[Flax] adding support for batch norm layers (#21581)

* [flax] adding support for batch norm layers

* fixing bugs related to pt+flax integration

* cleanup, batchnorm support in sharded pt to flax

* support for batchnorm tests in pt+flax integration

* simplifying checking batch norm layer

src/transformers/modeling_flax_pytorch_utils.py

@@ -83,6 +83,16 @@ def rename_key_and_reshape_tensor(
     if pt_tuple_key[-1] in ["weight", "gamma"] and is_key_or_prefix_key_in_dict(renamed_pt_tuple_key):
         return renamed_pt_tuple_key, pt_tensor
 
+    # batch norm layer mean
+    renamed_pt_tuple_key = pt_tuple_key[:-1] + ("mean",)
+    if pt_tuple_key[-1] == "running_mean" and not is_key_or_prefix_key_in_dict(pt_tuple_key):
+        return renamed_pt_tuple_key, pt_tensor
+
+    # batch norm layer var
+    renamed_pt_tuple_key = pt_tuple_key[:-1] + ("var",)
+    if pt_tuple_key[-1] == "running_var" and not is_key_or_prefix_key_in_dict(pt_tuple_key):
+        return renamed_pt_tuple_key, pt_tensor
+
     # embedding
     renamed_pt_tuple_key = pt_tuple_key[:-1] + ("embedding",)
     if pt_tuple_key[-1] == "weight" and is_key_or_prefix_key_in_dict(renamed_pt_tuple_key):
@@ -118,13 +128,25 @@ def convert_pytorch_state_dict_to_flax(pt_state_dict, flax_model):
     pt_state_dict = {k: v.numpy() for k, v in pt_state_dict.items()}
 
     model_prefix = flax_model.base_model_prefix
-    random_flax_state_dict = flatten_dict(flax_model.params)
+
+    # use params dict if the model contains batch norm layers
+    if "params" in flax_model.params:
+        flax_model_params = flax_model.params["params"]
+    else:
+        flax_model_params = flax_model.params
+    random_flax_state_dict = flatten_dict(flax_model_params)
+
+    # add batch_stats keys,values to dict
+    if "batch_stats" in flax_model.params:
+        flax_batch_stats = flatten_dict(flax_model.params["batch_stats"])
+        random_flax_state_dict.update(flax_batch_stats)
+
     flax_state_dict = {}
 
-    load_model_with_head_into_base_model = (model_prefix not in flax_model.params) and (
+    load_model_with_head_into_base_model = (model_prefix not in flax_model_params) and (
         model_prefix in {k.split(".")[0] for k in pt_state_dict.keys()}
     )
-    load_base_model_into_model_with_head = (model_prefix in flax_model.params) and (
+    load_base_model_into_model_with_head = (model_prefix in flax_model_params) and (
         model_prefix not in {k.split(".")[0] for k in pt_state_dict.keys()}
     )
 
@@ -154,8 +176,22 @@ def convert_pytorch_state_dict_to_flax(pt_state_dict, flax_model):
                     f"{random_flax_state_dict[flax_key].shape}, but is {flax_tensor.shape}."
                 )
 
-        # also add unexpected weight so that warning is thrown
-        flax_state_dict[flax_key] = jnp.asarray(flax_tensor)
+        # add batch stats if the model contains batchnorm layers
+        if "batch_stats" in flax_model.params:
+            if "mean" in flax_key[-1] or "var" in flax_key[-1]:
+                flax_state_dict[("batch_stats",) + flax_key] = jnp.asarray(flax_tensor)
+                continue
+            # remove num_batches_tracked key
+            if "num_batches_tracked" in flax_key[-1]:
+                flax_state_dict.pop(flax_key, None)
+                continue
+
+            # also add unexpected weight so that warning is thrown
+            flax_state_dict[("params",) + flax_key] = jnp.asarray(flax_tensor)
+
+        else:
+            # also add unexpected weight so that warning is thrown
+            flax_state_dict[flax_key] = jnp.asarray(flax_tensor)
 
     return unflatten_dict(flax_state_dict)
 
@@ -176,12 +212,21 @@ def convert_pytorch_sharded_state_dict_to_flax(shard_filenames, flax_model):
         pt_state_dict = {k: v.numpy() for k, v in pt_state_dict.items()}
 
         model_prefix = flax_model.base_model_prefix
-        random_flax_state_dict = flatten_dict(flax_model.params)
 
-        load_model_with_head_into_base_model = (model_prefix not in flax_model.params) and (
+        # use params dict if the model contains batch norm layers and then add batch_stats keys,values to dict
+        if "batch_stats" in flax_model.params:
+            flax_model_params = flax_model.params["params"]
+
+            random_flax_state_dict = flatten_dict(flax_model_params)
+            random_flax_state_dict.update(flatten_dict(flax_model.params["batch_stats"]))
+        else:
+            flax_model_params = flax_model.params
+            random_flax_state_dict = flatten_dict(flax_model_params)
+
+        load_model_with_head_into_base_model = (model_prefix not in flax_model_params) and (
             model_prefix in {k.split(".")[0] for k in pt_state_dict.keys()}
         )
-        load_base_model_into_model_with_head = (model_prefix in flax_model.params) and (
+        load_base_model_into_model_with_head = (model_prefix in flax_model_params) and (
             model_prefix not in {k.split(".")[0] for k in pt_state_dict.keys()}
         )
         # Need to change some parameters name to match Flax names
@@ -209,8 +254,25 @@ def convert_pytorch_sharded_state_dict_to_flax(shard_filenames, flax_model):
                         f"{random_flax_state_dict[flax_key].shape}, but is {flax_tensor.shape}."
                     )
 
-            # also add unexpected weight so that warning is thrown
-            flax_state_dict[flax_key] = jnp.asarray(flax_tensor)
+            # add batch stats if the model contains batchnorm layers
+            if "batch_stats" in flax_model.params:
+                if "mean" in flax_key[-1]:
+                    flax_state_dict[("batch_stats",) + flax_key] = jnp.asarray(flax_tensor)
+                    continue
+                if "var" in flax_key[-1]:
+                    flax_state_dict[("batch_stats",) + flax_key] = jnp.asarray(flax_tensor)
+                    continue
+                # remove num_batches_tracked key
+                if "num_batches_tracked" in flax_key[-1]:
+                    flax_state_dict.pop(flax_key, None)
+                    continue
+
+                # also add unexpected weight so that warning is thrown
+                flax_state_dict[("params",) + flax_key] = jnp.asarray(flax_tensor)
+
+            else:
+                # also add unexpected weight so that warning is thrown
+                flax_state_dict[flax_key] = jnp.asarray(flax_tensor)
     return unflatten_dict(flax_state_dict)
 
 
@@ -299,7 +361,16 @@ def load_flax_weights_in_pytorch_model(pt_model, flax_state):
         elif flax_key_tuple[-1] in ["scale", "embedding"]:
             flax_key_tuple = flax_key_tuple[:-1] + ("weight",)
 
-        flax_key = ".".join(flax_key_tuple)
+        # adding batch stats from flax batch norm to pt
+        elif "mean" in flax_key_tuple[-1]:
+            flax_key_tuple = flax_key_tuple[:-1] + ("running_mean",)
+        elif "var" in flax_key_tuple[-1]:
+            flax_key_tuple = flax_key_tuple[:-1] + ("running_var",)
+
+        if "batch_stats" in flax_state:
+            flax_key = ".".join(flax_key_tuple[1:])  # Remove the params/batch_stats header
+        else:
+            flax_key = ".".join(flax_key_tuple)
 
         if flax_key in pt_model_dict:
             if flax_tensor.shape != pt_model_dict[flax_key].shape:

src/transformers/modeling_flax_utils.py

@@ -837,14 +837,35 @@ class FlaxPreTrainedModel(PushToHubMixin, FlaxGenerationMixin):
                 # keep the params on CPU if we don't want to initialize
                 state = jax.tree_util.tree_map(lambda x: jax.device_put(x, jax.devices("cpu")[0]), state)
 
-        # if model is base model only use model_prefix key
-        if cls.base_model_prefix not in dict(model.params_shape_tree) and cls.base_model_prefix in state:
-            state = state[cls.base_model_prefix]
+        if "batch_stats" in state:  # if flax model contains batch norm layers
+            # if model is base model only use model_prefix key
+            if (
+                cls.base_model_prefix not in dict(model.params_shape_tree["params"])
+                and cls.base_model_prefix in state["params"]
+            ):
+                state["params"] = state["params"][cls.base_model_prefix]
+                state["batch_stats"] = state["batch_stats"][cls.base_model_prefix]
+
+            # if model is head model and we are loading weights from base model
+            # we initialize new params dict with base_model_prefix
+            if (
+                cls.base_model_prefix in dict(model.params_shape_tree["params"])
+                and cls.base_model_prefix not in state["params"]
+            ):
+                state = {
+                    "params": {cls.base_model_prefix: state["params"]},
+                    "batch_stats": {cls.base_model_prefix: state["batch_stats"]},
+                }
+
+        else:
+            # if model is base model only use model_prefix key
+            if cls.base_model_prefix not in dict(model.params_shape_tree) and cls.base_model_prefix in state:
+                state = state[cls.base_model_prefix]
 
-        # if model is head model and we are loading weights from base model
-        # we initialize new params dict with base_model_prefix
-        if cls.base_model_prefix in dict(model.params_shape_tree) and cls.base_model_prefix not in state:
-            state = {cls.base_model_prefix: state}
+            # if model is head model and we are loading weights from base model
+            # we initialize new params dict with base_model_prefix
+            if cls.base_model_prefix in dict(model.params_shape_tree) and cls.base_model_prefix not in state:
+                state = {cls.base_model_prefix: state}
 
         # flatten dicts
         state = flatten_dict(state)
@@ -854,6 +875,11 @@ class FlaxPreTrainedModel(PushToHubMixin, FlaxGenerationMixin):
         missing_keys = model.required_params - set(state.keys())
         unexpected_keys = set(state.keys()) - model.required_params
 
+        # Disabling warning when porting pytorch weights to flax, flax does not uses num_batches_tracked
+        for unexpected_key in unexpected_keys.copy():
+            if "num_batches_tracked" in unexpected_key[-1]:
+                unexpected_keys.remove(unexpected_key)
+
         if missing_keys and not _do_init:
             logger.warning(
                 f"The checkpoint {pretrained_model_name_or_path} is missing required keys: {missing_keys}. "

tests/test_modeling_flax_common.py

@@ -118,6 +118,30 @@ def random_attention_mask(shape, rng=None):
     return attn_mask
 
 
+def get_params(params, from_head_prefix=None):
+    """Function extracts relevant parameters into flatten dict from model params,
+    appends batch normalization statistics if present"""
+
+    # If Both parameters and batch normalization statistics are present
+    if "batch_stats" in params:
+        # Extract only parameters for the specified head prefix (if specified) and add batch statistics
+        if from_head_prefix is not None:
+            extracted_params = flatten_dict(unfreeze(params["params"][from_head_prefix]))
+            extracted_params.update(flatten_dict(params["batch_stats"][from_head_prefix]))
+        else:
+            extracted_params = flatten_dict(unfreeze(params["params"]))
+            extracted_params.update(flatten_dict(params["batch_stats"]))
+
+    # Only parameters are present
+    else:
+        if from_head_prefix is not None:
+            extracted_params = flatten_dict(unfreeze(params[from_head_prefix]))
+        else:
+            extracted_params = flatten_dict(unfreeze(params))
+
+    return extracted_params
+
+
 @require_flax
 class FlaxModelTesterMixin:
     model_tester = None
@@ -426,14 +450,14 @@ class FlaxModelTesterMixin:
                 continue
 
             model = base_class(config)
-            base_params = flatten_dict(unfreeze(model.params))
+            base_params = get_params(model.params)
 
             # check that all base model weights are loaded correctly
             with tempfile.TemporaryDirectory() as tmpdirname:
                 model.save_pretrained(tmpdirname)
                 head_model = model_class.from_pretrained(tmpdirname)
 
-                base_param_from_head = flatten_dict(unfreeze(head_model.params[head_model.base_model_prefix]))
+                base_param_from_head = get_params(head_model.params, from_head_prefix=head_model.base_model_prefix)
 
                 for key in base_param_from_head.keys():
                     max_diff = (base_params[key] - base_param_from_head[key]).sum().item()
@@ -448,14 +472,14 @@ class FlaxModelTesterMixin:
                 continue
 
             model = model_class(config)
-            base_params_from_head = flatten_dict(unfreeze(model.params[model.base_model_prefix]))
+            base_params_from_head = get_params(model.params, from_head_prefix=model.base_model_prefix)
 
             # check that all base model weights are loaded correctly
             with tempfile.TemporaryDirectory() as tmpdirname:
                 model.save_pretrained(tmpdirname)
                 base_model = base_class.from_pretrained(tmpdirname)
 
-                base_params = flatten_dict(unfreeze(base_model.params))
+                base_params = get_params(base_model.params)
 
                 for key in base_params_from_head.keys():
                     max_diff = (base_params[key] - base_params_from_head[key]).sum().item()
@@ -471,7 +495,7 @@ class FlaxModelTesterMixin:
                 continue
 
             model = base_class(config)
-            base_params = flatten_dict(unfreeze(model.params))
+            base_params = get_params(model.params)
 
             # convert Flax model to PyTorch model
             pt_model_class = getattr(transformers, base_class.__name__[4:])  # Skip the "Flax" at the beginning
@@ -484,7 +508,7 @@ class FlaxModelTesterMixin:
                 pt_model.save_pretrained(tmpdirname)
                 head_model = model_class.from_pretrained(tmpdirname, from_pt=True)
 
-                base_param_from_head = flatten_dict(unfreeze(head_model.params[head_model.base_model_prefix]))
+                base_param_from_head = get_params(head_model.params, from_head_prefix=head_model.base_model_prefix)
 
                 for key in base_param_from_head.keys():
                     max_diff = (base_params[key] - base_param_from_head[key]).sum().item()
@@ -500,7 +524,7 @@ class FlaxModelTesterMixin:
                 continue
 
             model = model_class(config)
-            base_params_from_head = flatten_dict(unfreeze(model.params[model.base_model_prefix]))
+            base_params_from_head = get_params(model.params, from_head_prefix=model.base_model_prefix)
 
             # convert Flax model to PyTorch model
             pt_model_class = getattr(transformers, model_class.__name__[4:])  # Skip the "Flax" at the beginning
@@ -512,7 +536,7 @@ class FlaxModelTesterMixin:
                 pt_model.save_pretrained(tmpdirname)
                 base_model = base_class.from_pretrained(tmpdirname, from_pt=True)
 
-                base_params = flatten_dict(unfreeze(base_model.params))
+                base_params = get_params(base_model.params)
 
                 for key in base_params_from_head.keys():
                     max_diff = (base_params[key] - base_params_from_head[key]).sum().item()
@@ -529,7 +553,7 @@ class FlaxModelTesterMixin:
 
             model = model_class(config)
             model.params = model.to_bf16(model.params)
-            base_params_from_head = flatten_dict(unfreeze(model.params[model.base_model_prefix]))
+            base_params_from_head = get_params(model.params, from_head_prefix=model.base_model_prefix)
 
             # convert Flax model to PyTorch model
             pt_model_class = getattr(transformers, model_class.__name__[4:])  # Skip the "Flax" at the beginning
@@ -541,7 +565,7 @@ class FlaxModelTesterMixin:
                 pt_model.save_pretrained(tmpdirname)
                 base_model = base_class.from_pretrained(tmpdirname, from_pt=True)
 
-                base_params = flatten_dict(unfreeze(base_model.params))
+                base_params = get_params(base_model.params)
 
                 for key in base_params_from_head.keys():
                     max_diff = (base_params[key] - base_params_from_head[key]).sum().item()