python类min()的实例源码

def lr_grad_norm_avg(self):
    # this is for enforcing lr * grad_norm not 
    # increasing dramatically in case of instability.
    #  Not necessary for basic use.
    global_state = self._global_state
    beta = self._beta
    if "lr_grad_norm_avg" not in global_state:
      global_state['grad_norm_squared_avg_log'] = 0.0
    global_state['grad_norm_squared_avg_log'] = \
      global_state['grad_norm_squared_avg_log'] * beta \
      + (1 - beta) * np.log(global_state['grad_norm_squared'] + eps)
    if "lr_grad_norm_avg" not in global_state:
      global_state["lr_grad_norm_avg"] = \
        0.0 * beta + (1 - beta) * np.log(self._lr * np.sqrt(global_state['grad_norm_squared'] ) + eps)
      # we monitor the minimal smoothed ||lr * grad||
      global_state["lr_grad_norm_avg_min"] = \
        np.exp(global_state["lr_grad_norm_avg"] / self.zero_debias_factor() )
    else:
      global_state["lr_grad_norm_avg"] = global_state["lr_grad_norm_avg"] * beta \
        + (1 - beta) * np.log(self._lr * np.sqrt(global_state['grad_norm_squared'] ) + eps)
      global_state["lr_grad_norm_avg_min"] = \
        min(global_state["lr_grad_norm_avg_min"], 
            np.exp(global_state["lr_grad_norm_avg"] / self.zero_debias_factor() ) )

def update_hyper_param(self):
    for group in self._optimizer.param_groups:
      group['momentum'] = self._mu_t
      #group['momentum'] = max(self._mu, self._mu_t)
      if self._force_non_inc_step == False:
        group['lr'] = self._lr_t * self._lr_factor
        # a loose clamping to prevent catastrophically large move. If the move
        # is too large, we set lr to 0 and only use the momentum to move
        if self._adapt_clip and (group['lr'] * np.sqrt(self._global_state['grad_norm_squared']) >= self._catastrophic_move_thresh):
          group['lr'] = self._catastrophic_move_thresh / np.sqrt(self._global_state['grad_norm_squared'] + eps)
          if self._verbose:
            logging.warning("clip catastropic move!")
      elif self._iter > self._curv_win_width:
        # force to guarantee lr * grad_norm not increasing dramatically. 
        # Not necessary for basic use. Please refer to the comments
        # in YFOptimizer.__init__ for more details
        self.lr_grad_norm_avg()
        debias_factor = self.zero_debias_factor()
        group['lr'] = min(self._lr * self._lr_factor,
          2.0 * self._global_state["lr_grad_norm_avg_min"] \
          / (np.sqrt(np.exp(self._global_state['grad_norm_squared_avg_log'] / debias_factor) ) + eps) )
    return

def update_hyper_param(self):
    for group in self._optimizer.param_groups:
      group['momentum'] = self._mu
      if self._force_non_inc_step == False:
        group['lr'] = min(self._lr * self._lr_factor, 
          self._lr_grad_norm_thresh / (math.sqrt(self._global_state["grad_norm_squared"] ) + eps) )
      elif self._iter > self._curv_win_width:
        # force to guarantee lr * grad_norm not increasing dramatically. 
        # Not necessary for basic use. Please refer to the comments
        # in YFOptimizer.__init__ for more details
        self.lr_grad_norm_avg()
        debias_factor = self.zero_debias_factor()
        group['lr'] = min(self._lr * self._lr_factor,
          2.0 * self._global_state["lr_grad_norm_avg_min"] \
          / (np.sqrt(np.exp(self._global_state['grad_norm_squared_avg_log'] / debias_factor) ) + eps) )
    return

def intersect(box_a, box_b):
    """ We resize both tensors to [A,B,2] without new malloc:
    [A,2] -> [A,1,2] -> [A,B,2]
    [B,2] -> [1,B,2] -> [A,B,2]
    Then we compute the area of intersect between box_a and box_b.
    Args:
      box_a: (tensor) bounding boxes, Shape: [A,4].
      box_b: (tensor) bounding boxes, Shape: [B,4].
    Return:
      (tensor) intersection area, Shape: [A,B].
    """
    A = box_a.size(0)
    B = box_b.size(0)
    max_xy = torch.min(box_a[:, 2:].unsqueeze(1).expand(A, B, 2),
                       box_b[:, 2:].unsqueeze(0).expand(A, B, 2))
    min_xy = torch.max(box_a[:, :2].unsqueeze(1).expand(A, B, 2),
                       box_b[:, :2].unsqueeze(0).expand(A, B, 2))
    inter = torch.clamp((max_xy - min_xy), min=0)
    return inter[:, :, 0] * inter[:, :, 1]

def intersect(box_a, box_b):
    """ We resize both tensors to [A,B,2] without new malloc:
    [A,2] -> [A,1,2] -> [A,B,2]
    [B,2] -> [1,B,2] -> [A,B,2]
    Then we compute the area of intersect between box_a and box_b.
    Args:
      box_a: (tensor) bounding boxes, Shape: [A,4].
      box_b: (tensor) bounding boxes, Shape: [B,4].
    Return:
      (tensor) intersection area, Shape: [A,B].
    """
    A = box_a.size(0)
    B = box_b.size(0)
    max_xy = torch.min(box_a[:, 2:].unsqueeze(1).expand(A, B, 2),
                       box_b[:, 2:].unsqueeze(0).expand(A, B, 2))
    min_xy = torch.max(box_a[:, :2].unsqueeze(1).expand(A, B, 2),
                       box_b[:, :2].unsqueeze(0).expand(A, B, 2))
    inter = torch.clamp((max_xy - min_xy), min=0)
    return inter[:, :, 0] * inter[:, :, 1]

def intersect(box_a, box_b):
    """ We resize both tensors to [A,B,2] without new malloc:
    [A,2] -> [A,1,2] -> [A,B,2]
    [B,2] -> [1,B,2] -> [A,B,2]
    Then we compute the area of intersect between box_a and box_b.
    Args:
      box_a: (tensor) bounding boxes, Shape: [A,4].
      box_b: (tensor) bounding boxes, Shape: [B,4].
    Return:
      (tensor) intersection area, Shape: [A,B].
    """
    A = box_a.size(0)
    B = box_b.size(0)
    max_xy = torch.min(box_a[:, 2:].unsqueeze(1).expand(A, B, 2),
                       box_b[:, 2:].unsqueeze(0).expand(A, B, 2))
    min_xy = torch.max(box_a[:, :2].unsqueeze(1).expand(A, B, 2),
                       box_b[:, :2].unsqueeze(0).expand(A, B, 2))
    inter = torch.clamp((max_xy - min_xy), min=0)
    return inter[:, :, 0] * inter[:, :, 1]

def intersect(box_a, box_b):
    """ We resize both tensors to [A,B,2] without new malloc:
    [A,2] -> [A,1,2] -> [A,B,2]
    [B,2] -> [1,B,2] -> [A,B,2]
    Then we compute the area of intersect between box_a and box_b.
    Args:
      box_a: (tensor) bounding boxes, Shape: [A,4].
      box_b: (tensor) bounding boxes, Shape: [B,4].
    Return:
      (tensor) intersection area, Shape: [A,B].
    """
    A = box_a.size(0)
    B = box_b.size(0)
    #pdb.set_trace()
    max_xy = torch.min(box_a[:, 2:].unsqueeze(1).expand(A, B, 2),
                       box_b[:, 2:].unsqueeze(0).expand(A, B, 2))
    min_xy = torch.max(box_a[:, :2].unsqueeze(1).expand(A, B, 2),
                       box_b[:, :2].unsqueeze(0).expand(A, B, 2))
    inter = torch.clamp((max_xy - min_xy), min=0)
    return inter[:, :, 0] * inter[:, :, 1]

def cosine_similarity(x1, x2, dim=1, eps=1e-8):
    r"""Returns cosine similarity between x1 and x2, computed along dim.

    Args:
        x1 (Variable): First input.
        x2 (Variable): Second input (of size matching x1).
        dim (int, optional): Dimension of vectors. Default: 1
        eps (float, optional): Small value to avoid division by zero. Default: 1e-8

    Shape:
        - Input: :math:`(\ast_1, D, \ast_2)` where D is at position `dim`.
        - Output: :math:`(\ast_1, \ast_2)` where 1 is at position `dim`.
    """
    w12 = torch.sum(x1 * x2, dim)
    w1 = torch.norm(x1, 2, dim)
    w2 = torch.norm(x2, 2, dim)
    return (w12 / (w1 * w2).clamp(min=eps)).squeeze()

def normalize(input, p=2, dim=1, eps=1e-12):
    r"""Performs :math:`L_p` normalization of inputs over specified dimension.

    Does:

    .. math::
        v = \frac{v}{\max(\lVert v \rVert_p, \epsilon)}

    for each subtensor v over dimension dim of input. Each subtensor is flattened into a vector,
    i.e. :math:`\lVert v \rVert_p` is not a matrix norm.

    With default arguments normalizes over the second dimension with Euclidean norm.

    Args:
        input: input tensor of any shape
        p (float): the exponent value in the norm formulation
        dim (int): the dimension to reduce
        eps (float): small value to avoid division by zero
    """
    return input / input.norm(p, dim, True).clamp(min=eps).expand_as(input)

def intersect(box_a, box_b):
    """ We resize both tensors to [A,B,2] without new malloc:
    [A,2] -> [A,1,2] -> [A,B,2]
    [B,2] -> [1,B,2] -> [A,B,2]
    Then we compute the area of intersect between box_a and box_b.
    Args:
      box_a: (tensor) bounding boxes, Shape: [A,4].
      box_b: (tensor) bounding boxes, Shape: [B,4].
    Return:
      (tensor) intersection area, Shape: [A,B].
    """
    A = box_a.size(0)
    B = box_b.size(0)
    max_xy = torch.min(box_a[:, 2:].unsqueeze(1).expand(A, B, 2),
                       box_b[:, 2:].unsqueeze(0).expand(A, B, 2))
    min_xy = torch.max(box_a[:, :2].unsqueeze(1).expand(A, B, 2),
                       box_b[:, :2].unsqueeze(0).expand(A, B, 2))
    inter = torch.clamp((max_xy - min_xy), min=0)
    return inter[:, :, 0] * inter[:, :, 1]

def normalize(input, p=2, dim=1, eps=1e-12):
    r"""Performs :math:`L_p` normalization of inputs over specified dimension.

    Does:

    .. math::
        v = \frac{v}{\max(\lVert v \rVert_p, \epsilon)}

    for each subtensor v over dimension dim of input. Each subtensor is
    flattened into a vector, i.e. :math:`\lVert v \rVert_p` is not a matrix
    norm.

    With default arguments normalizes over the second dimension with Euclidean
    norm.

    Args:
        input: input tensor of any shape
        p (float): the exponent value in the norm formulation. Default: 2
        dim (int): the dimension to reduce. Default: 1
        eps (float): small value to avoid division by zero. Default: 1e-12
    """
    return input / input.norm(p, dim, True).clamp(min=eps).expand_as(input)

def get_mean_and_std(dataset, max_load=10000):
    '''Compute the mean and std value of dataset.'''
    # dataloader = torch.utils.data.DataLoader(dataset, batch_size=1, shuffle=True, num_workers=2)
    mean = torch.zeros(3)
    std = torch.zeros(3)
    print('==> Computing mean and std..')
    N = min(max_load, len(dataset))
    for i in range(N):
        print(i)
        im,_,_ = dataset.load(1)
        for j in range(3):
            mean[j] += im[:,j,:,:].mean()
            std[j] += im[:,j,:,:].std()
    mean.div_(N)
    std.div_(N)
    return mean, std

def normalize(input, p=2, dim=1, eps=1e-12):
    r"""Performs :math:`L_p` normalization of inputs over specified dimension.

    Does:

    .. math::
        v = \frac{v}{\max(\lVert v \rVert_p, \epsilon)}

    for each subtensor v over dimension dim of input. Each subtensor is
    flattened into a vector, i.e. :math:`\lVert v \rVert_p` is not a matrix
    norm.

    With default arguments normalizes over the second dimension with Euclidean
    norm.

    Args:
        input: input tensor of any shape
        p (float): the exponent value in the norm formulation. Default: 2
        dim (int): the dimension to reduce. Default: 1
        eps (float): small value to avoid division by zero. Default: 1e-12
    """
    return input / input.norm(p, dim, True).clamp(min=eps).expand_as(input)

def min(x, axis=None, keepdims=False):
    def _min(x, axis, keepdims):
        y = torch.min(x, axis)[0]
        # Since keepdims argument of torch not functional
        return y if keepdims else torch.squeeze(y, axis)

    def _compute_output_shape(x, axis, keepdims):
        if axis is None:
            return ()

        shape = list(_get_shape(x))
        if keepdims:
            shape[axis] = 1
        else:
            del shape[axis]

        return tuple(shape)

    return get_op(_min, output_shape=_compute_output_shape, arguments=[axis, keepdims])(x)

def lr_grad_norm_avg(self):
    # this is for enforcing lr * grad_norm not
    # increasing dramatically in case of instability.
    #  Not necessary for basic use.
    global_state = self._global_state
    beta = self._beta
    if "lr_grad_norm_avg" not in global_state:
      global_state['grad_norm_squared_avg_log'] = 0.0
    global_state['grad_norm_squared_avg_log'] = \
      global_state['grad_norm_squared_avg_log'] * beta \
      + (1 - beta) * np.log(global_state['grad_norm_squared'] + eps)
    if "lr_grad_norm_avg" not in global_state:
      global_state["lr_grad_norm_avg"] = \
        0.0 * beta + (1 - beta) * np.log(self._lr * np.sqrt(global_state['grad_norm_squared'] ) + eps)
      # we monitor the minimal smoothed ||lr * grad||
      global_state["lr_grad_norm_avg_min"] = \
        np.exp(global_state["lr_grad_norm_avg"] / self.zero_debias_factor() )
    else:
      global_state["lr_grad_norm_avg"] = global_state["lr_grad_norm_avg"] * beta \
        + (1 - beta) * np.log(self._lr * np.sqrt(global_state['grad_norm_squared'] ) + eps)
      global_state["lr_grad_norm_avg_min"] = \
        min(global_state["lr_grad_norm_avg_min"],
            np.exp(global_state["lr_grad_norm_avg"] / self.zero_debias_factor() ) )

def logaddexp(x1: T.FloatTensor, x2: T.FloatTensor) -> T.FloatTensor:
    """
    Elementwise logaddexp function: log(exp(x1) + exp(x2))

    Args:
        x1: A tensor.
        x2: A tensor.

    Returns:
        tensor: Elementwise logaddexp.

    """
    # log(exp(x1) + exp(x2))
    # = log( exp(x1) (1 + exp(x2 - x1))) = x1 + log(1 + exp(x2 - x1))
    # = log( exp(x2) (exp(x1 - x2) + 1)) = x2 + log(1 + exp(x1 - x2))
    diff = torch.min(x2 - x1, x1 - x2)
    return torch.max(x1, x2) + torch.log1p(exp(diff))

def tmin(x: T.FloatTensor,
         axis: int = None,
         keepdims: bool = False) -> T.FloatingPoint:
    """
    Return the elementwise minimum of a tensor along the specified axis.

    Args:
        x: A float or tensor.
        axis (optional): The axis for taking the minimum.
        keepdims (optional): If this is set to true, the dimension of the tensor
                             is unchanged. Otherwise, the reduced axis is removed
                             and the dimension of the array is 1 less.

    Returns:
        if axis is None:
            float: The overall minimum of the elements in the tensor
        else:
            tensor: The minimum of the tensor along the specified axis.

    """
    if axis is not None:
        return x.min(dim=axis, keepdim=keepdims)[0]
    else:
        return x.min()

def get_update(self):
        actions, log_actions, rewards, critics, entropies, states, advantages = self._sample()
        # Compute auxiliary losses
        critics = self.critic(states)
        critic_loss = (rewards - critics).pow(2).mean()
        critic_loss = self.critic_weight * critic_loss
        entropy_loss = entropies.mean()
        entropy_loss = - self.entropy_weight * entropy_loss
        # Compute policy loss
        advantages = advantages.detach().view(-1, 1)
        new_actions = self.policy(states)
        log_probs = new_actions.compute_log_prob(actions)
        ratios = (log_probs - log_actions.detach()).exp()
        surr = ratios.view(-1, 1) * advantages
        clipped = th.clamp(ratios, 1.0 - self.clip, 1.0 + self.clip).view(-1, 1) * advantages
        policy_loss = - th.min(surr, clipped).mean()
        # Proceed to optimization
        loss = policy_loss + critic_loss + entropy_loss
        if self.epoch_optimized == self.num_epochs:
            loss.backward(retain_graph=False)
        else:
            loss.backward(retain_graph=True)
        if self.grad_clip > 0.0:
            th.nn.utils.clip_grad_norm(self.parameters(), self.grad_clip)

        # Store statistics
        self.stats['Num. Updates'] += 1.0
        self.stats['Critic Loss'] += critic_loss.data[0]
        self.stats['Entropy Loss'] += entropy_loss.data[0]
        self.stats['Policy Loss'] += policy_loss.data[0]
        self.stats['Total Loss'] += loss.data[0]
        return [p.grad.clone() for p in self.parameters()]

def get_lr(self):
    self._lr_t = (1.0 - math.sqrt(self._mu_t) )**2 / (self._h_min + eps)
    # slow start of lr to prevent huge lr when there is only a few iteration finished
    self._lr_t = min(self._lr_t, self._lr_t * (self._iter + 1) / float(10.0 * self._curv_win_width) )
    return

def updateOutput(self, input):
        self._lazyInit()
        dimension = self._getPositiveDimension(input)
        torch.min(self._output, self._indices, input, dimension)
        if input.dim() > 1:
          self.output.set_(self._output.select(dimension, 0))
        else:
          self.output.set_(self._output)

        return self.output

def type(self, type, tensorCache=None):
        # torch.min expects a LongTensor as indices, whereas cutorch.max expects a CudaTensor.
        if type == 'torch.cuda.FloatTensor':
            indices, self._indices = self._indices, None
            super(Min, self).type(type, tensorCache)
            self._indices = indices.type('torch.cuda.LongTensor') if indices else None
        else:
            # self._indices must be a LongTensor. Setting it to nil temporarily avoids
            # unnecessary memory allocations.
            indices, self._indices = self._indices, None
            super(Min, self).type(type, tensorCache)
            self._indices = indices.long() if indices else None

        return self

def test_min(self):
        self._testSelection(torch.min, min)

def test_cmin(self):
        self._testCSelection(torch.cmin, min)

def test_clamp(self):
        m1 = torch.rand(100).mul(5).add(-2.5)  # uniform in [-2.5, 2.5]
        # just in case we're extremely lucky.
        min_val = -1
        max_val = 1
        m1[1] = min_val
        m1[2] = max_val

        res1 = m1.clone()
        res1.clamp_(min_val, max_val)
        res2 = m1.clone()
        for i in iter_indices(res2):
            res2[i] = max(min_val, min(max_val, res2[i]))
        self.assertEqual(res1, res2)

def bbox_iou(box1, box2, x1y1x2y2=True):
    if x1y1x2y2:
        mx = min(box1[0], box2[0])
        Mx = max(box1[2], box2[2])
        my = min(box1[1], box2[1])
        My = max(box1[3], box2[3])
        w1 = box1[2] - box1[0]
        h1 = box1[3] - box1[1]
        w2 = box2[2] - box2[0]
        h2 = box2[3] - box2[1]
    else:
        mx = min(box1[0]-box1[2]/2.0, box2[0]-box2[2]/2.0)
        Mx = max(box1[0]+box1[2]/2.0, box2[0]+box2[2]/2.0)
        my = min(box1[1]-box1[3]/2.0, box2[1]-box2[3]/2.0)
        My = max(box1[1]+box1[3]/2.0, box2[1]+box2[3]/2.0)
        w1 = box1[2]
        h1 = box1[3]
        w2 = box2[2]
        h2 = box2[3]
    uw = Mx - mx
    uh = My - my
    cw = w1 + w2 - uw
    ch = h1 + h2 - uh
    carea = 0
    if cw <= 0 or ch <= 0:
        return 0.0

    area1 = w1 * h1
    area2 = w2 * h2
    carea = cw * ch
    uarea = area1 + area2 - carea
    return carea/uarea

def bbox_ious(boxes1, boxes2, x1y1x2y2=True):
    if x1y1x2y2:
        mx = torch.min(boxes1[0], boxes2[0])
        Mx = torch.max(boxes1[2], boxes2[2])
        my = torch.min(boxes1[1], boxes2[1])
        My = torch.max(boxes1[3], boxes2[3])
        w1 = boxes1[2] - boxes1[0]
        h1 = boxes1[3] - boxes1[1]
        w2 = boxes2[2] - boxes2[0]
        h2 = boxes2[3] - boxes2[1]
    else:
        mx = torch.min(boxes1[0]-boxes1[2]/2.0, boxes2[0]-boxes2[2]/2.0)
        Mx = torch.max(boxes1[0]+boxes1[2]/2.0, boxes2[0]+boxes2[2]/2.0)
        my = torch.min(boxes1[1]-boxes1[3]/2.0, boxes2[1]-boxes2[3]/2.0)
        My = torch.max(boxes1[1]+boxes1[3]/2.0, boxes2[1]+boxes2[3]/2.0)
        w1 = boxes1[2]
        h1 = boxes1[3]
        w2 = boxes2[2]
        h2 = boxes2[3]
    uw = Mx - mx
    uh = My - my
    cw = w1 + w2 - uw
    ch = h1 + h2 - uh
    mask = ((cw <= 0) + (ch <= 0) > 0)
    area1 = w1 * h1
    area2 = w2 * h2
    carea = cw * ch
    carea[mask] = 0
    uarea = area1 + area2 - carea
    return carea/uarea

def updateOutput(self, input):
        self._lazyInit()
        dimension = self._getPositiveDimension(input)
        torch.min(input, dimension, out=(self._output, self._indices))
        if input.dim() > 1:
            self.output.set_(self._output.select(dimension, 0))
        else:
            self.output.set_(self._output)

        return self.output

def type(self, type, tensorCache=None):
        # torch.min expects a LongTensor as indices, whereas cutorch.max expects a CudaTensor.
        if type == 'torch.cuda.FloatTensor':
            indices, self._indices = self._indices, None
            super(Min, self).type(type, tensorCache)
            self._indices = indices.type('torch.cuda.LongTensor') if indices is not None else None
        else:
            # self._indices must be a LongTensor. Setting it to nil temporarily avoids
            # unnecessary memory allocations.
            indices, self._indices = self._indices, None
            super(Min, self).type(type, tensorCache)
            self._indices = indices.long() if indices is not None else None

        return self

def test_min(self):
        self._testSelection(torch.min, min)

def test_min_elementwise(self):
        self._testCSelection(torch.min, min)