Load pytorch model got error DeserializeError("Candle pickle error: specified file not found in archive")

[zemelLeong]

I believe there may be a significant issue with my model definition. However, based on this error message, I can't determine exactly what adjustments are needed.

thread 'main' panicked at src\model_lore.rs:18:14:
called `Result::unwrap()` on an `Err` value: DeserializeError("Candle pickle error: specified file not found in archive")
note: run with `RUST_BACKTRACE=1` environment variable to display a backtrace

use crate::lore_detector::LoreDetectModel;
use crate::lore_processor::LoreProcessModel;
use burn::module::Module;
use burn::record::{FullPrecisionSettings, Recorder};
use burn::tensor::backend::Backend;
use burn_import::pytorch::PyTorchFileRecorder;

#[derive(Module, Debug)]
pub struct LoreModel<B: Backend> {
    model: LoreDetectModel<B>,
    processor: LoreProcessModel<B>,
}

impl<B: Backend> LoreModel<B> {
    pub fn new(model_path: &str, device: &B::Device) -> Self {
        let record: LoreModelRecord<B> = PyTorchFileRecorder::<FullPrecisionSettings>::default()
            .load(model_path.into(), device)
            .unwrap();

        let model = LoreDetectModel::new_with(record.model);
        let processor = LoreProcessModel::new_with(record.processor, device);

        Self { model, processor }
    }
}

model state weight keys

  "model.conv1.weight",
  "model.bn1.weight",
  "model.bn1.bias",
  "model.bn1.running_mean",
  "model.bn1.running_var",
  "model.bn1.num_batches_tracked",
  "model.layer1.0.conv1.weight",
  "model.layer1.0.conv1.bias",
  "model.layer1.0.bn1.weight",
  "model.layer1.0.bn1.bias",
  ...
  "processor.stacker.logi_encoder.0.weight",
  "processor.stacker.logi_encoder.0.bias",
  "processor.stacker.logi_encoder.2.weight",
  "processor.stacker.logi_encoder.2.bias",
  "processor.stacker.tsfm.linear.weight",
  "processor.stacker.tsfm.linear.bias",
  "processor.stacker.tsfm.encoder.pe.pe",
  "processor.stacker.tsfm.encoder.layers.0.norm_1.alpha",
  "processor.stacker.tsfm.encoder.layers.0.norm_1.bias",
  "processor.stacker.tsfm.encoder.layers.0.norm_2.alpha",
  ...

[zemelLeong]

I follow the official document test load pytorch model got same error too. My OS is Windows11.

use burn::{
    module::Module,
    nn::conv::{Conv2d, Conv2dConfig},
    tensor::{backend::Backend, Tensor},
};
use burn::record::{FullPrecisionSettings, Recorder};
use burn_import::pytorch::PyTorchFileRecorder;

#[derive(Module, Debug)]
pub struct Net<B: Backend> {
    conv1: Conv2d<B>,
    conv2: Conv2d<B>,
}

impl<B: Backend> Net<B> {
    /// Create a new model from the given record.
    pub fn new_with(record: NetRecord<B>) -> Self {
        let conv1 = Conv2dConfig::new([2, 2], [2, 2])
            .init_with(record.conv1);
        let conv2 = Conv2dConfig::new([2, 2], [2, 2])
            .with_bias(false)
            .init_with(record.conv2);
        Self { conv1, conv2 }
    }

    /// Forward pass of the model.
    pub fn forward(&self, x: Tensor<B, 4>) -> Tensor<B, 4> {
        let x = self.conv1.forward(x);
        self.conv2.forward(x)
    }
}

type Backend1 = burn_ndarray::NdArray<f32>;

fn main() {
    let device = Default::default();
    let record = PyTorchFileRecorder::<FullPrecisionSettings>::default()
        .load("J:/RustWorkspace/table-structure-recognition/files/conv2d.pt".into(), &device)
        .expect("Should decode state successfully");

    Net::<Backend1>::new_with(record);
}

C:/Users/11989/.cargo/bin/cargo.exe run --color=always --package table-structure-recognition --example simple-model-import
    Finished dev [unoptimized + debuginfo] target(s) in 0.42s
     Running `J:/cargo-target\debug\examples\simple-model-import.exe`
thread 'main' panicked at examples\simple-model-import.rs:39:10:
Should decode state successfully: DeserializeError("Candle pickle error: specified file not found in archive")
stack backtrace:
   0: std::panicking::begin_panic_handler
             at /rustc/82e1608dfa6e0b5569232559e3d385fea5a93112/library\std\src\panicking.rs:645
   1: core::panicking::panic_fmt
             at /rustc/82e1608dfa6e0b5569232559e3d385fea5a93112/library\core\src\panicking.rs:72
   2: core::result::unwrap_failed
             at /rustc/82e1608dfa6e0b5569232559e3d385fea5a93112/library\core\src\result.rs:1653
   3: enum2$<core::result::Result<simple_model_import::NetRecord<burn_ndarray::backend::NdArray<f32> >,enum2$<burn_core::record::recorder::RecorderError> > >::expect<simple_model_import::NetRecord<burn_ndarray::backend::NdArray<f32> >,enum2$<burn_core::record::
             at /rustc/82e1608dfa6e0b5569232559e3d385fea5a93112\library\core\src\result.rs:1034
   4: simple_model_import::main
             at .\examples\simple-model-import.rs:37
   5: core::ops::function::FnOnce::call_once<void (*)(),tuple$<> >
             at /rustc/82e1608dfa6e0b5569232559e3d385fea5a93112\library\core\src\ops\function.rs:250
   6: core::hint::black_box
             at /rustc/82e1608dfa6e0b5569232559e3d385fea5a93112\library\core\src\hint.rs:286
note: Some details are omitted, run with `RUST_BACKTRACE=full` for a verbose backtrace.
error: process didn't exit successfully: `J:/cargo-target\debug\examples\simple-model-import.exe` (exit code: 101)

[laggui]

Works on linux, but reproduced on windows. It is a known issue #1178 (also listed in the book).

It seems to have been fixed on the latest candle version (0.4.0), which I tested to confirm.

[laggui]

PR #1382 opened which should fix this windows issue by bumping to the candle-core 0.4.1

[antimora]

I will test after the upgrade and see if it works on Windows.

[antimora]

The PR has been merged! It should work on windows if you check out the main branch.