Upload decoder_config.json with huggingface_hub
Browse files- decoder_config.json +116 -0
decoder_config.json
ADDED
|
@@ -0,0 +1,116 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"seed": 1,
|
| 3 |
+
|
| 4 |
+
"decoder": {
|
| 5 |
+
"unets": [
|
| 6 |
+
{
|
| 7 |
+
"dim": 416,
|
| 8 |
+
"cond_dim": 512,
|
| 9 |
+
"image_embed_dim": 768,
|
| 10 |
+
"text_embed_dim": 768,
|
| 11 |
+
"cond_on_text_encodings": true,
|
| 12 |
+
"channels": 3,
|
| 13 |
+
"dim_mults": [1, 2, 3, 4],
|
| 14 |
+
"num_resnet_blocks": 4,
|
| 15 |
+
"attn_heads": 8,
|
| 16 |
+
"attn_dim_head": 64,
|
| 17 |
+
"sparse_attn": true,
|
| 18 |
+
"memory_efficient": true
|
| 19 |
+
}
|
| 20 |
+
],
|
| 21 |
+
"clip": {
|
| 22 |
+
"make": "openai",
|
| 23 |
+
"model": "ViT-L/14"
|
| 24 |
+
},
|
| 25 |
+
"image_sizes": [64],
|
| 26 |
+
"channels": 3,
|
| 27 |
+
"timesteps": 1000,
|
| 28 |
+
"loss_type": "l2",
|
| 29 |
+
"beta_schedule": ["cosine"],
|
| 30 |
+
"learned_variance": true
|
| 31 |
+
},
|
| 32 |
+
"data": {
|
| 33 |
+
"webdataset_base_url": "pipe:aws s3 cp --quiet s3://s-datasets/laion-aesthetic/data/laion2B-en-aesthetic/{}.tar -",
|
| 34 |
+
"embeddings_url": "s3://s-datasets/laion-aesthetic/ordered_embeddings/",
|
| 35 |
+
"num_workers": 12,
|
| 36 |
+
"batch_size": 22,
|
| 37 |
+
"start_shard": 0,
|
| 38 |
+
"end_shard": 5247,
|
| 39 |
+
"shard_width": 5,
|
| 40 |
+
"index_width": 4,
|
| 41 |
+
"splits": {
|
| 42 |
+
"train": 0.75,
|
| 43 |
+
"val": 0.15,
|
| 44 |
+
"test": 0.1
|
| 45 |
+
},
|
| 46 |
+
"shuffle_train": false,
|
| 47 |
+
"resample_train": true,
|
| 48 |
+
"preprocessing": {
|
| 49 |
+
"RandomResizedCrop": {
|
| 50 |
+
"size": [64, 64],
|
| 51 |
+
"scale": [0.75, 1.0],
|
| 52 |
+
"ratio": [1.0, 1.0]
|
| 53 |
+
},
|
| 54 |
+
"ToTensor": true
|
| 55 |
+
}
|
| 56 |
+
},
|
| 57 |
+
"train": {
|
| 58 |
+
"epochs": 1000,
|
| 59 |
+
"lr":1.2e-4,
|
| 60 |
+
"wd": 0.00,
|
| 61 |
+
"max_grad_norm": 0.5,
|
| 62 |
+
"save_every_n_samples": 200000,
|
| 63 |
+
"n_sample_images": 10,
|
| 64 |
+
"device": "cuda:0",
|
| 65 |
+
"epoch_samples": 2000000,
|
| 66 |
+
"validation_samples": 100000,
|
| 67 |
+
"use_ema": true,
|
| 68 |
+
"ema_beta": 0.9999,
|
| 69 |
+
"ema_inv_gamma":0.75,
|
| 70 |
+
"save_all": false,
|
| 71 |
+
"save_latest": true,
|
| 72 |
+
"save_best": true,
|
| 73 |
+
"unet_training_mask": [true]
|
| 74 |
+
},
|
| 75 |
+
"evaluate": {
|
| 76 |
+
"n_evaluation_samples": 30,
|
| 77 |
+
"FID": {
|
| 78 |
+
"feature": 64
|
| 79 |
+
},
|
| 80 |
+
"LPIPS": {
|
| 81 |
+
"net_type": "vgg",
|
| 82 |
+
"reduction": "mean"
|
| 83 |
+
}
|
| 84 |
+
},
|
| 85 |
+
"tracker": {
|
| 86 |
+
"data_path": ".tracker-data",
|
| 87 |
+
"overwrite_data_path": true,
|
| 88 |
+
|
| 89 |
+
"log": {
|
| 90 |
+
"log_type": "wandb",
|
| 91 |
+
|
| 92 |
+
"wandb_entity": "nousr_laion",
|
| 93 |
+
"wandb_project": "dalle2_train_decoder",
|
| 94 |
+
|
| 95 |
+
"verbose": true
|
| 96 |
+
},
|
| 97 |
+
|
| 98 |
+
"load": {
|
| 99 |
+
"load_from": null
|
| 100 |
+
},
|
| 101 |
+
|
| 102 |
+
"save": [{
|
| 103 |
+
"save_to": "wandb"
|
| 104 |
+
},
|
| 105 |
+
{
|
| 106 |
+
"save_to": "huggingface",
|
| 107 |
+
"huggingface_repo": "laion/DALLE2-PyTorch",
|
| 108 |
+
|
| 109 |
+
"save_all": false,
|
| 110 |
+
"save_latest": true,
|
| 111 |
+
"save_best": true,
|
| 112 |
+
|
| 113 |
+
"save_type": "model"
|
| 114 |
+
}]
|
| 115 |
+
}
|
| 116 |
+
}
|