axes=True,
legend=False,
labelsize=None,
normalize=False,
colorbar=False,
color=None,
label=None,
local_axes=None,
x_next=None,
alpha_next=None):
if plotting == False:
return None
if self.dim > 1:
return None
if local_axes is None and axes is None:
self.fig, (local_axes) = plt.subplots(1,
1,
sharex=True,
figsize=(10, 7))
elif local_axes is None:
local_axes = axes
elif axes is None:
pass # If the internal axes already have some value, and no new axes passed, do nothing
elif local_axes is not None and axes is not None:
local_axes = axes
local_pp = PlotProbability()
if x_next is not None and alpha_next is not None:
x_next_local = x_next
alpha_next_local = alpha_next
else:
x_next_local = None
alpha_next_local = 1.0
test_x_vec = torch.linspace(0.0, 1.0, Ndiv)[:, None]
test_x_vec = test_x_vec.unsqueeze(
1
) # Make this [Ntest x q x dim] = [n_batches x n_design_points x dim], with q=1 -> Double-check in the documentation!
var_vec = self.forward(X=test_x_vec).detach().cpu().numpy()
if self.dim == 1:
local_axes = local_pp.plot_acquisition_function(
var_vec=var_vec,
xpred_vec=test_x_vec.squeeze(1),
x_next=x_next_local,
acqui_next=alpha_next_local,
xlabel=xlabel,
ylabel=ylabel,
title=title,
legend=legend,
axes=local_axes,
clear_axes=clear_axes,
xlim=np.array([0., 1.]),
block=block,
labelsize=labelsize,
showtickslabels=showtickslabels,
showticks=showticks,
what2plot=self.iden,
color=color,
ylim=None)
plt.pause(0.25)
elif self.dim == 2:
if self.x_next is not None:
Xs = np.atleast_2d(self.x_next)
else:
Xs = self.x_next
local_axes = local_pp.plot_GP_2D_single(
var_vec=var_vec,
Ndiv_dim=Ndiv * np.ones(self.dim, dtype=np.int64),
Xs=Xs,
Ys=self.alpha_next,
x_label=xlabel,
y_label=ylabel,
title=title,
axes=local_axes,
clear_axes=clear_axes,
legend=legend,
block=block,
colorbar=colorbar,
color_Xs="gold")
plt.pause(0.25)
return local_axes
'''
use: comment out 2 out of 3 in the main codes and run
'''
import torch
import numpy as np
import torch.nn.functional as F
from torch.autograd import Variable
import matplotlib.pyplot as plt
batch_size = 200
x = torch.unsqueeze(torch.linspace(-5, 5, batch_size), dim=1)
y = x.pow(2) + torch.rand(x.size())
x, y = Variable(x), Variable(y)
def save():
model = torch.nn.Sequential(
torch.nn.Linear(1, 10),
torch.nn.ReLU(),
torch.nn.Linear(10, 1),
)
optimizer = torch.optim.SGD(model.parameters(), lr=0.01)
loss_fn = torch.nn.MSELoss()
plt.ion()
plt.show()
for epoch in range(1000 + 1):
y_pred = model(x)
def gen_encoder_output_proposals(self, memory, memory_padding_mask,
spatial_shapes):
"""Generate proposals from encoded memory.
Args:
memory (Tensor) : The output of encoder,
has shape (bs, num_key, embed_dim). num_key is
equal the number of points on feature map from
all level.
memory_padding_mask (Tensor): Padding mask for memory.
has shape (bs, num_key).
spatial_shapes (Tensor): The shape of all feature maps.
has shape (num_level, 2).
Returns:
tuple: A tuple of feature map and bbox prediction.
- output_memory (Tensor): The input of decoder, \
has shape (bs, num_key, embed_dim). num_key is \
equal the number of points on feature map from \
all levels.
- output_proposals (Tensor): The normalized proposal \
after a inverse sigmoid, has shape \
(bs, num_keys, 4).
"""
N, S, C = memory.shape
proposals = []
_cur = 0
for lvl, (H, W) in enumerate(spatial_shapes):
mask_flatten_ = memory_padding_mask[:, _cur:(_cur + H * W)].view(
N, H, W, 1)
valid_H = torch.sum(~mask_flatten_[:, :, 0, 0], 1)
valid_W = torch.sum(~mask_flatten_[:, 0, :, 0], 1)
grid_y, grid_x = torch.meshgrid(
torch.linspace(0,
H - 1,
H,
dtype=torch.float32,
device=memory.device),
torch.linspace(0,
W - 1,
W,
dtype=torch.float32,
device=memory.device))
grid = torch.cat([grid_x.unsqueeze(-1), grid_y.unsqueeze(-1)], -1)
scale = torch.cat([valid_W.unsqueeze(-1),
valid_H.unsqueeze(-1)], 1).view(N, 1, 1, 2)
grid = (grid.unsqueeze(0).expand(N, -1, -1, -1) + 0.5) / scale
wh = torch.ones_like(grid) * 0.05 * (2.0**lvl)
proposal = torch.cat((grid, wh), -1).view(N, -1, 4)
proposals.append(proposal)
_cur += (H * W)
output_proposals = torch.cat(proposals, 1)
output_proposals_valid = ((output_proposals > 0.01) &
(output_proposals < 0.99)).all(-1,
keepdim=True)
output_proposals = torch.log(output_proposals / (1 - output_proposals))
output_proposals = output_proposals.masked_fill(
memory_padding_mask.unsqueeze(-1), float('inf'))
output_proposals = output_proposals.masked_fill(
~output_proposals_valid, float('inf'))
output_memory = memory
output_memory = output_memory.masked_fill(
memory_padding_mask.unsqueeze(-1), float(0))
output_memory = output_memory.masked_fill(~output_proposals_valid,
float(0))
output_memory = self.enc_output_norm(self.enc_output(output_memory))
return output_memory, output_proposals
def forward(self, states):
ns = torch.linspace(0, 1, states.size(0)).view((-1, 1))
x = torch.cat((states, ns), 1)
return self.linear(x)
DemodFs_func = torch.zeros(model.N,
model.K,
device=device,
dtype=dtype,
requires_grad=False)
ModFs = torch.zeros(model.N,
1,
device=device,
dtype=dtype,
requires_grad=False)
DemodFs = torch.zeros(model.N,
model.K,
device=device,
dtype=dtype,
requires_grad=False)
p = torch.linspace(0, 2 * math.pi, model.N, device=device)
for order in range(0, model.order):
ModFs_func[:, 0] += model.alpha_mod[0, order] * torch.cos(
(p + model.phi_mod[0, order]) * (order + 1))
for k in range(0, model.K):
DemodFs_func[:, k] += model.alpha_demod[k, order] * torch.cos(
(p + model.phi_demod[k, order]) * (order + 1))
# Normalize ModFs and DemodFs
min_ModFs_func, _ = torch.min(ModFs_func, dim=0)
ModFs = ModFs_func - min_ModFs_func # ModFs can't be lower than zero (negative light)
min_DemodFs_func, _ = torch.min(DemodFs_func, dim=0)
max_DemodFs_func, _ = torch.max(DemodFs_func, dim=0)
DemodFs = (DemodFs_func - min_DemodFs_func) / (
max_DemodFs_func - min_DemodFs_func) # DemodFs can only be 0->1
ModFs = ModFs.repeat(1, model.K)
def test(
data,
weights=None,
batch_size=32,
imgsz=640,
conf_thres=0.001,
iou_thres=0.6, # for NMS
save_json=False,
single_cls=False,
augment=False,
verbose=False,
model=None,
dataloader=None,
save_dir=Path(''), # for saving images
save_txt=False, # for auto-labelling
save_hybrid=False, # for hybrid auto-labelling
save_conf=False, # save auto-label confidences
plots=True,
log_imgs=0): # number of logged images
# li = []
# temp_path = "/content/drive/MyDrive/00Colab Notebooks/07Datasets/COCO/cocomax/"
# for i in os.listdir(temp_path):
# li.append(cv2.imread(temp_path + i))
# logger.info("log:{}{}".format(len(li), li[0].shape))
# Initialize/load model and set device
training = model is not None
if training: # called by train.py
device = next(model.parameters()).device # get model device
else: # called directly
###############コメントアウト
# set_logging()
device = select_device(opt.device, batch_size=batch_size)
# Directories
save_dir = Path(
increment_path(Path(opt.project) / opt.name,
exist_ok=opt.exist_ok)) # increment run
(save_dir / 'labels' if save_txt else save_dir).mkdir(
parents=True, exist_ok=True) # make dir
# Load model
model = attempt_load(weights, map_location=device) # load FP32 model
imgsz = check_img_size(imgsz, s=model.stride.max()) # check img_size
# Multi-GPU disabled, incompatible with .half() https://github.com/ultralytics/yolov5/issues/99
# if device.type != 'cpu' and torch.cuda.device_count() > 1:
# model = nn.DataParallel(model)
# Half
half = device.type != 'cpu' # half precision only supported on CUDA
if half:
model.half()
logger.info("gpu_model:{}".format(gpu_used.gpuinfo()))
# Configure
model.eval()
###############コメントアウト
# is_coco = data.endswith('coco.yaml') # is COCO dataset
with open(data) as f:
data = yaml.load(f, Loader=yaml.FullLoader) # model dict
check_dataset(data) # check
nc = 1 if single_cls else int(data['nc']) # number of classes
iouv = torch.linspace(0.5, 0.95,
10).to(device) # iou vector for [email protected]:0.95
niou = iouv.numel()
# Logging
log_imgs, wandb = min(log_imgs, 100), None # ceil
try:
import wandb # Weights & Biases
except ImportError:
log_imgs = 0
# Dataloader
if not training:
#下二行、デバック
img = torch.zeros((1, 3, imgsz, imgsz), device=device) # init img
_ = model(img.half() if half else img
) if device.type != 'cpu' else None # run once
path = data['test'] if opt.task == 'test' else data[
'val'] # path to val/test images
##################追加
path = "/content/drive/MyDrive/00Colab Notebooks/07Datasets/COCO/cocomax"
logger.info(path)
# dataloader = create_dataloader(path, imgsz, batch_size, model.stride.max(), opt, pad=0.5, rect=True)[0]
dataloader = create_dataloader(path,
imgsz,
batch_size,
model.stride.max(),
opt,
pad=0.5,
rect=True)[0]
seen = 0
confusion_matrix = ConfusionMatrix(nc=nc)
names = {
k: v
for k, v in enumerate(
model.names if hasattr(model, 'names') else model.module.names)
}
# coco91class = coco80_to_coco91_class()
s = ('%20s' + '%12s' * 6) % ('Class', 'Images', 'Targets', 'P', 'R',
'[email protected]', '[email protected]:.95')
p, r, f1, mp, mr, map50, map, t0, t1, t_ex1, t_ex2,load_time = 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., []
loss = torch.zeros(3, device=device)
jdict, stats, ap, ap_class, wandb_images = [], [], [], [], []
time_veri.time_start()
for batch_i, (img, targets, paths,
shapes) in enumerate(tqdm(dataloader, desc=s)):
time_veri2.start_point(time_veri2.t_point0)
#前処理
time_veri.start_point(time_veri.t_point1)
gpu_used.gpu_clear()
img = img.to(device, non_blocking=True)
img = img.half() if half else img.float() # uint8 to fp16/32
img /= 255.0 # 0 - 255 to 0.0 - 1.0
targets = targets.to(device)
nb, _, height, width = img.shape # batch size, channels, height, width
time_veri.end_point(time_veri.t_point1)
# gpu_used.gpuinfo()
with torch.no_grad():
time_veri.start_point(time_veri.t_point2)
# Run model
inf_out, train_out = model(
img, augment=augment) # inference and training outputs
time_veri.end_point(time_veri.t_point2)
###################################コメントアウト
# # Compute loss
# if training:
# loss += compute_loss([x.float() for x in train_out], targets, model)[1][:3] # box, obj, cls
# Run NMS
time_veri.start_point(time_veri.t_point3)
targets[:, 2:] *= torch.Tensor([width, height, width,
height]).to(device) # to pixels
lb = [targets[targets[:, 0] == i, 1:] for i in range(nb)
] if save_hybrid else [] # for autolabelling
output = non_max_suppression(inf_out,
conf_thres=conf_thres,
iou_thres=iou_thres,
labels=lb)
time_veri.end_point(time_veri.t_point3)
# Statistics per image
time_veri.start_point(time_veri.t_point4)
for si, pred in enumerate(output):
##############################コメントアウト
# labels = targets[targets[:, 0] == si, 1:]
# nl = len(labels)
# tcls = labels[:, 0].tolist() if nl else [] # target class
# path = Path(paths[si])
# seen += 1
# if len(pred) == 0:
# if nl:
# stats.append((torch.zeros(0, niou, dtype=torch.bool), torch.Tensor(), torch.Tensor(), tcls))
# continue
# Predictions→元画像の大きさにリサイズ
predn = pred.clone()
# scale_coords(img[si].shape[1:], predn[:, :4], shapes[si][0], shapes[si][1]) # native-space pred
########################引数、元座標うけとり
ex_coords = scale_coords(img[si].shape[1:], predn[:, :4],
shapes[si][0],
shapes[si][1]) # native-space pred
###############################追加
# continue
time_veri.end_point(time_veri.t_point4)
time_veri2.end_point(time_veri2.t_point0)
################ ################ ################ ################ ################ ################ ################ ################ ################ ################ ################ ################ ################ ################ ################ ################ ################
# # Append to text file
# if save_txt:
# gn = torch.tensor(shapes[si][0])[[1, 0, 1, 0]] # normalization gain whwh
# for *xyxy, conf, cls in predn.tolist():
# xywh = (xyxy2xywh(torch.tensor(xyxy).view(1, 4)) / gn).view(-1).tolist() # normalized xywh
# line = (cls, *xywh, conf) if save_conf else (cls, *xywh) # label format
# with open(save_dir / 'labels' / (path.stem + '.txt'), 'a') as f:
# f.write(('%g ' * len(line)).rstrip() % line + '\n')
# # W&B logging
# if plots and len(wandb_images) < log_imgs:
# box_data = [{"position": {"minX": xyxy[0], "minY": xyxy[1], "maxX": xyxy[2], "maxY": xyxy[3]},
# "class_id": int(cls),
# "box_caption": "%s %.3f" % (names[cls], conf),
# "scores": {"class_score": conf},
# "domain": "pixel"} for *xyxy, conf, cls in pred.tolist()]
# boxes = {"predictions": {"box_data": box_data, "class_labels": names}} # inference-space
# wandb_images.append(wandb.Image(img[si], boxes=boxes, caption=path.name))
# # Append to pycocotools JSON dictionary
# if save_json:
# # [{"image_id": 42, "category_id": 18, "bbox": [258.15, 41.29, 348.26, 243.78], "score": 0.236}, ...
# image_id = int(path.stem) if path.stem.isnumeric() else path.stem
# box = xyxy2xywh(predn[:, :4]) # xywh
# box[:, :2] -= box[:, 2:] / 2 # xy center to top-left corner
# for p, b in zip(pred.tolist(), box.tolist()):
# jdict.append({'image_id': image_id,
# 'category_id': coco91class[int(p[5])] if is_coco else int(p[5]),
# 'bbox': [round(x, 3) for x in b],
# 'score': round(p[4], 5)})
# # Assign all predictions as incorrect
# correct = torch.zeros(pred.shape[0], niou, dtype=torch.bool, device=device)
# if nl:
# detected = [] # target indices
# tcls_tensor = labels[:, 0]
# # target boxes
# tbox = xywh2xyxy(labels[:, 1:5])
# scale_coords(img[si].shape[1:], tbox, shapes[si][0], shapes[si][1]) # native-space labels
# if plots:
# confusion_matrix.process_batch(pred, torch.cat((labels[:, 0:1], tbox), 1))
# # Per target class
# for cls in torch.unique(tcls_tensor):
# ti = (cls == tcls_tensor).nonzero(as_tuple=False).view(-1) # prediction indices
# pi = (cls == pred[:, 5]).nonzero(as_tuple=False).view(-1) # target indices
# # Search for detections
# if pi.shape[0]:
# # Prediction to target ious
# ious, i = box_iou(predn[pi, :4], tbox[ti]).max(1) # best ious, indices
# # Append detections
# detected_set = set()
# for j in (ious > iouv[0]).nonzero(as_tuple=False):
# d = ti[i[j]] # detected target
# if d.item() not in detected_set:
# detected_set.add(d.item())
# detected.append(d)
# correct[pi[j]] = ious[j] > iouv # iou_thres is 1xn
# if len(detected) == nl: # all targets already located in image
# break
# # Append statistics (correct, conf, pcls, tcls)
# stats.append((correct.cpu(), pred[:, 4].cpu(), pred[:, 5].cpu(), tcls))
################ ################ ################ ################ ################ ################ ################ ################ ################ ################ ################ ################ ################ ################
#####################追加
# time_veri.end_point(time_veri.t_point4)
# time_veri2.start_point(time_veri2.t_point0)
# continue
# # Plot images
# if plots and batch_i < 3:
# f = save_dir / f'test_batch{batch_i}_labels.jpg' # labels
# Thread(target=plot_images, args=(img, targets, paths, f, names), daemon=True).start()
# f = save_dir / f'test_batch{batch_i}_pred.jpg' # predictions
# Thread(target=plot_images, args=(img, output_to_target(output), paths, f, names), daemon=True).start()
#####################################追加##################################
time_veri.time_end()
logger.info("log:{}".format("処理中断終了"))
logger.info("for文全体:{}".format(time_veri.t_end - time_veri.t_start))
temp = (time_veri.t_end - time_veri.t_start)
logger.info("for文の中:{}".format(time_veri2.time_sum_list(return_all=True)))
logger.info("ミニバッチのロード時間:{}".format(temp - time_veri2.all))
logger.info("①～④の処理時間:{}sum：{}".format(time_veri.time_sum_list(),
time_veri.all))
logger.info("for文の中対fps:{}".format(1 / (time_veri2.all / 512)))
temp = time_veri.t_end - time_veri.t_start
logger.info("for文全体対fps_all:{}".format(1 / (temp / 512)))
logger.info("gpu_exit:{}".format(gpu_used.gpuinfo()))
logger.info("gpu_max,list:{}{}".format(gpu_used.used_max,
gpu_used.used_list))
import sys
sys.exit()
#####################################追加##################################
# Compute statistics
stats = [np.concatenate(x, 0) for x in zip(*stats)] # to numpy
if len(stats) and stats[0].any():
p, r, ap, f1, ap_class = ap_per_class(*stats,
plot=plots,
save_dir=save_dir,
names=names)
p, r, ap50, ap = p[:, 0], r[:, 0], ap[:, 0], ap.mean(
1) # [P, R, [email protected], [email protected]:0.95]
mp, mr, map50, map = p.mean(), r.mean(), ap50.mean(), ap.mean()
nt = np.bincount(stats[3].astype(np.int64),
minlength=nc) # number of targets per class
else:
nt = torch.zeros(1)
# Print results
pf = '%20s' + '%12.3g' * 6 # print format
print(pf % ('all', seen, nt.sum(), mp, mr, map50, map))
# Print results per class
if verbose and nc > 1 and len(stats):
for i, c in enumerate(ap_class):
print(pf % (names[c], seen, nt[c], p[i], r[i], ap50[i], ap[i]))
# Print speeds
t = tuple(x / seen * 1E3
for x in (t0, t1, t0 + t1)) + (imgsz, imgsz, batch_size) # tuple
if not training:
print(
'Speed: %.1f/%.1f/%.1f ms inference/NMS/total per %gx%g image at batch-size %g'
% t)
# Plots
if plots:
confusion_matrix.plot(save_dir=save_dir, names=list(names.values()))
if wandb and wandb.run:
wandb.log({"Images": wandb_images})
wandb.log({
"Validation": [
wandb.Image(str(f), caption=f.name)
for f in sorted(save_dir.glob('test*.jpg'))
]
})
# Save JSON
if save_json and len(jdict):
w = Path(weights[0] if isinstance(weights, list) else weights
).stem if weights is not None else '' # weights
anno_json = '../coco/annotations/instances_val2017.json' # annotations json
pred_json = str(save_dir / f"{w}_predictions.json") # predictions json
print('\nEvaluating pycocotools mAP... saving %s...' % pred_json)
with open(pred_json, 'w') as f:
json.dump(jdict, f)
try: # https://github.com/cocodataset/cocoapi/blob/master/PythonAPI/pycocoEvalDemo.ipynb
from pycocotools.coco import COCO
from pycocotools.cocoeval import COCOeval
anno = COCO(anno_json) # init annotations api
pred = anno.loadRes(pred_json) # init predictions api
eval = COCOeval(anno, pred, 'bbox')
if is_coco:
eval.params.imgIds = [
int(Path(x).stem) for x in dataloader.dataset.img_files
] # image IDs to evaluate
eval.evaluate()
eval.accumulate()
eval.summarize()
map, map50 = eval.stats[:
2] # update results ([email protected]:0.95, [email protected])
except Exception as e:
print(f'pycocotools unable to run: {e}')
# Return results
if not training:
s = f"\n{len(list(save_dir.glob('labels/*.txt')))} labels saved to {save_dir / 'labels'}" if save_txt else ''
print(f"Results saved to {save_dir}{s}")
model.float() # for training
maps = np.zeros(nc) + map
for i, c in enumerate(ap_class):
maps[c] = ap[i]
return (mp, mr, map50, map,
*(loss.cpu() / len(dataloader)).tolist()), maps, t
#coding: utf-8
import torch
from torch.autograd import Variable
import matplotlib.pyplot as plt
x = torch.unsqueeze(torch.linspace(-1, 1, 100), dim=1) # x data (tensor), shape=(100, 1)
y = x.pow(2) + 0.2*torch.rand(x.size()) # noisy y data (tensor), shape=(100, 1)
x, y = Variable(x, requires_grad=False), Variable(y, requires_grad=False)
def save():
net = torch.nn.Sequential(
torch.nn.Linear(1,30),
torch.nn.ReLU(),
torch.nn.Linear(30,30),
torch.nn.ReLU(),
torch.nn.Linear(30,1)
)
# optm = torch.optim.SGD(net.parameters(),lr=5e-2)
optm = torch.optim.Adam(net.parameters())
loss_func = torch.nn.MSELoss()
for i in range(100):
pred = net(x)
loss = loss_func(pred,y)
optm.zero_grad()
loss.backward()
optm.step()
torch.save(net, 'net.pkl') # 保存整个网络
def __init__(self, model_config: AttrDict, model_name: str):
super().__init__()
assert model_config.INPUT_TYPE in ["rgb",
"bgr"], "Input type not supported"
trunk_config = copy.deepcopy(model_config.TRUNK.VISION_TRANSFORMERS)
logging.info("Building model: Vision Transformer from yaml config")
# Hacky workaround
trunk_config = AttrDict(
{k.lower(): v
for k, v in trunk_config.items()})
img_size = trunk_config.image_size
patch_size = trunk_config.patch_size
in_chans = 3
embed_dim = trunk_config.hidden_dim
depth = trunk_config.num_layers
num_heads = trunk_config.num_heads
mlp_ratio = 4.0
qkv_bias = trunk_config.qkv_bias
qk_scale = trunk_config.qk_scale
drop_rate = trunk_config.dropout_rate
attn_drop_rate = trunk_config.attention_dropout_rate
drop_path_rate = trunk_config.drop_path_rate
hybrid_backbone_string = None
# TODO Implement hybrid backbones
if "HYBRID" in trunk_config.keys():
hybrid_backbone_string = trunk_config.HYBRID
norm_layer = nn.LayerNorm
self.num_features = (
self.embed_dim
) = embed_dim # num_features for consistency with other models
# TODO : Enable Hybrid Backbones
if hybrid_backbone_string:
self.patch_embed = globals()[hybrid_backbone_string](
out_dim=embed_dim, img_size=img_size)
# if hybrid_backbone is not None:
# self.patch_embed = HybridEmbed(
# hybrid_backbone,
# img_size=img_size,
# in_chans=in_chans,
# embed_dim=embed_dim,
# )
else:
self.patch_embed = PatchEmbed(
img_size=img_size,
patch_size=patch_size,
in_chans=in_chans,
embed_dim=embed_dim,
)
num_patches = self.patch_embed.num_patches
self.class_token = nn.Parameter(torch.zeros(1, 1, embed_dim))
self.pos_embedding = nn.Parameter(
torch.zeros(1, num_patches + 1, embed_dim))
self.pos_drop = nn.Dropout(p=drop_rate)
dpr = [x.item() for x in torch.linspace(0, drop_path_rate, depth)
] # stochastic depth decay rule
self.blocks = nn.ModuleList([
Block(
dim=embed_dim,
num_heads=num_heads,
mlp_ratio=mlp_ratio,
qkv_bias=qkv_bias,
qk_scale=qk_scale,
drop=drop_rate,
attn_drop=attn_drop_rate,
drop_path=dpr[i],
norm_layer=norm_layer,
) for i in range(depth)
])
self.norm = norm_layer(embed_dim)
# NOTE as per official impl, we could have a pre-logits
# representation dense layer + tanh here
# self.repr = nn.Linear(embed_dim, representation_size)
# self.repr_act = nn.Tanh()
trunc_normal_(self.pos_embedding, std=0.02)
trunc_normal_(self.class_token, std=0.02)
self.apply(self._init_weights)
def __init__(self, model_config, model_name):
super().__init__()
trunk_config = copy.deepcopy(model_config.TRUNK.CONVIT)
trunk_config.update(model_config.TRUNK.VISION_TRANSFORMERS)
logging.info("Building model: ConViT from yaml config")
# Hacky workaround
trunk_config = AttrDict(
{k.lower(): v
for k, v in trunk_config.items()})
image_size = trunk_config.image_size
patch_size = trunk_config.patch_size
classifier = trunk_config.classifier
assert image_size % patch_size == 0, "Input shape indivisible by patch size"
assert classifier in ["token", "gap"], "Unexpected classifier mode"
n_gpsa_layers = trunk_config.n_gpsa_layers
class_token_in_local_layers = trunk_config.class_token_in_local_layers
mlp_dim = trunk_config.mlp_dim
embed_dim = trunk_config.hidden_dim
locality_dim = trunk_config.locality_dim
attention_dropout_rate = trunk_config.attention_dropout_rate
dropout_rate = trunk_config.dropout_rate
drop_path_rate = trunk_config.drop_path_rate
num_layers = trunk_config.num_layers
locality_strength = trunk_config.locality_strength
num_heads = trunk_config.num_heads
qkv_bias = trunk_config.qkv_bias
qk_scale = trunk_config.qk_scale
use_local_init = trunk_config.use_local_init
hybrid_backbone = None
if "hybrid" in trunk_config.keys():
hybrid_backbone = trunk_config.hybrid
in_chans = 3
# TODO: Make this configurable
norm_layer = nn.LayerNorm
self.classifier = classifier
self.n_gpsa_layers = n_gpsa_layers
self.class_token_in_local_layers = class_token_in_local_layers
# For consistency with other models
self.num_features = self.embed_dim = self.hidden_dim = embed_dim
self.locality_dim = locality_dim
# Hybrid backbones not tested
if hybrid_backbone is not None:
self.patch_embed = HybridEmbed(
hybrid_backbone,
img_size=image_size,
in_chans=in_chans,
embed_dim=embed_dim,
)
else:
self.patch_embed = PatchEmbed(
img_size=image_size,
patch_size=patch_size,
in_chans=in_chans,
embed_dim=embed_dim,
)
seq_length = (image_size // patch_size)**2
self.seq_length = seq_length
self.class_token = nn.Parameter(torch.zeros(1, 1, embed_dim))
self.pos_embedding = nn.Parameter(torch.zeros(1, seq_length,
embed_dim))
self.pos_drop = nn.Dropout(p=dropout_rate)
if class_token_in_local_layers:
seq_length += 1
# stochastic depth decay rule
dpr = [x.item() for x in torch.linspace(0, drop_path_rate, num_layers)]
layers = []
for i in range(num_layers):
if i < self.n_gpsa_layers:
if locality_strength > 0:
layer_locality_strength = locality_strength
else:
layer_locality_strength = 1 / (i + 1)
layers.append(
AttentionBlock(
attention_module=GPSA,
embed_dim=embed_dim,
num_heads=num_heads,
mlp_dim=mlp_dim,
qkv_bias=qkv_bias,
qk_scale=qk_scale,
dropout_rate=dropout_rate,
attention_dropout_rate=attention_dropout_rate,
drop_path_rate=dpr[i],
norm_layer=norm_layer,
locality_strength=layer_locality_strength,
locality_dim=self.locality_dim,
use_local_init=use_local_init,
))
else:
layers.append(
AttentionBlock(
attention_module=SelfAttention,
embed_dim=embed_dim,
num_heads=num_heads,
mlp_dim=mlp_dim,
qkv_bias=qkv_bias,
qk_scale=qk_scale,
dropout_rate=dropout_rate,
attention_dropout_rate=attention_dropout_rate,
drop_path_rate=dpr[i],
norm_layer=norm_layer,
))
self.blocks = nn.ModuleList(layers)
self.norm = norm_layer(embed_dim)
trunc_normal_(self.pos_embedding, std=0.02)
trunc_normal_(self.class_token, std=0.02)
self.apply(self._init_weights)
"""
View more, visit my tutorial page: https://morvanzhou.github.io/tutorials/
My Youtube Channel: https://www.youtube.com/user/MorvanZhou
Dependencies:
torch: 0.1.11
"""
import torch
import torch.utils.data as Data
torch.manual_seed(1) # reproducible
BATCH_SIZE = 5
# BATCH_SIZE = 8
x = torch.linspace(1, 10, 10) # this is x data (torch tensor)
y = torch.linspace(10, 1, 10) # this is y data (torch tensor)
torch_dataset = Data.TensorDataset(x, y)
loader = Data.DataLoader(
dataset=torch_dataset, # torch TensorDataset format
batch_size=BATCH_SIZE, # mini batch size
shuffle=True, # random shuffle for training
num_workers=2, # subprocesses for loading data
)
for epoch in range(3): # train entire dataset 3 times
for step, (batch_x, batch_y) in enumerate(loader): # for each training step
# train your data...
print('Epoch: ', epoch, '| Step: ', step, '| batch x: ',
batch_x.numpy(), '| batch y: ', batch_y.numpy())
import torch
import torch.utils.data as Data
BATCH_SIZE = 5
x = torch.linspace(1, 10, 10) # this is x data (torch tensor)
y = torch.linspace(10, 1, 10) # this is y data (torch tensor)
torch_dataset = Data.TensorDataset(x, y)
loader = Data.DataLoader(
dataset=torch_dataset,
batch_size=BATCH_SIZE,
shuffle=True,
num_workers=2,
)
if __name__ == '__main__':
for epoch in range(3):
for step, (batch_x, batch_y) in enumerate(loader):
# training....
print('Epoch: ', epoch, '| Step: ', step, '| batch x: ',
batch_x.numpy(), '| batch y: ', batch_y.numpy())
def get_region_boxes_v2(output,
n_models,
conf_thresh,
num_classes,
anchors,
num_anchors,
only_objectness=1,
validation=False):
cs = n_models
nA = num_anchors
nC = num_classes
anchor_step = int(len(anchors) / num_anchors)
if len(output.shape) == 3:
output = output.unsqueeze(0)
batch = output.shape[0]
assert (output.shape[1] == (5 + num_classes) * num_anchors)
nH = h = output.shape[2]
nW = w = output.shape[3]
assert (batch % n_models == 0)
bs = batch // n_models
t0 = time.time()
all_boxes = []
# import pdb; pdb.set_trace()
cls = output.view(output.shape[0], nA, (5 + nC), nH, nW)
cls = cls.index_select(2,
torch.linspace(5, 5 + nC - 1, nC).long()).squeeze()
cls = cls.view(bs, cs,
nA * nC * nH * nW).transpose(1, 2).contiguous().view(
bs * nA * nC * nH * nW, cs)
normfn = torch.nn.Softmax(dim=1)
print("=======time:", float(round((time.time() - t0) * 1000)))
# cls = torch.nn.Softmax(dim=1)(Variable(cls)).data
cls = normfn(Variable(cls)).data
cls_confs = cls.view(bs, nA * nC * nH * nW,
cs).transpose(1, 2).contiguous().view(
bs * cs * nA, nC, nH * nW).transpose(1, 2).view(
bs * cs * nA * nH * nW, nC)
output = output.view(batch * num_anchors, 5 + num_classes,
h * w).transpose(0, 1).contiguous().view(
5 + num_classes, batch * num_anchors * h * w)
grid_x = torch.linspace(0, w - 1, w).repeat(h, 1).repeat(
batch * num_anchors, 1, 1).view(batch * num_anchors * h * w)
grid_y = torch.linspace(0, h - 1, h).repeat(w, 1).t().repeat(
batch * num_anchors, 1, 1).view(batch * num_anchors * h * w)
xs = torch.sigmoid(output[0]) + grid_x
ys = torch.sigmoid(output[1]) + grid_y
anchor_w = torch.Tensor(anchors).view(num_anchors,
anchor_step).index_select(
1, torch.LongTensor([0]))
anchor_h = torch.Tensor(anchors).view(num_anchors,
anchor_step).index_select(
1, torch.LongTensor([1]))
anchor_w = anchor_w.repeat(batch, 1).repeat(1, 1, h * w).view(
batch * num_anchors * h * w)
anchor_h = anchor_h.repeat(batch, 1).repeat(1, 1, h * w).view(
batch * num_anchors * h * w)
ws = torch.exp(output[2]) * anchor_w
hs = torch.exp(output[3]) * anchor_h
det_confs = torch.sigmoid(output[4])
# cls_confs = torch.nn.Softmax()(Variable(output[5:5+num_classes].transpose(0,1))).data
#print("==========",cls_confs.shape)
#cls_confs=torch.sigmoid(cls_confs)
print("==========", cls_confs.shape)
cls_max_confs, cls_max_ids = torch.max(cls_confs, 1) #33800L, 1L
cls_max_confs = cls_max_confs.view(-1)
cls_max_ids = cls_max_ids.view(-1)
t1 = time.time()
sz_hw = h * w
sz_hwa = sz_hw * num_anchors
det_confs = convert2cpu(det_confs)
cls_max_confs = convert2cpu(cls_max_confs)
cls_max_ids = convert2cpu_long(cls_max_ids)
xs = convert2cpu(xs)
ys = convert2cpu(ys)
ws = convert2cpu(ws)
hs = convert2cpu(hs)
if validation:
cls_confs = convert2cpu(cls_confs.view(-1, num_classes))
t2 = time.time()
#det_confs shape:[1,batch*num_anchors*h*w]
#print(batch,h,w,det_confs.shape,cls_max_confs.shape,cls_confs.shape,num_classes)
#(1280, 13, 13, (1081600L,), (1081600L,), (1081600L, 1L), 1)
for b in range(batch): #40
boxes = []
for cy in range(h): #13
for cx in range(w): #13
for i in range(num_anchors): #5
ind = b * sz_hwa + i * sz_hw + cy * w + cx
det_conf = det_confs[ind]
if only_objectness:
conf = det_confs[ind]
else: ##object confidence * cls confidence
conf = det_confs[ind] * cls_max_confs[ind]
#print(b,cy,cx,i,conf)
if conf > conf_thresh:
bcx = xs[ind]
bcy = ys[ind]
bw = ws[ind]
bh = hs[ind]
cls_max_conf = cls_max_confs[ind]
cls_max_id = cls_max_ids[ind]
box = [
bcx / w, bcy / h, bw / w, bh / h, det_conf,
cls_max_conf, cls_max_id
]
if (not only_objectness) and validation:
for c in range(num_classes):
tmp_conf = cls_confs[ind][c]
if c != cls_max_id and tmp_conf > conf_thresh: #det_confs[ind]*tmp_conf > conf_thresh:
box.append(tmp_conf)
box.append(c)
boxes.append(box)
#print("====================================",len(boxes))
all_boxes.append(boxes)
t3 = time.time()
if False:
print('---------------------------------')
print('matrix computation : %f' % (t1 - t0))
print(' gpu to cpu : %f' % (t2 - t1))
print(' boxes filter : %f' % (t3 - t2))
print('---------------------------------')
return all_boxes
""" Know more, visit 莫烦Python: https://morvanzhou.github.io/tutorials/ My Youtube Channel: https://www.youtube.com/user/MorvanZhou Dependencies: torch: 0.1.11 """ import torch import torch.nn.functional as F from torch.autograd import Variable import matplotlib.pyplot as plt # fake data x = torch.linspace(-5, 5, 200) # x data (tensor), shape=(100, 1) x = Variable(x) x_np = x.data.numpy() # following are popular activation functions y_relu = F.relu(x).data.numpy() y_sigmoid = F.sigmoid(x).data.numpy() y_tanh = F.tanh(x).data.numpy() y_softplus = F.softplus(x).data.numpy() # y_softmax = F.softmax(x) softmax is a special kind of activation function, it is about probability # plt to visualize these activation function plt.figure(1, figsize=(8, 6)) plt.subplot(221) plt.plot(x_np, y_relu, c='red', label='relu') plt.ylim((-1, 5)) plt.legend(loc='best')
matplotlib
"""
import torch
import torch.utils.data as Data
import torch.nn.functional as F
from torch.autograd import Variable
import matplotlib.pyplot as plt
# torch.manual_seed(1) # reproducible
LR = 0.01
BATCH_SIZE = 32
EPOCH = 12
# fake dataset
x = torch.unsqueeze(torch.linspace(-1, 1, 1000), dim=1)
y = x.pow(2) + 0.1*torch.normal(torch.zeros(*x.size()))
# plot dataset
plt.scatter(x.numpy(), y.numpy())
plt.show()
# put dateset into torch dataset
torch_dataset = Data.TensorDataset(data_tensor=x, target_tensor=y)
loader = Data.DataLoader(dataset=torch_dataset, batch_size=BATCH_SIZE, shuffle=True, num_workers=2,)
# default network
class Net(torch.nn.Module):
def __init__(self):
super(Net, self).__init__()
G = torch.bmm(grid[:, :, 0, 1].unsqueeze(2), grid[:, 0, :, 0].unsqueeze(1))
G.size()
plt.imshow(G.squeeze())
'''
Grid created by Jason for cropping
'''
# the grid is created by an inner product across a linspace in x & y
N = 1
W, H = 128, 128
scale = 0.7
theta2 = torch.from_numpy(np.array([[0.0, 0.5, 0.2]])).type(torch.float32)
base_grid = [torch.linspace(-1, 1, W).expand(N, W),
torch.linspace(-1, 1, H).expand(N, H)]
m_x = torch.max(torch.zeros_like(base_grid[0]),
1 - torch.abs(theta2[:, 1].unsqueeze(1) - base_grid[0])).unsqueeze(2)
m_y = torch.max(torch.zeros_like(base_grid[1]),
1 - torch.abs(theta2[:, 2].unsqueeze(1) - base_grid[1])).unsqueeze(1)
# the final mask is a batch-matrix-multiple of each inner product
# the dimensions expected here are m_x = [N, W, 1] & m_y = [N, 1, H]
M = torch.bmm(m_x, m_y)
M_clip = torch.clamp(M, M.max() * scale, M.max())
plt.rcParams['figure.figsize'] = 8, 8
fig, axis = plt.subplots(nrows=1, ncols=2)
def __init__(self,
image_size,
border_size,
input_nc=3,
output_nc=3,
ngf=64,
norm_layer=nn.BatchNorm2d,
use_dropout=False,
n_blocks=6,
padding_type='reflect'):
assert (n_blocks >= 0)
super(FDS, self).__init__()
self.input_nc = input_nc
self.output_nc = output_nc
self.ngf = ngf
if type(norm_layer) == functools.partial:
use_bias = norm_layer.func == nn.InstanceNorm2d
else:
use_bias = norm_layer == nn.InstanceNorm2d
# Encoder
encoder = [
nn.ReflectionPad2d(3),
nn.Conv2d(input_nc, ngf, kernel_size=7, padding=0, bias=use_bias),
norm_layer(ngf),
nn.ReLU(True)
]
n_downsampling = 2
for i in range(n_downsampling):
mult = 2**i
encoder += [
nn.Conv2d(ngf * mult,
ngf * mult * 2,
kernel_size=3,
stride=2,
padding=1,
bias=use_bias),
norm_layer(ngf * mult * 2),
nn.ReLU(True)
]
mult = 2**n_downsampling
for i in range(int(n_blocks / 2)):
encoder += [
ResnetBlock(ngf * mult,
padding_type=padding_type,
norm_layer=norm_layer,
use_dropout=use_dropout,
use_bias=use_bias)
]
# 1x1 conv to reduce dimensionality
encoder += [
nn.Conv2d(256, 16, kernel_size=1, padding=0),
nn.BatchNorm2d(16)
]
self.encoder = nn.Sequential(*encoder)
# Decoder
decoder = []
decoder += [
nn.Conv2d(16, 256, kernel_size=1, padding=0),
nn.BatchNorm2d(256)
]
for i in range(int(n_blocks / 2)):
decoder += [
ResnetBlock(ngf * mult,
padding_type=padding_type,
norm_layer=norm_layer,
use_dropout=use_dropout,
use_bias=use_bias)
]
for i in range(n_downsampling):
mult = 2**(n_downsampling - i)
decoder += [ #nn.ConvTranspose2d(ngf * mult, int(ngf * mult / 2),kernel_size=3, stride=2,padding=1, output_padding=1,bias=use_bias),
nn.Upsample(scale_factor=2, mode='nearest'),
nn.Conv2d(ngf * mult,
int(ngf * mult / 2),
kernel_size=3,
stride=1,
padding=1,
bias=use_bias), # output_padding=1,
norm_layer(int(ngf * mult / 2)),
nn.ReLU(True)
]
decoder += [nn.ReflectionPad2d(3)]
decoder += [nn.Conv2d(ngf, output_nc, kernel_size=7, padding=0)]
decoder += [nn.Tanh()]
self.decoder = nn.Sequential(*decoder)
self.drop_out = nn.Dropout(p=0.1)
self.drop_out50 = nn.Dropout(p=0.5)
# small grid to apply warp on the feature maps
yy = torch.linspace(-1, 1, int(image_size[0] / 4))
xx = torch.linspace(-1, 1, int(
(image_size[1] + (border_size * 2)) / 4))
grid_x, grid_y = torch.meshgrid(yy, xx)
self.grid_small = torch.cat(
[grid_y.unsqueeze(2), grid_x.unsqueeze(2)], dim=2).cuda()
def test(model=None, args=None, params=None):
if model is not None:
device = next(model.parameters()).device
else:
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
model = torch.load(os.path.join('weights', 'best.pt'),
device)['model'].float().fuse().eval()
half = device.type != 'cpu'
if half:
model.half()
model.eval()
iou_v = torch.linspace(0.5, 0.95, 10).to(device)
n_iou = iou_v.numel()
file_names = []
with open('../Dataset/Dubai/test.txt') as f:
for file_name in f.readlines():
file_names.append('../Dataset/Dubai/images/' + file_name.rstrip() +
'.jpg')
loader, _ = input_fn(file_names, args.image_size, args.batch_size // 2,
model.head.stride.max(), names, params)
seen = 0
s = ('%10s' * 3) % ('precision', 'recall', 'mAP')
p, r, f1, mp, mr, map50, mean_ap, t0, t1 = 0., 0., 0., 0., 0., 0., 0., 0., 0.
stats, ap, ap_class = [], [], []
for img, targets, paths, shapes in tqdm.tqdm(loader, desc=s):
img = img.to(device, non_blocking=True)
img = img.half() if half else img.float()
img /= 255.0
targets = targets.to(device)
nb, _, height, width = img.shape
wh_wh = torch.Tensor([width, height, width, height]).to(device)
with torch.no_grad():
t = util.time_synchronized()
inf_out, train_out = model(img)
t0 += util.time_synchronized() - t
t = util.time_synchronized()
output = util.non_max_suppression(inf_out, 0.001)
t1 += util.time_synchronized() - t
for si, pred in enumerate(output):
labels = targets[targets[:, 0] == si, 1:]
nl = len(labels)
t_cls = labels[:, 0].tolist() if nl else []
seen += 1
if len(pred) == 0:
if nl:
stats.append((torch.zeros(0, n_iou, dtype=torch.bool),
torch.Tensor(), torch.Tensor(), t_cls))
continue
pred_n = pred.clone()
util.scale_coords(img[si].shape[1:], pred_n[:, :4], shapes[si][0],
shapes[si][1])
correct = torch.zeros(pred.shape[0],
n_iou,
dtype=torch.bool,
device=device)
if nl:
detected = []
t_cls_tensor = labels[:, 0]
t_box = util.wh2xy(labels[:, 1:5]) * wh_wh
util.scale_coords(img[si].shape[1:], t_box, shapes[si][0],
shapes[si][1])
for cls in torch.unique(t_cls_tensor):
ti = (cls == t_cls_tensor).nonzero(as_tuple=False).view(-1)
pi = (cls == pred[:, 5]).nonzero(as_tuple=False).view(-1)
if pi.shape[0]:
iou_list, i = util.box_iou(pred_n[pi, :4],
t_box[ti]).max(1)
detected_set = set()
for j in (iou_list > iou_v[0]).nonzero(as_tuple=False):
d = ti[i[j]]
if d.item() not in detected_set:
detected_set.add(d.item())
detected.append(d)
correct[pi[j]] = iou_list[j] > iou_v
if len(detected) == nl:
break
stats.append(
(correct.cpu(), pred[:, 4].cpu(), pred[:, 5].cpu(), t_cls))
stats = [numpy.concatenate(x, 0) for x in zip(*stats)]
if len(stats) and stats[0].any():
p, r, ap, f1, ap_class = util.ap_per_class(*stats)
p, r, ap50, ap = p[:, 0], r[:, 0], ap[:, 0], ap.mean(1)
mp, mr, map50, mean_ap = p.mean(), r.mean(), ap50.mean(), ap.mean()
print('%10.3g' * 3 % (mp, mr, mean_ap))
if model is None:
t = tuple(x / seen * 1E3 for x in (t0, t1, t0 + t1))
s = f'Speed: {t[0]:.1f}/{t[1]:.1f}/{t[2]:.1f} ms inference/nms/total'
print(
f'{s} per {args.image_size}x{args.image_size} image at batch-size {args.batch_size}'
)
model.float()
return mp, mr, map50, mean_ap
def read_sdf(file_path, target_grid_res, target_bounding_box_min,
target_bounding_box_max, target_voxel_size):
with open(file_path) as file:
line = file.readline()
# Get grid resolutions
grid_res = line.split()
grid_res_x = int(grid_res[0])
grid_res_y = int(grid_res[1])
grid_res_z = int(grid_res[2])
# Get bounding box min
line = file.readline()
bounding_box_min = line.split()
bounding_box_min_x = float(bounding_box_min[0])
bounding_box_min_y = float(bounding_box_min[1])
bounding_box_min_z = float(bounding_box_min[2])
line = file.readline()
voxel_size = float(line)
# max bounding box (we need to plus 0.0001 to avoid round error)
bounding_box_max_x = bounding_box_min_x + voxel_size * (grid_res_x - 1
) # + 0.0001
bounding_box_max_y = bounding_box_min_y + voxel_size * (grid_res_y - 1
) #+ 0.0001
bounding_box_max_z = bounding_box_min_z + voxel_size * (grid_res_z - 1
) #+ 0.0001
min_bounding_box_min = min(bounding_box_min_x, bounding_box_min_y,
bounding_box_min_z)
print(bounding_box_min_x, bounding_box_min_y, bounding_box_min_z)
max_bounding_box_max = max(bounding_box_max_x, bounding_box_max_y,
bounding_box_max_z)
print(bounding_box_max_x, bounding_box_max_y, bounding_box_max_z)
max_dist = max(bounding_box_max_x - bounding_box_min_x,
bounding_box_max_y - bounding_box_min_y,
bounding_box_max_z - bounding_box_min_z)
max_grid_res = max(grid_res_x, grid_res_y, grid_res_z)
grid = []
for i in range(grid_res_x):
grid.append([])
for j in range(grid_res_y):
grid[i].append([])
for k in range(grid_res_z):
grid[i][j].append(2)
for i in range(grid_res_z):
for j in range(grid_res_y):
for k in range(grid_res_x):
grid_value = float(file.readline())
grid[k][j][i] = grid_value
grid = Tensor(grid)
target_grid = Tensor(target_grid_res, target_grid_res, target_grid_res)
linear_space_x = torch.linspace(0, target_grid_res - 1,
target_grid_res)
linear_space_y = torch.linspace(0, target_grid_res - 1,
target_grid_res)
linear_space_z = torch.linspace(0, target_grid_res - 1,
target_grid_res)
first_loop = linear_space_x.repeat(
target_grid_res * target_grid_res,
1).t().contiguous().view(-1).unsqueeze_(1)
second_loop = linear_space_y.repeat(
target_grid_res,
target_grid_res).t().contiguous().view(-1).unsqueeze_(1)
third_loop = linear_space_z.repeat(target_grid_res *
target_grid_res).unsqueeze_(1)
loop = torch.cat((first_loop, second_loop, third_loop), 1).cuda()
min_x = Tensor([bounding_box_min_x]).repeat(
target_grid_res * target_grid_res * target_grid_res, 1)
min_y = Tensor([bounding_box_min_y]).repeat(
target_grid_res * target_grid_res * target_grid_res, 1)
min_z = Tensor([bounding_box_min_z]).repeat(
target_grid_res * target_grid_res * target_grid_res, 1)
bounding_min_matrix = torch.cat((min_x, min_y, min_z), 1)
move_to_center_x = Tensor([
(max_dist - (bounding_box_max_x - bounding_box_min_x)) / 2
]).repeat(target_grid_res * target_grid_res * target_grid_res, 1)
move_to_center_y = Tensor([
(max_dist - (bounding_box_max_y - bounding_box_min_y)) / 2
]).repeat(target_grid_res * target_grid_res * target_grid_res, 1)
move_to_center_z = Tensor([
(max_dist - (bounding_box_max_z - bounding_box_min_z)) / 2
]).repeat(target_grid_res * target_grid_res * target_grid_res, 1)
move_to_center_matrix = torch.cat(
(move_to_center_x, move_to_center_y, move_to_center_z), 1)
# Get the position of the grid points in the refined grid
points = bounding_min_matrix + target_voxel_size * max_dist / (
target_bounding_box_max -
target_bounding_box_min) * loop - move_to_center_matrix
if points[(points[:, 0] < bounding_box_min_x)].shape[0] != 0:
points[(points[:, 0] < bounding_box_min_x)] = Tensor(
[bounding_box_max_x, bounding_box_max_y,
bounding_box_max_z]).view(1, 3)
if points[(points[:, 1] < bounding_box_min_y)].shape[0] != 0:
points[(points[:, 1] < bounding_box_min_y)] = Tensor(
[bounding_box_max_x, bounding_box_min_y,
bounding_box_min_z]).view(1, 3)
if points[(points[:, 2] < bounding_box_min_z)].shape[0] != 0:
points[(points[:, 2] < bounding_box_min_z)] = Tensor(
[bounding_box_max_x, bounding_box_min_y,
bounding_box_min_z]).view(1, 3)
if points[(points[:, 0] > bounding_box_max_x)].shape[0] != 0:
points[(points[:, 0] > bounding_box_max_x)] = Tensor(
[bounding_box_max_x, bounding_box_min_y,
bounding_box_min_z]).view(1, 3)
if points[(points[:, 1] > bounding_box_max_y)].shape[0] != 0:
points[(points[:, 1] > bounding_box_max_y)] = Tensor(
[bounding_box_max_x, bounding_box_min_y,
bounding_box_min_z]).view(1, 3)
if points[(points[:, 2] > bounding_box_max_z)].shape[0] != 0:
points[(points[:, 2] > bounding_box_max_z)] = Tensor(
[bounding_box_max_x, bounding_box_min_y,
bounding_box_min_z]).view(1, 3)
voxel_min_point_index_x = torch.floor(
(points[:, 0].unsqueeze_(1) - min_x) /
voxel_size).clamp(max=grid_res_x - 2)
voxel_min_point_index_y = torch.floor(
(points[:, 1].unsqueeze_(1) - min_y) /
voxel_size).clamp(max=grid_res_y - 2)
voxel_min_point_index_z = torch.floor(
(points[:, 2].unsqueeze_(1) - min_z) /
voxel_size).clamp(max=grid_res_z - 2)
voxel_min_point_index = torch.cat(
(voxel_min_point_index_x, voxel_min_point_index_y,
voxel_min_point_index_z), 1)
voxel_min_point = bounding_min_matrix + voxel_min_point_index * voxel_size
# Compute the sdf value of the grid points in the refined grid
target_grid = calculate_sdf_value(
grid, points, voxel_min_point, voxel_min_point_index, voxel_size,
grid_res_x, grid_res_y,
grid_res_z).view(target_grid_res, target_grid_res, target_grid_res)
return target_grid
matplotlib
"""
import torch
import torch.utils.data as Data
import torch.nn.functional as F
from torch.autograd import Variable
import matplotlib.pyplot as plt
# torch.manual_seed(1) # reproducible
LR = 0.01
BATCH_SIZE = 32
EPOCH = 12
# fake dataset
x = torch.unsqueeze(torch.linspace(-1, 1, 1000), dim=1)
y = x.pow(2) + 0.1 * torch.normal(torch.zeros(*x.size()))
# plot dataset
plt.scatter(x.numpy(), y.numpy())
plt.show()
# put dateset into torch dataset
torch_dataset = Data.TensorDataset(data_tensor=x, target_tensor=y)
loader = Data.DataLoader(
dataset=torch_dataset,
batch_size=BATCH_SIZE,
shuffle=True,
num_workers=2,
)
def make_mesh():
domain = torch.linspace(-2.5, 2.5, 50)
z1, z2 = torch.meshgrid((domain, domain))
z = torch.stack((z1, z2), dim=-1)
return z1, z2, z
c -= learning_rate * grad_c
d -= learning_rate * grad_d
t1 = time.time()
print(f'Result: y = {a} + {b} x + {c} x^2 + {d} x^3')
y1 = np.array([a + b * x + c * x**2 + d * x**3 for x in x])
fig, axes = plot_make()
sns.scatterplot(x=x, y=y, ax=axes, linewidth=0, s=0.1)
sns.scatterplot(x=x, y=y1, ax=axes, linewidth=0, s=0.1)
#%%
dtype = torch.float
device = torch.device("cuda:0")
x = torch.linspace(-math.pi, math.pi, 2000, device=device, dtype=dtype)
y = torch.sin(x)
a = torch.randn((), device=device, dtype=dtype)
b = torch.randn((), device=device, dtype=dtype)
c = torch.randn((), device=device, dtype=dtype)
d = torch.randn((), device=device, dtype=dtype)
learning_rate = 1e-6
for t in range(2000):
# Forward pass: compute predicted y
y_pred = a + b * x + c * x**2 + d * x**3
# Compute and print loss
loss = (y_pred - y).pow(2).sum().item()
def __init__(self, reg_max=16):
super(Integral, self).__init__()
self.reg_max = reg_max
self.register_buffer('project',
torch.linspace(0, self.reg_max, self.reg_max + 1))
# PyTorch分批数据
import torch
import torch.utils.data as Data #将数据分批次需要用到
# 设置随机数种子
torch.manual_seed(33)
#设置批次大小
BATCH_SIZE = 8
x = torch.linspace(1, 16, 16) # 1到16共16个点
y = torch.linspace(16, 1, 16) # 65到1共16个点
print(x)
print(y)
#将x,y读取，转换成Tensor格式
torch_dataset = Data.TensorDataset(x, y)
loader = Data.DataLoader(
dataset=torch_dataset, # torch TensorDataset format
batch_size=BATCH_SIZE, # 最新批数据
shuffle=True, # 是否随机打乱数据
num_workers=2, # 用于加载数据的子进程
)
def make_fields(protein_dict,
channels=['CA'],
bin_size=2.0,
num_bins=50,
return_bins=False):
"""
This function takes a protein dict (from load_pdb function) and outputs a
large tensor containing many atomic "fields" for the protein.
The fields describe the atomic "density" (an exponentially decaying function
of number of atoms in a voxel) of any particular atom type.
Parameters:
protein_dict (dict, requred): dictionary from the load_pdb function
channels (list-like, optional): the different atomic densities we want fields for
theoretically these different fields provide different chemical information
full list of available channels is in protein_dict['atom_type_set']
bin_size (float, optional): the side-length (angstrom) of a given voxel in the box
that atomic densities are placed in
num_bins (int, optional): how big is the cubic field tensor side length
(i.e., num_bins is box side length)
Returns:
dictionary: A list of atomic density tensors (50x50x50), one for each
channel in channels
"""
# sets of allowed filters to build channels with
residue_filters = protein_dict['residue_set']
atom_filters = protein_dict['atom_type_set']
residue_property_filters = np.array(['acidic', 'basic', 'polar', 'nonpolar',\
'charged', 'amphipathic'])
other_filters = np.array(['backbone', 'sidechains'])
# consolidate into one set of filters
filter_set = {'atom':atom_filters, 'residue':residue_filters,\
'residue_property':residue_property_filters, 'other':other_filters}
# construct a single empty field, then initialize a dictionary with one
# empty field for every channel we are going to calculate the density for
empty_field = torch.zeros((num_bins, num_bins, num_bins))
fields = {channel: empty_field for channel in channels}
# create linearly spaced grid (default is -49 to 49 in steps of 2)
grid_1d = torch.linspace(start=-num_bins / 2 * bin_size + bin_size / 2,
end=num_bins / 2 * bin_size - bin_size / 2,
steps=num_bins)
# This makes three 3D meshgrids in for the x, y, and z positions
# These cubes will be flattened, then used as a reference coordinate system
# to place the actual channel densities into
xgrid, ygrid, zgrid = grid_positions(grid_1d)
for channel_index, channel in enumerate(channels):
# no illegal channels allowed, assume the channel sucks
channel_allowed = check_channel(channel, filter_set)
if channel_allowed:
pass
else:
#err_string = 'Allowed channels are: in a protein\'s atom_type_set,
# residue_set',or the \'sidechains\' and \'backbone\' channels.'
raise ValueError('The channel ', channel,
' is not allowed for this protein.')
# Extract positions of atoms that are part of the current channel
atom_positions = find_channel_atoms(channel, protein_dict, filter_set)
# print('This is channel ', channel)
atom_positions = torch.FloatTensor(atom_positions)
# xgrid.view(-1, 1) is 125,000 long, because it's viewing a 50x50x50 cube in one column
# then you repeat that column horizontally for each atom
xx_xx = xgrid.view(-1, 1).repeat(1, len(atom_positions))
yy_yy = ygrid.view(-1, 1).repeat(1, len(atom_positions))
zz_zz = zgrid.view(-1, 1).repeat(1, len(atom_positions))
# at this point we've created 3 arrays that are 125,000 long
# and as wide as the number of atoms that are the current channel type
# these 3 arrays just contain the flattened x,y,z positions of our 50x50x50 box
# now do the same thing as above, just with the ACTUAL atomic position data
posx_posx = atom_positions[:, 0].contiguous().view(1, -1).repeat(
len(xgrid.view(-1)), 1)
posy_posy = atom_positions[:, 1].contiguous().view(1, -1).repeat(
len(ygrid.view(-1)), 1)
posz_posz = atom_positions[:, 2].contiguous().view(1, -1).repeat(
len(zgrid.view(-1)), 1)
# three tensors of the same size, with actual atomic coordinates
# normalizes the atomic positions with respect to the center of the box
# and calculates density of atoms in each voxel
sigma = 0.5 * bin_size
density = torch.exp(-((xx_xx - posx_posx)**2 + (yy_yy - posy_posy)**2 +
(zz_zz - posz_posz)**2) / (2 * (sigma)**2))
# Normalize so each atom density sums to one
density /= torch.sum(density, dim=0)
# Sum densities and reshape to original shape
sum_densities = torch.sum(density, dim=1).view(xgrid.shape)
# set all nans to 0
sum_densities[sum_densities != sum_densities] = 0
# add two empty dimmensions to make it 1x1x50x50x50, needed for CNN
# sum_densities = sum_densities.unsqueeze(0)
# sum_densities = sum_densities.unsqueeze(0)
#fields[atom_type_index] = sum_densities
fields[channel] = sum_densities.numpy()
if return_bins:
return fields, num_bins
else:
return fields
Dependencies: torch: 0.1.11 matplotlib """ import torch from torch.autograd import Variable import matplotlib.pyplot as plt # torch.manual_seed(1) # reproducible N_SAMPLES = 20 N_HIDDEN = 300 # training data x = torch.unsqueeze(torch.linspace(-1, 1, N_SAMPLES), 1) y = x + 0.3*torch.normal(torch.zeros(N_SAMPLES, 1), torch.ones(N_SAMPLES, 1)) x, y = Variable(x), Variable(y) # test data test_x = torch.unsqueeze(torch.linspace(-1, 1, N_SAMPLES), 1) test_y = test_x + 0.3*torch.normal(torch.zeros(N_SAMPLES, 1), torch.ones(N_SAMPLES, 1)) test_x, test_y = Variable(test_x, volatile=True), Variable(test_y, volatile=True) # show data plt.scatter(x.data.numpy(), y.data.numpy(), c='magenta', s=50, alpha=0.5, label='train') plt.scatter(test_x.data.numpy(), test_y.data.numpy(), c='cyan', s=50, alpha=0.5, label='test') plt.legend(loc='upper left') plt.ylim((-2.5, 2.5)) plt.show()
def get_region_boxes(output,
netshape,
conf_thresh,
num_classes,
anchors,
num_anchors,
only_objectness=1,
validation=False,
use_cuda=True):
device = torch.device("cuda" if use_cuda else "cpu")
anchors = anchors.to(device)
anchor_step = anchors.size(0) // num_anchors
if output.dim() == 3:
output = output.unsqueeze(0)
batch = output.size(0)
assert (output.size(1) == (5 + num_classes) * num_anchors)
h = output.size(2)
w = output.size(3)
cls_anchor_dim = batch * num_anchors * h * w
if netshape[0] != 0:
nw, nh = netshape
else:
nw, nh = w, h
t0 = time.time()
all_boxes = []
output = output.view(batch * num_anchors, 5 + num_classes,
h * w).transpose(0, 1).contiguous().view(
5 + num_classes, cls_anchor_dim)
grid_x = torch.linspace(0, w - 1,
w).repeat(batch * num_anchors, h,
1).view(cls_anchor_dim).to(device)
grid_y = torch.linspace(0, h - 1, h).repeat(w, 1).t().repeat(
batch * num_anchors, 1, 1).view(cls_anchor_dim).to(device)
ix = torch.LongTensor(range(0, 2)).to(device)
anchor_w = anchors.view(num_anchors,
anchor_step).index_select(1, ix[0]).repeat(
batch, h * w).view(cls_anchor_dim)
anchor_h = anchors.view(num_anchors,
anchor_step).index_select(1, ix[1]).repeat(
batch, h * w).view(cls_anchor_dim)
xs, ys = output[0].sigmoid() + grid_x, output[1].sigmoid() + grid_y
ws, hs = output[2].exp() * anchor_w.detach(), output[3].exp(
) * anchor_h.detach()
det_confs = output[4].sigmoid()
# by ysyun, dim=1 means input is 2D or even dimension else dim=0
cls_confs = torch.nn.Softmax(dim=1)(output[5:5 + num_classes].transpose(
0, 1)).detach()
cls_max_confs, cls_max_ids = torch.max(cls_confs, 1)
cls_max_confs = cls_max_confs.view(-1)
cls_max_ids = cls_max_ids.view(-1)
t1 = time.time()
sz_hw = h * w
sz_hwa = sz_hw * num_anchors
det_confs = convert2cpu(det_confs)
cls_max_confs = convert2cpu(cls_max_confs)
cls_max_ids = convert2cpu_long(cls_max_ids)
xs, ys = convert2cpu(xs), convert2cpu(ys)
ws, hs = convert2cpu(ws), convert2cpu(hs)
if validation:
cls_confs = convert2cpu(cls_confs.view(-1, num_classes))
t2 = time.time()
for b in range(batch):
boxes = []
for cy in range(h):
for cx in range(w):
for i in range(num_anchors):
ind = b * sz_hwa + i * sz_hw + cy * w + cx
det_conf = det_confs[ind]
conf = det_conf * (cls_max_confs[ind]
if not only_objectness else 1.0)
if conf > conf_thresh:
bcx = xs[ind]
bcy = ys[ind]
bw = ws[ind]
bh = hs[ind]
cls_max_conf = cls_max_confs[ind]
cls_max_id = cls_max_ids[ind]
box = [
bcx / w, bcy / h, bw / nw, bh / nh, det_conf,
cls_max_conf, cls_max_id
]
if (not only_objectness) and validation:
for c in range(num_classes):
tmp_conf = cls_confs[ind][c]
if c != cls_max_id and det_confs[
ind] * tmp_conf > conf_thresh:
box.append(tmp_conf)
box.append(c)
boxes.append(box)
all_boxes.append(boxes)
t3 = time.time()
if False:
print('---------------------------------')
print('matrix computation : %f' % (t1 - t0))
print(' gpu to cpu : %f' % (t2 - t1))
print(' boxes filter : %f' % (t3 - t2))
print('---------------------------------')
return all_boxes
""" View more, visit my tutorial page: https://morvanzhou.github.io/tutorials/ My Youtube Channel: https://www.youtube.com/user/MorvanZhou Dependencies: torch: 0.1.11 matplotlib """ import torch import torch.nn.functional as F from torch.autograd import Variable import matplotlib.pyplot as plt # fake data x = torch.linspace(-5, 5, 200) # x data (tensor), shape=(100, 1) x = Variable(x) x_np = x.data.numpy() # numpy array for plotting # following are popular activation functions y_relu = F.relu(x).data.numpy() y_sigmoid = F.sigmoid(x).data.numpy() y_tanh = F.tanh(x).data.numpy() y_softplus = F.softplus(x).data.numpy() # y_softmax = F.softmax(x) softmax is a special kind of activation function, it is about probability # plt to visualize these activation function plt.figure(1, figsize=(8, 6)) plt.subplot(221) plt.plot(x_np, y_relu, c='red', label='relu') plt.ylim((-1, 5))
import torch
import torch.nn.functional as F
import numpy as np
import matplotlib.pyplot as plt
from model import AE
from calc_uncertainty import uncertainty
import random
import math
### Make meshgrid
end, d = 5, 101
x1 = torch.linspace(0, end, d)
y1 = torch.linspace(0, end, d)
xx, yy = np.meshgrid(np.arange(0, end, d), np.arange(0, end, d))
uncertainty_data = torch.from_numpy(np.dstack(np.meshgrid(x1, y1)))
### Make training data
z = random.randint(1, 10)
ind_x = list(range(z + 80, z + 90, 1))
# ind_y = [int((a-50)*(a-50)*0.2 + a + 1) for a in ind_x]
ind_y = [int(-1 * a + 101)
for a in ind_x] # Change these things to make shape of training data
index = [a + b * d for (a, b) in zip(ind_x, ind_y)]
print(ind_x, ind_y, index)
# index = torch.randperm(d*d)[:20] --> If you want to select random location for training data
index = torch.randperm(d * d)[:16]
uncertainty_data = uncertainty_data.reshape((d * d, 2))
train_data = uncertainty_data[index, :]
