CGIL#
Arguments#
Options
- --clip_backbonestr
Help: Clip backbone
Default:
ViT-L/14
- --learning_rate_alignmentfloat
Help: Learning rate for GR.
Default:
0.05
- --optim_alignmentstr
Help: Optimizer for GR.
Default:
adamwChoices:
sgd, adam, adamw
- --optim_alignment_wdfloat
Help: Weight decay for GR.
Default:
0
- --lambda_ortho_first_stagefloat
Help: Orthogonality loss coefficient for coop
Default:
1
- --num_epochs_alignmentint
Help: Num. of epochs for GR.
Default:
30
- --batch_size_alignmentint
Help: Batch size for alignment.
Default:
128
- --gr_mog_n_componentsint
Help: Number of components for GR with MOG.
Default:
5
- --gr_mog_n_itersint
Help: Number of EM iterations during fit for GR with MOG.
Default:
500
- --gr_vae_hidden_dimint
Help: Hidden dimension for GR with VAE.
Default:
512
- --gr_vae_latent_dimint
Help: Latent dimension for GR with VAE.
Default:
256
- --gr_vae_n_itersint
Help: Number of iterations for GR with VAE.
Default:
500
- --train_only_current_promptsint
Help: Train only current prompts.
Default:
0Choices:
0, 1
- --align_with_ortholossint
Help: Align with orthogonality loss.
Default:
0Choices:
0, 1
- --lr_vaefloat
Help: Learning rate for VAE.
Default:
0.0002
- --general_contextint
Help: Use general context (number of contexts created).
Default:
0
- --generated_contextint
Help: Use generated context.
Default:
0
- --cocoopint
Help: Use image embedding to generate context.
Default:
0
- --combo_contextint
Help: Use both generated and prompt context.
Default:
1
- --n_contextint
Help: Use both generated and prompt context.
Default:
1
- --g_modelsstr
Help: Generative model to use for alignment
Default:
vaeChoices:
vae, mog, gauss, diffusion
Classes#
- class models.cgil.CGIL(backbone, loss, args, transform, dataset=None)[source]#
Bases:
FutureModel