micro

micro/config.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:31acced0f5ecc3a484fd3c358dabaf3e5a7fce777b0472846ef774f4a734a839
+size 1893

micro/diffusion_pytorch_model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:24bc2a898a6829689cc3fcd3d2c1d4e4a757aa157243d127316ca0d8e3267372
+size 2078471360

sdxl/config.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1fc6f3f339d56f6e5fc44b2b62cd0ceb46616137961a105f2579298631328f21
+size 1768

sdxl/diffusion_pytorch_model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f14e079e3164d58155294204b475f1782d48e4364a1487000a28b0a16e6a281f
+size 3944400016

src/micro_create.ipynb

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:eeb3c833ffba08a44cf7dce894a67310ad1b7570d7401259e71bc6c770785e75
+size 33978

src/sdxl_create.ipynb

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:11a356e99139cd42df2110a02820b7d2c8cdd1fbe86011be16b27fa862efb4a6
+size 27389

src/sdxs_create_unet-2b.ipynb

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:11f42016cce7ee2d56f5bcff4649008491222a23a70d144dbb87babccc6beaa7
+oid sha256:1d2d6a5077aad9a1e416c69627411951a897959b0bfb026caba89634be0ec158
 size 146107

train.py

@@ -26,13 +26,13 @@ import torch.nn.functional as F
 
 # --------------------------- Параметры ---------------------------
 ds_path = "datasets/384"
-project = "2b"
-batch_size = 64
-base_learning_rate = 8e-5
-min_learning_rate = 4e-5
-num_epochs = 25
+project = "micro"
+batch_size = 5
+base_learning_rate = 1e-4
+min_learning_rate = 5e-5
+num_epochs = 5
 # samples/save per epoch
-sample_interval_share = 5
+sample_interval_share = 40
 use_wandb = True
 save_model = True
 use_decay = True
@@ -49,13 +49,14 @@ torch.backends.cudnn.allow_tf32 = True
 torch.backends.cuda.enable_mem_efficient_sdp(False)
 dtype = torch.float32
 save_barrier = 1.03
+warmup_percent = 0.01
 dispersive_temperature=0.5
 dispersive_weight= 0.05
 percentile_clipping = 95 # 8bit optim
 betta2 = 0.97
 eps = 1e-5
-clip_grad_norm = .75
-steps_offset = 1 # Scheduler
+clip_grad_norm = 1.0
+steps_offset = 0 # Scheduler
 limit = 0
 checkpoints_folder = ""
 mixed_precision = "bf16" #"fp16"
@@ -510,7 +511,7 @@ else:
     
     def lr_schedule(step):
         x = step / (total_training_steps * world_size)
-        warmup = 0.05
+        warmup = warmup_percent
         
         if not use_decay:
             return base_learning_rate