Merge pull request #750 from drjinying/master

Specify "interpolation" mode in vision_transformer's resize_pos_embed

Author: rwightman

timm/models/nest.py

@@ -377,7 +377,7 @@ def resize_pos_embed(posemb, posemb_new):
     size_new = int(math.sqrt(num_blocks_new*seq_length_new))
     # First change to (1, C, H, W)
     posemb = deblockify(posemb, int(math.sqrt(seq_length_old))).permute(0, 3, 1, 2)
-    posemb = F.interpolate(posemb, size=[size_new, size_new], mode='bilinear')
+    posemb = F.interpolate(posemb, size=[size_new, size_new], mode='bicubic', align_corners=False)
     # Now change to new (1, T, N, C)
     posemb = blockify(posemb.permute(0, 2, 3, 1), int(math.sqrt(seq_length_new)))
     return posemb

timm/models/vision_transformer.py

@@ -494,7 +494,7 @@ def resize_pos_embed(posemb, posemb_new, num_tokens=1, gs_new=()):
     assert len(gs_new) >= 2
     _logger.info('Position embedding grid-size from %s to %s', [gs_old, gs_old], gs_new)
     posemb_grid = posemb_grid.reshape(1, gs_old, gs_old, -1).permute(0, 3, 1, 2)
-    posemb_grid = F.interpolate(posemb_grid, size=gs_new, mode='bilinear')
+    posemb_grid = F.interpolate(posemb_grid, size=gs_new, mode='bicubic', align_corners=False)
     posemb_grid = posemb_grid.permute(0, 2, 3, 1).reshape(1, gs_new[0] * gs_new[1], -1)
     posemb = torch.cat([posemb_tok, posemb_grid], dim=1)
     return posemb