2907

micro/config.json

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:31acced0f5ecc3a484fd3c358dabaf3e5a7fce777b0472846ef774f4a734a839
-size 1893
+oid sha256:18891ea8f81c705422b12d7575493f5abb949d60f2814a989a436fde65e84c82
+size 1873

micro/diffusion_pytorch_model.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:24bc2a898a6829689cc3fcd3d2c1d4e4a757aa157243d127316ca0d8e3267372
-size 2078471360
+oid sha256:9fa21e51b39a8f8f44036b0e271cddc65914ea3591cce6e0a35506425026f183
+size 3929714960

src/micro_create.ipynb

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:eeb3c833ffba08a44cf7dce894a67310ad1b7570d7401259e71bc6c770785e75
-size 33978
+oid sha256:856a34f139aad0b40359836b62caa7a005ee898990e416a7a04a9ded0881496a
+size 33982

train.py

@@ -27,12 +27,12 @@ import torch.nn.functional as F
 # --------------------------- Параметры ---------------------------
 ds_path = "datasets/384"
 project = "micro"
-batch_size = 5
+batch_size = 64
 base_learning_rate = 1e-4
 min_learning_rate = 5e-5
-num_epochs = 5
+num_epochs = 50
 # samples/save per epoch
-sample_interval_share = 40
+sample_interval_share = 10
 use_wandb = True
 save_model = True
 use_decay = True
@@ -54,12 +54,12 @@ dispersive_temperature=0.5
 dispersive_weight= 0.05
 percentile_clipping = 95 # 8bit optim
 betta2 = 0.97
-eps = 1e-5
+eps = 1e-6
 clip_grad_norm = 1.0
 steps_offset = 0 # Scheduler
 limit = 0
 checkpoints_folder = ""
-mixed_precision = "bf16" #"fp16"
+mixed_precision = "no" #"fp16"
 gradient_accumulation_steps = 1
 accelerator = Accelerator(
     mixed_precision=mixed_precision,
@@ -807,7 +807,7 @@ for epoch in range(start_epoch, start_epoch + num_epochs):
                     avg_loss = np.mean(batch_losses[-sample_interval:])
                     avg_tloss = np.mean(batch_tlosses[-sample_interval:])
                     avg_grad = torch.mean(torch.stack(batch_grads[-sample_interval:])).cpu().item()
-                    print(f"Эпоха {epoch}, шаг {global_step}, средний лосс: {avg_loss:.6f}")
+                    print(f"Эпоха {epoch}, шаг {global_step}, средний лосс: {avg_loss:.6f}, grad: {avg_grad:.6f}")
 
                     if save_model:
                         print("saving:",avg_loss < min_loss*save_barrier)