onnx2tf 1.29.16__py3-none-any.whl → 1.29.18__py3-none-any.whl

onnx2tf/__init__.py

@@ -1,3 +1,3 @@
 from onnx2tf.onnx2tf import convert, main
 
-__version__ = '1.29.16'
+__version__ = '1.29.18'

onnx2tf/ops/Add.py

@@ -21,6 +21,7 @@ from onnx2tf.utils.common_functions import (
     disable_unnecessary_transpose,
     shape_unmatched_special_avoidance_workaround,
     merge_two_consecutive_identical_ops_into_one,
+    transpose_with_flexing_deterrence,
     deterring_shape_corruption_due_to_broadcast,
     acquisition_of_validation_data,
     onnx_tf_tensor_validation,
@@ -297,6 +298,117 @@ def make_node(
             )
         tf_type = tf.identity
 
+    def _normalize_dim(dim):
+        return int(dim) if isinstance(dim, (int, np.integer)) else None
+
+    def _get_static_shape(tensor):
+        shape = getattr(tensor, 'shape', None)
+        if shape is None or shape == tf.TensorShape(None):
+            return None
+        return [_normalize_dim(dim) for dim in list(shape)]
+
+    def _shape_match_with_none(expected, actual):
+        if expected is None or actual is None:
+            return False
+        if len(expected) != len(actual):
+            return False
+        for e_dim, a_dim in zip(expected, actual):
+            e_dim = _normalize_dim(e_dim)
+            a_dim = _normalize_dim(a_dim)
+            if e_dim is None or a_dim is None:
+                continue
+            if e_dim != a_dim:
+                return False
+        return True
+
+    def _perm_shape(shape, perm):
+        return [shape[i] for i in perm] if shape is not None else None
+
+    def _limited_perms(rank):
+        identity = list(range(rank))
+        perms = [identity]
+        if rank == 3:
+            perms.append([0, 2, 1])
+        elif rank == 4:
+            perms.extend([[0, 2, 3, 1], [0, 3, 1, 2]])
+        elif rank == 5:
+            perms.extend([[0, 2, 3, 4, 1], [0, 4, 1, 2, 3]])
+        return perms
+
+    def _ranked_perms(perms, input_shape, onnx_shape):
+        if input_shape is None or onnx_shape is None:
+            return perms
+        scored = []
+        for perm in perms:
+            score = 0
+            for out_idx, in_idx in enumerate(perm):
+                if out_idx >= len(onnx_shape) or in_idx >= len(input_shape):
+                    continue
+                o_dim = _normalize_dim(onnx_shape[out_idx])
+                i_dim = input_shape[in_idx]
+                if isinstance(o_dim, int) and isinstance(i_dim, int) and o_dim == i_dim:
+                    score += o_dim
+            scored.append((score, 1 if perm == list(range(len(perm))) else 0, perm))
+        scored.sort(key=lambda x: (x[0], x[1]), reverse=True)
+        return [p for _, _, p in scored]
+
+    # Rescue guard for unexpected broadcasted shapes
+    if not enable_gelu:
+        expected_shape = None
+        if graph_node_output_shape is not None:
+            expected_shape = [_normalize_dim(dim) for dim in list(graph_node_output_shape)]
+        output_shape = _get_static_shape(tf_layers_dict[graph_node_output.name]['tf_node'])
+        input_shape_1 = _get_static_shape(input_tensor_1)
+        input_shape_2 = _get_static_shape(input_tensor_2)
+        if expected_shape is not None \
+            and output_shape is not None \
+            and not _shape_match_with_none(expected_shape, output_shape) \
+            and input_shape_1 is not None \
+            and input_shape_2 is not None \
+            and len(input_shape_1) == len(expected_shape) \
+            and len(input_shape_2) == len(expected_shape):
+
+            rank = len(expected_shape)
+            perms = _limited_perms(rank)
+            perm_list_1 = _ranked_perms(perms, input_shape_1, expected_shape)
+            perm_list_2 = _ranked_perms(perms, input_shape_2, expected_shape)
+            rescue_done = False
+            for perm_1 in perm_list_1:
+                for perm_2 in perm_list_2:
+                    try_input_1 = transpose_with_flexing_deterrence(
+                        input_tensor=input_tensor_1,
+                        perm=perm_1,
+                        **kwargs,
+                    )
+                    try_input_2 = transpose_with_flexing_deterrence(
+                        input_tensor=input_tensor_2,
+                        perm=perm_2,
+                        **kwargs,
+                    )
+                    try:
+                        rescue_tensor = tf.math.add(
+                            x=try_input_1 \
+                                if not isinstance(try_input_1, np.ndarray) \
+                                    else tf.convert_to_tensor(try_input_1),
+                            y=try_input_2 \
+                                if not isinstance(try_input_2, np.ndarray) \
+                                    else tf.convert_to_tensor(try_input_2),
+                            name=graph_node.name,
+                        )
+                    except Exception as ex:
+                        continue
+
+                    rescue_shape = _get_static_shape(rescue_tensor)
+                    if _shape_match_with_none(expected_shape, rescue_shape):
+                        input_tensor_1 = try_input_1
+                        input_tensor_2 = try_input_2
+                        tf_layers_dict[graph_node_output.name]['tf_node'] = rescue_tensor
+                        tf_type = tf.math.add
+                        rescue_done = True
+                        break
+                if rescue_done:
+                    break
+
     # Post-process transpose
     tf_layers_dict[graph_node_output.name]['tf_node'] = \
         post_process_transpose(

onnx2tf/ops/Concat.py

@@ -291,6 +291,78 @@ def make_node(
         tf_type = tf.constant
 
     else:
+        def _normalize_dim(dim):
+            return int(dim) if isinstance(dim, (int, np.integer)) else None
+
+        def _get_static_shape(tensor):
+            shape = getattr(tensor, 'shape', None)
+            if shape is None or shape == tf.TensorShape(None):
+                return None
+            return [_normalize_dim(dim) for dim in list(shape)]
+
+        def _shape_match_with_none(onnx_shape, tf_shape):
+            if onnx_shape is None or tf_shape is None:
+                return False
+            if len(onnx_shape) != len(tf_shape):
+                return False
+            for o_dim, t_dim in zip(onnx_shape, tf_shape):
+                o_dim = _normalize_dim(o_dim)
+                t_dim = _normalize_dim(t_dim)
+                if o_dim is None or t_dim is None:
+                    continue
+                if o_dim != t_dim:
+                    return False
+            return True
+
+        def _can_concat_shapes(shapes, axis):
+            if shapes is None or any(s is None for s in shapes):
+                return True
+            rank = len(shapes[0])
+            for idx in range(rank):
+                if idx == axis:
+                    continue
+                dims = [s[idx] for s in shapes]
+                known = [d for d in dims if isinstance(d, int)]
+                if len(known) >= 2 and len(set(known)) != 1:
+                    return False
+            return True
+
+        def _perm_shape(shape, perm):
+            return [shape[i] for i in perm] if shape is not None else None
+
+        def _limited_perms(rank):
+            identity = list(range(rank))
+            perms = [identity]
+            if rank == 3:
+                perms.append([0, 2, 1])
+            elif rank == 4:
+                perms.extend([[0, 2, 3, 1], [0, 3, 1, 2]])
+            elif rank == 5:
+                perms.extend([[0, 2, 3, 4, 1], [0, 4, 1, 2, 3]])
+            return perms
+
+        def _base_perms(rank):
+            if rank <= 1:
+                return [list(range(rank))]
+            return [list(p) for p in itertools.permutations(range(rank))]
+
+        def _ranked_perms(perms, input_shape, axis, onnx_shape):
+            identity = list(range(len(perms[0]))) if perms else []
+            scored = []
+            for perm in perms:
+                score = 0
+                if input_shape is not None and onnx_shape is not None:
+                    for out_idx, in_idx in enumerate(perm):
+                        if out_idx == axis:
+                            continue
+                        o_dim = _normalize_dim(onnx_shape[out_idx]) if out_idx < len(onnx_shape) else None
+                        i_dim = input_shape[in_idx] if in_idx < len(input_shape) else None
+                        if isinstance(o_dim, int) and isinstance(i_dim, int) and o_dim == i_dim:
+                            score += o_dim
+                scored.append((score, 1 if perm == identity else 0, perm))
+            scored.sort(key=lambda x: (x[0], x[1]), reverse=True)
+            return [p for _, _, p in scored]
+
         try:
             # normal concat attempt
             tf_layers_dict[graph_node_output.name]['tf_node'] = \
@@ -301,35 +373,109 @@ def make_node(
                 )
         except:
             # Workaround to reduce error rate when merging tensors with undefined dimensions
-            try:
-                # Attempts to bind with the axis specified in ONNX
-                tf_layers_dict[graph_node_output.name]['tf_node'] = \
-                    tf.concat(
-                        values=values,
-                        axis=int(graph_node.attrs.get('axis', 0)),
-                        name=graph_node.name,
-                    )
-                axis = int(graph_node.attrs.get('axis', 0))
-            except:
-                # If not successful with the same axis as ONNX, try to combine by other axes
-                # Trial in reverse order from the axis at the end
-                value_rank = len(values[0].shape)
-                succeed = False
-                for idx in reversed(range(value_rank)):
+            original_values = values
+            original_shapes = [_get_static_shape(v) for v in original_values]
+            value_rank = getattr(original_values[0].shape, 'rank', None)
+            if value_rank is None:
+                value_rank = len(original_values[0].shape)
+
+            onnx_shape_list = None
+            if onnx_output_shape is not None:
+                onnx_shape_list = [_normalize_dim(dim) for dim in list(onnx_output_shape)]
+
+            onnx_axis = int(graph_node.attrs.get('axis', 0))
+            onnx_axis = onnx_axis + value_rank if onnx_axis < 0 else onnx_axis
+
+            def _axis_score(axis_idx):
+                if onnx_shape_list is not None and axis_idx < len(onnx_shape_list):
+                    onnx_dim = onnx_shape_list[axis_idx]
+                    if isinstance(onnx_dim, int):
+                        return onnx_dim
+                score = 0
+                for shape in original_shapes:
+                    if shape is None or axis_idx >= len(shape):
+                        continue
+                    dim = shape[axis_idx]
+                    if isinstance(dim, int):
+                        score += dim
+                return score
+
+            axis_candidates = list(range(value_rank))
+            axis_candidates.sort(
+                key=lambda a: (_axis_score(a), 1 if a == onnx_axis else 0),
+                reverse=True,
+            )
+
+            base_perms = _base_perms(value_rank)
+            max_combo = 20000
+            if len(base_perms) ** len(original_values) > max_combo:
+                base_perms = _limited_perms(value_rank)
+
+            succeed = False
+            matched = False
+            chosen_axis = None
+            chosen_values = None
+            chosen_tensor = None
+
+            for axis_idx in axis_candidates:
+                perm_lists = [
+                    _ranked_perms(base_perms, shape, axis_idx, onnx_shape_list)
+                    for shape in original_shapes
+                ]
+                for perm_combo in itertools.product(*perm_lists):
+                    permuted_shapes = [
+                        _perm_shape(shape, perm) for shape, perm in zip(original_shapes, perm_combo)
+                    ]
+                    if not _can_concat_shapes(permuted_shapes, axis_idx):
+                        continue
+                    try_values = [
+                        value if perm == list(range(value_rank)) else
+                        transpose_with_flexing_deterrence(
+                            input_tensor=value,
+                            perm=perm,
+                            **kwargs,
+                        )
+                        for value, perm in zip(original_values, perm_combo)
+                    ]
                     try:
-                        tf_layers_dict[graph_node_output.name]['tf_node'] = \
+                        concat_tensor = \
                             tf.concat(
-                                values=values,
-                                axis=idx,
+                                values=try_values,
+                                axis=axis_idx,
                                 name=graph_node.name,
                             )
-                        axis = idx
+                    except:
+                        continue
+
+                    if not succeed:
                         succeed = True
+                        chosen_axis = axis_idx
+                        chosen_values = try_values
+                        chosen_tensor = concat_tensor
+
+                    if onnx_shape_list is None:
+                        matched = True
+                        chosen_axis = axis_idx
+                        chosen_values = try_values
+                        chosen_tensor = concat_tensor
                         break
-                    except:
-                        pass
-                if not succeed:
-                    raise
+
+                    output_shape = _get_static_shape(concat_tensor)
+                    if _shape_match_with_none(onnx_shape_list, output_shape):
+                        matched = True
+                        chosen_axis = axis_idx
+                        chosen_values = try_values
+                        chosen_tensor = concat_tensor
+                        break
+                if matched:
+                    break
+
+            if succeed:
+                tf_layers_dict[graph_node_output.name]['tf_node'] = chosen_tensor
+                axis = chosen_axis
+                values = chosen_values
+            else:
+                raise
 
         # Attempts to force axis correction when the number of axes in the combined tensor do not exactly match.
         # However, if more than 2 patterns of correct answers exist, give up the correction.

onnx2tf/ops/ImageDecoder.py

@@ -0,0 +1,147 @@
+import random
+random.seed(0)
+import numpy as np
+np.random.seed(0)
+import tensorflow as tf
+import onnx_graphsurgeon as gs
+import cv2
+from onnx2tf.utils.common_functions import (
+    get_constant_or_variable,
+    print_node_info,
+    inverted_operation_enable_disable,
+    make_tf_node_info,
+    get_replacement_parameter,
+    pre_process_transpose,
+    post_process_transpose,
+)
+from onnx2tf.utils.logging import *
+
+
+def _as_tensor(value):
+    if isinstance(value, np.ndarray):
+        return tf.convert_to_tensor(value)
+    if isinstance(value, (np.generic, int, float, bool, str, bytes)):
+        return tf.convert_to_tensor(value)
+    return value
+
+
+def _decode_image_np(encoded_stream, pixel_format):
+    if encoded_stream is None:
+        return np.zeros((0, 0, 0), dtype=np.uint8)
+    if encoded_stream.dtype != np.uint8:
+        encoded_stream = encoded_stream.astype(np.uint8)
+    if encoded_stream.size == 0:
+        return np.zeros((0, 0, 0), dtype=np.uint8)
+    if encoded_stream.ndim != 1:
+        encoded_stream = encoded_stream.reshape(-1)
+    try:
+        if pixel_format == 'Grayscale':
+            flag = cv2.IMREAD_GRAYSCALE
+        else:
+            flag = cv2.IMREAD_COLOR
+        decoded = cv2.imdecode(encoded_stream, flag)
+        if decoded is None:
+            raise ValueError('cv2.imdecode failed')
+        if pixel_format == 'RGB':
+            decoded = cv2.cvtColor(decoded, cv2.COLOR_BGR2RGB)
+        if pixel_format == 'Grayscale' and decoded.ndim == 2:
+            decoded = decoded[..., np.newaxis]
+        return decoded.astype(np.uint8)
+    except Exception:
+        if pixel_format == 'Grayscale':
+            return np.zeros((0, 0, 1), dtype=np.uint8)
+        return np.zeros((0, 0, 3), dtype=np.uint8)
+
+
+@print_node_info
+@inverted_operation_enable_disable
+@get_replacement_parameter
+def make_node(
+    *,
+    graph_node: gs.Node,
+    tf_layers_dict: dict,
+    **kwargs: dict,
+):
+    """ImageDecoder
+
+    Parameters
+    ----------
+    graph_node: gs.Node
+        graph_surgeon Node
+
+    tf_layers_dict: dict
+        optype, shape, dtype, tensorflow graph
+    """
+    before_op_output_shape_trans_1 = \
+        tf_layers_dict.get(graph_node.inputs[0].name, {}).get('before_op_output_shape_trans', True)
+    before_op_output_shape_trans = \
+        before_op_output_shape_trans_1
+
+    graph_node_input = get_constant_or_variable(
+        graph_node.inputs[0],
+        before_op_output_shape_trans,
+    )
+    graph_node_output: gs.Variable = graph_node.outputs[0]
+    shape = graph_node_output.shape
+    dtype = graph_node_output.dtype
+
+    input_tensor = tf_layers_dict[graph_node_input.name]['tf_node'] \
+        if isinstance(graph_node_input, gs.Variable) else graph_node_input
+
+    # Preserving Graph Structure (Dict)
+    tf_layers_dict[graph_node_output.name] = {
+        'optype': graph_node.op,
+        'shape': shape,
+        'dtype': dtype,
+    }
+
+    # Pre-process transpose
+    input_tensor = pre_process_transpose(
+        value_before_transpose=input_tensor,
+        param_target='inputs',
+        param_name=graph_node.inputs[0].name,
+        **kwargs,
+    )
+
+    # Generation of TF OP
+    input_tensor = _as_tensor(input_tensor)
+    pixel_format = graph_node.attrs.get('pixel_format', 'RGB')
+    if pixel_format not in ['RGB', 'BGR', 'Grayscale']:
+        error(
+            f'ImageDecoder pixel_format={pixel_format} is not supported.\n' +
+            f'graph_node.name: {graph_node.name}'
+        )
+        pixel_format = 'RGB'
+
+    decoded = tf.numpy_function(
+        func=lambda x: _decode_image_np(x, pixel_format),
+        inp=[input_tensor],
+        Tout=tf.uint8,
+        name=graph_node.name,
+    )
+    channels = 1 if pixel_format == 'Grayscale' else 3
+    decoded = tf.ensure_shape(decoded, [None, None, channels])
+    tf_layers_dict[graph_node_output.name]['tf_node'] = decoded
+
+    # Post-process transpose
+    tf_layers_dict[graph_node_output.name]['tf_node'] = post_process_transpose(
+        value_before_transpose=tf_layers_dict[graph_node_output.name]['tf_node'],
+        param_target='outputs',
+        param_name=graph_node.outputs[0].name,
+        **kwargs,
+    )
+
+    # Generation of Debug Info
+    tf_layers_dict[graph_node_output.name]['tf_node_info'] = \
+        make_tf_node_info(
+            node_info={
+                'tf_op_type': 'ImageDecoder',
+                'tf_inputs': {
+                    'encoded_stream': input_tensor,
+                    'pixel_format': pixel_format,
+                },
+                'tf_outputs': {
+                    'output': tf_layers_dict[graph_node_output.name]['tf_node'],
+                },
+            }
+        )