Given initial keypoint predictions, we identify missing joints :param det: numpy.ndarray of size (17, 128, 128) :param tag: n

associative embedding loss for one image

Copy the ``.grad`` attribute from stored references to fp16 parameters to the ``.grad`` attribute of the fp32 master parameters that

Zero fp32 and fp16 parameter grads.

Args: index (int): Index Returns: tuple: Tuple (image, target). target is the object returned by ``coco.load

Args: index (int): Index Returns: tuple: Tuple (image, target). target is the object returned by ``coco.load

:attr:`backward` performs the following conceptual steps: 1. fp32_loss = loss.float() (see first Note below) 2. scaled_loss =

Clips fp32 master gradients via ``torch.nn.utils.clip_grad_norm``. Args: max_norm (float or int): max norm of the gradien

ouput_flipped: numpy.ndarray(batch_size, num_joints, height, width)

flip coords

accumulate the tag loss for each image in the batch

When running with :class:`FP16_Optimizer`, ``.grad`` attributes of a model's fp16 leaves should not be regarded as truthful,

Creates a list of FP32 master parameters for a given model, as in `Training Neural Networks with Mixed Precision: Real Examples`_. Args:

batch_image: [batch_size, channel, height, width] batch_joints: [batch_size, num_joints, 3], batch_joints_vis: [batch_size, num_joints, 1