Add usage to README.md

README.md

@@ -20,19 +20,19 @@ Here we rewrite a neural network to a form that makes it possible to produce zk-
 
 So imagine there is a neural network that classifies images: looks at an image of a shopping bag and a receipt and verifies if the receipt matches the shopping bag. The neural network is public, the grocery store has shared it with clients and sets the rule: you can leave our store if the neural network allows. What we can now do with this library is to make ourselves a photo and then create a proof certifying that the neural network accepts our photo. We can share the proof with the grocery store, but keep our photo (and with it our shopping list) private.
 
-The above example has an obvious problem (when did i make the photo?), but even more importantly - this solution is not at all enough efficient computationaly to make such usecase possible. Therefore this project should be treated more like a proof of concept - but with computational advancements and generalization to multi-party computations, this presents exciting possibilities.
+The above example has an obvious problem (when did i make the photo?), but even more importantly -- this solution is not at all enough efficient computationally to make such use case possible. Therefore this project should be treated more like a proof of concept - but with computational advancements and generalization to multi-party computations, this presents exciting possibilities.
 
 Here, either the weights or the inputs or any combination of the two can be made private/public.
 
 ### Demo
 
-To see how it's used and how it works, check out the tests in [lib::test_trained_into_snark](https://github.com/przyjacielpkp/zkml/blob/main/lib/src/lib.rs#L76). Or run them with `cargo test --profile=test`. Tests take about a minute to run on my laptop.
+The usage of the app consists of three steps:
 
-These demonstrate the full functionality: 
- - a trained model is taken, that is a computation graph with the weight assignments (and some more bookeeping, we admit the abstraction is leaky here)
- - the snark public key gets calculated
- - prover provides the private argument and calculates the proof, together with evaluation of the circuit that is being claimed
- - the evaluation when mapped back to float matches the evaluation of the original non-snark computation
+- Use `zkml model` subcommand to train the model and save all necessary setup data to files. The exact description of all the arguments of this subcommand is available under `zkml model --help`,
+- Use `zkml server` subcommand to run a verifying server. The server must have access to the files generated by `zkml model`.
+- Use `zkml client` subcommand to run a prover. The prover must have access to the files generated by `zkml model` and a file with private input.
+
+The client will communicate with the server to perform the proof of knowledge of the secret input it has access to.
 
 ### How does it work
 
@@ -41,12 +41,11 @@ The native model of computations which are possible to be made into a snark way
 
 This means two things:
 
- - Need to encode floats as prime field elements. This is done by scaling, for details see [[Note: floats as ints]](https://github.com/przyjacielpkp/zkml/blob/c678d410adc3de188ce439b94ad4b9edba7785cf/lib/src/snark/snark.rs#L124).
- - Need to rewrite vectorized computations to significantly more scalar computions. This is done and described in [the scalar module](https://github.com/przyjacielpkp/zkml/blob/c678d410adc3de188ce439b94ad4b9edba7785cf/lib/src/scalar.rs#L131).
+- Need to encode floats as prime field elements. This is done by scaling, for details see [[Note: floats as ints]](https://github.com/przyjacielpkp/zkml/blob/c678d410adc3de188ce439b94ad4b9edba7785cf/lib/src/snark/snark.rs#L124).
+- Need to rewrite vectorized computations to significantly more scalar computions. This is done and described in [the scalar module](https://github.com/przyjacielpkp/zkml/blob/c678d410adc3de188ce439b94ad4b9edba7785cf/lib/src/scalar.rs#L131).
 
 For a description of what we're dealing with initially as the input ml computation, see [[Note: graph representation]](https://github.com/przyjacielpkp/zkml/blob/c678d410adc3de188ce439b94ad4b9edba7785cf/lib/src/scalar.rs#L31). This is a general and simple representation, similar abstraction level to tinygrad, onnx, pytorch or tensorflow models map onto it (provided these're defined statically, see [pytorch docs](https://pytorch.org/tutorials/intermediate/torch_export_tutorial.html#graph-breaks)).
 
-
 ## Building
 
 Build with cargo:
@@ -79,4 +78,4 @@ which can be installed systemwide, or provided in a developer shell by running `
 rustflags = ["-Zlinker-features=-lld"]
 ```
 
-in `.cargo/config.toml`.
+in `.cargo/config.toml`.