File size: 4,652 Bytes
d69b6f3
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
import torch
import torch.nn as nn
import torch.nn.functional as F
import math
from huggingface_hub import PyTorchModelHubMixin


class ResBlock1D(nn.Module):
    """
    Residual Block for extracting rhythmic features from audio spectrograms.
    Maintains temporal resolution while increasing receptive field.
    """

    def __init__(self, channels, kernel_size=3, dilation=1):
        super().__init__()
        padding = (kernel_size - 1) * dilation // 2
        self.conv1 = nn.Conv1d(
            channels, channels, kernel_size, padding=padding, dilation=dilation
        )
        self.bn1 = nn.BatchNorm1d(channels)
        self.conv2 = nn.Conv1d(
            channels, channels, kernel_size, padding=padding, dilation=dilation
        )
        self.bn2 = nn.BatchNorm1d(channels)

    def forward(self, x):
        res = x
        x = F.gelu(self.bn1(self.conv1(x)))
        x = self.bn2(self.conv2(x))
        return F.gelu(x + res)


class GameChartEvaluator(nn.Module, PyTorchModelHubMixin):
    def __init__(self, input_dim=80, d_model=128, n_layers=4):
        super().__init__()

        # --- Early Fusion ---
        # Input is (Batch, 80 * 2, Time)
        # We stack Music (80) + Chart (80) = 160 channels
        self.input_proj = nn.Conv1d(
            input_dim * 2, d_model, kernel_size=3, stride=1, padding=1
        )

        # --- STRICT TEMPORAL ENCODER ---
        # No Pooling (stride=1) to preserve 11ms resolution
        # Dilations allow seeing context without losing resolution
        self.encoder = nn.Sequential(
            ResBlock1D(d_model, kernel_size=3, dilation=1),
            ResBlock1D(d_model, kernel_size=3, dilation=2),
            ResBlock1D(d_model, kernel_size=3, dilation=4),
            ResBlock1D(d_model, kernel_size=3, dilation=8),
            # Add more layers if you need wider context (e.g. 16, 32)
        )

        # --- SCORING HEAD ---
        # Simple projection to scalar
        self.quality_proj = nn.Linear(d_model, 1)

        # Learnable Mixing
        self.raw_severity = nn.Parameter(torch.tensor(0.0))

    def forward(self, music_mels, chart_mels):
        """
        music_mels: (Batch, 80, Time)
        chart_mels: (Batch, 80, Time)
        """
        # 1. Early Fusion: Concatenate along Channel dimension
        # Shape becomes (Batch, 160, Time)
        x = torch.cat([music_mels, chart_mels], dim=1)

        # 2. Extract Features (Strictly Local + Context)
        x = F.gelu(self.input_proj(x))
        x = self.encoder(x)

        # 3. Predict Score per Frame
        # (Batch, Dim, Time) -> (Batch, Time, Dim)
        x = x.permute(0, 2, 1)
        local_scores = torch.sigmoid(self.quality_proj(x))  # (Batch, Time, 1)

        # 4. Error-Sensitive Pooling
        avg_score = local_scores.mean(dim=1)

        k = max(1, int(local_scores.size(1) * 0.1))
        min_vals, _ = torch.topk(local_scores, k, dim=1, largest=False)
        worst_score = min_vals.mean(dim=1)

        alpha = torch.sigmoid(self.raw_severity)
        final_score = (alpha * worst_score) + ((1 - alpha) * avg_score)

        return final_score.squeeze(1)

    def predict_trace(self, music_mels, chart_mels):
        """
        Explainability Method: Returns the second-by-second quality curve.

        Returns:
            local_scores: (Batch, Time) - The quality score at every timestep.
        """
        with torch.no_grad():
            # 1. Early Fusion: Concatenate along Channel dimension
            # Shape becomes (Batch, 160, Time)
            x = torch.cat([music_mels, chart_mels], dim=1)

            # 2. Extract Features (Strictly Local + Context)
            x = F.gelu(self.input_proj(x))
            x = self.encoder(x)

            # 3. Predict Score per Frame
            # (Batch, Dim, Time) -> (Batch, Time, Dim)
            x = x.permute(0, 2, 1)
            local_scores = torch.sigmoid(self.quality_proj(x))  # (Batch, Time, 1)
        return local_scores.squeeze(2)


if __name__ == "__main__":
    # Sanity Check
    from torchinfo import summary

    model = GameChartEvaluator()
    print(
        f"Model initialized. Learnable Severity: {torch.sigmoid(model.raw_severity).item():.2f}"
    )

    # Dummy data (Batch=2, Freq=80, Time=1000)
    m = torch.randn(2, 80, 1000)
    c = torch.randn(2, 80, 1000)

    output = model(m, c)
    print(f"Output shape: {output.shape}")  # Should be torch.Size([2])
    print(f"Scores: {output}")

    # Trace check
    trace = model.predict_trace(m, c)
    print(
        f"Trace shape: {trace.shape}"
    )  # Should be torch.Size([2, 500]) (due to MaxPool1d(2))

    summary(model, input_data=[m, c])