Provide deterministic builds #427
Merged
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Contributor
josevalim
commented
Dec 13, 2022
•
josevalim
commented
- Parameter IDs were removed
- Dropouts are completely removed from the network via a new :mode option
- Freezing traverse the nodes directly without relying on IDs
- Removed almost all usage of backend_copy(Nx.Defn.Expr)
- We use integers as cache keys after the cache is built
josevalim
commented
Dec 13, 2022
josevalim
changed the title
josevalim
commented
Dec 13, 2022
lib/axon/compiler.ex
Outdated
| ) | ||
|
|
||
| # Names are computed lazily, so compute name from current | ||
| # op and aggregate op_counts. | ||
| name = name_fn.(op_name, op_counts) | ||
| op_counts = Map.update(op_counts, op_name, 1, fn x -> x + 1 end) | ||
|
|
||
| # TODO: Hack for dropout with key, fix with a better implementation |
There was a problem hiding this comment.
Perhaps it should not be based on the name but I would say it should be based on the key. Thoughts?
josevalim
commented
Dec 13, 2022
| # Compute arguments to be forwarded and ensure `:mode` is included | ||
| # for inference/training behavior dependent functions | ||
| args = Enum.reverse(tensor_inputs) ++ [Keyword.put(opts, :mode, mode)] | ||
| args = Enum.reverse(tensor_inputs, [Keyword.put(opts, :mode, mode)]) |
There was a problem hiding this comment.
Enum.reverse(list, tail) is an efficient version of Enum.reverse(list) ++ tail.
josevalim
commented
Dec 13, 2022
lib/axon/layers.ex
Outdated
| {_, out, :train} -> | ||
| out | ||
| end) | ||
| Nx.select(mask, input / keep_prob, Nx.tensor(0, type: Nx.type(input))) |
There was a problem hiding this comment.
I removed the mode check from dropout because it is no longer relevant.
josevalim
commented
Dec 13, 2022
josevalim
commented
Dec 13, 2022
| @@ -114,7 +114,7 @@ defmodule CompilerTest do | |||
| x2 = Axon.dense(input, 64) | |||
| model = Axon.add(x1, x2) | |||
|
|
|||
| {init_fn, _predict_fn} = Axon.build(model) | |||
| {init_fn, _predict_fn} = Axon.build(model, debug: true) | |||
There was a problem hiding this comment.
We only get fn stacktraces with debug: true now.
josevalim
commented
Dec 13, 2022
lib/axon.ex
Outdated
| @@ -2722,7 +2721,8 @@ defmodule Axon do | |||
|
|
|||
| defp rnn_state(x, units, rnn_type, parent_name, state_name, initializer) do | |||
| initializer = initializer || :glorot_uniform | |||
| key = Nx.Random.key(:erlang.system_time()) |> Nx.backend_copy(Nx.Defn.Expr) | |||
| # TODO: This key should be managed by the compiler | |||
There was a problem hiding this comment.
Similar to dropout.
|
This is good to go! |
josevalim
commented
Dec 14, 2022
josevalim
commented
Dec 14, 2022
seanmor5
force-pushed
the
jv-deterministic
branch
from
66c75bf to
1da7d56
Compare
This was referenced Dec 21, 2022
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.