multi scale samper Args: data_source(dataset) scales(list): several scales for image resolution first_bs(int): batch size for the first scale in scales divided_factor(list[w, h]): ImageNet models down-sample ima
(self,
data_source,
scales,
first_bs=512,
fix_bs=True,
divided_factor=[8, 16],
is_training=True,
max_ratio=10,
max_bs=1024,
seed=None)
| 10 | class RatioSampler(Sampler): |
| 11 | |
| 12 | def __init__(self, |
| 13 | data_source, |
| 14 | scales, |
| 15 | first_bs=512, |
| 16 | fix_bs=True, |
| 17 | divided_factor=[8, 16], |
| 18 | is_training=True, |
| 19 | max_ratio=10, |
| 20 | max_bs=1024, |
| 21 | seed=None): |
| 22 | """ |
| 23 | multi scale samper |
| 24 | Args: |
| 25 | data_source(dataset) |
| 26 | scales(list): several scales for image resolution |
| 27 | first_bs(int): batch size for the first scale in scales |
| 28 | divided_factor(list[w, h]): ImageNet models down-sample images by a factor, ensure that width and height dimensions are multiples are multiple of devided_factor. |
| 29 | is_training(boolean): mode |
| 30 | """ |
| 31 | # min. and max. spatial dimensions |
| 32 | self.data_source = data_source |
| 33 | # self.data_idx_order_list = np.array(data_source.data_idx_order_list) |
| 34 | self.ds_width = data_source.ds_width |
| 35 | self.seed = data_source.seed |
| 36 | if self.ds_width: |
| 37 | self.wh_ratio = data_source.wh_ratio |
| 38 | self.wh_ratio_sort = data_source.wh_ratio_sort |
| 39 | self.n_data_samples = len(self.data_source) |
| 40 | self.max_ratio = max_ratio |
| 41 | self.max_bs = max_bs |
| 42 | |
| 43 | if isinstance(scales[0], list): |
| 44 | width_dims = [i[0] for i in scales] |
| 45 | height_dims = [i[1] for i in scales] |
| 46 | elif isinstance(scales[0], int): |
| 47 | width_dims = scales |
| 48 | height_dims = scales |
| 49 | base_im_w = width_dims[0] |
| 50 | base_im_h = height_dims[0] |
| 51 | base_batch_size = first_bs |
| 52 | base_elements = base_im_w * base_im_h * base_batch_size |
| 53 | self.base_elements = base_elements |
| 54 | self.base_batch_size = base_batch_size |
| 55 | self.base_im_h = base_im_h |
| 56 | self.base_im_w = base_im_w |
| 57 | |
| 58 | # Get the GPU and node related information |
| 59 | num_replicas = torch.cuda.device_count() if torch.cuda.is_available( |
| 60 | ) else 1 |
| 61 | # rank = dist.get_rank() |
| 62 | rank = (int(os.environ['LOCAL_RANK']) |
| 63 | if 'LOCAL_RANK' in os.environ else 0) |
| 64 | # self.rank = rank |
| 65 | # adjust the total samples to avoid batch dropping |
| 66 | num_samples_per_replica = int( |
| 67 | math.ceil(self.n_data_samples * 1.0 / num_replicas)) |
| 68 | |
| 69 | img_indices = [idx for idx in range(self.n_data_samples)] |
nothing calls this directly
no test coverage detected