Upload checkpoints/phase2_e04/config.json with huggingface_hub
Browse files
checkpoints/phase2_e04/config.json
ADDED
|
@@ -0,0 +1,89 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"model": {
|
| 3 |
+
"clip_model": "openai/clip-vit-large-patch14",
|
| 4 |
+
"clip_hidden": 768,
|
| 5 |
+
"clip_layers": 12,
|
| 6 |
+
"clip_max_tokens": 77,
|
| 7 |
+
"freeze_clip": true,
|
| 8 |
+
"n_memory_tokens": 8,
|
| 9 |
+
"bank_size": 64,
|
| 10 |
+
"anchor_dim": 768,
|
| 11 |
+
"n_bank_heads": 8,
|
| 12 |
+
"bank_cross_layers": 2,
|
| 13 |
+
"gate_type": "gru",
|
| 14 |
+
"extract_layers": [
|
| 15 |
+
1,
|
| 16 |
+
3,
|
| 17 |
+
5,
|
| 18 |
+
7,
|
| 19 |
+
9,
|
| 20 |
+
11
|
| 21 |
+
],
|
| 22 |
+
"layer_fusion": "learned",
|
| 23 |
+
"max_content_tokens": 18,
|
| 24 |
+
"segment_overlap": 4,
|
| 25 |
+
"max_segments": 32,
|
| 26 |
+
"cv_target": 0.2,
|
| 27 |
+
"sequence_output": true,
|
| 28 |
+
"sequence_len": 77,
|
| 29 |
+
"sequence_recon_layers": 2,
|
| 30 |
+
"sequence_recon_heads": 8,
|
| 31 |
+
"collect_content_tokens": true,
|
| 32 |
+
"max_content_positions": 256,
|
| 33 |
+
"teacher_model": "answerdotai/ModernBERT-large",
|
| 34 |
+
"teacher_hidden": 1024,
|
| 35 |
+
"return_dict": true,
|
| 36 |
+
"output_hidden_states": false,
|
| 37 |
+
"dtype": null,
|
| 38 |
+
"chunk_size_feed_forward": 0,
|
| 39 |
+
"is_encoder_decoder": false,
|
| 40 |
+
"architectures": null,
|
| 41 |
+
"id2label": {
|
| 42 |
+
"0": "LABEL_0",
|
| 43 |
+
"1": "LABEL_1"
|
| 44 |
+
},
|
| 45 |
+
"label2id": {
|
| 46 |
+
"LABEL_0": 0,
|
| 47 |
+
"LABEL_1": 1
|
| 48 |
+
},
|
| 49 |
+
"problem_type": null,
|
| 50 |
+
"_name_or_path": "",
|
| 51 |
+
"transformers_version": "5.0.0",
|
| 52 |
+
"model_type": "memory_clip_seq",
|
| 53 |
+
"output_attentions": false
|
| 54 |
+
},
|
| 55 |
+
"training": {
|
| 56 |
+
"max_train_samples": 50000,
|
| 57 |
+
"max_val_samples": 2000,
|
| 58 |
+
"min_caption_length": 100,
|
| 59 |
+
"phase1_epochs": 5,
|
| 60 |
+
"phase1_lr_seq": 0.002,
|
| 61 |
+
"phase1_lr_proj": 0.001,
|
| 62 |
+
"phase2_epochs": 5,
|
| 63 |
+
"phase2_lr_bank": 0.0005,
|
| 64 |
+
"phase2_lr_output": 0.0002,
|
| 65 |
+
"phase2_lr_proj": 0.0005,
|
| 66 |
+
"phase2_lr_seq": 0.001,
|
| 67 |
+
"batch_size": 64,
|
| 68 |
+
"min_lr": 1e-06,
|
| 69 |
+
"weight_decay": 0.01,
|
| 70 |
+
"grad_clip": 1.0,
|
| 71 |
+
"warmup_steps": 200,
|
| 72 |
+
"modern_weight": 1.0,
|
| 73 |
+
"procrustes_weight": 0.3,
|
| 74 |
+
"cv_weight": 0.05,
|
| 75 |
+
"temperature": 0.07,
|
| 76 |
+
"sequence_weight": 1.0,
|
| 77 |
+
"sequence_cosine_weight": 0.5,
|
| 78 |
+
"modern_max_len": 4096,
|
| 79 |
+
"procrustes_n_samples": 300,
|
| 80 |
+
"v1_checkpoint": "",
|
| 81 |
+
"v1_repo_id": "AbstractPhil/geolip-clip-vit-large-patch14-ctx576",
|
| 82 |
+
"v1_filename": "model.safetensors",
|
| 83 |
+
"checkpoint_dir": "/home/claude/memory_clip_seq_checkpoints",
|
| 84 |
+
"tensorboard_dir": "/home/claude/memory_clip_seq_tb",
|
| 85 |
+
"metrics_file": "/home/claude/memory_clip_seq_checkpoints/metrics.json",
|
| 86 |
+
"log_every": 20,
|
| 87 |
+
"eval_every": 200
|
| 88 |
+
}
|
| 89 |
+
}
|