Saving weights and logs of epoch 6

config.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:67ab6b43f4bacd25ccb5e78e065aa2c118535865f9621645b9f0caad1249e47c
+size 1360

flax_model.msgpack

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:69324e2320f6e0c2619bce081b9a703fb4f3dadd403c7b960875a5a8c61d1f39
+size 241981002

run_translation_t5_flax.py

@@ -260,8 +260,10 @@ def write_metric(summary_writer, train_metrics, eval_metrics, train_time, step):
         for i, val in enumerate(vals):
             summary_writer.scalar(tag, val, step - len(vals) + i + 1)
 
+    #Pere - dropping all values that are not float
     for metric_name, value in eval_metrics.items():
-        summary_writer.scalar(f"eval_{metric_name}", value, step)
+        if isinstance(value,float):
+            summary_writer.scalar(f"eval_{metric_name}", value, step)
 
 
 def create_learning_rate_fn(
@@ -499,7 +501,7 @@ def main():
         )
 
     # Metric
-    metric = load_metric("bleu")
+    metric = load_metric("sacrebleu")
 
     def postprocess_text(preds, labels):
         preds = [pred.strip() for pred in preds]
@@ -519,14 +521,22 @@ def main():
         #Probably not needed for bleu - pere
         #decoded_preds, decoded_labels = postprocess_text(decoded_preds, decoded_labels)
         
-        breakpoint()
-        result = metric.compute(predictions=decoded_preds, references=decoded_labels)
+        #breakpoint()
+        #result = metric.compute(predictions=decoded_preds, references=decoded_labels)
+        decoded_labels_list = [[d] for d in decoded_labels]
+        result = metric.compute(predictions=decoded_preds, references=decoded_labels_list)
+        
+        #Debug stuff - pere
+        print("Example translations")
+        for i in range(0,5):
+            print(f'{decoded_preds[i]} - {decoded_labels_list[i]}')
+        #breakpoint()
         # Extract a few results from ROUGE
-        result = {key: value.mid.fmeasure * 100 for key, value in result.items()}
+        #result = {key: value.mid.fmeasure * 100 for key, value in result.items()}
 
-        prediction_lens = [np.count_nonzero(pred != tokenizer.pad_token_id) for pred in preds]
-        result["gen_len"] = np.mean(prediction_lens)
-        result = {k: round(v, 4) for k, v in result.items()}
+        #prediction_lens = [np.count_nonzero(pred != tokenizer.pad_token_id) for pred in preds]
+        #result["gen_len"] = np.mean(prediction_lens)
+        #result = {k: round(v, 4) for k, v in result.items()}
         return result
 
     # Enable tensorboard only on the master node