| import torch |
| import torch.nn as nn |
| from einops.layers.torch import Rearrange |
|
|
| from .cbramod import CBraMod |
|
|
|
|
| class Model(nn.Module): |
| def __init__(self, param): |
| super(Model, self).__init__() |
| self.backbone = CBraMod( |
| in_dim=200, out_dim=200, d_model=200, |
| dim_feedforward=800, seq_len=30, |
| n_layer=12, nhead=8 |
| ) |
| if param.use_pretrained_weights: |
| map_location = torch.device(f'cuda:{param.cuda}') |
| self.backbone.load_state_dict(torch.load(param.foundation_dir, map_location=map_location)) |
| self.backbone.proj_out = nn.Identity() |
| if param.classifier == 'avgpooling_patch_reps': |
| self.classifier = nn.Sequential( |
| Rearrange('b c s d -> b d c s'), |
| nn.AdaptiveAvgPool2d((1, 1)), |
| nn.Flatten(), |
| nn.Linear(200, param.num_of_classes), |
| ) |
| elif param.classifier == 'all_patch_reps_onelayer': |
| self.classifier = nn.Sequential( |
| Rearrange('b c s d -> b (c s d)'), |
| nn.Linear(64 * 4 * 200, param.num_of_classes), |
| ) |
| elif param.classifier == 'all_patch_reps_twolayer': |
| self.classifier = nn.Sequential( |
| Rearrange('b c s d -> b (c s d)'), |
| nn.Linear(64 * 4 * 200, 200), |
| nn.ELU(), |
| nn.Dropout(param.dropout), |
| nn.Linear(200, param.num_of_classes), |
| ) |
| elif param.classifier == 'all_patch_reps': |
| self.classifier = nn.Sequential( |
| Rearrange('b c s d -> b (c s d)'), |
| nn.Linear(64 * 4 * 200, 4 * 200), |
| nn.ELU(), |
| nn.Dropout(param.dropout), |
| nn.Linear(4 * 200, 200), |
| nn.ELU(), |
| nn.Dropout(param.dropout), |
| nn.Linear(200, param.num_of_classes), |
| ) |
|
|
|
|
| def forward(self, x): |
| bz, ch_num, seq_len, patch_size = x.shape |
| feats = self.backbone(x) |
| out = self.classifier(feats) |
| return out |
|
|
