torch model ensembling is integrated

src/eynollah/training/train.py

@@ -825,7 +825,7 @@ def run(_config,
             usable_checkpoints = [os.path.join(dir_output, 'model_{epoch:02d}'.format(epoch=epoch + 1))
                                   for epoch in usable_checkpoints]
             ens_path = os.path.join(dir_output, 'model_ens_avg')
-            run_ensembling(usable_checkpoints, ens_path)
+            run_ensembling(usable_checkpoints, ens_path, framework='tensorflow')
             _log.info("ensemble model saved under '%s'", ens_path)
 
     elif task=='reading_order':

src/eynollah/training/weights_ensembling.py

@@ -1,4 +1,5 @@
 import os
+from typing import Optional
 from warnings import catch_warnings, simplefilter
 
 import click
@@ -11,33 +12,56 @@ from ocrd_utils import tf_disable_interactive_logs
 tf_disable_interactive_logs()
 import tensorflow as tf
 from tensorflow.keras.models import load_model
+import torch
+from transformers import VisionEncoderDecoderModel
 
 from ..patch_encoder import (
     PatchEncoder,
     Patches,
 )
     
-def run_ensembling(model_dirs, out_dir):
+def run_ensembling(dir_models, out, framework):
-    all_weights = []
+    ls_models = os.listdir(dir_models)
+    # model: Optional[VisionEncoderDecoderModel] = None
+    # model_name: Optional[str] = None
+    if framework=="torch":
+        models = []
+        sd_models = []
         
-    for model_dir in model_dirs:
+        for model_name in ls_models:
-        assert os.path.isdir(model_dir), model_dir
+            model = VisionEncoderDecoderModel.from_pretrained(os.path.join(dir_models, model_name))
-        model = load_model(model_dir, compile=False,
+            models.append(model)
-                           custom_objects=dict(PatchEncoder=PatchEncoder,
+            sd_models.append(model.state_dict())
-                                               Patches=Patches))
+        for key in sd_models[0]:
-        all_weights.append(model.get_weights())
+            sd_models[0][key] = sum(sd[key] for sd in sd_models) / len(sd_models)
             
-    new_weights = []
+        model.load_state_dict(sd_models[0])
-    for layer_weights in zip(*all_weights):
+        os.system("mkdir "+out)
-        layer_weights = np.array([np.array(weights).mean(axis=0)
+        torch.save(model.state_dict(), os.path.join(out, "pytorch_model.bin"))
-                                  for weights in zip(*layer_weights)])
+        os.system('cp ' + os.path.join(os.path.join(dir_models, model_name), "config.json") + " " + out)
-        new_weights.append(layer_weights)
+        
+    else:
+        weights=[]
+
+        for model_name in ls_models:
+            model =  load_model(os.path.join(dir_models, model_name), compile=False, custom_objects={'PatchEncoder':PatchEncoder, 'Patches': Patches})
+            weights.append(model.get_weights())
+            
+        new_weights = list()
+
+        for weights_list_tuple in zip(*weights):
+            new_weights.append(
+                [np.array(weights_).mean(axis=0)\
+                    for weights_ in zip(*weights_list_tuple)])
+            
+
+
+        new_weights = [np.array(x) for x in new_weights]
             
-    #model = tf.keras.models.clone_model(model)
         model.set_weights(new_weights)
-
+        model.save(out)
-    model.save(out_dir)
+        os.system('cp '+os.path.join(os.path.join(dir_models, model_name), "config.json") + " " + out)
-    os.system('cp ' + os.path.join(model_dirs[0], "config.json ") + out_dir + "/")
+        os.system('cp '+os.path.join(os.path.join(dir_models, model_name), "characters_org.txt") + " " + out)
 
 @click.command()
 @click.option(
@@ -56,12 +80,19 @@ def run_ensembling(model_dirs, out_dir):
     required=True,
     type=click.Path(exists=False, file_okay=False),
 )
-def ensemble_cli(in_, out):
+@click.option(
+    "--framework",
+    "-fw",
+    help="this parameter gets tensorflow or torch as model framework",
+    type=click.Choice(['torch', 'tensorflow']),
+    default="tensorflow"
+)
+
+def ensemble_cli(in_, out, framework):
     """
     mix multiple model weights
 
     Load a sequence of models and mix them into a single ensemble model
     by averaging their weights. Write the resulting model.
     """
-    run_ensembling(in_, out)
+    run_ensembling(in_, out, framework)
-