Train (neural_nets_lib.Ocannl.Train)

module Ops = Arrayjit.Ops

module Tn = Arrayjit.Tnode

module Nd = Arrayjit.Ndarray

module NTDSL = Operation.NTDSL

module Asgns = Arrayjit.Assignments

module Idx = Arrayjit.Indexing

module Task = Arrayjit.Task

module Utils = Arrayjit.Utils

module Rand = Arrayjit.Rand.Lib

module BT = Arrayjit.Backend_intf

module type Backend = Arrayjit.Backend_intf.Backend

val _get_local_debug_runtime : unit -> (module Minidebug_runtime.Debug_runtime)

module CDSL : sig ... end

module IDX : sig ... end

val run : 'a BT.routine -> Base.unit

val is_param : Tensor.t -> Base.bool

val get_params : Tensor.t -> (Tensor.t, Tensor.comparator_witness) Base.Set.t

val set_on_host : ?from_device:bool -> Tn.t -> unit

val set_materialized : Tn.t -> unit

val set_hosted : Tn.t -> unit

val forward : 
  ?disable_rootness_check:bool ->
  Tensor.t ->
  Arrayjit.Assignments.comp

Sets the tensor's value as "fully on host", returns the tensor's forward code with a label-derived comment.

type updaten = {

loss : Tensor.t;
params : (Tensor.t, Tensor.comparator_witness) Base.Set.t;
fwd_bprop : Asgns.comp;

}

val diff_or_error : Tensor.t -> Base.String.t -> Tensor.diff

val grad_update_nochecks : Tensor.t -> updaten

val grad_update : 
  ?disable_rootness_check:bool ->
  ?setup_for_parallel:bool ->
  Tensor.t ->
  updaten

Returns the tensor's forward, zeroing gradients, and backprop code wrapped with label-derived comments. Sets the tensor's value as "fully on host". If setup_for_parallel is true (false by default), sets the parameters and their gradients as "non-local" (on-device).

val sgd_one : 
  learning_rate:Tensor.t ->
  ?momentum:Base.Float.t ->
  ?weight_decay:Base.float ->
  ?nesterov:bool ->
  Tensor.t ->
  Arrayjit.Assignments.comp

See: https://github.com/tinygrad/tinygrad/blob/master/tinygrad/nn/optim.py

val sgd_update : 
  learning_rate:Tensor.t ->
  ?momentum:Base.Float.t ->
  ?weight_decay:Base.float ->
  ?nesterov:bool ->
  updaten ->
  Arrayjit.Assignments.comp

val sequential_loop : 
  f:(unit -> Base.unit) ->
  (Idx.static_symbol * Base.int Base.ref) list ->
  Base.unit

All and only bindings with associated ranges are iterated, with the binding's initial value lost. Bindings without ranges remain at their initial values.

val round_robin : 
  (unit -> unit) Base.Array.t ->
  Idx.lowered_bindings Base.Array.t ->
  (Idx.static_symbol * Base.int Base.ref) list ->
  sync:(Base.int -> unit) ->
  Base.unit

Distributes iterated indices to workers in a round-robin fashion. All and only bindings with associated ranges are iterated, with the binding's initial value lost. Bindings without ranges remain at their initial values. sync is called after each round of calling all workers, and at the end if needed, with the number of workers called during the round.

val round_robin_dry_run : 
  num_streams:Base__Int.t ->
  (Idx.static_symbol * Base.int Base.ref) list ->
  dry_sync:(Base__Int.t -> unit) ->
  Base.unit

val set_virtual : Tn.t -> unit

val every_non_literal_on_host : Tensor.t -> Base.unit

module Lazy = Utils.Lazy

val parallel_update : 
  (module Backend
    with type buffer_ptr = 'buffer_ptr
     and type dev = 'dev
     and type event = 'event
     and type runner = 'runner) ->
  grad_updates:
    ('buffer_ptr,
      ('buffer_ptr, 'dev, 'runner, 'event) Arrayjit__Backend_intf.stream_ref)
      Arrayjit__Backend_intf.context
      BT.routine
      Base.array ->
  sgd_update:
    ('buffer_ptr,
      ('buffer_ptr, 'dev, 'runner, 'event) Arrayjit__Backend_intf.stream_ref)
      Arrayjit__Backend_intf.context
      BT.routine ->
  copy_to_merge:bool ->
  post_sync:(num_synced_devices:Base.int -> Base.unit) ->
  updaten ->
  Base.unit ->
  Base.unit

Performs one optimization step, potentially in parallel (if grad_updates are linked with different streams or devices). All jitted code must have the same bindings. Iterates over bindings with ranges, calling one of grad_updates in a round-robin fashion, and performs the following synchronization each time all grad_updates have been called:

merges all gradients into the device of grad_updates.(0),
calls sgd_update,
copies all parameters from the grad_updates.(0) device to the other devices, if needed,
calls post_sync with the number of devices synced since the previous sync.

All and only bindings with associated ranges are iterated, with the binding's initial value lost. Bindings without ranges remain at their initial values.

val get_all_suggested_streams : 
  ?max_num_streams:Base.int ->
  (module Backend
    with type buffer_ptr = 'buffer_ptr
     and type dev = 'dev
     and type event = 'event
     and type runner = 'runner) ->
  ('buffer_ptr, 'dev, 'runner, 'event) Arrayjit__Backend_intf.device_ref
    Base.Array.t
  * ('buffer_ptr, 'dev, 'runner, 'event) Arrayjit__Backend_intf.stream_ref
      Base.Array.t

val to_routine : 
  (module Backend
    with type buffer_ptr = 'buffer_ptr
     and type dev = 'dev
     and type event = 'event
     and type runner = 'runner) ->
  ('buffer_ptr,
    ('buffer_ptr, 'dev, 'runner, 'event) Arrayjit__Backend_intf.stream_ref)
    Arrayjit__Backend_intf.context ->
  ?name:Base.string ->
  Arrayjit.Indexing.unit_bindings ->
  Arrayjit.Assignments.comp ->
  ('buffer_ptr,
    ('buffer_ptr, 'dev, 'runner, 'event) Arrayjit__Backend_intf.stream_ref)
    Arrayjit__Backend_intf.context
    Arrayjit__Backend_intf.routine

type example_train_result = {

inputs : Tensor.t;
outputs : Tensor.t;
model_result : Tensor.t;
infer_callback : Base.float Base.array -> Base.float Base.array;
(*
Note: infer_callback is significantly less efficient than using the model via arrayjit.
*)
rev_batch_losses : Base.float Base.list;
rev_epoch_losses : Base.float Base.list;
learning_rates : Base.float Base.list;
used_memory : Base.int;

}

val example_train_loop : 
  ?disable_rootness_check:Base.bool ->
  seed:Base.int ->
  batch_size:Base__Int.t ->
  init_lr:Base.float ->
  ?lr_schedule:
    (batch_n:Idx.static_symbol -> step_n:Idx.static_symbol -> Tensor.t) ->
  ?copy_to_merge:bool ->
  ?max_num_streams:Base.int ->
  data_len:Base__Int.t ->
  epochs:Base__Int.t ->
  inputs:(b:Base__Int.t list -> Tensor.t) ->
  outputs:(b:Base__Int.t list -> Tensor.t) ->
  model:(Tensor.t -> Tensor.t) ->
  loss_fn:(output:Tensor.t -> expectation:Tensor.t -> Tensor.t) ->
  weight_decay:Base.float ->
  ?per_batch_callback:
    (at_batch:Base.int ->
      at_step:Base.int ->
      learning_rate:Base.Float.t ->
      batch_loss:Base.Float.t ->
      epoch_loss:float ->
      unit) ->
  ?per_epoch_callback:
    (at_step:Base.int ->
      at_epoch:int ->
      learning_rate:Base.Float.t ->
      epoch_loss:float ->
      unit) ->
  ?per_epoch_debug_streams:bool ->
  (module Backend) ->
  unit ->
  example_train_result

val forward_and_ctx : 
  ?disable_rootness_check:Base.bool ->
  (module Backend
    with type buffer_ptr = 'buffer_ptr
     and type dev = 'dev
     and type event = 'event
     and type runner = 'runner) ->
  ('buffer_ptr,
    ('buffer_ptr, 'dev, 'runner, 'event) Arrayjit__Backend_intf.stream_ref)
    Arrayjit__Backend_intf.context ->
  ?bindings:(Base.unit -> Base.unit) Idx.bindings ->
  Tensor.t ->
  ('buffer_ptr,
    ('buffer_ptr, 'dev, 'runner, 'event) Arrayjit__Backend_intf.stream_ref)
    Arrayjit__Backend_intf.context

val forward_and_forget : 
  ?disable_rootness_check:Base.bool ->
  (module Backend
    with type buffer_ptr = 'a
     and type dev = 'b
     and type event = 'c
     and type runner = 'd) ->
  ('a, ('a, 'b, 'd, 'c) Arrayjit__Backend_intf.stream_ref)
    Arrayjit__Backend_intf.context ->
  ?bindings:(Base.unit -> Base.unit) Idx.bindings ->
  Tensor.t ->
  Base.unit