onnx2tf 1.29.15__py3-none-any.whl → 1.29.17__py3-none-any.whl

onnx2tf/ops/DequantizeLinear.py

@@ -15,6 +15,43 @@ from onnx2tf.utils.common_functions import (
     post_process_transpose,
 )
 
+def _expand_scale_or_zero_point(
+    *,
+    value,
+    input_tensor,
+    axis: int,
+    block_size: int,
+):
+    value_rank = len(value.shape)
+    input_rank = len(input_tensor.shape)
+
+    if value_rank == 0:
+        return value
+
+    if input_rank <= 0:
+        return value
+
+    if axis < 0 or axis >= input_rank:
+        axis = 0
+
+    # Blocked quantization: expand along axis then slice to input shape
+    if block_size > 0 and value_rank == input_rank:
+        if value.shape[axis] is None \
+            or input_tensor.shape[axis] is None \
+            or value.shape[axis] != input_tensor.shape[axis]:
+            expanded = tf.repeat(value, repeats=block_size, axis=axis)
+            expanded = tf.slice(expanded, [0] * input_rank, tf.shape(input_tensor))
+            return expanded
+        return value
+
+    # Per-axis quantization: reshape 1-D to broadcast
+    if value_rank == 1 and input_rank is not None:
+        shape = [1] * input_rank
+        shape[axis] = -1
+        return tf.reshape(value, shape)
+
+    return value
+
 
 @print_node_info
 @inverted_operation_enable_disable
@@ -63,6 +100,11 @@ def make_node(
 
     input_tensor = tf_layers_dict[graph_node_input_1.name]['tf_node'] \
         if isinstance(graph_node_input_1, gs.Variable) else graph_node_input_1
+    input_is_dequantized = False
+    input_nhwc = False
+    if isinstance(graph_node_input_1, gs.Variable):
+        input_is_dequantized = tf_layers_dict.get(graph_node_input_1.name, {}).get('is_dequantized', False)
+        input_nhwc = tf_layers_dict.get(graph_node_input_1.name, {}).get('nhwc', False)
 
     # Pre-process transpose
     input_tensor = pre_process_transpose(
@@ -72,12 +114,10 @@ def make_node(
         **kwargs,
     )
 
-    input_tensor_shape = input_tensor.shape
-    input_tensor_rank = len(input_tensor_shape)
+    input_tensor_rank = len(input_tensor.shape)
+    input_tensor_dtype = input_tensor.dtype
     x_scale = tf_layers_dict[graph_node_input_2.name]['tf_node'] \
         if isinstance(graph_node_input_2, gs.Variable) else graph_node_input_2
-    x_scale_shape = x_scale.shape
-    x_scale_rank = len(x_scale_shape)
     x_zero_point = tf_layers_dict[graph_node_input_3.name]['tf_node'] \
         if isinstance(graph_node_input_3, gs.Variable) else graph_node_input_3
 
@@ -87,48 +127,50 @@ def make_node(
         tensor_rank=input_tensor_rank,
         before_op_output_shape_trans=before_op_output_shape_trans,
     )
+    if input_tensor_rank == 1:
+        axis = 0
 
     # Preserving Graph Structure (Dict)
     tf_layers_dict[graph_node_output.name] = {
         'optype': graph_node.op,
         'shape': shape,
         'dtype': dtype,
+        'is_dequantized': True,
+        'nhwc': input_nhwc,
     }
 
     # Generation of TF OP
 
     input_tensor = tf.cast(input_tensor, tf.float32)
+    x_scale = tf.cast(x_scale, tf.float32)
 
-    # Reshape process is needed for per-axis dequantization
-    # when scale is a 1-D tensor
-    if x_scale_rank == 1 and x_scale_shape[0] != 1:
-        shape_broadcast = list([1 for _ in range(axis)] + [input_tensor_shape[axis]] + [1 for _ in range(axis + 1, input_tensor_rank)])
-        x_scale = tf.reshape(
-            tensor=x_scale,
-            shape=shape_broadcast,
-        )
-    elif x_scale_rank == 1 and x_scale_shape[0] == 1:
-        shape_broadcast = [1 for i in range(input_tensor_rank)]
-
-    subed_tensor = input_tensor
-    if len(graph_node.inputs) >= 3 and input_tensor.dtype != tf.int32:
-        x_zero_point = tf.cast(
-            x=x_zero_point,
-            dtype=tf.float32,
-        )
-        x_zero_point = tf.reshape(
-            tensor=x_zero_point,
-            shape=shape_broadcast,
-        ) if x_scale_rank == 1 else x_zero_point
-        subed_tensor = tf.subtract(
-            x=input_tensor,
-            y=x_zero_point,
-        )
-    tf_layers_dict[graph_node_output.name]['tf_node'] = \
-        tf.multiply(
-            x=subed_tensor,
-            y=x_scale,
-        )
+    block_size = int(graph_node.attrs.get('block_size', 0))
+    x_scale = _expand_scale_or_zero_point(
+        value=x_scale,
+        input_tensor=input_tensor,
+        axis=axis,
+        block_size=block_size,
+    )
+
+    if input_is_dequantized:
+        tf_layers_dict[graph_node_output.name]['tf_node'] = input_tensor
+    else:
+        if x_zero_point is None or input_tensor_dtype == tf.int32:
+            x_zero_point = tf.zeros_like(x_scale)
+        else:
+            x_zero_point = tf.cast(x_zero_point, tf.float32)
+            x_zero_point = _expand_scale_or_zero_point(
+                value=x_zero_point,
+                input_tensor=input_tensor,
+                axis=axis,
+                block_size=block_size,
+            )
+
+        tf_layers_dict[graph_node_output.name]['tf_node'] = \
+            tf.multiply(
+                x=tf.subtract(input_tensor, x_zero_point),
+                y=x_scale,
+            )
 
     if hasattr(tf_layers_dict[graph_node_output.name]['tf_node'], 'numpy'):
         tf_layers_dict[graph_node_output.name]['tf_node'] = \

onnx2tf/ops/DynamicQuantizeLinear.py

@@ -43,6 +43,9 @@ def make_node(
         graph_node.inputs[0],
         before_op_output_shape_trans,
     )
+    input_nhwc = False
+    if isinstance(graph_node_input_1, gs.Variable):
+        input_nhwc = tf_layers_dict.get(graph_node_input_1.name, {}).get('nhwc', False)
     graph_node_output_1: gs.Variable = graph_node.outputs[0]
     o1_shape = graph_node_output_1.shape
     o1_dtype = graph_node_output_1.dtype
@@ -58,6 +61,8 @@ def make_node(
         'optype': graph_node.op,
         'shape': o1_shape,
         'dtype': o1_dtype,
+        'is_dequantized': True,
+        'nhwc': input_nhwc,
     }
     tf_layers_dict[graph_node_output_2.name] = {
         'optype': graph_node.op,
@@ -82,35 +87,31 @@ def make_node(
     )
 
     # Generation of TF OP
-    dtype = tf.uint8
-    qmin = dtype.min
-    qmax = dtype.max
-    min_x = tf.math.minimum(0., tf.math.reduce_min(input_tensor_1))
-    max_x = tf.math.maximum(0., tf.math.reduce_max(input_tensor_1))
+    qmin = 0.0
+    qmax = 255.0
+    min_x = tf.math.minimum(0.0, tf.math.reduce_min(input_tensor_1))
+    max_x = tf.math.maximum(0.0, tf.math.reduce_max(input_tensor_1))
     y_scale = (max_x - min_x) / (qmax - qmin)
     intermediate_zero_point = qmin - (min_x / y_scale)
-    y_zero_point = tf.clip_by_value(
-        tf.round(
-            x=intermediate_zero_point
-        ),
+    clipped_zero_point = tf.clip_by_value(
+        intermediate_zero_point,
         clip_value_min=qmin,
         clip_value_max=qmax,
     )
-    y = tf.cast(
-        tf.clip_by_value(
-            (tf.round(input_tensor_1 / y_scale) + y_zero_point),
-            clip_value_min=qmin,
-            clip_value_max=qmax,
-        ),
-        dtype=dtype,
+    y_zero_point = tf.round(clipped_zero_point)
+    y_quant = tf.clip_by_value(
+        tf.round(input_tensor_1 / y_scale) + y_zero_point,
+        clip_value_min=qmin,
+        clip_value_max=qmax,
     )
+    y = (y_quant - y_zero_point) * y_scale
 
     tf_layers_dict[graph_node_output_1.name]['tf_node'] = y
     tf_layers_dict[graph_node_output_2.name]['tf_node'] = y_scale
     tf_layers_dict[graph_node_output_3.name]['tf_node'] = \
         tf.cast(
             x=y_zero_point,
-            dtype=dtype,
+            dtype=tf.uint8,
         )
 
     # Post-process transpose

onnx2tf/ops/QLinearConcat.py

@@ -10,6 +10,10 @@ from onnx2tf.utils.common_functions import (
     print_node_info,
     inverted_operation_enable_disable,
     make_tf_node_info,
+    pre_process_transpose,
+    replace_parameter,
+    shape_is_equal_ignore_order,
+    transpose_with_flexing_deterrence,
 )
 
 
@@ -35,8 +39,7 @@ def make_node(
     y_zero_point_list = [i for i in graph_node.inputs[1::3]]
     input_list = [i for i in graph_node.inputs[2::3]]
 
-    input_tensor_shape = input_list[0].shape
-    input_tensor_rank = len(input_tensor_shape)
+    input_tensor_rank = len(input_list[0].shape)
 
     before_op_output_shape_trans = True
     for graph_node_input in input_list:
@@ -46,6 +49,9 @@ def make_node(
             before_op_output_shape_trans and before_op_output_shape_trans_n
 
     got_values = []
+    nhwc_flags = []
+    same_input_shape_as_onnxs = []
+    input_is_dequantized_list = []
     got_y_scale_list = []
     got_y_zero_point_list = []
     for input, y_scale, y_zero_point  in zip(input_list, y_scale_list, y_zero_point_list):
@@ -55,8 +61,24 @@ def make_node(
         )
         if isinstance(const_or_var, gs.Variable):
             got_values.append(tf_layers_dict[const_or_var.name]['tf_node'])
+            nhwc_flags.append(
+                tf_layers_dict[const_or_var.name].get('nhwc', False)
+            )
+            same_input_shape_as_onnxs.append(
+                True if input.shape is not None and len(input.shape) > 0 \
+                    and input.shape == tf_layers_dict[const_or_var.name]['tf_node'].shape else False
+            )
+            input_is_dequantized_list.append(
+                tf_layers_dict[const_or_var.name].get('is_dequantized', False)
+            )
         else:
             got_values.append(const_or_var)
+            nhwc_flags.append(False)
+            same_input_shape_as_onnxs.append(
+                True if input.shape is not None and len(input.shape) > 0 \
+                    and input.shape == const_or_var.shape else False
+            )
+            input_is_dequantized_list.append(False)
 
         const_or_var = get_constant_or_variable(
             y_scale,
@@ -82,50 +104,247 @@ def make_node(
     dtype = graph_node_output.dtype
 
     axis = graph_node.attrs.get('axis', 0)
-    # NCHW->NHWC, NCDHW->NDHWC
-    axis = convert_axis(
-        axis=axis,
-        tensor_rank=len(shape) if shape is not None else input_tensor_rank,
-        before_op_output_shape_trans=before_op_output_shape_trans,
-    )
+
+    # Shape Unmatched Special Avoidance Workaround
+    if True in same_input_shape_as_onnxs and True in nhwc_flags:
+        before_op_output_shape_trans = True
+        new_values = []
+        for same_input_shape_as_onnx, nhwc_flag, value in zip(same_input_shape_as_onnxs, nhwc_flags, got_values):
+            if same_input_shape_as_onnx and not nhwc_flag:
+                if len(value.shape) == 3:
+                    new_values.append(
+                        transpose_with_flexing_deterrence(
+                            input_tensor=value,
+                            perm=[0, 2, 1],
+                            **kwargs,
+                        )
+                    )
+                elif len(value.shape) == 4:
+                    new_values.append(
+                        transpose_with_flexing_deterrence(
+                            input_tensor=value,
+                            perm=[0, 2, 3, 1],
+                            **kwargs,
+                        )
+                    )
+                elif len(value.shape) == 5:
+                    new_values.append(
+                        transpose_with_flexing_deterrence(
+                            input_tensor=value,
+                            perm=[0, 2, 3, 4, 1],
+                            **kwargs,
+                        )
+                    )
+                else:
+                    new_values.append(value)
+            else:
+                new_values.append(value)
+        got_values = new_values
 
     # Preserving Graph Structure (Dict)
+    nhwc_judge = True
+    for graph_node_input in input_list:
+        if isinstance(graph_node_input, gs.Variable) \
+            and tf_layers_dict.get(graph_node_input.name, {}).get('nhwc', False):
+            nhwc_judge = nhwc_judge and True
+        elif isinstance(graph_node_input, gs.Constant) \
+            and hasattr(graph_node_input, 'values') \
+            and isinstance(graph_node_input.values, np.ndarray):
+            nhwc_judge = nhwc_judge or False
+        else:
+            nhwc_judge = nhwc_judge and False
+
     tf_layers_dict[graph_node_output.name] = {
         'optype': graph_node.op,
         'shape': shape,
         'dtype': dtype,
+        'is_dequantized': True,
     }
+    if nhwc_judge:
+        tf_layers_dict[graph_node_output.name]['nhwc'] = True
 
     # Generation of TF OP
 
+    # NCHW->NHWC, NCDHW->NDHWC
+    axis = convert_axis(
+        axis=axis,
+        tensor_rank=len(shape) if shape is not None else input_tensor_rank,
+        before_op_output_shape_trans=before_op_output_shape_trans,
+    )
+
+    # Param replacement
+    before_axis = axis
+    axis = replace_parameter(
+        value_before_replacement=axis,
+        param_target='attributes',
+        param_name='axis',
+        **kwargs,
+    )
+
     # TensorFlow does not support Concat for scalar values, so convert to tensor
-    values = [
-        value if len(value.shape) > 0 else tf.reshape(value, [1]) for value in got_values
-    ]
+    values = []
+    for graph_node_input, value in zip(input_list, got_values):
+        value = pre_process_transpose(
+            value_before_transpose=value,
+            param_target='inputs',
+            param_name=graph_node_input.name,
+            **kwargs,
+        )
+        values.append(value if len(value.shape) > 0 else tf.reshape(value, [1]))
+
+    def _infer_concat_axis(values, output_shape):
+        if not values:
+            return None
+        ranks = []
+        shapes = []
+        for val in values:
+            if val.shape is None or val.shape == tf.TensorShape(None):
+                return None
+            shape_list = list(val.shape)
+            ranks.append(len(shape_list))
+            shapes.append(shape_list)
+        if len(set(ranks)) != 1:
+            return None
+        rank = ranks[0]
+        candidates = []
+        for ax in range(rank):
+            ok = True
+            for dim in range(rank):
+                if dim == ax:
+                    continue
+                base = shapes[0][dim]
+                for s in shapes[1:]:
+                    if base is None or s[dim] is None:
+                        continue
+                    if base != s[dim]:
+                        ok = False
+                        break
+                if not ok:
+                    break
+            if not ok:
+                continue
+            if output_shape is not None and len(output_shape) == rank:
+                out_dim = output_shape[ax]
+                if out_dim is not None:
+                    sum_dim = 0
+                    for s in shapes:
+                        if s[ax] is None:
+                            sum_dim = None
+                            break
+                        sum_dim += s[ax]
+                    if sum_dim is None or sum_dim != out_dim:
+                        continue
+            candidates.append(ax)
+        if len(candidates) == 1:
+            return candidates[0]
+        return None
+
+    inferred_axis = _infer_concat_axis(values, shape if shape is not None else None)
+    if inferred_axis is not None:
+        axis = inferred_axis
     # cast all inputs to float32
     casted_x_list = []
     casted_y_zero_point_list = []
-    for x, y_zero_point in zip(values, got_y_zero_point_list):
+    casted_y_scale_list = []
+    for x, y_scale, y_zero_point in zip(values, got_y_scale_list, got_y_zero_point_list):
         casted_x_list.append(tf.cast(x, tf.float32))
+        casted_y_scale_list.append(tf.cast(y_scale, tf.float32))
         casted_y_zero_point_list.append(tf.cast(y_zero_point, tf.float32))
     # dequantize x with y_scale, y_zero_point
     dequantized_x_list = []
-    for x, y_scale, y_zero_point in zip(casted_x_list, got_y_scale_list, casted_y_zero_point_list):
-        dequantized_value = tf.add(
-            x=tf.divide(
-                x=x,
+    for x, y_scale, y_zero_point, is_dequantized in zip(
+        casted_x_list,
+        casted_y_scale_list,
+        casted_y_zero_point_list,
+        input_is_dequantized_list,
+    ):
+        if is_dequantized:
+            dequantized_x_list.append(x)
+        else:
+            dequantized_value = tf.multiply(
+                x=tf.subtract(x, y_zero_point),
                 y=y_scale,
-            ),
-            y=y_zero_point,
-        )
-        dequantized_x_list.append(dequantized_value)
+            )
+            dequantized_x_list.append(dequantized_value)
 
-    tf_layers_dict[graph_node_output.name]['tf_node'] = \
-        tf.concat(
-            values=dequantized_x_list,
-            axis=axis,
-            name=graph_node.name,
-        )
+    try:
+        tf_layers_dict[graph_node_output.name]['tf_node'] = \
+            tf.concat(
+                values=dequantized_x_list,
+                axis=axis,
+                name=graph_node.name,
+            )
+    except:
+        try:
+            onnx_axis = int(graph_node.attrs.get('axis', 0))
+            tf_layers_dict[graph_node_output.name]['tf_node'] = \
+                tf.concat(
+                    values=dequantized_x_list,
+                    axis=onnx_axis,
+                    name=graph_node.name,
+                )
+            axis = onnx_axis
+        except:
+            value_rank = len(dequantized_x_list[0].shape)
+            succeed = False
+            for idx in reversed(range(value_rank)):
+                try:
+                    tf_layers_dict[graph_node_output.name]['tf_node'] = \
+                        tf.concat(
+                            values=dequantized_x_list,
+                            axis=idx,
+                            name=graph_node.name,
+                        )
+                    axis = idx
+                    succeed = True
+                    break
+                except:
+                    pass
+            if not succeed:
+                raise
+
+    output_tensor_shape = tf_layers_dict[graph_node_output.name]['tf_node'].shape
+    if output_tensor_shape != tf.TensorShape(None):
+        output_tensor_rank = len(output_tensor_shape)
+        if graph_node.outputs[0].shape is not None \
+            and axis != 0 \
+            and output_tensor_rank >= 2 \
+            and before_axis == axis:
+            if not shape_is_equal_ignore_order(list(graph_node.outputs[0].shape), list(output_tensor_shape)):
+                matched_axes = []
+                for dummy_axis in range(1, output_tensor_rank):
+                    try:
+                        dummy_concat_tensor = \
+                            tf.concat(
+                                values=dequantized_x_list,
+                                axis=dummy_axis,
+                                name=graph_node.name,
+                            )
+                        dummy_output_shape = dummy_concat_tensor.shape
+                        if shape_is_equal_ignore_order(list(graph_node.outputs[0].shape), list(dummy_output_shape)):
+                            matched_axes.append(dummy_axis)
+                    except:
+                        pass
+                if len(matched_axes) == 1:
+                    tf_layers_dict[graph_node_output.name]['tf_node'] = \
+                        tf.concat(
+                            values=dequantized_x_list,
+                            axis=matched_axes[0],
+                            name=graph_node.name,
+                        )
+                    axis = matched_axes[0]
+                elif not nhwc_judge:
+                    onnx_axis = int(graph_node.attrs.get('axis', 0))
+                    onnx_axis = output_tensor_rank - 1 if onnx_axis == -1 else onnx_axis
+                    if onnx_axis == output_tensor_rank - 1 \
+                        and onnx_axis in matched_axes:
+                        tf_layers_dict[graph_node_output.name]['tf_node'] = \
+                            tf.concat(
+                                values=dequantized_x_list,
+                                axis=onnx_axis,
+                                name=graph_node.name,
+                            )
+                        axis = onnx_axis
 
     # Generation of Debug Info
     tf_inputs = {f"input{idx}": dequantized_x for idx, dequantized_x in enumerate(dequantized_x_list)}