Oriskany/acdc-mds-512px

ACDC 512px Consolidated MDS Dataset Shards

Dataset Summary

This dataset contains consolidated 512MB MosaicML Streaming (.mds) shards for ACDC, preprocessed for high-throughput monocular depth distillation (ZipDepth).

• Domain: acdc
• Shortest Edge: Strictly preserved at 512px ($\min(H, W) = 512\text{px}$).
• Teacher Supervision: Pseudo-depth distilled with Depth-Anything-V2-Large-hf ($100%$ dense pixel coverage).
• Compression: RGB JPEG (Quality 95) + 16-bit Unsigned PNG Depth (uint16, range $[0, 65535]$).
• Shard Size Limit: 512MB per shard (high throughput, optimal for py1e block shuffling).

Loading with MosaicML Streaming

from streaming import StreamingDataset

dataset = StreamingDataset(
    remote="hf://datasets/Oriskany/acdc-mds-512px/train",
    local="/tmp/stream_cache/acdc/train",
    shuffle=True,
    shuffle_algo="py1e",
    shuffle_block_size=16384,
    batch_size=64
)

for sample in dataset:
    rgb_bytes = sample["rgb"]
    depth_bytes = sample["depth"]
    meta = sample["meta"]