def to_np_img(t: torch.tensor):
if len(t.shape) == 4:
t = t[0]
if t.shape[-1] == 3 or t.shape[-1] == 2:
if type(t) is not np.ndarray:
return t.cpu().detach().numpy()
return t
return t.cpu().detach().numpy().transpose(1, 2, 0)
def add(self, output: torch.tensor, target: torch.tensor):
output = output.cpu().detach().squeeze().numpy()[np.newaxis]
target = target.cpu().detach().squeeze().numpy()[np.newaxis]
assert output.ndim == target.ndim, 'target and output do not match'
assert output.ndim == 1
self.output = np.hstack([self.output, output])
self.target = np.hstack([self.target, target])
def imshow(img: Tensor) -> None:
if len(img.shape) > 3:
img = img.squeeze()
img = torch.clamp(img / 2 + 0.5, 0, 1)
npimg = img.cpu().numpy()
plt.imshow(np.transpose(npimg, (1, 2, 0)))
plt.show()
def act(model: Policy, state: torch.tensor):
with torch.no_grad():
p = F.softmax(model.forward(state)).cpu().numpy()
valid_moves = (state.cpu().numpy().reshape(
3, 3, 3).argmax(axis=2).reshape(-1) == 0)
p = valid_moves * p
return p.argmax()
def _get_recalls(y_true: torch.tensor, pred_labels: List) -> List[float]:
y_true = y_true.cpu().numpy()
return [
metrics.recall_score(pred_labels[idx],
y_true[:, idx],
average='macro') for idx in range(3)
]
def frequency_based_masking(tokens: torch.tensor,
token_sampling_weights: np.ndarray,
mask_prob: float) -> torch.Tensor:
"""
Function to mask tokens based on frequency.
Inputs:
1) tokens: Tensor with token ids of shape (batch_size x seq_len)
2) token_sampling_weights: numpy array with shape (batch_size x seq_len)
and each element representing the sampling weight assicated with
the corresponding token in tokens
3) mask_prob: Probability of masking a particular token
Outputs:
mask: Tensor with same shape as input tokens (batch_size x seq_len)
with masked tokens represented by a 1 and everything else as 0.
"""
batch_size, seq_len = tokens.size()
num_masked_per_batch = int(batch_size * seq_len * mask_prob)
indices = tokens.cpu().numpy().flatten()
# get the weights associated with each token
weights = np.take(token_sampling_weights, indices)
# sample tokens based on the computed weights
tokens_to_mask = np.random.choice(len(weights),
num_masked_per_batch,
replace=False,
p=weights / weights.sum())
mask = torch.zeros(batch_size * seq_len)
mask[tokens_to_mask] = 1
mask = mask.view(batch_size, seq_len).long().to(tokens.device)
return mask
def un_normalize(img: torch.tensor, mean: list, std: list):
size = img.size()
n_img = img.cpu().view(size[0], size[2], size[3], size[1]).detach().numpy()
m = np.concatenate(
(mean[0] *
np.ones(shape=[n_img.shape[0], n_img.shape[1], n_img.shape[2], 1]),
mean[1] *
np.ones(shape=[n_img.shape[0], n_img.shape[1], n_img.shape[2], 1]),
mean[2] *
np.ones(shape=[n_img.shape[0], n_img.shape[1], n_img.shape[2], 1])),
-1)
s = np.concatenate(
(std[0] *
np.ones(shape=[n_img.shape[0], n_img.shape[1], n_img.shape[2], 1]),
std[1] *
np.ones(shape=[n_img.shape[0], n_img.shape[1], n_img.shape[2], 1]),
std[2] *
np.ones(shape=[n_img.shape[0], n_img.shape[1], n_img.shape[2], 1])),
-1)
res_img = np.multiply(n_img + m, s) * 255
return res_img
def rolling_cumulation(data_ts: torch.tensor, window: int) -> torch.tensor:
"""the input is a variation(ratio) of something, like returns"""
tensor_np = data_ts.cpu().detach().numpy()
tensor_df = pd.DataFrame(tensor_np)
output_df = (1 + tensor_df).rolling(window).apply(np.prod, raw=False) - 1
output_np = np.array(output_df)
output_ts = torch.tensor(output_np).squeeze()
return output_ts
def rolling_max_drawdown(data_ts: torch.tensor, window: int) -> torch.tensor:
''' input is price '''
tensor_np = data_ts.cpu().detach().numpy()
tensor_df = pd.DataFrame(tensor_np)
output_df = tensor_df.rolling(window).apply(_max_drawdown, raw=False)
output_np = np.array(output_df)
output_ts = torch.tensor(output_np).squeeze()
return output_ts
def postprocess(img: torch.tensor):
# img = img[0][2]
img = torch.argmax(img, dim=1)
img = img.cpu().numpy()
img = np.squeeze(img)
img = re_normalize(img)
# img = util.invert(img)
return img
def rolling_max_drawdown_from_returns(data_ts: torch.tensor,
window: int) -> torch.tensor:
''' input is returns '''
returns_np = data_ts.cpu().detach().numpy()
returns_df = pd.DataFrame(returns_np)
price = (1 + returns_df).cumprod()
output_df = price.rolling(window).apply(_max_drawdown, raw=False)
output_np = np.array(output_df)
output_ts = torch.tensor(output_np).squeeze()
return output_ts
def torch_to_frame(img: torch.tensor) -> np.array:
img = (img + 1) / 2
img = torch.clamp(img, 0, 1)
img = torch.squeeze(img, dim=0)
img = img.cpu().detach().numpy()
img = img.transpose(1, 2, 0)
img = (img * 255).astype(np.uint8)
img = cv2.cvtColor(img, cv2.COLOR_RGB2BGR)
img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)
return img
def torch2numpy(tensor: torch.tensor, nptype: str = 'float32'):
cpu_tensor = tensor.cpu().detach()
arr = cpu_tensor.numpy().astype(nptype)
if len(arr.shape) == 4:
transpose = (0, 2, 3, 1)
elif len(arr.shape) == 3:
transpose = (1, 2, 0)
else:
transpose = (0, 1)
arr = arr.transpose(transpose)
return arr
def fix_features(features: torch.tensor) -> torch.tensor:
features = features.cpu()
features = features.numpy()
for i in range(0, features.shape[0]):
for j in range(0, features.shape[1]):
if features[i][j] <= -100.0:
features[i][j] = -99.9999
if features[i][j] >= 100.0:
features[i][j] = 99.9999
features = torch.from_numpy(features)
features = features.to(device)
return features
def __init__(
self,
ref: torch.tensor, # [H] torch.tensor
num_actions: int,
u_min: float,
u_max: float):
super(Linear_Actions, self).__init__()
# us = np.random.uniform(low=u_min, high=u_max, size=(H, num_actions))
us = np.reshape(ref.cpu().detach().numpy(), (-1, 1))
us = np.repeat(us, num_actions, axis=1)
us = np.clip(us, u_min, u_max)
self.us = torch.nn.Parameter(torch.from_numpy(us).float())
def __init__(
self,
ref: torch.tensor, # [H] torch.tensor
num_actions: int,
action_dim: int,
u_min: float,
u_max: float):
super(Actions, self).__init__()
us = np.reshape(ref.cpu().detach().numpy(), (1, -1, 1))
us = np.repeat(us, num_actions, axis=0)
us = np.repeat(us, action_dim, axis=2)
us = np.clip(us, u_min, u_max)
self.us = torch.nn.Parameter(torch.from_numpy(us).float())
def encode_torch(obj: torch.tensor) -> str:
"""
Encode a torch tensor as a bytestring.
Parameters
----------
obj : ``torch.tensor``
Object to be encoded
Returns
-------
bytestr : ``str``
bytestring representation of the passed object
"""
return encode_numpy(obj.cpu().detach().numpy())
def get_template_texture(vertices: torch.tensor, faces: torch.tensor,
texture_map: torch.tensor):
device = vertices.device
verts_uv = convert_3d_to_uv_coordinates(vertices)
vertex_rgb = torch.nn.functional.grid_sample(
texture_map.unsqueeze(0), 2 * verts_uv.unsqueeze(0).unsqueeze(0) - 1)
vertex_rgb = vertex_rgb.squeeze(2).permute(0, 2, 1) * 255
texture = Textures([texture_map.cpu().permute(1, 2, 0)],
faces_uvs=faces.unsqueeze(0),
verts_uvs=verts_uv.unsqueeze(0),
verts_rgb=vertex_rgb).to(device)
return texture
def onnx_inference(model_path: str, inputs: tensor) -> tensor:
"""Run ONNX inference
Args:
model_path (str): Path to the ONNX model.
inputs (Tensor): Batch of input image.
Returns:
Tensor: Network output.
"""
inputs = np.array(inputs.cpu())
ort_session = onnxruntime.InferenceSession(model_path)
ort_inputs = {ort_session.get_inputs()[0].name: inputs}
ort_outs = ort_session.run(None, ort_inputs)
return torch.tensor(ort_outs[0])
def drawHeatMapV1(hm:torch.tensor,box:torch.tensor,device="cpu"):
"""
:param hm: torch.tensor shape [128,128]
:param box: torch.tensor [x1,y1,x2,y2] 缩减到 hm 大小上
:return:
"""
hm = hm.cpu().numpy()
x1,y1,x2,y2 = box
h,w = y2-y1,x2-x1
cx,cy = (x2+x1)/2.,(y2+y1)/2.
cx,cy = cx.int().item(),cy.int().item()
h,w = h.item(),w.item()
radius = gaussian_radius((h,w))
# radius = math.sqrt(h*w) # 不推荐
radius = max(1, int(radius))
# hm = torch.from_numpy(draw_msra_gaussian(hm, (cx,cy), radius)).to(device) # 等价于 drawHeatMapV2
hm = torch.from_numpy(draw_umich_gaussian(hm, (cx,cy), radius)).to(device)
return hm
def transit(self, a: torch.tensor, ref: torch.tensor, threshold: float):
"""
MDP transition function
:param a: list of actions, [t1, t2, t3], shape=(3,)
:param ref: list of references, [ref_q1, ref_q2, ref_q3], shape=(3,)
:param threshold: tolerance for derivation
:return:
"""
self.step(a.cpu())
r = reward(self.state, ref, threshold)
if r == 100: # terminal state
self.terminate = True
if self.show:
self.show_data.append([[self.l1_x, self.l1_y],
[self.l2_x, self.l2_y],
[self.l3_x, self.l3_y]])
self.show_a.append(
[a[0][0].item(), a[0][1].item(), a[0][2].item()])
return r
def tensor2list(cudatensor: torch.tensor) -> List:
return list(cudatensor.cpu().numpy())
def decode(self, ids: torch.tensor) -> list:
ids = np.array(torch.squeeze(ids.cpu())).tolist() # 2D
return list(map(lambda x: self.tokenizer.convert_ids_to_tokens(x), ids))
def to_numpy(self, t:tensor): return t.detach.cpu().numpy() if t.requires_grad else t.cpu().numpy()
def predict(self, inps):
def to_numpy(torch_tensor: torch.tensor) -> np.array:
return torch_tensor.cpu().detach().numpy()
def imshow(img: Tensor) -> None:
img = img / 2 + 0.5
npimg = img.cpu().numpy()
plt.imshow(np.transpose(npimg, (1, 2, 0)))
plt.show()
def to_numpy(t: tensor):
return t.detach.cpu().numpy() if t.requires_grad else t.cpu().numpy()
def postprocess(img: torch.tensor):
img = torch.argmax(img, dim=1) # perform argmax to generate 1 channel
img = img.cpu().numpy() # send to cpu and transform to numpy.ndarray
img = np.squeeze(img) # remove batch dim and channel dim -> [H, W]
img = re_normalize(img) # scale it to the range [0-255]
return img
def evaluate(self,
gt_points: torch.tensor,
source_points: torch.tensor,
gt_normals=None):
"""computes the chamfer distance between 2 point clouds
Arguments:
gt_points {torch.tensor} -- [description]
source_points {torch.tensor} -- [description]
Returns:
[dict] -- [description]
"""
##### Computing Chamfer Distances ######
gt_points = gt_points.cuda().detach()
source_points = source_points.cuda().detach()
d_gt2source, d_source2gt, idx3, idx4 = self.cham_loss(
gt_points.unsqueeze(0), source_points.unsqueeze(0))
idx3 = idx3.long().squeeze()
idx4 = idx4.long().squeeze()
# mean(squared_d(gt->source)) + mean(squared_d(source->gt))
chamfer_dist = (d_gt2source.mean() + d_source2gt.mean()) / 2
chamfer_dist_abs = (d_gt2source.sqrt().mean() +
d_source2gt.sqrt().mean()) / 2
out_dict = {}
out_dict['chamfer_dist'] = chamfer_dist.cpu().item()
self.eval_results['chamfer_dist'].append(out_dict['chamfer_dist'])
out_dict['chamfer_dist_abs'] = chamfer_dist_abs.cpu().item()
self.eval_results['chamfer_dist_abs'].append(
out_dict['chamfer_dist_abs'])
############ PSNR ##############
if gt_normals is not None: # Computing PSNR if we have normals
gt_normals = gt_normals.cuda().detach()
d_plane_gt2source = torch.sum(
(gt_points - source_points[idx3, :]) * gt_normals, dim=1)
d_plane_source2gt = torch.sum(
(source_points - gt_points[idx4, :]) * gt_normals[idx4, :],
dim=1)
chamfer_plane = (d_plane_gt2source.abs().mean() +
d_plane_source2gt.abs().mean()) / 2
out_dict['chamfer_dist_plane'] = chamfer_plane.cpu().item()
self.eval_results['chamfer_dist_plane'].append(
out_dict['chamfer_dist_plane'])
###### IOU #######
gt_points_np = gt_points.cpu().numpy()
source_points_np = source_points.cpu().numpy()
# print('gt_points shape',g)
center = (np.max(gt_points_np, axis=0, keepdims=True) +
np.min(gt_points_np, axis=0, keepdims=True)) / 2
resolution = np.array(
[self.config['evaluation']['iou_grid']['resolution']])
size_meter = np.array([self.config['grid']['size']])
gt_grid = occupancy_grid.OccupancyGrid(center=center,
resolution=resolution,
size_meter=size_meter)
gt_grid.addPoints(gt_points_np)
source_grid = occupancy_grid.OccupancyGrid(center=center,
resolution=resolution,
size_meter=size_meter)
source_grid.addPoints(source_points_np)
out_dict['iou'] = occupancy_grid.gridIOU(gt_grid.grid,
source_grid.grid)
self.eval_results['iou'].append(out_dict['iou'])
return out_dict
def to_np(t: torch.tensor) -> np.array:
"""Converts a PyTorch tensor to a Numpy array.
"""
return t.cpu().detach().numpy()
