Trim config to inference-relevant settings
Browse files- config.yaml +1 -33
config.yaml
CHANGED
|
@@ -1,8 +1,6 @@
|
|
| 1 |
project:
|
| 2 |
name: hyper3-clip
|
| 3 |
-
experiment: hyper3_vitb_clip_uncha_hier_beta_argent_mp5_paper_scratch_8x500k_s31
|
| 4 |
seed: 31
|
| 5 |
-
output_dir: /sc/projects/sci-aisc/matin.mahmood/runs/hyper3_vitb_hierbeta_argent_mp5_paper_scratch_500k_v1/hyper3_vitb_clip_uncha_hier_beta_argent_mp5_paper_scratch_8x500k_s31
|
| 6 |
model:
|
| 7 |
objective: uncha
|
| 8 |
vision_backbone: vit_base_patch16_224
|
|
@@ -16,58 +14,31 @@ model:
|
|
| 16 |
embed_dim: 512
|
| 17 |
curv_init: 1.0
|
| 18 |
learn_curv: true
|
| 19 |
-
entail_weight: 0.2
|
| 20 |
inter_aperture_scale: 0.7
|
| 21 |
intra_aperture_scale: 1.2
|
| 22 |
-
uncha_piecewise_factor: 0.1
|
| 23 |
-
uncha_calibration_alpha: 10.0
|
| 24 |
-
uncha_stop_grad_calibration: true
|
| 25 |
uncha_entailment_geometry: lorentz
|
| 26 |
-
uncha_aggregate_weight: 0.0
|
| 27 |
-
uncha_entailment_loss: hier_beta_argent
|
| 28 |
-
uncha_argent_beta: 1.0
|
| 29 |
-
uncha_argent_norm_weight: 0.1
|
| 30 |
-
uncha_argent_aux_weight: 0.5
|
| 31 |
-
uncha_argent_aggregation: uncha
|
| 32 |
-
uncha_part_weight_power: 0.0
|
| 33 |
uncha_contrastive_loss: ce
|
| 34 |
-
uncha_sigmoid_bias_init: -10.0
|
| 35 |
-
uncha_sigmoid_negative_weight: 1.0
|
| 36 |
-
uncha_part_quality_mode: none
|
| 37 |
-
uncha_part_quality_topk: 5
|
| 38 |
-
uncha_part_quality_temperature: 4.0
|
| 39 |
-
uncha_entailment_warmup_steps: 0
|
| 40 |
-
uncha_global_local_mode: repeat
|
| 41 |
-
beta_clip_global_weight: 0.0
|
| 42 |
-
beta_clip_weight: 0.0
|
| 43 |
beta_clip_beta: 0.5
|
| 44 |
beta_clip_similarity: dot
|
| 45 |
beta_clip_num_heads: 8
|
| 46 |
beta_clip_mlp_ratio: 4.0
|
| 47 |
beta_clip_drop_cls_token: true
|
|
|
|
| 48 |
fuse_beta_query_encoder_forwards: true
|
| 49 |
group_beta_query_pooling: true
|
| 50 |
-
beta_clip_variant: ce
|
| 51 |
phyclip_product_metric: l1
|
| 52 |
training:
|
| 53 |
total_steps: 500000
|
| 54 |
global_batch_size: 768
|
| 55 |
grad_accum_steps: 1
|
| 56 |
-
num_workers: 8
|
| 57 |
lr: 0.0005
|
| 58 |
weight_decay: 0.2
|
| 59 |
betas:
|
| 60 |
- 0.9
|
| 61 |
- 0.98
|
| 62 |
warmup_steps: 4000
|
| 63 |
-
log_interval: 20
|
| 64 |
-
ckpt_interval: 10000
|
| 65 |
amp: true
|
| 66 |
max_grad_norm: 1.0
|
| 67 |
-
resume: true
|
| 68 |
-
resume_from: null
|
| 69 |
-
resume_from_env: RESUME_FROM_CHECKPOINT
|
| 70 |
-
find_unused_parameters: true
|
| 71 |
optimizer:
|
| 72 |
no_decay_params:
|
| 73 |
- logit_scale
|
|
@@ -85,12 +56,9 @@ data:
|
|
| 85 |
part_sampling: all
|
| 86 |
max_parts: 5
|
| 87 |
train_transform: tight_crop_color_jitter_gray
|
| 88 |
-
tarfiles:
|
| 89 |
-
- /sc/projects/sci-aisc/matin.mahmood/datasets/hycoclip/train/GRIT/processed/*.tar
|
| 90 |
shuffle_buffer: 4000
|
| 91 |
image_size: 224
|
| 92 |
max_text_length: 77
|
| 93 |
-
num_workers: 8
|
| 94 |
image_normalization: imagenet
|
| 95 |
beta_clip:
|
| 96 |
enabled: true
|
|
|
|
| 1 |
project:
|
| 2 |
name: hyper3-clip
|
|
|
|
| 3 |
seed: 31
|
|
|
|
| 4 |
model:
|
| 5 |
objective: uncha
|
| 6 |
vision_backbone: vit_base_patch16_224
|
|
|
|
| 14 |
embed_dim: 512
|
| 15 |
curv_init: 1.0
|
| 16 |
learn_curv: true
|
|
|
|
| 17 |
inter_aperture_scale: 0.7
|
| 18 |
intra_aperture_scale: 1.2
|
|
|
|
|
|
|
|
|
|
| 19 |
uncha_entailment_geometry: lorentz
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 20 |
uncha_contrastive_loss: ce
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 21 |
beta_clip_beta: 0.5
|
| 22 |
beta_clip_similarity: dot
|
| 23 |
beta_clip_num_heads: 8
|
| 24 |
beta_clip_mlp_ratio: 4.0
|
| 25 |
beta_clip_drop_cls_token: true
|
| 26 |
+
beta_clip_variant: ce
|
| 27 |
fuse_beta_query_encoder_forwards: true
|
| 28 |
group_beta_query_pooling: true
|
|
|
|
| 29 |
phyclip_product_metric: l1
|
| 30 |
training:
|
| 31 |
total_steps: 500000
|
| 32 |
global_batch_size: 768
|
| 33 |
grad_accum_steps: 1
|
|
|
|
| 34 |
lr: 0.0005
|
| 35 |
weight_decay: 0.2
|
| 36 |
betas:
|
| 37 |
- 0.9
|
| 38 |
- 0.98
|
| 39 |
warmup_steps: 4000
|
|
|
|
|
|
|
| 40 |
amp: true
|
| 41 |
max_grad_norm: 1.0
|
|
|
|
|
|
|
|
|
|
|
|
|
| 42 |
optimizer:
|
| 43 |
no_decay_params:
|
| 44 |
- logit_scale
|
|
|
|
| 56 |
part_sampling: all
|
| 57 |
max_parts: 5
|
| 58 |
train_transform: tight_crop_color_jitter_gray
|
|
|
|
|
|
|
| 59 |
shuffle_buffer: 4000
|
| 60 |
image_size: 224
|
| 61 |
max_text_length: 77
|
|
|
|
| 62 |
image_normalization: imagenet
|
| 63 |
beta_clip:
|
| 64 |
enabled: true
|