Spaces:
Sleeping
Sleeping
| import itertools | |
| from omegaconf import OmegaConf | |
| def bifpn_config(min_level, max_level, weight_method=None): | |
| """BiFPN config. | |
| Adapted from https://github.com/google/automl/blob/56815c9986ffd4b508fe1d68508e268d129715c1/efficientdet/keras/fpn_configs.py | |
| """ | |
| p = OmegaConf.create() | |
| weight_method = weight_method or 'fastattn' | |
| num_levels = max_level - min_level + 1 | |
| node_ids = {min_level + i: [i] for i in range(num_levels)} | |
| level_last_id = lambda level: node_ids[level][-1] | |
| level_all_ids = lambda level: node_ids[level] | |
| id_cnt = itertools.count(num_levels) | |
| p.nodes = [] | |
| for i in range(max_level - 1, min_level - 1, -1): | |
| # top-down path. | |
| p.nodes.append({ | |
| 'reduction': 1 << i, | |
| 'inputs_offsets': [level_last_id(i), level_last_id(i + 1)], | |
| 'weight_method': weight_method, | |
| }) | |
| node_ids[i].append(next(id_cnt)) | |
| for i in range(min_level + 1, max_level + 1): | |
| # bottom-up path. | |
| p.nodes.append({ | |
| 'reduction': 1 << i, | |
| 'inputs_offsets': level_all_ids(i) + [level_last_id(i - 1)], | |
| 'weight_method': weight_method, | |
| }) | |
| node_ids[i].append(next(id_cnt)) | |
| return p | |
| def panfpn_config(min_level, max_level, weight_method=None): | |
| """PAN FPN config. | |
| This defines FPN layout from Path Aggregation Networks as an alternate to | |
| BiFPN, it does not implement the full PAN spec. | |
| Paper: https://arxiv.org/abs/1803.01534 | |
| """ | |
| p = OmegaConf.create() | |
| weight_method = weight_method or 'fastattn' | |
| num_levels = max_level - min_level + 1 | |
| node_ids = {min_level + i: [i] for i in range(num_levels)} | |
| level_last_id = lambda level: node_ids[level][-1] | |
| id_cnt = itertools.count(num_levels) | |
| p.nodes = [] | |
| for i in range(max_level, min_level - 1, -1): | |
| # top-down path. | |
| offsets = [level_last_id(i), level_last_id(i + 1)] if i != max_level else [level_last_id(i)] | |
| p.nodes.append({ | |
| 'reduction': 1 << i, | |
| 'inputs_offsets': offsets, | |
| 'weight_method': weight_method, | |
| }) | |
| node_ids[i].append(next(id_cnt)) | |
| for i in range(min_level, max_level + 1): | |
| # bottom-up path. | |
| offsets = [level_last_id(i), level_last_id(i - 1)] if i != min_level else [level_last_id(i)] | |
| p.nodes.append({ | |
| 'reduction': 1 << i, | |
| 'inputs_offsets': offsets, | |
| 'weight_method': weight_method, | |
| }) | |
| node_ids[i].append(next(id_cnt)) | |
| return p | |
| def qufpn_config(min_level, max_level, weight_method=None): | |
| """A dynamic quad fpn config that can adapt to different min/max levels. | |
| It extends the idea of BiFPN, and has four paths: | |
| (up_down -> bottom_up) + (bottom_up -> up_down). | |
| Paper: https://ieeexplore.ieee.org/document/9225379 | |
| Ref code: From contribution to TF EfficientDet | |
| https://github.com/google/automl/blob/eb74c6739382e9444817d2ad97c4582dbe9a9020/efficientdet/keras/fpn_configs.py | |
| """ | |
| p = OmegaConf.create() | |
| weight_method = weight_method or 'fastattn' | |
| quad_method = 'fastattn' | |
| num_levels = max_level - min_level + 1 | |
| node_ids = {min_level + i: [i] for i in range(num_levels)} | |
| level_last_id = lambda level: node_ids[level][-1] | |
| level_all_ids = lambda level: node_ids[level] | |
| level_first_id = lambda level: node_ids[level][0] | |
| id_cnt = itertools.count(num_levels) | |
| p.nodes = [] | |
| for i in range(max_level - 1, min_level - 1, -1): | |
| # top-down path 1. | |
| p.nodes.append({ | |
| 'reduction': 1 << i, | |
| 'inputs_offsets': [level_last_id(i), level_last_id(i + 1)], | |
| 'weight_method': weight_method | |
| }) | |
| node_ids[i].append(next(id_cnt)) | |
| node_ids[max_level].append(node_ids[max_level][-1]) | |
| for i in range(min_level + 1, max_level): | |
| # bottom-up path 2. | |
| p.nodes.append({ | |
| 'reduction': 1 << i, | |
| 'inputs_offsets': level_all_ids(i) + [level_last_id(i - 1)], | |
| 'weight_method': weight_method | |
| }) | |
| node_ids[i].append(next(id_cnt)) | |
| i = max_level | |
| p.nodes.append({ | |
| 'reduction': 1 << i, | |
| 'inputs_offsets': [level_first_id(i)] + [level_last_id(i - 1)], | |
| 'weight_method': weight_method | |
| }) | |
| node_ids[i].append(next(id_cnt)) | |
| node_ids[min_level].append(node_ids[min_level][-1]) | |
| for i in range(min_level + 1, max_level + 1, 1): | |
| # bottom-up path 3. | |
| p.nodes.append({ | |
| 'reduction': 1 << i, | |
| 'inputs_offsets': [ | |
| level_first_id(i), level_last_id(i - 1) if i != min_level + 1 else level_first_id(i - 1)], | |
| 'weight_method': weight_method | |
| }) | |
| node_ids[i].append(next(id_cnt)) | |
| node_ids[min_level].append(node_ids[min_level][-1]) | |
| for i in range(max_level - 1, min_level, -1): | |
| # top-down path 4. | |
| p.nodes.append({ | |
| 'reduction': 1 << i, | |
| 'inputs_offsets': [node_ids[i][0]] + [node_ids[i][-1]] + [level_last_id(i + 1)], | |
| 'weight_method': weight_method | |
| }) | |
| node_ids[i].append(next(id_cnt)) | |
| i = min_level | |
| p.nodes.append({ | |
| 'reduction': 1 << i, | |
| 'inputs_offsets': [node_ids[i][0]] + [level_last_id(i + 1)], | |
| 'weight_method': weight_method | |
| }) | |
| node_ids[i].append(next(id_cnt)) | |
| node_ids[max_level].append(node_ids[max_level][-1]) | |
| # NOTE: the order of the quad path is reversed from the original, my code expects the output of | |
| # each FPN repeat to be same as input from backbone, in order of increasing reductions | |
| for i in range(min_level, max_level + 1): | |
| # quad-add path. | |
| p.nodes.append({ | |
| 'reduction': 1 << i, | |
| 'inputs_offsets': [node_ids[i][2], node_ids[i][4]], | |
| 'weight_method': quad_method | |
| }) | |
| node_ids[i].append(next(id_cnt)) | |
| return p | |
| def get_fpn_config(fpn_name, min_level=3, max_level=7): | |
| if not fpn_name: | |
| fpn_name = 'bifpn_fa' | |
| name_to_config = { | |
| 'bifpn_sum': bifpn_config(min_level=min_level, max_level=max_level, weight_method='sum'), | |
| 'bifpn_attn': bifpn_config(min_level=min_level, max_level=max_level, weight_method='attn'), | |
| 'bifpn_fa': bifpn_config(min_level=min_level, max_level=max_level, weight_method='fastattn'), | |
| 'pan_sum': panfpn_config(min_level=min_level, max_level=max_level, weight_method='sum'), | |
| 'pan_fa': panfpn_config(min_level=min_level, max_level=max_level, weight_method='fastattn'), | |
| 'qufpn_sum': qufpn_config(min_level=min_level, max_level=max_level, weight_method='sum'), | |
| 'qufpn_fa': qufpn_config(min_level=min_level, max_level=max_level, weight_method='fastattn'), | |
| } | |
| return name_to_config[fpn_name] | |