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

use std::{borrow::Cow, convert::TryInto, ffi::CStr, sync::Mutex};

use anyhow::{Context, Error};
use hotg_rune_core::{ElementType as RuneElementType, Shape, TFLITE_MIMETYPE};
use hotg_runecoral::{
    AccelerationBackend, ElementType, InferenceContext, Tensor,
    TensorDescriptor, TensorMut,
};

use crate::callbacks::Model;

/// Create a new [`Model`] backed by [`hotg_runecoral`].
pub fn load_tflite(
    model: &[u8],
    inputs: &[Shape<'_>],
    outputs: &[Shape<'_>],
) -> Result<Box<dyn Model>, Error> {
    let input_descriptors = inputs
        .iter()
        .map(descriptor)
        .collect::<Result<Vec<_>, Error>>()
        .context("Invalid input")?;
    let output_descriptors = outputs
        .iter()
        .map(descriptor)
        .collect::<Result<Vec<_>, Error>>()
        .context("Invalid output")?;

    let ctx = InferenceContext::create_context(
        TFLITE_MIMETYPE,
        model,
        AccelerationBackend::NONE,
    )
    .context("Unable to create the inference context")?;

    let model_input_descriptors: Vec<_> = ctx.inputs().collect();
    ensure_shapes_equal(&input_descriptors, &model_input_descriptors)?;
    let model_output_descriptors: Vec<_> = ctx.outputs().collect();
    ensure_shapes_equal(&output_descriptors, &model_output_descriptors)?;

    Ok(Box::new(RuneCoralModel {
        ctx: Mutex::new(ctx),
        inputs: inputs.iter().map(|s| s.to_owned()).collect(),
        input_descriptors,
        outputs: outputs.iter().map(|s| s.to_owned()).collect(),
        output_descriptors,
    }))
}

fn descriptor(s: &Shape) -> Result<TensorDescriptor<'static>, Error> {
    let dimensions: Vec<i32> = s
        .dimensions()
        .iter()
        .copied()
        .map(|d| d.try_into().unwrap())
        .collect();

    Ok(TensorDescriptor {
        name: CStr::from_bytes_with_nul(b"\0").unwrap(),
        element_type: element_type(s.element_type())?,
        shape: Cow::Owned(dimensions),
    })
}

struct RuneCoralModel {
    ctx: Mutex<InferenceContext>,
    inputs: Vec<Shape<'static>>,
    input_descriptors: Vec<TensorDescriptor<'static>>,
    outputs: Vec<Shape<'static>>,
    output_descriptors: Vec<TensorDescriptor<'static>>,
}

impl Model for RuneCoralModel {
    fn infer(
        &mut self,
        inputs: &[&[u8]],
        outputs: &mut [&mut [u8]],
    ) -> Result<(), Error> {
        let mut ctx = self.ctx.lock().expect("Lock was poisoned");

        let inputs: Vec<Tensor<'_>> = self
            .input_descriptors
            .iter()
            .zip(inputs)
            .map(|(desc, data)| Tensor {
                element_type: desc.element_type,
                shape: Cow::Borrowed(&desc.shape),
                buffer: *data,
            })
            .collect();

        let mut outputs: Vec<TensorMut<'_>> = self
            .output_descriptors
            .iter()
            .zip(outputs)
            .map(|(desc, data)| TensorMut {
                element_type: desc.element_type,
                shape: Cow::Borrowed(&desc.shape),
                buffer: *data,
            })
            .collect();

        ctx.infer(&inputs, &mut outputs)
            .context("Inference failed")?;

        Ok(())
    }

    fn input_shapes(&self) -> &[Shape<'_>] { &self.inputs }

    fn output_shapes(&self) -> &[Shape<'_>] { &self.outputs }
}

fn element_type(rune_type: RuneElementType) -> Result<ElementType, Error> {
    Ok(match rune_type {
        RuneElementType::I8 => ElementType::Int8,
        RuneElementType::U8 => ElementType::UInt8,
        RuneElementType::I16 => ElementType::Int16,
        RuneElementType::I32 => ElementType::Int32,
        RuneElementType::I64 => ElementType::Int64,
        RuneElementType::F32 => ElementType::Float32,
        RuneElementType::F64 => ElementType::Float64,
        RuneElementType::String => ElementType::String,
        _ => {
            anyhow::bail!(
                "librunecoral doesn't support {:?} tensors",
                rune_type
            )
        },
    })
}

fn ensure_shapes_equal(
    from_rune: &[TensorDescriptor<'_>],
    from_model: &[TensorDescriptor<'_>],
) -> Result<(), Error> {
    if from_rune.len() == from_model.len()
        && from_rune.iter().zip(from_model.iter()).all(|(x, y)| {
            x.element_type == y.element_type && x.shape == y.shape
        })
    {
        return Ok(());
    }

    fn pretty_shapes(descriptors: &[TensorDescriptor<'_>]) -> String {
        format!(
            "[{}]",
            descriptors
                .iter()
                .map(|d| format!("{}", d))
                .collect::<Vec<_>>()
                .join(", ")
        )
    }

    anyhow::bail!(
        "The Rune said tensors would be {}, but the model said they would be \
         {}",
        pretty_shapes(from_rune),
        pretty_shapes(from_model),
    );
}