Updated training results

Training.md

@@ -13,6 +13,24 @@ an improvement or fix was done to the conversion.
 
 The training process itself takes depending on the hardware you use a 2-3 days.
 
+## Run 101
+
+Commit: ba12ebef4606948816a06f4a011248d07a6f06da
+Date: 10 Sep 2024
+SER: 9%
+Validation result: 6.4
+
+Training runs now also pick the last iteration and not the one with the lowest validation loss.
+
+## Run 100
+
+Commit: e317d1ba4452798036d2b24a20f37061b8441bae
+Date: 10 Sep 2024
+SER: 14%
+Validation result: 7.6
+
+Increased model depth from 4 to 8.
+
 ## Run 70
 
 Commit: 11c1eeaf5760d617f09678e276866d31253a5ace
@@ -67,7 +85,6 @@ SER: 53%
 
 First training run within the `homr` repo.
 
-
 ## Previous runs
 
 These runs where performed inside the [Polyphonic-TrOMR](https://raw.githubusercontent.com/liebharc/Polyphonic-TrOMR/master/Training.md) repo.

training/transformer/train.py

@@ -2,6 +2,7 @@
 import shutil
 import sys
 
+import safetensors
 import torch
 import torch._dynamo
 from transformers import Trainer, TrainingArguments  # type: ignore
@@ -150,8 +151,15 @@ def train_transformer(fp32: bool = False, pretrained: bool = False, resume: str
     if pretrained:
         eprint("Loading pretrained model")
         model = TrOMR(config)
-        tr_omr_pretrained = config.filepaths.checkpoint
-        model.load_state_dict(torch.load(tr_omr_pretrained), strict=False)
+        checkpoint_file_path = config.filepaths.checkpoint
+        if ".safetensors" in checkpoint_file_path:
+            tensors = {}
+            with safetensors.safe_open(checkpoint_file_path, framework="pt", device=0) as f:  # type: ignore
+                for k in f.keys():
+                    tensors[k] = f.get_tensor(k)
+            model.load_state_dict(tensors, strict=False)
+        else:
+            model.load_state_dict(torch.load(checkpoint_file_path), strict=False)
     else:
         model = TrOMR(config)