erickfm/frame-checkpoints

MIMIC Checkpoints

Trained model checkpoints from MIMIC experiments. Each checkpoint contains the model weights, optimizer state, scheduler state, normalization stats, and full model config.

Trained on erickfm/frame-melee (~95k tournament replays).

Naming Convention

Filenames encode the full config:

d{d_model}_L{layers}_ff{feedforward}_{encoder}_pos-{pos_enc}_seq{seq_len}_sl-{stick_loss}_bl-{btn_loss}_step{step}.pt

Checkpoints

Checkpoint	d_model	Layers	Encoder	Pos Enc	Seq Len	Stick Loss	Size
d768_L4_..._seq180_sl-mse_..._step15625	768	4	hybrid16	learned	180	mse	391MB
d768_L4_..._seq60_sl-huber_..._step5208	768	4	hybrid16	learned	60	huber	389MB
d768_L4_..._seq60_sl-mse_..._step18000	768	4	hybrid16	learned	60	mse	389MB
d768_L4_..._seq90_..._step7812	768	4	hybrid16	learned	90	mse	390MB
d768_L4_..._seq120_..._step10416	768	4	hybrid16	learned	120	mse	390MB
d768_L4_..._seq30_..._step2604	768	4	hybrid16	learned	30	mse	389MB
d768_L4_..._sl-discrete_..._step5208	768	4	hybrid16	learned	60	discrete	393MB
d768_L8_..._step10416	768	8	hybrid16	learned	60	mse	730MB
d1024_L4_..._hybrid16_..._step24000	1024	4	hybrid16	learned	60	mse	659MB
d1024_L4_..._default_..._step30000	1024	4	default	learned	60	mse	708MB
d1024_L8_..._step34375	1024	8	hybrid16	learned	60	mse	1.3GB
d1536_L2_..._step62500	1536	2	hybrid16	learned	60	mse	744MB
d1536_L8_..._step37500	1536	8	hybrid16	learned	60	mse	2.8GB
d512_L4_..._pos-alibi_..._step15625	512	4	hybrid16	alibi	60	mse	195MB
d512_L4_..._pos-rope_..._step15625	512	4	hybrid16	rope	60	mse	195MB
d512_L4_..._seq60_..._step30000	512	4	hybrid16	learned	60	mse	196MB
d512_L4_..._seq120_..._step31240	512	4	hybrid16	learned	120	mse	196MB
d512_L4_..._seq180_..._step41666	512	4	hybrid16	learned	180	mse	196MB
d512_L4_..._seq240_..._step62500	512	4	hybrid16	learned	240	mse	197MB
d512_L4_..._seq360_..._step62500	512	4	hybrid16	learned	360	mse	197MB
d512_L8_..._hybrid16_..._step31250	512	8	hybrid16	learned	60	mse	347MB
d512_L8_..._composite8_..._step30000	512	8	composite8	learned	60	mse	341MB

Loading

import torch
from model import FramePredictor, ModelConfig

ckpt = torch.load("checkpoint.pt", map_location="cpu")
cfg = ModelConfig(**ckpt["config"])
model = FramePredictor(cfg)
model.load_state_dict(ckpt["model_state_dict"])
model.eval()

Note on Missing Checkpoints

Multiple runs per machine shared the same checkpoints/ directory with step-number filenames. Runs that finished at the same step count overwrote each other. These 23 checkpoints are the surviving unique configs. Some notable runs (depth-2L, depth-6L, focal-btn, baseline at step 5208) were overwritten.

Related

• MIMIC -- Training code and inference
• erickfm/frame-melee -- Training dataset
• RESULTS.md -- Full experiment results