Skip to content

metric

Source code in tinybig/koala/statistics/metric.py 
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
def metric(
    metric_name: str,
    x: torch.Tensor,
    *args, **kwargs
):
    assert x is not None and metric_name is not None
    match metric_name:
        case 'mean': return mean(x=x)
        case 'batch_mean': return batch_mean(x=x, *args, **kwargs)
        case 'weighted_mean' | 'wmean': return weighted_mean(x=x, *args, **kwargs)
        case 'batch_weighted_mean' | 'batch_wmean': return batch_weighted_mean(x=x, *args, **kwargs)
        case 'geometric_mean' | 'gmean': return geometric_mean(x=x)
        case 'batch_geometric_mean' | 'batch_gmean': return batch_geometric_mean(x=x, *args, **kwargs)
        case 'harmonic_mean' | 'hmean': return harmonic_mean(x=x, *args, **kwargs)
        case 'batch_harmonic_mean' | 'batch_hmean': return batch_harmonic_mean(x=x, *args, **kwargs)
        case 'median': return median(x=x)
        case 'batch_median': return batch_median(x=x, *args, **kwargs)
        case 'mode': return mode(x=x)
        case 'batch_mode': return batch_mode(x=x, *args, **kwargs)
        case 'entropy': return entropy(x=x)
        case 'batch_entropy': return batch_entropy(x=x, *args, **kwargs)
        case 'variance' | 'var': return variance(x=x)
        case 'batch_variance' | 'batch_var': return batch_variance(x=x, *args, **kwargs)
        case 'std' | 'standard_deviation': return std(x=x)
        case 'batch_std' | 'batch_standard_deviation': return batch_std(x=x, *args, **kwargs)
        case 'skewness' | 'skew': return skewness(x=x)
        case 'batch_skewness' | 'batch_skew': return batch_skewness(x=x, *args, **kwargs)
        case _: raise ValueError(f'Unknown metric name: {metric_name}...')
Source code in tinybig/koala/statistics/metric.py 
42
43
44
def mean(x: torch.Tensor):
    assert x.ndim == 1
    return torch.mean(x)
Source code in tinybig/koala/statistics/metric.py 
47
48
49
def batch_mean(x: torch.Tensor, dim: int = 1):
    assert x.ndim == 2 and dim in [0, 1]
    return torch.mean(x, dim=dim)
Source code in tinybig/koala/statistics/metric.py 
69
70
71
72
73
def geometric_mean(x: torch.Tensor):
    assert x.ndim == 1 and torch.all(x > 0)
    log_x = torch.log(x)
    mean_log_x = torch.mean(log_x)
    return torch.exp(mean_log_x)
Source code in tinybig/koala/statistics/metric.py 
76
77
78
79
80
81
def batch_geometric_mean(x: torch.Tensor, dim: int = 1):
    assert x.ndim == 2 and dim in [0, 1] and torch.all(x > 0)

    log_x = torch.log(x)
    mean_log_x = torch.mean(log_x, dim=dim)
    return torch.exp(mean_log_x)
Source code in tinybig/koala/statistics/metric.py 
84
85
86
87
88
89
90
91
92
93
def harmonic_mean(x: torch.Tensor, weights=None):
    assert x.ndim == 1 and torch.all(x > 0)

    if weights is None:
        return x.numel() / torch.sum(1 / x)
    else:
        assert weights.ndim == 1 and x.shape == weights.shape
        weighted_sum = torch.sum(weights)
        weighted_reciprocal_sum = torch.sum(weights / x)
        return weighted_sum / weighted_reciprocal_sum
Source code in tinybig/koala/statistics/metric.py 
 96
 97
 98
 99
100
101
102
103
104
105
106
107
def batch_harmonic_mean(x: torch.Tensor, weights=None, dim: int = 1):
    assert x.ndim == 2 and dim in [0, 1] and torch.all(x > 0)

    if weights is not None:
        assert weights.ndim == 1 and x.shape[dim] == weights.shape[0]

    if weights is None:
        reciprocal_sum = torch.sum(1 / x, dim=dim)
        return x.shape[dim] / reciprocal_sum
    else:
        weighted_reciprocal_sum = torch.sum(weights / x, dim=dim)
        return torch.sum(weights, dim=dim) / weighted_reciprocal_sum
Source code in tinybig/koala/statistics/metric.py 
52
53
54
55
56
def weighted_mean(x: torch.Tensor, weights: torch.Tensor):
    assert x.ndim == 1 and x.shape == weights.shape
    weighted_sum = torch.sum(x * weights)
    sum_weights = torch.sum(weights)
    return weighted_sum / sum_weights
Source code in tinybig/koala/statistics/metric.py 
59
60
61
62
63
64
65
66
def batch_weighted_mean(x: torch.Tensor, weights: torch.Tensor, dim: int = 1):
    assert x.ndim == 2 and dim in [0, 1] and x.shape[1] == weights.shape[0]

    weights = weights.unsqueeze(0) if dim == 1 else weights.unsqueeze(1)
    weighted_sum = torch.sum(x * weights, dim=dim)
    sum_weights = torch.sum(weights, dim=dim)

    return weighted_sum / sum_weights
Source code in tinybig/koala/statistics/metric.py 
110
111
112
def median(x: torch.Tensor):
    assert x.ndim == 1
    return torch.median(x)
Source code in tinybig/koala/statistics/metric.py 
115
116
117
def batch_median(x: torch.Tensor, dim: int = 1):
    assert x.ndim == 2 and dim in [0, 1]
    return torch.median(x, dim=dim).values
Source code in tinybig/koala/statistics/metric.py 
120
121
122
def mode(x: torch.Tensor):
    assert x.ndim == 1
    return torch.mode(x)
Source code in tinybig/koala/statistics/metric.py 
125
126
127
def batch_mode(x: torch.Tensor, dim: int = 1):
    assert x.ndim == 2 and dim in [0, 1]
    return torch.mode(x, dim=dim)
Source code in tinybig/koala/statistics/metric.py 
153
154
155
def std(x: torch.Tensor):
    assert x.ndim == 1
    return torch.std(x)
Source code in tinybig/koala/statistics/metric.py 
158
159
160
def batch_std(x: torch.Tensor, dim: int = 1):
    assert x.ndim == 2 and dim in [0, 1]
    return torch.std(x, dim=dim)
Source code in tinybig/koala/statistics/metric.py 
130
131
132
133
def entropy(x: torch.Tensor):
    assert x.ndim == 1 and torch.all(x >= 0) and torch.isclose(torch.sum(x), torch.tensor(1.0)), "The tensor values must sum to 1."
    entropy_value = -torch.sum(x * torch.log(x + 1e-12))
    return entropy_value
Source code in tinybig/koala/statistics/metric.py 
136
137
138
139
140
def batch_entropy(x: torch.Tensor, dim: int = 1):
    assert x.ndim == 2 and dim in [0, 1]
    assert torch.all(x >= 0) and torch.allclose(torch.sum(x, dim=dim), torch.tensor(1.0))
    entropy_values = -torch.sum((x+1e-12) * torch.log(x), dim=dim)
    return entropy_values
Source code in tinybig/koala/statistics/metric.py 
143
144
145
def variance(x: torch.Tensor):
    assert x.ndim == 1
    return torch.var(x)
Source code in tinybig/koala/statistics/metric.py 
148
149
150
def batch_variance(x: torch.Tensor, dim: int = 1):
    assert x.ndim == 2 and dim in [0, 1]
    return torch.var(x, dim=dim)
Source code in tinybig/koala/statistics/metric.py 
163
164
165
166
def skewness(x: torch.Tensor):
    assert x.ndim == 1
    skewness_value = torch.mean((x - torch.mean(x)) ** 3) / (torch.std(x, unbiased=False) ** 3 + 1e-12)
    return skewness_value
Source code in tinybig/koala/statistics/metric.py 
169
170
171
172
def batch_skewness(x: torch.Tensor, dim: int = 1):
    assert x.ndim == 2 and dim in [0, 1]
    skewness_value = torch.mean((x - torch.mean(x, dim=dim, keepdim=True)) ** 3, dim=dim) / (torch.std(x, dim=dim, unbiased=False, keepdim=True) ** 3 + 1e-12)
    return skewness_value.squeeze(dim)