Getting nan loss on ema test samples while using open clip models for prior training #325
Unanswered
RicBriChri
asked this question in
Q&A
Replies: 0 comments
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
-
I'm getting nan loss on the ema text samples while I'm using pretrained and my own open clip models, I'm not sure what mistake I'm doing. The training losses are converging well but during the testing phase, the loss is nan. I'm using the following config file:
{
"prior": {
"clip": {
"make": "open_clip",
"model": "ViT-B-32",
"force_quick_gelu": true
},
"net": {
"dim": 512,
"depth": 12,
"num_timesteps": 1000,
"max_text_len": 77,
"num_time_embeds": 1,
"num_image_embeds": 1,
"num_text_embeds": 1,
"dim_head": 64,
"heads": 12,
"ff_mult": 4,
"norm_out": true,
"attn_dropout": 0.05,
"ff_dropout": 0.05,
"final_proj": true,
"normformer": true,
"rotary_emb": true
},
"image_embed_dim": 512,
"image_size": 256,
"image_channels": 3,
"timesteps": 1000,
"sample_timesteps": 64,
"cond_drop_prob": 0.1,
"loss_type": "l2",
"predict_x_start": true,
"beta_schedule": "cosine",
"condition_on_text_encodings": true
},
"data": {
"batch_size": 128,
"num_data_points": 30000,
"eval_every_seconds": 1600,
"image_url": " ",
"meta_url": " ",
"splits": {
"train": 0.9,
"val": 0.05,
"test": 0.05
}
},
"train": {
"epochs": 1,
"lr": 1.1e-4,
"wd": 6.02e-2,
"max_grad_norm": 0.5,
"use_ema": true,
"ema_beta": 0.9999,
"ema_update_after_step": 50,
"warmup_steps": 50,
"amp": false,
"save_every_seconds": 3600,
"eval_timesteps": [64],
"random_seed": 84513
},
}
Beta Was this translation helpful? Give feedback.
All reactions