State serialization/deserialization overhaul #247
Conversation
nathanielsimard
changed the title
| [`bincode`](https://github.com/bincode-org/bincode) (for compactness) and included as part of the | ||
| final wasm output. The MNIST model is initialized with trained weights from memory during the | ||
| runtime. | ||
|
|
||
| The inference API for JavaScript is exposed with the help of | ||
| [`wasm-bindgen`](https://github.com/rustwasm/wasm-bindgen)'s library and tools. | ||
|
|
There was a problem hiding this comment.
Under the future improvements you can now remove #201 and #202
There was a problem hiding this comment.
Probably you can also remove references to the wasm file since it is smaller, so that we don't have to keep updating it.
There was a problem hiding this comment.
The state is not saved with f16 anymore, since it doesn't compile correctly without std, so I'll keep the link to issue 202 as potential improvement.
There was a problem hiding this comment.
Probably you can also remove references to the wasm file since it is smaller, so that we don't have to keep updating it.
I'm not sure what you are refering to.
There was a problem hiding this comment.
Sorry I meant in the readme file, the document talks about the file size of the wasm output under comparison section (e.g. 1,509,747 bytes).
It's okay. I'll update the readme file not to talk about the file sizes.
There was a problem hiding this comment.
It's the same size because it's still in f32... for now :)
| [`bincode`](https://github.com/bincode-org/bincode) (for compactness) and included as part of the | ||
| final wasm output. The MNIST model is initialized with trained weights from memory during the | ||
| runtime. | ||
|
|
||
| The inference API for JavaScript is exposed with the help of | ||
| [`wasm-bindgen`](https://github.com/rustwasm/wasm-bindgen)'s library and tools. | ||
|
|
There was a problem hiding this comment.
Sorry I meant in the readme file, the document talks about the file size of the wasm output under comparison section (e.g. 1,509,747 bytes).
It's okay. I'll update the readme file not to talk about the file sizes.
| @@ -64,14 +65,21 @@ pub fn run<B: ADBackend>(device: B::Device) { | |||
| .metric_valid_plot(AccuracyMetric::new()) | |||
| .metric_train_plot(LossMetric::new()) | |||
| .metric_valid_plot(LossMetric::new()) | |||
| .with_file_checkpointer::<f32>(2) | |||
| .with_file_checkpointer::<burn::tensor::f16>(2, StateFormat::default()) | |||
There was a problem hiding this comment.
Is it still f16, since you reverted back your changes about f16.
There was a problem hiding this comment.
The checkpoint during training are f16 with compressed bin format (bin.gz) but we now save the final model somewhere else, we don't use the checkpoints.
| @@ -78,17 +78,18 @@ pub fn train<B: ADBackend, D: TextClassificationDataset + 'static>( | |||
| .metric_valid(AccuracyMetric::new()) | |||
| .metric_train_plot(LossMetric::new()) | |||
| .metric_valid_plot(LossMetric::new()) | |||
| .with_file_checkpointer::<f32>(2) | |||
| .with_file_checkpointer::<burn::tensor::f16>(2, StateFormat::default()) | |||
There was a problem hiding this comment.
Do you need to revert back this change or does it work because serialization works for std?
There was a problem hiding this comment.
serialization works with std
fix #202 I saved models with f16, but It's also possible to save weights with bf16. I don't expect much difference in performance for both.
fix #201 by default we save states in compressed bincode, which is extremelly small :)