| import torch |
|
|
| from model.wan_kv_mmd import WanKVMMD |
| from video_fd.encoders.vjepa2_dense_metadata import ( |
| canonical_vjepa2_dense_feature_metadata, |
| uniform_dense_token_indices, |
| ) |
|
|
|
|
| def test_triple_vjepa_metadata_matches_bank_last_64_contract(): |
| model = object.__new__(WanKVMMD) |
| model.feature_type = "triple_dinov3_vjepa2_videomae_v2_dense_latent" |
| model.dense_feature_prefix = "triple" |
| model.dense_encoder_name = "triple" |
| model.dense_log_prefix = "triple_dense" |
| model.vjepa2_reference_model_id = "vjepa2-vitl-fpc64-256" |
| model.vjepa2_model_id = "/unused/local/model" |
| model.vjepa2_dense_layer = 15 |
| model.vjepa2_dense_layers = (15,) |
| model.vjepa2_dense_num_frames = 64 |
| model.vjepa2_dense_source_num_frames = 81 |
| model.vjepa2_dense_frame_sampling = "last_contiguous" |
| model.vjepa2_dense_resize_short_side = 292 |
| model.vjepa2_dense_crop_size = 256 |
| model.vjepa2_dense_raw_tokens = 2048 |
| model.vjepa2_dense_num_tokens = 1024 |
| model.vjepa2_dense_feature_dim = 1024 |
| indices = uniform_dense_token_indices(2048, 1024) |
| actual = model._vjepa2_feature_metadata(indices) |
| expected = canonical_vjepa2_dense_feature_metadata( |
| model_id="vjepa2-vitl-fpc64-256", |
| layer_index=15, |
| num_frames=64, |
| resize_short_side=292, |
| crop_size=256, |
| raw_tokens=2048, |
| num_tokens=1024, |
| feature_dim=1024, |
| token_indices=indices, |
| frame_sampling="last_contiguous", |
| source_num_frames=81, |
| ) |
| assert actual == expected |
| assert actual["selected_frame_start"] == 17 |
| assert actual["selected_frame_end_inclusive"] == 80 |
|
|
|
|
| def test_dual_dinov3_vjepa_metadata_uses_explicit_uniform_indices_by_default(): |
| model = object.__new__(WanKVMMD) |
| model.feature_type = "dual_dinov3_vjepa2_dense_latent" |
| model.dense_feature_prefix = "dinov3_vjepa2" |
| model.dense_encoder_name = "dinov3_vjepa2" |
| model.dense_log_prefix = "dinov3_vjepa2_dense" |
| model.vjepa2_reference_model_id = "vjepa2-vitl-fpc64-256" |
| model.vjepa2_model_id = "/unused/local/model" |
| model.vjepa2_dense_layer = 15 |
| model.vjepa2_dense_layers = (15,) |
| model.vjepa2_dense_num_frames = 64 |
| model.vjepa2_dense_source_num_frames = 81 |
| model.vjepa2_dense_frame_sampling = "center_contiguous" |
| model.vjepa2_dense_resize_short_side = 292 |
| model.vjepa2_dense_crop_size = 256 |
| model.vjepa2_dense_raw_tokens = 2048 |
| model.vjepa2_dense_num_tokens = 1024 |
| model.vjepa2_dense_feature_dim = 1024 |
| actual = model._vjepa2_feature_metadata() |
| indices = uniform_dense_token_indices(2048, 1024) |
| expected = canonical_vjepa2_dense_feature_metadata( |
| model_id="vjepa2-vitl-fpc64-256", |
| layer_index=15, |
| num_frames=64, |
| resize_short_side=292, |
| crop_size=256, |
| raw_tokens=2048, |
| num_tokens=1024, |
| feature_dim=1024, |
| token_indices=indices, |
| frame_sampling="last_contiguous", |
| source_num_frames=81, |
| ) |
| assert actual == expected |
| assert actual["token_alignment"] == "same_explicit_indices_and_order" |
|
|
