| { | |
| "ckpt_path": null, | |
| "cond_stage_config": { | |
| "params": { | |
| "max_length": 77, | |
| "version": "openai/clip-vit-large-patch14", | |
| "zero_embedding_radio": 0.1 | |
| }, | |
| "target": "michelangelo.models.conditional_encoders.encoder_factory.FrozenAlignedCLIPTextEmbedder" | |
| }, | |
| "cond_stage_key": "text", | |
| "denoiser_cfg": { | |
| "params": { | |
| "context_dim": 768, | |
| "heads": 12, | |
| "init_scale": 1.0, | |
| "input_channels": 64, | |
| "layers": 8, | |
| "n_ctx": 256, | |
| "output_channels": 64, | |
| "skip_ln": true, | |
| "use_checkpoint": true, | |
| "width": 768 | |
| }, | |
| "target": "michelangelo.models.asl_diffusion.asl_udt.ConditionalASLUDTDenoiser" | |
| }, | |
| "first_stage_config": { | |
| "params": { | |
| "aligned_module_cfg": { | |
| "params": { | |
| "clip_model_version": "openai/clip-vit-large-patch14" | |
| }, | |
| "target": "michelangelo.models.tsal.clip_asl_module.CLIPAlignedShapeAsLatentModule" | |
| }, | |
| "loss_cfg": { | |
| "target": "torch.nn.Identity" | |
| }, | |
| "shape_module_cfg": { | |
| "params": { | |
| "embed_dim": 64, | |
| "heads": 12, | |
| "include_pi": false, | |
| "init_scale": 0.25, | |
| "num_decoder_layers": 16, | |
| "num_encoder_layers": 8, | |
| "num_freqs": 8, | |
| "num_latents": 256, | |
| "point_feats": 3, | |
| "qkv_bias": false, | |
| "use_checkpoint": true, | |
| "use_ln_post": true, | |
| "width": 768 | |
| }, | |
| "target": "michelangelo.models.tsal.sal_perceiver.AlignedShapeLatentPerceiver" | |
| } | |
| }, | |
| "target": "michelangelo.models.tsal.asl_pl_module.AlignedShapeAsLatentPLModule" | |
| }, | |
| "first_stage_key": "surface", | |
| "ignore_keys": [], | |
| "loss_cfg": { | |
| "loss_type": "mse" | |
| }, | |
| "optimizer_cfg": { | |
| "optimizer": { | |
| "params": { | |
| "betas": [ | |
| 0.9, | |
| 0.99 | |
| ], | |
| "eps": 1e-06, | |
| "weight_decay": 0.01 | |
| }, | |
| "target": "torch.optim.AdamW" | |
| }, | |
| "scheduler": { | |
| "params": { | |
| "f_max": 1.0, | |
| "f_min": 0.001, | |
| "f_start": 1e-06, | |
| "warm_up_steps": 5000 | |
| }, | |
| "target": "michelangelo.utils.trainings.lr_scheduler.LambdaWarmUpCosineFactorScheduler" | |
| } | |
| }, | |
| "scale_by_std": false, | |
| "scheduler_cfg": { | |
| "denoise": { | |
| "params": { | |
| "beta_end": 0.012, | |
| "beta_schedule": "scaled_linear", | |
| "beta_start": 0.00085, | |
| "clip_sample": false, | |
| "num_train_timesteps": 1000, | |
| "set_alpha_to_one": false, | |
| "steps_offset": 1 | |
| }, | |
| "target": "diffusers.schedulers.DDIMScheduler" | |
| }, | |
| "eta": 0.0, | |
| "guidance_scale": 7.5, | |
| "noise": { | |
| "params": { | |
| "beta_end": 0.012, | |
| "beta_schedule": "scaled_linear", | |
| "beta_start": 0.00085, | |
| "clip_sample": false, | |
| "num_train_timesteps": 1000, | |
| "variance_type": "fixed_small" | |
| }, | |
| "target": "diffusers.schedulers.DDPMScheduler" | |
| }, | |
| "num_inference_steps": 50 | |
| }, | |
| "z_scale_factor": 1.0 | |
| } |