mct-nightly 2.1.0.20240717.444__py3-none-any.whl → 2.1.0.20240719.444__py3-none-any.whl

{mct_nightly-2.1.0.20240717.444.dist-info → mct_nightly-2.1.0.20240719.444.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: mct-nightly
-Version: 2.1.0.20240717.444
+Version: 2.1.0.20240719.444
 Summary: A Model Compression Toolkit for neural networks
 Home-page: UNKNOWN
 License: UNKNOWN

{mct_nightly-2.1.0.20240717.444.dist-info → mct_nightly-2.1.0.20240719.444.dist-info}/RECORD

@@ -1,4 +1,4 @@
-model_compression_toolkit/__init__.py,sha256=lgLwXN1Xe39gig854ixa-q2WUH5g4yl5eKb4XC_EKsg,1573
+model_compression_toolkit/__init__.py,sha256=JshHkyFrgHVzyqpaLEB0DIouHvcS77kBPMhPQ58BgKQ,1573
 model_compression_toolkit/constants.py,sha256=9pVleMwnhlM4QwIL2HcEq42I1uF4rlSw63RUjkxOF4w,3923
 model_compression_toolkit/defaultdict.py,sha256=LSc-sbZYXENMCw3U9F4GiXuv67IKpdn0Qm7Fr11jy-4,2277
 model_compression_toolkit/logger.py,sha256=3DByV41XHRR3kLTJNbpaMmikL8icd9e1N-nkQAY9oDk,4567
@@ -31,9 +31,9 @@ model_compression_toolkit/core/common/fusion/__init__.py,sha256=Rf1RcYmelmdZmBV5
 model_compression_toolkit/core/common/fusion/layer_fusing.py,sha256=lOubqpc18TslhXZijWUJQAa1c3jIB2S-M-5HK78wJPQ,5548
 model_compression_toolkit/core/common/graph/__init__.py,sha256=Xr-Lt_qXMdrCnnOaUS_OJP_3iTTGfPCLf8_vSrQgCs0,773
 model_compression_toolkit/core/common/graph/base_graph.py,sha256=lmIw0srKiwCvz7KWqfwKTxyQHDy3s6rWMIXzFAa1UMo,38326
-model_compression_toolkit/core/common/graph/base_node.py,sha256=X_0zqHrKYAsmnj9tAKjVYasbFcZD8OHpjdiMj9ugQs0,29436
+model_compression_toolkit/core/common/graph/base_node.py,sha256=Hqv5lsEYT2uz5FYAX44Rsps_Ax73_kVoXy_DNaXIddU,29448
 model_compression_toolkit/core/common/graph/edge.py,sha256=buoSEUZwilWBK3WeBKpJ-GeDaUA1SDdOHxDpxU_bGpk,3784
-model_compression_toolkit/core/common/graph/functional_node.py,sha256=BbxQ-WRk4R-5hbpQDBANkhRRTkaG7eogeiJwLfLb_EU,3950
+model_compression_toolkit/core/common/graph/functional_node.py,sha256=XvzydBSRxgpYdKS-aYVaWtH3FDzJPKGad3bai9wF3BI,3956
 model_compression_toolkit/core/common/graph/graph_matchers.py,sha256=CrDoHYq4iPaflgJWmoJ1K4ziLrRogJvFTVWg8P0UcDU,4744
 model_compression_toolkit/core/common/graph/graph_searches.py,sha256=2oKuW6L8hP-oL0lFO9PhQFt9fEFgVJwpc1u4fHExAtE,5128
 model_compression_toolkit/core/common/graph/virtual_activation_weights_node.py,sha256=3el-A7j1oyoo1_9zq3faQp7IeRsFXFCvnrb3zZFXpU0,9803
@@ -159,7 +159,7 @@ model_compression_toolkit/core/keras/back2framework/__init__.py,sha256=rhIiXg_nB
 model_compression_toolkit/core/keras/back2framework/factory_model_builder.py,sha256=urpfyHvIzD08QzPBWusVBT_dKZ8ZUf1I1zIQNb4qe5Y,2233
 model_compression_toolkit/core/keras/back2framework/float_model_builder.py,sha256=9SFHhX-JnkB8PvYIIHRYlReBDI_RkZY9LditzW_ElLk,2444
 model_compression_toolkit/core/keras/back2framework/instance_builder.py,sha256=fBj13c6zkVoWX4JJG18_uXPptiEJqXClE_zFbaFB6Q8,4517
-model_compression_toolkit/core/keras/back2framework/keras_model_builder.py,sha256=KXA5rik1cvCSgIbybPfX3tsMlzoILDboVONGyqzXGh0,16290
+model_compression_toolkit/core/keras/back2framework/keras_model_builder.py,sha256=XFSSaET4oPWB_cx-Q_c9pDJfWyQ1qXT9JXBl5FJCTa4,18137
 model_compression_toolkit/core/keras/back2framework/mixed_precision_model_builder.py,sha256=ygIS1WIiftF1VC3oGhc8N6j7MryKtWgEg8nr50p7f4U,15587
 model_compression_toolkit/core/keras/back2framework/quantized_model_builder.py,sha256=5wFb4nx_F0Wu4c8pLf6n6OzxOHtpOJ6_3mQsNSXIudU,2481
 model_compression_toolkit/core/keras/graph_substitutions/__init__.py,sha256=mjbqLD-KcG3eNeCYpu1GBS7VclGVOQ63x2p6mAAuba4,698
@@ -199,7 +199,7 @@ model_compression_toolkit/core/keras/quantizer/lut_fake_quant.py,sha256=Up3-sbuA
 model_compression_toolkit/core/keras/reader/__init__.py,sha256=mjbqLD-KcG3eNeCYpu1GBS7VclGVOQ63x2p6mAAuba4,698
 model_compression_toolkit/core/keras/reader/common.py,sha256=eZWjBcvTDUX7fCWmy1OAH4lYLFTh59_UQ_nP_Gjp4yw,2594
 model_compression_toolkit/core/keras/reader/connectivity_handler.py,sha256=AgF6qXZOJMeXvc-pBnGY23BJz7wPBx2aTYxHiO8efec,11303
-model_compression_toolkit/core/keras/reader/node_builder.py,sha256=SAPkgL8aqJjnB6eCucU2D4m50WACCzWC8wjCVtFnwp8,10424
+model_compression_toolkit/core/keras/reader/node_builder.py,sha256=H4ZvaRKt7W7zKzmEDlhveHWf_rLlUCxM0G1s7lnLzn0,13953
 model_compression_toolkit/core/keras/reader/reader.py,sha256=wS9UQ2wJKnkZYe9JHwQp7ygDr6CRlzrxmIyLDv1Qz6U,8109
 model_compression_toolkit/core/keras/reader/nested_model/__init__.py,sha256=mjbqLD-KcG3eNeCYpu1GBS7VclGVOQ63x2p6mAAuba4,698
 model_compression_toolkit/core/keras/reader/nested_model/edges_merger.py,sha256=K6KAH9o8KSG6baLmhKoCrYK-i-wb6gRKiZmoijFqEYA,7906
@@ -517,8 +517,8 @@ model_compression_toolkit/xquant/pytorch/model_analyzer.py,sha256=b93o800yVB3Z-i
 model_compression_toolkit/xquant/pytorch/pytorch_report_utils.py,sha256=yrZNVRm2IRU7r7R-hjS2lOQ6wvEEvbeunvf2jKoWjXk,3277
 model_compression_toolkit/xquant/pytorch/similarity_functions.py,sha256=CERxq5K8rqaiE-DlwhZBTUd9x69dtYJlkHOPLB54vm8,2354
 model_compression_toolkit/xquant/pytorch/tensorboard_utils.py,sha256=eyMoXt5o5EnMr6d-rpCwQdX5mAiYiymvbgKv4tf7-a0,4576
-mct_nightly-2.1.0.20240717.444.dist-info/LICENSE.md,sha256=aYSSIb-5AFPeITTvXm1UAoe0uYBiMmSS8flvXaaFUks,10174
-mct_nightly-2.1.0.20240717.444.dist-info/METADATA,sha256=-s4IptD94mb3pm2pxywOC5RpgZ1D0NULKBxTKOz-zNg,19719
-mct_nightly-2.1.0.20240717.444.dist-info/WHEEL,sha256=GJ7t_kWBFywbagK5eo9IoUwLW6oyOeTKmQ-9iHFVNxQ,92
-mct_nightly-2.1.0.20240717.444.dist-info/top_level.txt,sha256=gsYA8juk0Z-ZmQRKULkb3JLGdOdz8jW_cMRjisn9ga4,26
-mct_nightly-2.1.0.20240717.444.dist-info/RECORD,,
+mct_nightly-2.1.0.20240719.444.dist-info/LICENSE.md,sha256=aYSSIb-5AFPeITTvXm1UAoe0uYBiMmSS8flvXaaFUks,10174
+mct_nightly-2.1.0.20240719.444.dist-info/METADATA,sha256=fhA5w_OA1BHS2PndN7W4uW0KXhv7MM7IPKkTeJxpzLU,19719
+mct_nightly-2.1.0.20240719.444.dist-info/WHEEL,sha256=GJ7t_kWBFywbagK5eo9IoUwLW6oyOeTKmQ-9iHFVNxQ,92
+mct_nightly-2.1.0.20240719.444.dist-info/top_level.txt,sha256=gsYA8juk0Z-ZmQRKULkb3JLGdOdz8jW_cMRjisn9ga4,26
+mct_nightly-2.1.0.20240719.444.dist-info/RECORD,,

model_compression_toolkit/__init__.py

@@ -27,4 +27,4 @@ from model_compression_toolkit import data_generation
 from model_compression_toolkit import pruning
 from model_compression_toolkit.trainable_infrastructure.keras.load_model import keras_load_quantized_model
 
-__version__ = "2.1.0.20240717.000444"
+__version__ = "2.1.0.20240719.000444"

model_compression_toolkit/core/common/graph/base_node.py

@@ -36,7 +36,7 @@ class BaseNode:
                  framework_attr: Dict[str, Any],
                  input_shape: Tuple[Any],
                  output_shape: Tuple[Any],
-                 weights: Dict[str, np.ndarray],
+                 weights: Dict[Union[str, int], np.ndarray],
                  layer_class: type,
                  reuse: bool = False,
                  reuse_group: str = None,

model_compression_toolkit/core/common/graph/functional_node.py

@@ -59,7 +59,7 @@ class FunctionalNode(BaseNode):
                          has_activation=has_activation)
 
         self.op_call_kwargs = op_call_kwargs
-        self.op_call_args = op_call_args
+        self.op_call_args = list(op_call_args)
         self.functional_op = functional_op
         self.inputs_as_list = inputs_as_list
         self.tensor_input_allocs = [] if tensor_input_allocs is None else tensor_input_allocs

model_compression_toolkit/core/keras/back2framework/keras_model_builder.py

@@ -12,6 +12,7 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 # ==============================================================================
+from copy import copy
 
 import tensorflow as tf
 from keras.models import Model
@@ -19,6 +20,7 @@ from packaging import version
 
 from model_compression_toolkit.core.common.back2framework.base_model_builder import BaseModelBuilder
 from model_compression_toolkit.core.common.user_info import UserInformation
+from model_compression_toolkit.logger import Logger
 
 if version.parse(tf.__version__) >= version.parse("2.13"):
     from keras import Input
@@ -271,15 +273,38 @@ class KerasModelBuilder(BaseModelBuilder):
                                                                            out_tensors_of_n_float)
         else:
             input_tensors = [tensor for tensor_list in input_tensors for tensor in tensor_list]  # flat list of lists
+            if isinstance(n, FunctionalNode):
+                op_call_kwargs = {} if n.op_call_kwargs is None else copy(n.op_call_kwargs)
             if not isinstance(op_func, KerasQuantizationWrapper):
                 # The KerasQuantizationWrapper will insert the quantized positional weights internally.
-                input_tensors = n.insert_positional_weights_to_input_list(input_tensors)
+                if isinstance(n, FunctionalNode):
+                    if n.tensor_input_allocs is not None:
+                        if n.inputs_as_list:
+                            input_tensors = n.insert_positional_weights_to_input_list(input_tensors)
+                        else:
+                            # If the were any const attributes in the layer's inputs, we retrieve them as kwargs
+                            # for the operator call.
+                            for pos, k in enumerate(n.tensor_input_allocs):
+                                if k not in op_call_kwargs:  # op_call_kwargs is initialized because we are under FunctionalNode
+                                    # If the argument is saved in tensor_input_allocs but does not exists in the node kwargs
+                                    # then it is expected to be either an input tensor or a positional weight of the node.
+                                    arg = n.weights.get(pos)
+                                    if arg is None:
+                                        if len(input_tensors) == 0:
+                                            Logger.critical(f"Couldn't find a weight or input tensor matching operator's "
+                                                            f"argument name '{k}' in location {pos} for node {n.name}.")
+                                        arg = input_tensors.pop(0)
+                                    op_call_kwargs.update({k: arg})
+                else:
+                    # If the operator is not a functional node then positional weights should be inserted
+                    # into the inputs list.
+                    input_tensors = n.insert_positional_weights_to_input_list(input_tensors)
             # Build a functional node using its args
             if isinstance(n, FunctionalNode):
                 if n.inputs_as_list:  # If the first argument should be a list of tensors:
-                    out_tensors_of_n_float = op_func(input_tensors, *n.op_call_args, **n.op_call_kwargs)
+                    out_tensors_of_n_float = op_func(input_tensors, *n.op_call_args, **op_call_kwargs)
                 else:  # If the input tensors should not be a list but iterated:
-                    out_tensors_of_n_float = op_func(*input_tensors, *n.op_call_args, **n.op_call_kwargs)
+                    out_tensors_of_n_float = op_func(*input_tensors, *n.op_call_args, **op_call_kwargs)
             else:
                 # If operator expects a single input tensor, it cannot be a list as it should
                 # have a dtype field.

model_compression_toolkit/core/keras/reader/node_builder.py

@@ -12,7 +12,9 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 # ==============================================================================
-from typing import Any, List, Dict
+from copy import copy
+
+from typing import Any, List, Dict, Union, Tuple
 
 import tensorflow as tf
 from tensorflow.python.util import tf_inspect
@@ -41,7 +43,7 @@ layers = keras.layers
 
 REUSED_IDENTIFIER = '_reused_'
 
-is_const = lambda x: isinstance(x, (tf.Variable, tf.Tensor, np.ndarray, float))
+is_const = lambda x: isinstance(x, (tf.Variable, tf.Tensor, np.ndarray, tuple, list))
 is_tensor = lambda x: isinstance(x, KerasTensor)
 
 
@@ -62,35 +64,139 @@ def get_kwargs2index(tfoplambda_layer: TFOpLambda) -> Dict[str, int]:
     Positional weights are saved according to their index in the node's call arguments, so
     need to know the function arguments' names in case the weights are in the kwargs.
 
-    Note: the kwargs2index dictionary is initialized manually (and not with tf_inspect) so
-    it will only include the arguments that may contain constants. For example, we don't
-    want the transpose_a attribute of tf.matmul to be saved as a constant.
-
-    Every operation we add support to, needs to be added here.
-
     Args:
         tfoplambda_layer: TFOpLambda layer.
 
     Returns:
         A dictionary with argument number and index: {arg_name: arg_index}.
     """
-    kwargs2index = {tf.add: {'x': 0, 'y': 1},
-                    tf.subtract: {'x': 0, 'y': 1},
-                    tf.divide: {'x': 0, 'y': 1},
-                    tf.truediv: {'x': 0, 'y': 1},
-                    tf.multiply: {'x': 0, 'y': 1},
-                    tf.pow: {'x': 0, 'y': 1},
-                    tf.matmul: {'a': 0, 'b': 1}}.get(tfoplambda_layer.function)
-    if not kwargs2index:
-        # In TF 2.15 the function attribute is different and doesn't match the original
-        # operation object we use. Therefore, we extract kwargs2index with the symbol.
-        kwargs2index = {'__operators__.add': {'x': 0, 'y': 1},
-                        'math.add': {'x': 0, 'y': 1},
-                        'math.multiply': {'x': 0, 'y': 1},
-                        'linalg.matmul': {'a': 0, 'b': 1},
-                        'concat': {'values': 0}}.get(tfoplambda_layer.symbol, {})
-
-    return kwargs2index
+
+    full_args = tf_inspect.getfullargspec(tfoplambda_layer.function).args
+
+    return {arg_name: i for i, arg_name in enumerate(full_args)}
+
+
+def _extract_const_attrs_from_kwargs(op_call_kwargs: Dict[str, Any],
+                                     kwarg2index: Dict[str, int],
+                                     weights: Dict[Union[str, int], Any]) -> Dict[str, Any]:
+    """
+    Extract const weights of the layer from the operator's key arguments dictionary.
+    This function extracts the attributes, updates the nodes weights dictionary and removes them from the original
+    kwargs mapping.
+
+    Args:
+        op_call_kwargs: A mapping of the operator key arguments.
+        kwarg2index: A dictionary with argument number and index: {arg_name: arg_index}.
+        weights: Node weights mapping. This dictionary is modified by this function.
+
+    Returns: A modified operator key arguments mapping.
+
+    """
+
+    # read weights from call kwargs
+    for k, v in op_call_kwargs.items():
+        if is_const(v):
+            # if k in kwarg2index:
+            weights.update({kwarg2index[k]: to_numpy(v, is_single_tensor=True)})
+
+    # remove weights and KerasTensors from op_call_kwargs
+    op_call_kwargs = {k: v for k, v in op_call_kwargs.items()
+                      if not (kwarg2index.get(k) in weights or is_tensor(v))}
+
+    return op_call_kwargs
+
+
+def _build_arguments_alloc(n: KerasNode, inputs_as_list: bool, kwarg2index: Dict[str, int]) -> List:
+    """
+    Builds arguments allocation list.
+    In Keras, if there is any argument that is a constant, we convert all arguments and inputs to be
+    considered as op kwargs for simpler reconstruction of the model from the graph later.
+    Therefore, we build a location list that includes the argument names (keys).
+    If the input is a list, then we don't need to save the keys, since we can assume that all possible constant
+    arguments are within the first argument (the list) and are stored by their position in the list.
+
+    Args:
+        n: fx node.
+        inputs_as_list: Is node's inputs are a list.
+
+    Returns:
+        A list of argument allocations in the node's inputs.
+
+    """
+
+    tensor_input_alloc = []
+    op_call_args = list(n.call_args)
+    if not inputs_as_list:
+        sorted_kwargs_pos = sorted(kwarg2index.items(), key=lambda x: x[1])
+        tensor_input_alloc = [k for k, _ in sorted_kwargs_pos[:len(op_call_args)]]
+        for k, idx in sorted_kwargs_pos[len(op_call_args):]:
+            if k in n.call_kwargs:
+                tensor_input_alloc.append(k)
+
+    return tensor_input_alloc
+
+def _extract_const_attrs_from_args(op_call_args: List[Any],
+                                   op_call_kwargs: Dict[str, Any],
+                                   inputs_as_list: bool,
+                                   tensor_inputs_alloc: List,
+                                   weights: Dict[Union[str, int], Any]) -> Tuple:
+    """
+    Extract const weights of the layer from the operator's arguments list.
+    This function extracts the attributes, updates the nodes weights dictionary and removes them from the original
+    arguments list.
+
+    Args:
+        op_call_args: A list of the operator arguments.
+        op_call_kwargs: A mapping of key-arguments of the operator.
+        inputs_as_list: Whether the input of the layer is a list.
+        tensor_inputs_alloc: Allocation of argument inputs to the operator (if there are const inputs, otherwise None).
+        weights: Node weights mapping. This dictionary is modified by this function.
+
+    Returns: A modified operator arguments list.
+
+    """
+
+    move_args_to_kwargs = tensor_inputs_alloc is not None and len(tensor_inputs_alloc) > 0
+
+    # read weights from call args
+    for i, arg in enumerate(op_call_args[0] if inputs_as_list else op_call_args):
+        if is_const(arg):
+            weights.update({i: to_numpy(arg, is_single_tensor=True)})
+        else:
+            if not inputs_as_list:
+                if move_args_to_kwargs:
+                    # In this case we move all arguments and inputs to the kwargs
+                    op_call_kwargs.update({tensor_inputs_alloc[i]: arg})
+
+    # remove weights and KerasTensors from op_call_args
+    if inputs_as_list:
+        op_call_args = tuple(op_call_args[1:])
+    else:
+        op_call_args = tuple([a for i, a in enumerate(op_call_args)
+                              if not (i in weights or is_tensor(a) or (move_args_to_kwargs and tensor_inputs_alloc[i]
+                                                                       in op_call_kwargs))])
+
+    return op_call_args
+
+
+def _has_const_attributes(op_call_args: List, op_call_kwargs: Dict, input_as_list: bool) -> bool:
+    """
+    Returns whether the layer's input include a constant tensor (that we might want to quantize).
+
+    Args:
+        op_call_args: A list of arguments to the layer.
+        op_call_kwargs: A dictionary of key-arguments to the layer.
+        input_as_list: Whether the input to the layer is a list of tensors.
+
+    Returns: True if the input arguments include a constant tensor, False otherwise.
+
+    """
+    if input_as_list:
+        return any([is_const(a) for a in op_call_args[0]])
+    const_args = [a for a in op_call_args if is_const(a)]
+    const_kwargs = [k for k, v in op_call_kwargs.items() if is_const(v)]
+
+    return len(const_args) > 0 or len(const_kwargs) > 0
 
 
 def build_node(node: KerasNode,
@@ -110,8 +216,8 @@ def build_node(node: KerasNode,
     """
     keras_layer = node.layer  # get the layer the node represents.
     layer_config = keras_layer.get_config()  # layer configuration to reconstruct it.
-    op_call_args = node.call_args
-    op_call_kwargs = node.call_kwargs
+    op_call_args = copy(node.call_args)
+    op_call_kwargs = copy(node.call_kwargs)
     layer_class = type(keras_layer)  # class path to instantiating it in back2framework.
     weights = {v.name: v.numpy() for v in keras_layer.weights}  # layer's weights
 
@@ -152,32 +258,14 @@ def build_node(node: KerasNode,
         if len(weights) > 0:
             Logger.critical('Functional nodes are not expected to have weights in this framework.')
 
-        # read weights from call args
-        tf_function_symbols = get_tf_function_symbols()
-        for i, arg in enumerate(op_call_args[0] if inputs_as_list else op_call_args):
-            if is_const(arg) or (
-                    keras_layer.symbol in tf_function_symbols and
-                    isinstance(arg, (tuple, list))):
-                if inputs_as_list or i in kwarg2index.values():
-                    weights.update({i: to_numpy(arg, is_single_tensor=True)})
-        # remove weights and KerasTensors and weights from op_call_args
-        if inputs_as_list:
-            op_call_args = tuple(op_call_args[1:])
-        else:
-            op_call_args = tuple([a for i, a in enumerate(op_call_args)
-                                  if not (i in weights or is_tensor(a))])
-
-        # read weights from call kwargs
-        weight_keys = []
-        for k, v in op_call_kwargs.items():
-            if is_const(v) or (keras_layer.symbol in tf_function_symbols and
-                               isinstance(v, (tuple, list))):
-                if k in kwarg2index:
-                    weights.update({kwarg2index[k]: to_numpy(v, is_single_tensor=True)})
-                    weight_keys.append(k)
-        # remove weights and KerasTensors and weights from op_call_kwargs
-        op_call_kwargs = {k: v for k, v in op_call_kwargs.items()
-                          if not (kwarg2index.get(k) in weights or is_tensor(v))}
+        # Build tensor_input_alloc required for the model builder. All inputs are received as a list in the builder,
+        # so tensor_input_alloc is used to allocate each input in the correct place in the node's args & kwargs.
+        tensor_input_alloc = None if not _has_const_attributes(op_call_args, op_call_kwargs, inputs_as_list) \
+            else _build_arguments_alloc(node, inputs_as_list, kwarg2index)
+
+        op_call_args = _extract_const_attrs_from_args(op_call_args, op_call_kwargs, inputs_as_list,
+                                                      tensor_input_alloc, weights)
+        op_call_kwargs = _extract_const_attrs_from_kwargs(op_call_kwargs, kwarg2index, weights)
 
         node = FunctionalNode(node_name,
                               layer_config,
@@ -190,7 +278,8 @@ def build_node(node: KerasNode,
                               is_reused,
                               reuse_group,
                               functional_op=keras_layer.function,
-                              inputs_as_list=inputs_as_list)
+                              inputs_as_list=inputs_as_list,
+                              tensor_input_allocs=tensor_input_alloc)
     else:
         # Read constant weights from layers such as layers.Add
         if len(op_call_args) > 0 and isinstance(op_call_args[0], (list, tuple)):