mmdiff / configs /nyu_depth_config.yaml
multimodalart's picture
multimodalart HF Staff
Upload folder using huggingface_hub
2267636 verified
Raw
History Blame Contribute Delete
3.45 kB
# NYU Depth Training Configuration - Native Resolution (Variable Resolution with FluxResizer)
# Uses FLUX.1-dev and FluxResizer for optimal native resolutions (no padding)
# Experiment settings
experiment:
name: "nyu_depth_native_resolution"
description: "NYU Depth with FLUX.1-dev at native resolutions using FluxResizer"
# Concept configurations (same as original)
concepts:
basic: ["depth", "surface", "object", "background"]
basic_2: ["depth", "surface", "object", "near", "far"]
indoor: ["wall", "floor", "furniture", "object", "depth"]
spatial: ["near", "far", "depth", "distance", "surface"]
detailed: ["wall", "floor", "ceiling", "furniture", "object", "depth", "surface"]
comprehensive: ["wall", "floor", "ceiling", "furniture", "object", "depth", "near", "far", "surface", "indoor", "outdoor", "structure"]
semantic: ["bedroom", "kitchen", "bathroom", "living_room", "depth", "furniture", "wall", "floor"]
geometric: ["depth", "surface", "edge", "plane", "corner", "boundary", "gradient", "distance"]
# Training settings
training:
concept_config: "basic_2" # Which concept configuration to use
epochs: 20
learning_rate: 1e-4
batch_size: 1 # REQUIRED: Must be 1 for variable resolution
accumulate_grad_batches: 4 # Effective batch size = 4
gradient_clip_val: 1.0
precision: "16-mixed"
# Data settings
data:
dataset: "nyu_depth_v2"
# NYU Depth V2 standard train/test split. Point the trainer at it via CLI args
# (--data_path / --filenames_path); see shell_scripts/train_nyu.sh.
num_classes: 1 # Depth is regression, but num_classes used for decoder
min_depth: 0.1 # Minimum valid depth (meters)
max_depth: 10.0 # Maximum valid depth (meters)
# NOTE: No target_size! FluxResizer selects optimal resolution per image
# Model settings
model:
flux_model: "black-forest-labs/FLUX.1-dev" # Changed from schnell to dev
dtype: "float16"
dino_model: "dinov3_vitb16" # DINOv3 base model
feature_locations:
transformer_blocks: [4, 9, 13, 18]
single_transformer_blocks: [4, 15, 26, 37]
decoder:
features: 256
hyperfeature_fusion:
num_timesteps: 4
fusion_type: "transformer"
hidden_dim: 768
num_transformer_layers: 3
layer_scale_init: 1e-4
# FLUX settings
flux:
timesteps: 28
guidance_scale: 3.5
num_inference_steps: 1
concept_timesteps: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27]
concept_layers: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17]
# Hardware settings
hardware:
devices: 1 # Number of GPUs (DDP for multi-GPU)
#strategy: "ddp_find_unused_parameters_true" # Required for DDP with conditional logic
num_sanity_val_steps: 1
detect_anomaly: true
accelerator: "gpu"
num_workers: 8
# Paths (timestamps will be automatically added)
paths:
# All outputs live under ${MMDIFF_OUTPUT} (set it to a writable dir, e.g. /work/<user>/mmdiff_out).
cache_base_dir: "${MMDIFF_OUTPUT}/cache"
log_base_dir: "${MMDIFF_OUTPUT}/logs"
checkpoint_base_dir: "${MMDIFF_OUTPUT}/checkpoints"
use_timestamp: true
permanent_cache_dir: "${MMDIFF_OUTPUT}/cache/nyu_native_feature_cache" # Native resolution feature cache
# Logging
logging:
log_every_n_steps: 10
save_top_k: 3
monitor: "val_loss"
mode: "min"
# HuggingFace
huggingface:
token: "" # Leave empty and authenticate via `huggingface-cli login` or the HF_TOKEN env var