File size: 6,832 Bytes
1f1b5fd
 
 
 
650fcdf
1f1b5fd
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
650fcdf
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1f1b5fd
 
 
 
 
 
 
 
650fcdf
 
 
 
 
1f1b5fd
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
650fcdf
 
 
 
 
 
 
 
 
 
 
 
 
1f1b5fd
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
import onnxscript
import onnx_ir as ir
import onnx_ir.passes.common
import numpy as np
import onnxslim


class ReplaceDftWithMatMulRule(onnxscript.rewriter.RewriteRuleClassBase):
    def pattern(self, op, x, dft_length):
        x = op.Reshape(x, _allow_other_inputs=True)
        dft = op.DFT(x, dft_length, _outputs=["dft_output"])
        real_part = op.Slice(dft, [0], [1], [-1])
        return op.Squeeze(real_part, [-1])

    def rewrite(self, op, x: ir.Value, dft_length: ir.Value, dft_output: ir.Value):
        # Get the DFT node attributes
        dft_node = dft_output.producer()
        assert dft_node is not None

        dft_size = ir.convenience.get_const_tensor(dft_length).numpy().item()

        # Create one-sided DFT matrix (only real part, DC to Nyquist)
        # The real part of DFT is: Re(DFT[k]) = sum(x[n] * cos(2*pi*k*n/N))
        # For one-sided DFT, we only need frequencies from 0 to Nyquist (dft_size//2 + 1)
        num_freqs = dft_size // 2 + 1

        # Vectorized creation of DFT matrix
        n = np.arange(dft_size, dtype=np.float32)[:, np.newaxis]  # Shape: (dft_size, 1)
        k = np.arange(num_freqs, dtype=np.float32)[
            np.newaxis, :
        ]  # Shape: (1, num_freqs)
        dft_matrix = np.cos(
            2 * np.pi * k * n / dft_size
        )  # Shape: (dft_size, num_freqs)

        # Create constant node for the DFT matrix
        dft_matrix = op.initializer(ir.tensor(dft_matrix), name=f"{x.name}_dft_matrix")

        # DFT axis is already at the end, direct matrix multiplication
        result = op.MatMul(x, dft_matrix)

        return result


class ReplaceSplit(onnxscript.rewriter.RewriteRuleClassBase):
    def pattern(self, op, x):
        return op.Split(
            x, _allow_other_inputs=True, _outputs=["split_out_1", "split_out_2"]
        )

    def rewrite(self, op, x: ir.Value, **kwargs):
        zero = op.initializer(ir.tensor(np.array([0], dtype=np.int64)), "zero")
        batch_size = op.Gather(x, zero)
        sample_size = op.initializer(
            ir.tensor(np.array([144000], dtype=np.int32)), "sample_size"
        )
        return batch_size, sample_size


class RemoveCast(onnxscript.rewriter.RewriteRuleClassBase):
    def pattern(self, op, x):
        return op.Cast(x)

    def rewrite(self, op, x: ir.Value, **kwargs):
        return op.Identity(x)


class RemoveReversedSequenceFork(onnxscript.rewriter.RewriteRuleClassBase):
    def pattern(self, op, x, y, scale, bias):
        x = op.Transpose(x)
        y = op.Transpose(y)
        x = op.ReverseSequence(x, _allow_other_inputs=True)
        y = op.ReverseSequence(y, _allow_other_inputs=True)
        x = op.Unsqueeze(x, _allow_other_inputs=True)
        y = op.Unsqueeze(y, _allow_other_inputs=True)
        concat = op.Concat(x, y)
        mul = op.Mul(concat, scale)
        add = op.Add(mul, bias)
        return op.Transpose(add)

    def rewrite(self, op, x, y, scale, bias, **kwargs):
        # x: batch, 511, 96
        neg_one = op.initializer(ir.tensor(np.array([-1], dtype=np.int64)), "neg_one")
        int_64_min = op.initializer(
            ir.tensor(np.array([-9223372036854775808], dtype=np.int64)), "int_64_min"
        )
        # slice
        x = op.Slice(x, neg_one, int_64_min, neg_one, neg_one)
        y = op.Slice(y, neg_one, int_64_min, neg_one, neg_one)
        x = op.Unsqueeze(x, neg_one)
        y = op.Unsqueeze(y, neg_one)
        concat = op.Concat(x, y, axis=3)
        # batch, 511, 96, 2
        mul = op.Mul(concat, scale)
        add = op.Add(mul, bias)
        return op.Transpose(add, perm=[0, 3, 2, 1])  # batch, 2, 96, 511


model = ir.load("model.onnx")

# Set dynamic axes
model.graph.inputs[0].shape = ir.Shape(["batch", 144000])
model.graph.outputs[0].shape = ir.Shape(["batch", 6522])

onnxscript.rewriter.rewrite(
    model,
    [
        ReplaceDftWithMatMulRule().rule(),
        ReplaceSplit().rule(),
        RemoveCast().rule(),
    ],
)

# Change all int32 initializers to int64
initializers = list(model.graph.initializers.values())
for initializer in initializers:
    if initializer.dtype == ir.DataType.INT32:
        int32_array = initializer.const_value.numpy()
        int64_array = int32_array.astype(np.int64)
        new_initializer = ir.val(initializer.name, const_value=ir.tensor(int64_array))
        model.graph.initializers.pop(initializer.name)
        model.graph.initializers.add(new_initializer)
        initializer.replace_all_uses_with(new_initializer)

onnxscript.optimizer.optimize(
    model, input_size_limit=1024 * 1024 * 1024, output_size_limit=1024 * 1024 * 1024
)


# Remove Slice-Reshape
def remove_slice_reshape(model: ir.Model):
    mul_node = model.graph.node("model/MEL_SPEC1/Mul")
    first_reshape = model.graph.node("model/MEL_SPEC1/stft/frame/Reshape_1")
    first_shape = ir.val(
        "first_shape", const_value=ir.tensor([-1, 72000, 2], dtype=ir.DataType.INT64)
    )
    model.graph.initializers.add(first_shape)
    second_reshape = model.graph.node("model/MEL_SPEC2/stft/frame/Reshape_1")
    second_shape = ir.val(
        "second_shape", const_value=ir.tensor([-1, 18000, 8], dtype=ir.DataType.INT64)
    )
    model.graph.initializers.add(second_shape)

    third_reshape = model.graph.node("model/MEL_SPEC1/stft/frame/Reshape_4")
    third_shape = ir.val(
        "third_shape", const_value=ir.tensor([-1, 511, 2048], dtype=ir.DataType.INT64)
    )
    model.graph.initializers.add(third_shape)
    fourth_reshape = model.graph.node("model/MEL_SPEC2/stft/frame/Reshape_4")
    fourth_shape = ir.val(
        "fourth_shape", const_value=ir.tensor([-1, 511, 1024], dtype=ir.DataType.INT64)
    )
    model.graph.initializers.add(fourth_shape)

    # Replace with Mul-Reshape-Gather
    first_reshape.replace_input_with(0, mul_node.outputs[0])
    first_reshape.replace_input_with(1, first_shape)
    second_reshape.replace_input_with(0, mul_node.outputs[0])
    second_reshape.replace_input_with(1, second_shape)
    third_reshape.replace_input_with(1, third_shape)
    fourth_reshape.replace_input_with(1, fourth_shape)


remove_slice_reshape(model)
# Run DCE again
onnxscript.optimizer.optimize(
    model, input_size_limit=1024 * 1024 * 1024, output_size_limit=1024 * 1024 * 1024
)

print("Slimming model...")
model = ir.from_proto(onnxslim.slim(ir.to_proto(model)))

print("Removing reversed sequence fork...")
onnxscript.rewriter.rewrite(
    model,
    [
        RemoveReversedSequenceFork.rule(),
    ],
)

# Use onnxslim to do shape inference
model = ir.from_proto(onnxslim.slim(ir.to_proto(model)))

onnx_ir.passes.common.ClearMetadataAndDocStringPass()(model)
model.graph.inputs[0].name = "input"
model.graph.outputs[0].name = "output"
model.ir_version = 10
model.producer_name = "onnx-ir"
model.graph.name = "BirdNET-v2.4"

ir.save(model, "birdnet.onnx")