Compresser Encoder (Perceiver Resampler)
Phase 0 pretrained Perceiver for the Mamba-3 Semantic Video Compressor.
Architecture
- Type: Perceiver Resampler (cross-attention compressor)
- Input: [B, 576, 1664] — V-JEPA 2.1 ViT-Gigantic patch latents
- Output: [B, 64, 512] — compressed tokens
- Params: ~20.6M
- Details: 64 learned queries, 6 cross-attention layers, 16 heads, FFN 512→2048
Training
- Dataset: Vjepa_mamba_dataset_v2 (50 hours video, 384×384, 8fps)
- V-JEPA: Frozen vjepa2_1_vit_gigantic_384 (2.2B params)
- Loss: MSE reconstruction via autoencoder (Perceiver → Decoder → V-JEPA latent)
- Optimizer: AdamW, lr=1e-4, cosine to 1e-6
- Hardware: RTX 4090 (48 GB), bf16
Usage
from model.models.perceiver import PerceiverResampler
model = PerceiverResampler(input_dim=1664, output_dim=512, num_queries=64)
model.load_state_dict(torch.load("perceiver_stepX_hrsY.pt"))
# Input: [B, 576, 1664] V-JEPA latents → Output: [B, 64, 512]
Part of the Mamba-3 Semantic Video Compressor pipeline.