brainstate 0.1.10__py2.py3-none-any.whl → 0.2.1__py2.py3-none-any.whl

brainstate/transform/_util.py

@@ -0,0 +1,286 @@
+# Copyright 2024 BrainX Ecosystem Limited. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+# ==============================================================================
+
+from functools import wraps
+from typing import Sequence, Tuple, Hashable
+
+from brainstate._state import StateTraceStack
+from brainstate.typing import PyTree
+from ._make_jaxpr import StatefulFunction
+
+
+def wrap_single_fun_in_multi_branches(
+    stateful_fun: StatefulFunction,
+    merged_state_trace: StateTraceStack,
+    read_state_vals: Sequence[PyTree | None],
+    return_states: bool = True,
+    cache_key: Hashable = None,
+):
+    """
+    Wrap a stateful function for use in multi-branch control flow.
+
+    This function creates a wrapper that allows a stateful function to be used
+    in control flow operations where multiple functions share state. It manages
+    state values by extracting only the states needed by this specific function
+    from a merged state trace.
+
+    Parameters
+    ----------
+    stateful_fun : StatefulFunction
+        The stateful function to be wrapped.
+    merged_state_trace : StateTraceStack
+        The merged state trace containing all states from multiple functions.
+    read_state_vals : sequence of PyTree or None
+        The original read state values for all states in the merged trace.
+    return_states : bool, default True
+        Whether to return updated state values along with the function output.
+
+    Returns
+    -------
+    callable
+        A wrapped function that can be used in multi-branch control flow.
+
+    Examples
+    --------
+    Usage in conditional execution:
+
+    .. code-block:: python
+
+        >>> import brainstate
+        >>> import jax.numpy as jnp
+        >>>
+        >>> # Create states
+        >>> state1 = brainstate.State(jnp.array([1.0]))
+        >>> state2 = brainstate.State(jnp.array([2.0]))
+        >>>
+        >>> def branch_fn(x):
+        ...     state1.value *= x
+        ...     return state1.value + state2.value
+        >>>
+        >>> # During compilation, this wrapper allows the function
+        >>> # to work with merged state traces from multiple branches
+        >>> sf = brainstate.transform.StatefulFunction(branch_fn)
+        >>> # wrapped_fn = wrap_single_fun_in_multi_branches(sf, merged_trace, read_vals)
+    """
+    state_ids_belong_to_this_fun = {id(st): st for st in stateful_fun.get_states_by_cache(cache_key)}
+
+    @wraps(stateful_fun.fun)
+    def wrapped_branch(write_state_vals, *operands):
+        # "write_state_vals" should have the same length as "merged_state_trace.states"
+        assert len(merged_state_trace.states) == len(write_state_vals) == len(read_state_vals)
+
+        # get all state values needed for this function, which is a subset of "write_state_vals"
+        st_vals_for_this_fun = []
+        for write, st, val_w, val_r in zip(merged_state_trace.been_writen,
+                                           merged_state_trace.states,
+                                           write_state_vals,
+                                           read_state_vals):
+            if id(st) in state_ids_belong_to_this_fun:
+                st_vals_for_this_fun.append(val_w if write else val_r)
+
+        # call this function
+        new_state_vals, out = stateful_fun.jaxpr_call(st_vals_for_this_fun, *operands)
+        assert len(new_state_vals) == len(st_vals_for_this_fun)
+
+        if return_states:
+            # get all written state values
+            new_state_vals = {id(st): val for st, val in zip(stateful_fun.get_states_by_cache(cache_key), new_state_vals)}
+            write_state_vals = tuple([
+                (new_state_vals[id(st)] if id(st) in state_ids_belong_to_this_fun else w_val)
+                if write else None
+                for write, st, w_val in zip(merged_state_trace.been_writen,
+                                            merged_state_trace.states,
+                                            write_state_vals)
+            ])
+            return write_state_vals, out
+        return out
+
+    return wrapped_branch
+
+
+def wrap_single_fun_in_multi_branches_while_loop(
+    stateful_fun: StatefulFunction,
+    merged_state_trace: StateTraceStack,
+    read_state_vals: Sequence[PyTree | None],
+    return_states: bool = True,
+    cache_key: Hashable = None,
+):
+    """
+    Wrap a stateful function for use in while loop control flow.
+
+    This function creates a wrapper specifically designed for while loop operations
+    where multiple functions share state. It manages state values by extracting only
+    the states needed by this specific function from a merged state trace, with
+    special handling for the loop's init_val structure.
+
+    Parameters
+    ----------
+    stateful_fun : StatefulFunction
+        The stateful function to be wrapped.
+    merged_state_trace : StateTraceStack
+        The merged state trace containing all states from multiple functions.
+    read_state_vals : sequence of PyTree or None
+        The original read state values for all states in the merged trace.
+    return_states : bool, default True
+        Whether to return updated state values along with the function output.
+
+    Returns
+    -------
+    callable
+        A wrapped function that can be used in while loop control flow.
+
+    Examples
+    --------
+    Usage in while loop operations:
+
+    .. code-block:: python
+
+        >>> import brainstate
+        >>> import jax.numpy as jnp
+        >>>
+        >>> # Create states
+        >>> counter = brainstate.State(jnp.array([0]))
+        >>> accumulator = brainstate.State(jnp.array([0.0]))
+        >>>
+        >>> def cond_fn(val):
+        ...     return counter.value < 10
+        >>>
+        >>> def body_fn(val):
+        ...     counter.value += 1
+        ...     accumulator.value += val
+        ...     return val * 2
+        >>>
+        >>> # During compilation, this wrapper allows the functions
+        >>> # to work with merged state traces in while loops
+        >>> sf_cond = brainstate.transform.StatefulFunction(cond_fn)
+        >>> sf_body = brainstate.transform.StatefulFunction(body_fn)
+        >>> # wrapped_cond = wrap_single_fun_in_multi_branches_while_loop(sf_cond, ...)
+        >>> # wrapped_body = wrap_single_fun_in_multi_branches_while_loop(sf_body, ...)
+    """
+    state_ids_belong_to_this_fun = {id(st): st for st in stateful_fun.get_states_by_cache(cache_key)}
+
+    @wraps(stateful_fun.fun)
+    def wrapped_branch(init_val):
+        write_state_vals, init_val = init_val
+        # "write_state_vals" should have the same length as "merged_state_trace.states"
+        assert len(merged_state_trace.states) == len(write_state_vals) == len(read_state_vals)
+
+        # get all state values needed for this function, which is a subset of "write_state_vals"
+        st_vals_for_this_fun = []
+        for write, st, val_w, val_r in zip(merged_state_trace.been_writen,
+                                           merged_state_trace.states,
+                                           write_state_vals,
+                                           read_state_vals):
+            if id(st) in state_ids_belong_to_this_fun:
+                st_vals_for_this_fun.append(val_w if write else val_r)
+
+        # call this function
+        new_state_vals, out = stateful_fun.jaxpr_call(st_vals_for_this_fun, init_val)
+        assert len(new_state_vals) == len(st_vals_for_this_fun)
+
+        if return_states:
+            # get all written state values
+            new_state_vals = {id(st): val for st, val in zip(stateful_fun.get_states_by_cache(cache_key), new_state_vals)}
+            write_state_vals = tuple([
+                (new_state_vals[id(st)] if id(st) in state_ids_belong_to_this_fun else w_val)
+                if write else None
+                for write, st, w_val in zip(merged_state_trace.been_writen,
+                                            merged_state_trace.states,
+                                            write_state_vals)
+            ])
+            return write_state_vals, out
+        return out
+
+    return wrapped_branch
+
+
+def wrap_single_fun(
+    stateful_fun: StatefulFunction,
+    been_writen: Sequence[bool],
+    read_state_vals: Tuple[PyTree | None],
+):
+    """
+    Wrap a stateful function for use in scan operations.
+
+    This function creates a wrapper specifically designed for scan operations.
+    It manages state values by combining written and read states, calls the
+    stateful function, and returns only the written states along with the
+    carry and output values.
+
+    Parameters
+    ----------
+    stateful_fun : StatefulFunction
+        The stateful function to be wrapped for scan operations.
+    been_writen : sequence of bool
+        Boolean flags indicating which states have been written to.
+    read_state_vals : tuple of PyTree or None
+        The original read state values for all states.
+
+    Returns
+    -------
+    callable
+        A wrapped function that can be used in scan operations with proper
+        state management.
+
+    Examples
+    --------
+    Usage in scan operations:
+
+    .. code-block:: python
+
+        >>> import brainstate
+        >>> import jax.numpy as jnp
+        >>>
+        >>> # Create states
+        >>> state1 = brainstate.State(jnp.array([0.0]))
+        >>> state2 = brainstate.State(jnp.array([1.0]))
+        >>>
+        >>> def scan_fn(carry, x):
+        ...     state1.value += x  # This state will be written
+        ...     result = carry + state1.value + state2.value  # state2 is only read
+        ...     return result, result ** 2
+        >>>
+        >>> # During compilation, this wrapper allows the function
+        >>> # to work properly in scan operations
+        >>> sf = brainstate.transform.StatefulFunction(scan_fn)
+        >>> # wrapped_fn = wrap_single_fun(sf, been_written_flags, read_values)
+        >>>
+        >>> # The wrapped function handles state management automatically
+        >>> xs = jnp.arange(5.0)
+        >>> init_carry = 0.0
+        final_carry, ys = brainstate.transform.scan(scan_fn, init_carry, xs)
+    """
+
+    @wraps(stateful_fun.fun)
+    def wrapped_fun(new_carry, inputs):
+        writen_state_vals, carry = new_carry
+        assert len(been_writen) == len(writen_state_vals) == len(read_state_vals)
+
+        # collect all written and read states
+        state_vals = [
+            written_val if written else read_val
+            for written, written_val, read_val in zip(been_writen, writen_state_vals, read_state_vals)
+        ]
+
+        # call the jaxpr
+        state_vals, (carry, out) = stateful_fun.jaxpr_call(state_vals, carry, inputs)
+
+        # only return the written states
+        writen_state_vals = tuple([val if written else None for written, val in zip(been_writen, state_vals)])
+
+        # return
+        return (writen_state_vals, carry), out
+
+    return wrapped_fun