Saving weights and logs of step 8

flax_model.msgpack

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:79e9400819ada0aa172374be8a0a62667546a8db83f8483f337944a9eaf9cb19
+oid sha256:7b4812d4f42c82ee8a963f90fe45ce25a75b0283d71e093ff578ba3c65de9d6e
 size 498796983

run.sh

@@ -7,10 +7,13 @@
     --dataset_name="oscar" \
     --dataset_config_name="unshuffled_deduplicated_als" \
     --max_seq_length="128" \
-    --per_device_train_batch_size="4" \
-    --per_device_eval_batch_size="4" \
+    --per_device_train_batch_size="1" \
+    --per_device_eval_batch_size="1" \
     --learning_rate="3e-4" \
     --warmup_steps="1000" \
     --overwrite_output_dir \
     --num_train_epochs="8" \
+		--logging_steps="10" \
+		--save_steps="8" \
+		--eval_steps="15" \
     --push_to_hub

run_mlm_flax.py

@@ -606,7 +606,7 @@ if __name__ == "__main__":
             state, train_metric, dropout_rngs = p_train_step(state, model_inputs, dropout_rngs)
             train_metrics.append(train_metric)
 
-            cur_step = epoch * num_train_samples + step
+            cur_step = epoch * (num_train_samples // train_batch_size) + step
 
             if cur_step % training_args.logging_steps == 0 and cur_step > 0:
                 # Save metrics
@@ -621,43 +621,43 @@ if __name__ == "__main__":
 
                 train_metrics = []
 
-        # ======================== Evaluating ==============================
-        num_eval_samples = len(tokenized_datasets["validation"])
-        eval_samples_idx = jnp.arange(num_eval_samples)
-        eval_batch_idx = generate_batch_splits(eval_samples_idx, eval_batch_size)
+            if cur_step % training_args.eval_steps == 0 and step > 0:
+                # ======================== Evaluating ==============================
+                num_eval_samples = len(tokenized_datasets["validation"])
+                eval_samples_idx = jnp.arange(num_eval_samples)
+                eval_batch_idx = generate_batch_splits(eval_samples_idx, eval_batch_size)
 
-        eval_metrics = []
-        for i, batch_idx in enumerate(tqdm(eval_batch_idx, desc="Evaluating ...", position=2)):
-            samples = [tokenized_datasets["validation"][int(idx)] for idx in batch_idx]
-            model_inputs = data_collator(samples, pad_to_multiple_of=16)
+                eval_metrics = []
+                for i, batch_idx in enumerate(tqdm(eval_batch_idx, desc="Evaluating ...", position=2)):
+                    samples = [tokenized_datasets["validation"][int(idx)] for idx in batch_idx]
+                    model_inputs = data_collator(samples, pad_to_multiple_of=16)
 
-            # Model forward
-            model_inputs = shard(model_inputs.data)
-            metrics = p_eval_step(state.params, model_inputs)
-            eval_metrics.append(metrics)
-
-        # normalize eval metrics
-        eval_metrics = get_metrics(eval_metrics)
-        eval_metrics = jax.tree_map(jnp.sum, eval_metrics)
-        eval_normalizer = eval_metrics.pop("normalizer")
-        eval_metrics = jax.tree_map(lambda x: x / eval_normalizer, eval_metrics)
-
-        # Update progress bar
-        epochs.desc = (
-            f"Epoch... ({epoch + 1}/{num_epochs} | Loss: {eval_metrics['loss']}, Acc: {eval_metrics['accuracy']})"
-        )
+                    # Model forward
+                    model_inputs = shard(model_inputs.data)
+                    metrics = p_eval_step(state.params, model_inputs)
+                    eval_metrics.append(metrics)
 
-        # Save metrics
-        if has_tensorboard and jax.process_index() == 0:
-            cur_step = epoch * (len(tokenized_datasets["train"]) // train_batch_size)
-            write_eval_metric(summary_writer, eval_metrics, cur_step)
-
-        # save checkpoint after each epoch and push checkpoint to the hub
-        if jax.process_index() == 0:
-            params = jax.device_get(jax.tree_map(lambda x: x[0], state.params))
-            model.save_pretrained(
-                training_args.output_dir,
-                params=params,
-                push_to_hub=training_args.push_to_hub,
-                commit_message=f"Saving weights and logs of epoch {epoch+1}",
-            )
+                # normalize eval metrics
+                eval_metrics = get_metrics(eval_metrics)
+                eval_metrics = jax.tree_map(jnp.sum, eval_metrics)
+                eval_normalizer = eval_metrics.pop("normalizer")
+                eval_metrics = jax.tree_map(lambda x: x / eval_normalizer, eval_metrics)
+
+                # Update progress bar
+                epochs.desc = f"Step... ({cur_step} | Loss: {eval_metrics['loss']}, Acc: {eval_metrics['accuracy']})"
+
+                # Save metrics
+                if has_tensorboard and jax.process_index() == 0:
+                    cur_step = epoch * (len(tokenized_datasets["train"]) // train_batch_size)
+                    write_eval_metric(summary_writer, eval_metrics, cur_step)
+
+            if cur_step % training_args.save_steps == 0 and step > 0:
+                # save checkpoint after each epoch and push checkpoint to the hub
+                if jax.process_index() == 0:
+                    params = jax.device_get(jax.tree_map(lambda x: x[0], state.params))
+                    model.save_pretrained(
+                        training_args.output_dir,
+                        params=params,
+                        push_to_hub=training_args.push_to_hub,
+                        commit_message=f"Saving weights and logs of step {cur_step}",
+                    )