simoswish/Reinforcment-Learning-Project-2526-checkpoints

Reinforcment-Learning-Project-2526 — training checkpoints

Checkpoint archive for simoswish02/Reinforcment-Learning-Project-2526: multi-drone cooperative search on a 32×32 grid with a shared Dueling Double DQN (ConvNeXt + MHSA global encoder, local CNN, dueling head).

The checkpoints were previously stored in the GitHub repository via Git LFS and moved here to free LFS storage.

Layout

Files are grouped by the Git branch they came from, keeping their original path:

<branch>/checkpoints/...

• convnext_attn_net/checkpoints/ — delivery run reported in the project (best.pt, last.pt, phase2_ConvNeXT_10k_dr_ep{1000..10000}.pt), plus earlier Phase0/ and Phase3/ runs from that branch's history.
• Other branch folders (32x32_Grid, 64x64_Grid, FixedMapSize_*, GUI_upgrade, feat/map-drone-position, main) — best.pt/last.pt of earlier experimental architectures. Their periodic snapshots were not archived.
• When a path held several versions over time, older versions carry a __<sha256[:8]> suffix (e.g. best__7bb8381b.pt).

manifest.csv lists every file with its SHA-256, size, source branch, original path and the commit (hash, date, message) that introduced it. The SHA-256 equals the former Git LFS object id.

Usage

git clone https://github.com/simoswish02/Reinforcment-Learning-Project-2526.git
cd Reinforcment-Learning-Project-2526
hf download simoswish/Reinforcment-Learning-Project-2526-checkpoints --include "convnext_attn_net/checkpoints/*.pt" --local-dir hf_ckpt
mkdir -p checkpoints && cp hf_ckpt/convnext_attn_net/checkpoints/*.pt checkpoints/

python main.py --mode eval --checkpoint checkpoints/best.pt

Checkpoints are torch.save dicts written by DQNAgent.save (keys include q_net, optimizer/scheduler state and training metadata). Checkpoints from other branches match the network code of that branch, not the delivery branch.