Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Added batching to regression metrics #297

Merged
merged 4 commits into from
Sep 13, 2024
Merged

Added batching to regression metrics #297

Changes from 1 commit
Commits
Show all changes
4 commits
Select commit Hold shift + click to select a range
355daec
Added batching to regression metrcs
norm4nn Sep 2, 2024
9cd51ac
Migrated from multioutput to axes option
norm4nn Sep 11, 2024
eaf3204
mix format
norm4nn Sep 11, 2024
8444407
trigger build
norm4nn Sep 12, 2024
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
176 changes: 150 additions & 26 deletions lib/scholar/metrics/regression.ex
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -15,6 +15,16 @@ defmodule Scholar.Metrics.Regression do
import Scholar.Shared
import Scholar.Metrics.Distance

general_schema = [
axes: [
type: {:custom, Scholar.Options, :axes, []},
doc: """
Axes to calculate the distance over. By default the distance
is calculated between the whole tensors.
"""
]
]

r2_schema = [
force_finite: [
type: :boolean,
Expand All @@ -26,7 +36,8 @@ defmodule Scholar.Metrics.Regression do
]
]

@r2_schema NimbleOptions.new!(r2_schema)
@general_schema NimbleOptions.new!(general_schema)
@r2_schema NimbleOptions.new!(general_schema ++ r2_schema)

# Standard Metrics

Expand All @@ -36,6 +47,10 @@ defmodule Scholar.Metrics.Regression do

$$MAE = \frac{\sum_{i=1}^{n} |\hat{y_i} - y_i|}{n}$$

## Options

#{NimbleOptions.docs(@general_schema)}

## Examples

iex> y_true = Nx.tensor([[0.0, 1.0], [0.0, 0.0]])
Expand All @@ -45,13 +60,22 @@ defmodule Scholar.Metrics.Regression do
f32
0.5
>
iex> Scholar.Metrics.Regression.mean_absolute_error(y_true, y_pred, axes: [0])
#Nx.Tensor<
f32[2]
[1.0, 0.0]
>
"""
defn mean_absolute_error(y_true, y_pred) do
deftransform mean_absolute_error(y_true, y_pred, opts \\ []) do
mean_absolute_error_n(y_true, y_pred, NimbleOptions.validate!(opts, @general_schema))
end

defn mean_absolute_error_n(y_true, y_pred, opts \\ []) do
assert_same_shape!(y_true, y_pred)

(y_true - y_pred)
|> Nx.abs()
|> Nx.mean()
|> Nx.mean(axes: opts[:axes])
end

@doc ~S"""
Expand All @@ -60,6 +84,10 @@ defmodule Scholar.Metrics.Regression do

$$MSE = \frac{\sum_{i=1}^{n} (\hat{y_i} - y_i)^2}{n}$$

## Options

#{NimbleOptions.docs(@general_schema)}

## Examples

iex> y_true = Nx.tensor([[0.0, 2.0], [0.5, 0.0]])
Expand All @@ -69,9 +97,14 @@ defmodule Scholar.Metrics.Regression do
f32
0.5625
>
iex> Scholar.Metrics.Regression.mean_square_error(y_true, y_pred, axes: [0])
#Nx.Tensor<
f32[2]
[0.625, 0.5]
>
"""
defn mean_square_error(y_true, y_pred) do
mean_tweedie_deviance_n(y_true, y_pred, 0)
deftransform mean_square_error(y_true, y_pred, opts \\ []) do
mean_tweedie_deviance_n(y_true, y_pred, 0, NimbleOptions.validate!(opts, @general_schema))
end

@doc ~S"""
Expand All @@ -80,6 +113,10 @@ defmodule Scholar.Metrics.Regression do

$$MSLE = \frac{\sum_{i=1}^{n} (\log(\hat{y_i} + 1) - \log(y_i + 1))^2}{n}$$

## Options

#{NimbleOptions.docs(@general_schema)}

## Examples

iex> y_true = Nx.tensor([[0.0, 1.0], [0.0, 0.0]])
Expand All @@ -89,9 +126,16 @@ defmodule Scholar.Metrics.Regression do
f32
0.24022650718688965
>
iex> Scholar.Metrics.Regression.mean_square_log_error(y_true, y_pred, axes: [0])
#Nx.Tensor<
f32[2]
[0.4804530143737793, 0.0]
>
"""
defn mean_square_log_error(y_true, y_pred) do
mean_square_error(Nx.log(y_true + 1), Nx.log(y_pred + 1))
deftransform mean_square_log_error(y_true, y_pred, opts \\ []) do
y_true = Nx.log(Nx.add(y_true, 1))
y_pred = Nx.log(Nx.add(y_pred, 1))
mean_square_error(y_true, y_pred, NimbleOptions.validate!(opts, @general_schema))
end

@doc ~S"""
Expand All @@ -101,6 +145,10 @@ defmodule Scholar.Metrics.Regression do

$$MAPE = \frac{\sum_{i=1}^{n} \frac{|\hat{y_i} - y_i|}{max(\epsilon, \hat{y_i})}}{n}$$

## Options

#{NimbleOptions.docs(@general_schema)}

## Examples

iex> y_true = Nx.tensor([3, -0.5, 2, 7])
Expand All @@ -118,18 +166,33 @@ defmodule Scholar.Metrics.Regression do
f32
209715.28125
>
iex> y_true = Nx.tensor([[0.5, 1], [-1, 1], [7, -6]])
iex> y_pred = Nx.tensor([[0, 2], [-1, 2], [8, -5]])
iex> Scholar.Metrics.Regression.mean_absolute_percentage_error(y_true, y_pred, axes: [0])
#Nx.Tensor<
f32[2]
[0.380952388048172, 0.7222222685813904]
>
"""
defn mean_absolute_percentage_error(y_true, y_pred) do
deftransform mean_absolute_percentage_error(y_true, y_pred, opts \\ []) do
assert_same_shape!(y_true, y_pred)

mean_absolute_percentage_error_n(
y_true,
y_pred,
NimbleOptions.validate!(opts, @general_schema)
)
end

defnp mean_absolute_percentage_error_n(y_true, y_pred, opts \\ []) do
eps =
Nx.type(y_true)
|> Nx.Type.merge(Nx.type(y_pred))
|> Nx.Type.to_floating()
|> Nx.Constants.epsilon()

(Nx.abs(y_true - y_pred) / Nx.max(eps, Nx.abs(y_true)))
|> Nx.mean()
|> Nx.mean(axes: opts[:axes])
end

@doc """
Expand All @@ -147,6 +210,10 @@ defmodule Scholar.Metrics.Regression do
\end{cases}$$
'''}

## Options

#{NimbleOptions.docs(@general_schema)}

## Examples

iex> y_true = Nx.tensor([1, 1, 1, 1, 1, 2, 2, 1, 3, 1], type: :u32)
Expand All @@ -156,9 +223,16 @@ defmodule Scholar.Metrics.Regression do
f32
0.18411168456077576
>
iex> y_true = Nx.tensor([[1, 1, 1, 1], [1, 2, 2, 1]], type: :u32)
iex> y_pred = Nx.tensor([[2, 2, 1, 1], [2, 2, 2, 1]], type: :u32)
iex> Scholar.Metrics.Regression.mean_tweedie_deviance(y_true, y_pred, 1, axes: [0])
#Nx.Tensor<
f32[4]
[0.6137056350708008, 0.3068528175354004, 0.0, 0.0]
>
"""
defn mean_tweedie_deviance(y_true, y_pred, power) do
mean_tweedie_deviance_n(y_true, y_pred, power)
deftransform mean_tweedie_deviance(y_true, y_pred, power, opts \\ []) do
mean_tweedie_deviance_n(y_true, y_pred, power, NimbleOptions.validate!(opts, @general_schema))
end

@doc """
Expand All @@ -167,6 +241,10 @@ defmodule Scholar.Metrics.Regression do

Note: This function cannot be used in `defn`.

## Options

#{NimbleOptions.docs(@general_schema)}

## Examples

iex> y_true = Nx.tensor([1, 1, 1, 1, 1, 2, 2, 1, 3, 1], type: :u32)
Expand All @@ -177,7 +255,7 @@ defmodule Scholar.Metrics.Regression do
0.18411168456077576
>
"""
def mean_tweedie_deviance!(y_true, y_pred, power) do
def mean_tweedie_deviance!(y_true, y_pred, power, opts \\ []) do
message = "mean Tweedie deviance with power=#{power} can only be used on "

case check_tweedie_deviance_power(y_true, y_pred, power) |> Nx.to_number() do
Expand All @@ -188,10 +266,10 @@ defmodule Scholar.Metrics.Regression do
100 -> raise "something went wrong, branch should never appear"
end

mean_tweedie_deviance_n(y_true, y_pred, power)
mean_tweedie_deviance(y_true, y_pred, power, opts)
end

defnp mean_tweedie_deviance_n(y_true, y_pred, power) do
defnp mean_tweedie_deviance_n(y_true, y_pred, power, opts \\ []) do
deviance =
cond do
power < 0 ->
Expand Down Expand Up @@ -226,7 +304,7 @@ defmodule Scholar.Metrics.Regression do
)
end

Nx.mean(deviance)
Nx.mean(deviance, axes: opts[:axes])
end

defnp check_tweedie_deviance_power(y_true, y_pred, power) do
Expand Down Expand Up @@ -264,6 +342,10 @@ defmodule Scholar.Metrics.Regression do
Calculates the mean Poisson deviance of predictions
with respect to targets.

## Options

#{NimbleOptions.docs(@general_schema)}

## Examples

iex> y_true = Nx.tensor([1, 1, 1, 1, 1, 2, 2, 1, 3, 1], type: :u32)
Expand All @@ -273,15 +355,27 @@ defmodule Scholar.Metrics.Regression do
f32
0.18411168456077576
>

iex> y_true = Nx.tensor([[1, 1, 1, 1], [1, 2, 2, 1]], type: :u32)
iex> y_pred = Nx.tensor([[2, 2, 1, 1], [2, 2, 2, 1]], type: :u32)
iex> Scholar.Metrics.Regression.mean_poisson_deviance(y_true, y_pred, axes: [1])
#Nx.Tensor<
f32[2]
[0.3068528175354004, 0.1534264087677002]
>
"""
defn mean_poisson_deviance(y_true, y_pred) do
mean_tweedie_deviance_n(y_true, y_pred, 1)
deftransform mean_poisson_deviance(y_true, y_pred, opts \\ []) do
mean_tweedie_deviance_n(y_true, y_pred, 1, NimbleOptions.validate!(opts, @general_schema))
end

@doc """
Calculates the mean Gamma deviance of predictions
with respect to targets.

## Options

#{NimbleOptions.docs(@general_schema)}

## Examples

iex> y_true = Nx.tensor([1, 1, 1, 1, 1, 2, 2, 1, 3, 1], type: :u32)
Expand All @@ -291,9 +385,16 @@ defmodule Scholar.Metrics.Regression do
f32
0.115888312458992
>
iex> y_true = Nx.tensor([[1, 1, 1, 1, 1], [2, 2, 1, 3, 1]], type: :u32)
iex> y_pred = Nx.tensor([[2, 2, 1, 1, 2], [2, 2, 1, 3, 1]], type: :u32)
iex> Scholar.Metrics.Regression.mean_gamma_deviance(y_true, y_pred, axes: [0])
#Nx.Tensor<
f32[5]
[0.1931471824645996, 0.1931471824645996, 0.0, 0.0, 0.1931471824645996]
>
"""
defn mean_gamma_deviance(y_true, y_pred) do
mean_tweedie_deviance_n(y_true, y_pred, 2)
deftransform mean_gamma_deviance(y_true, y_pred, opts \\ []) do
mean_tweedie_deviance_n(y_true, y_pred, 2, NimbleOptions.validate!(opts, @general_schema))
end

@doc """
Expand All @@ -317,6 +418,14 @@ defmodule Scholar.Metrics.Regression do
0.9486081600189209
>

iex> y_true = Nx.tensor([[3, -0.5], [2, 7]], type: {:f, 32})
iex>y_pred = Nx.tensor([[2.5, 0.0], [2, 8]], type: {:f, 32})
iex> Scholar.Metrics.Regression.r2_score(y_true, y_pred, axes: [0])
#Nx.Tensor<
f32[2]
[0.6800000071525574, 0.9559956192970276]
>

iex> y_true = Nx.tensor([-2.0, -2.0, -2.0], type: :f64)
iex> y_pred = Nx.tensor([-2.0, -2.0, -2.0 + 1.0e-8], type: :f64)
iex> Scholar.Metrics.Regression.r2_score(y_true, y_pred, force_finite: true)
Expand Down Expand Up @@ -354,11 +463,11 @@ defmodule Scholar.Metrics.Regression do
end

defnp r2_score_n(y_true, y_pred, opts) do
check_shape(y_true, y_pred)
ssr = squared_euclidean(y_true, y_pred)
assert_same_shape!(y_true, y_pred)
ssr = squared_euclidean(y_true, y_pred, axes: opts[:axes])

y_mean = Nx.broadcast(Nx.mean(y_true), Nx.shape(y_true))
sst = squared_euclidean(y_true, y_mean)
sst = squared_euclidean(y_true, y_mean, axes: opts[:axes])

handle_non_finite(ssr, sst, opts)
end
Expand Down Expand Up @@ -413,6 +522,14 @@ defmodule Scholar.Metrics.Regression do
f32
1.0
>

iex> y_true = Nx.tensor([[3, -0.5], [2, 7]], type: {:f, 32})
iex> y_pred = Nx.tensor([[2.5, 0.0], [2, 8]], type: {:f, 32})
iex> Scholar.Metrics.Regression.explained_variance_score(y_true, y_pred, axes: [0])
#Nx.Tensor<
f32[2]
[0.75, 0.995555579662323]
>
"""
deftransform explained_variance_score(y_true, y_pred, opts \\ []) do
explained_variance_score_n(
Expand All @@ -423,13 +540,14 @@ defmodule Scholar.Metrics.Regression do
end

defnp explained_variance_score_n(y_true, y_pred, opts) do
y_diff_avg = Nx.mean(y_true - y_pred, axes: [0])
axes = opts[:axes]
y_diff_avg = Nx.mean(y_true - y_pred, axes: axes)
sample_size = Nx.axis_size(y_true, 0)

numerator = squared_euclidean(y_true, y_pred + y_diff_avg, axes: [0]) / sample_size
numerator = squared_euclidean(y_true, y_pred + y_diff_avg, axes: axes) / sample_size

y_true_avg = Nx.mean(y_true, axes: [0])
denominator = Nx.mean((y_true - y_true_avg) ** 2, axes: [0])
y_true_avg = Nx.mean(y_true, axes: axes)
denominator = Nx.mean((y_true - y_true_avg) ** 2, axes: axes)
handle_non_finite(numerator, denominator, opts)
end

Expand Down Expand Up @@ -516,6 +634,10 @@ defmodule Scholar.Metrics.Regression do
"""
defn max_residual_error(y_true, y_pred) do
check_shape(y_true, y_pred)
max_residual_error_n(y_true, y_pred)
end

defnp max_residual_error_n(y_true, y_pred) do
Nx.reduce_max(Nx.abs(y_true - y_pred))
end

Expand Down Expand Up @@ -577,6 +699,8 @@ defmodule Scholar.Metrics.Regression do
The residual error is defined as $$|y - \hat{y}|$$ where $y$ is a true value
and $\hat{y}$ is a predicted value.

## Options

#{NimbleOptions.docs(@mean_pinball_loss_schema)}

## Examples
Expand Down
Loading