Upload params.txt with huggingface_hub

params.txt

@@ -7,7 +7,7 @@ beta2: 0.98
 beta_decay_epochs: 16
 cache_dir: None
 calculate_full: True
-checkpoint_path: ./logs/fit_w_prev-bias-10.0sample-negFalse-scaleTrue-pos1.0-eps1e-06-stocFalselr0.0004-ViT-B-32-1990_0124032041/checkpoints
+checkpoint_path: ./logs/emn_fix-bias-10.0sample-negFalse-scaleFalse-pos1.0-eps1e-06-stocFalselr0.0004-ViT-B-32-1990_0124032307/checkpoints
 coca_caption_loss_weight: 2.0
 coca_contrastive_loss_weight: 1.0
 contrast_neg_only: False
@@ -28,12 +28,12 @@ distill: False
 distill_model: None
 distill_pretrained: None
 distributed: True
-ema_z_mode: ema_only
+ema_z_mode: ema_n_sample
 epochs: 33
 epochs_cooldown: None
 eps: 1e-06
 fit_neg_only: True
-fit_w_prev: True
+fit_w_prev: False
 force_custom_text: False
 force_image_size: None
 force_patch_dropout: None
@@ -53,11 +53,11 @@ imagenet_val: /scratch/imagenet/val/
 init_lambda: 1.0
 init_logit_bias: -10.0
 init_logit_scale: None
-lambda_eps: 1e-07
+lambda_eps: 1e-12
 lambda_lr: 0.001
-lambda_tolerance: 0.005
-lambda_update_frequency: [5]
-learn_logit_bias: True
+lambda_tolerance: 0.0005
+lambda_update_frequency: [20, 10, 1]
+learn_logit_bias: False
 learn_logit_scale: True
 local_loss: True
 local_rank: 0
@@ -70,7 +70,7 @@ lock_text_unlocked_layers: 0
 log_every_n_steps: 100
 log_level: 20
 log_local: False
-log_path: ./logs/fit_w_prev-bias-10.0sample-negFalse-scaleTrue-pos1.0-eps1e-06-stocFalselr0.0004-ViT-B-32-1990_0124032041/out.log
+log_path: ./logs/emn_fix-bias-10.0sample-negFalse-scaleFalse-pos1.0-eps1e-06-stocFalselr0.0004-ViT-B-32-1990_0124032307/out.log
 logit_scale_clamp: 100
 logs: ./logs/
 loss_type: COND_EXP
@@ -84,11 +84,11 @@ model_beta_init: 0.0
 model_ema_beta: 0.99
 model_update_type: ONE_STEP
 n_class_tokens: -1
-name: fit_w_prev-bias-10.0sample-negFalse-scaleTrue-pos1.0-eps1e-06-stocFalselr0.0004-ViT-B-32-1990_0124032041
+name: emn_fix-bias-10.0sample-negFalse-scaleFalse-pos1.0-eps1e-06-stocFalselr0.0004-ViT-B-32-1990_0124032307
 no_set_device_rank: False
 norm_cap: 1.0
 normalize_type: L2
-note: fit_w_prev
+note: emn_fix
 pos_coef: 1.0
 precision: amp
 pretrained: 
@@ -101,7 +101,7 @@ report_to: wandb
 resume: None
 save_frequency: 1
 save_most_recent: False
-scale_loss: True
+scale_loss: False
 seed: 42
 siglip: False
 skip_scheduler: False
@@ -114,7 +114,7 @@ trace: False
 train_data: /scratch/cc12m/{00000..01240}.tar
 train_data_upsampling_factors: None
 train_num_samples: 9187328
-update_ema_mlps_every_n_steps: 1
+update_ema_mlps_every_n_steps: 2000
 use_bn_sync: False
 use_bnb_linear: None
 use_feature_diff: False