Delete console_out_needs_retrain_for_full_benchmark.txt
Browse files
console_out_needs_retrain_for_full_benchmark.txt
DELETED
|
@@ -1,118 +0,0 @@
|
|
| 1 |
-
======================================================================
|
| 2 |
-
TRAINING: MEMORY-EXTENDED CLIP-L TEXT ENCODER
|
| 3 |
-
======================================================================
|
| 4 |
-
Device: cuda
|
| 5 |
-
GPU: NVIDIA RTX PRO 6000 Blackwell Server Edition
|
| 6 |
-
VRAM: 102.0 GB
|
| 7 |
-
Loading CLIP text encoder: openai/clip-vit-large-patch14
|
| 8 |
-
Loading weights: 100%
|
| 9 |
-
196/196 [00:00<00:00, 4354.51it/s, Materializing param=text_model.final_layer_norm.weight]
|
| 10 |
-
CLIPTextModel LOAD REPORT from: openai/clip-vit-large-patch14
|
| 11 |
-
Key | Status | |
|
| 12 |
-
---------------------------------------------------------------+------------+--+-
|
| 13 |
-
vision_model.encoder.layers.{0...23}.mlp.fc2.bias | UNEXPECTED | |
|
| 14 |
-
vision_model.encoder.layers.{0...23}.self_attn.out_proj.bias | UNEXPECTED | |
|
| 15 |
-
vision_model.encoder.layers.{0...23}.layer_norm2.bias | UNEXPECTED | |
|
| 16 |
-
vision_model.encoder.layers.{0...23}.self_attn.v_proj.weight | UNEXPECTED | |
|
| 17 |
-
vision_model.encoder.layers.{0...23}.self_attn.v_proj.bias | UNEXPECTED | |
|
| 18 |
-
vision_model.encoder.layers.{0...23}.self_attn.out_proj.weight | UNEXPECTED | |
|
| 19 |
-
vision_model.encoder.layers.{0...23}.self_attn.k_proj.bias | UNEXPECTED | |
|
| 20 |
-
vision_model.encoder.layers.{0...23}.self_attn.q_proj.weight | UNEXPECTED | |
|
| 21 |
-
vision_model.encoder.layers.{0...23}.layer_norm1.bias | UNEXPECTED | |
|
| 22 |
-
vision_model.encoder.layers.{0...23}.self_attn.k_proj.weight | UNEXPECTED | |
|
| 23 |
-
vision_model.encoder.layers.{0...23}.self_attn.q_proj.bias | UNEXPECTED | |
|
| 24 |
-
vision_model.encoder.layers.{0...23}.mlp.fc1.bias | UNEXPECTED | |
|
| 25 |
-
vision_model.encoder.layers.{0...23}.mlp.fc2.weight | UNEXPECTED | |
|
| 26 |
-
vision_model.encoder.layers.{0...23}.mlp.fc1.weight | UNEXPECTED | |
|
| 27 |
-
vision_model.encoder.layers.{0...23}.layer_norm1.weight | UNEXPECTED | |
|
| 28 |
-
vision_model.encoder.layers.{0...23}.layer_norm2.weight | UNEXPECTED | |
|
| 29 |
-
visual_projection.weight | UNEXPECTED | |
|
| 30 |
-
text_model.embeddings.position_ids | UNEXPECTED | |
|
| 31 |
-
vision_model.embeddings.position_embedding.weight | UNEXPECTED | |
|
| 32 |
-
vision_model.embeddings.patch_embedding.weight | UNEXPECTED | |
|
| 33 |
-
vision_model.post_layernorm.weight | UNEXPECTED | |
|
| 34 |
-
vision_model.embeddings.position_ids | UNEXPECTED | |
|
| 35 |
-
vision_model.pre_layrnorm.weight | UNEXPECTED | |
|
| 36 |
-
logit_scale | UNEXPECTED | |
|
| 37 |
-
vision_model.pre_layrnorm.bias | UNEXPECTED | |
|
| 38 |
-
vision_model.embeddings.class_embedding | UNEXPECTED | |
|
| 39 |
-
vision_model.post_layernorm.bias | UNEXPECTED | |
|
| 40 |
-
text_projection.weight | UNEXPECTED | |
|
| 41 |
-
|
| 42 |
-
Notes:
|
| 43 |
-
- UNEXPECTED :can be ignored when loading from different task/architecture; not ok if you expect identical arch.
|
| 44 |
-
`torch_dtype` is deprecated! Use `dtype` instead!
|
| 45 |
-
CLIP: 123,060,480 frozen
|
| 46 |
-
Memory + projector: 34,463,494 trainable
|
| 47 |
-
Extract: (1, 3, 5, 7, 9, 11)
|
| 48 |
-
Bank: 64 anchors, 2 cross-attn
|
| 49 |
-
Memory: 8 tokens
|
| 50 |
-
Effective context: 32 × 18 = 576 tokens
|
| 51 |
-
|
| 52 |
-
Loading ModernBERT-large...
|
| 53 |
-
Loading weights: 100%
|
| 54 |
-
170/170 [00:00<00:00, 2622.21it/s, Materializing param=layers.27.mlp_norm.weight]
|
| 55 |
-
ModernBertModel LOAD REPORT from: answerdotai/ModernBERT-large
|
| 56 |
-
Key | Status | |
|
| 57 |
-
------------------+------------+--+-
|
| 58 |
-
decoder.bias | UNEXPECTED | |
|
| 59 |
-
head.norm.weight | UNEXPECTED | |
|
| 60 |
-
head.dense.weight | UNEXPECTED | |
|
| 61 |
-
|
| 62 |
-
Notes:
|
| 63 |
-
- UNEXPECTED :can be ignored when loading from different task/architecture; not ok if you expect identical arch.
|
| 64 |
-
394,781,696 params (frozen)
|
| 65 |
-
|
| 66 |
-
Loading long captions...
|
| 67 |
-
Loading CaptionEmporium/conceptual-captions-cc12m-llavanext...
|
| 68 |
-
Got 50000 long captions from CC12M-llavanext
|
| 69 |
-
Token indices sequence length is longer than the specified maximum sequence length for this model (127 > 77). Running this sequence through the model will result in indexing errors
|
| 70 |
-
Caption token lengths (CLIP tokenizer, sample of 500):
|
| 71 |
-
mean=96 median=98 max=251 >77: 67.0%
|
| 72 |
-
50000 captions loaded
|
| 73 |
-
Example: Two loaves of golden brown Cuban bread, one slightly overlapping the other, resting on a white surfa...
|
| 74 |
-
|
| 75 |
-
Computing static Procrustes on 300 captions...
|
| 76 |
-
Student: torch.Size([300, 768]), Teacher: torch.Size([300, 1024])
|
| 77 |
-
Procrustes: cos 0.0008 → 0.8160
|
| 78 |
-
[ModernBERT] Procrustes init: |R|=27.713
|
| 79 |
-
bank : 27,171,840 params @ lr=0.002
|
| 80 |
-
proj : 787,456 params @ lr=0.001
|
| 81 |
-
output : 6,504,198 params @ lr=0.0005
|
| 82 |
-
|
| 83 |
-
Training: 34,463,494 params
|
| 84 |
-
781 batches/epoch × 64
|
| 85 |
-
Losses: modern(1.0) + procrustes(0.3) + cv(0.05)
|
| 86 |
-
Epoch 1/10: 0%| | 0/782 [00:00<?, ?it/s]Token indices sequence length is longer than the specified maximum sequence length for this model (90 > 77). Running this sequence through the model will result in indexing errors
|
| 87 |
-
Epoch 1/10: 100%|██████████| 782/782 [09:17<00:00, 1.40it/s, cv=0.181, loss=2.754, m_acc=0.645]
|
| 88 |
-
|
| 89 |
-
Epoch 1: 557s loss=2.7538 m_acc=0.645 cv=0.181
|
| 90 |
-
Epoch 2/10: 100%|██████████| 782/782 [09:19<00:00, 1.40it/s, cv=0.204, loss=2.423, m_acc=0.846]
|
| 91 |
-
|
| 92 |
-
Epoch 2: 559s loss=2.4232 m_acc=0.846 cv=0.204
|
| 93 |
-
Epoch 3/10: 100%|██████████| 782/782 [09:19<00:00, 1.40it/s, cv=0.166, loss=2.383, m_acc=0.862]
|
| 94 |
-
|
| 95 |
-
Epoch 3: 560s loss=2.3831 m_acc=0.862 cv=0.166
|
| 96 |
-
Epoch 4/10: 100%|██████████| 782/782 [09:20<00:00, 1.40it/s, cv=0.164, loss=2.375, m_acc=0.864]
|
| 97 |
-
|
| 98 |
-
Epoch 4: 561s loss=2.3754 m_acc=0.864 cv=0.164
|
| 99 |
-
Epoch 5/10: 100%|██████████| 782/782 [09:19<00:00, 1.40it/s, cv=0.165, loss=2.331, m_acc=0.886]
|
| 100 |
-
|
| 101 |
-
Epoch 5: 559s loss=2.3307 m_acc=0.886 cv=0.165
|
| 102 |
-
Epoch 6/10: 100%|██████████| 782/782 [09:19<00:00, 1.40it/s, cv=0.166, loss=2.331, m_acc=0.888]
|
| 103 |
-
|
| 104 |
-
Epoch 6: 559s loss=2.3311 m_acc=0.888 cv=0.166
|
| 105 |
-
Epoch 7/10: 100%|██████████| 782/782 [09:19<00:00, 1.40it/s, cv=0.171, loss=2.338, m_acc=0.882]
|
| 106 |
-
|
| 107 |
-
Epoch 7: 560s loss=2.3381 m_acc=0.882 cv=0.171
|
| 108 |
-
Epoch 8/10: 100%|██████████| 782/782 [09:20<00:00, 1.39it/s, cv=0.163, loss=2.253, m_acc=0.920]
|
| 109 |
-
|
| 110 |
-
Epoch 8: 561s loss=2.2533 m_acc=0.920 cv=0.163
|
| 111 |
-
Epoch 9/10: 100%|██████████| 782/782 [09:18<00:00, 1.40it/s, cv=0.163, loss=2.222, m_acc=0.932]
|
| 112 |
-
|
| 113 |
-
Epoch 9: 559s loss=2.2215 m_acc=0.932 cv=0.163
|
| 114 |
-
Epoch 10/10: 100%|██████████| 782/782 [09:15<00:00, 1.41it/s, cv=0.162, loss=2.206, m_acc=0.934]
|
| 115 |
-
|
| 116 |
-
Epoch 10: 556s loss=2.2060 m_acc=0.934 cv=0.162
|
| 117 |
-
|
| 118 |
-
Done.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|