From 5de0e2c2da9fedb47ecc59d5dcf5cfd0c4475944 Mon Sep 17 00:00:00 2001
From: Ming Ding <dm_thu@qq.com>
Date: Mon, 13 Dec 2021 15:41:33 +0000
Subject: [PATCH] v0.1.4
---
CHANGE_LOG.md | 21 +++++++++++++++++++++
setup.py | 2 +-
2 files changed, 22 insertions(+), 1 deletion(-)
diff --git a/CHANGE_LOG.md b/CHANGE_LOG.md
index 0daa8ae..9e599c8 100644
--- a/CHANGE_LOG.md
+++ b/CHANGE_LOG.md
@@ -32,4 +32,25 @@ del old['module']['word_embeddings.weight']
1. Add generation.autoregressive_sampling.evalute_perplexity
2. fix Runtime Error in skipping Nan Loss
+# 2021.12.13 v0.1.4
+1. Add non_conflict attention_fn
+2. Add Prefix-Tuning
+3. Now, you can use `kw_args['output_this_layer']` (any hooks in the transformer layers) to return values to final outputs and `kw_args['output_cross_layer']` to pass values to `kw_args` in the next layer.
+
+Examples:
+```
+def attention_fn(...some_args):
+ ...
+ kw_args['output_this_layer']['mem_kv'] = cache_kv
+ ...
+```
+This will let the key `'mem_kv'` appear in the `outputs_per_layers[i]` of `logits, *outputs_per_layers = model(...)`.
+
+```
+def attention_fn(...some_args, **kw_args):
+ ...
+ kw_args['output_cross_layer']['last_attention_map'] = attention_map
+ ...
+```
+This will let the key `'last_attention_map'` appear in the next layer's `kw_args` (all hooks).
diff --git a/setup.py b/setup.py
index f106e9f..8e38426 100644
--- a/setup.py
+++ b/setup.py
@@ -16,7 +16,7 @@ def _requirements():
setup(
name="SwissArmyTransformer",
- version='0.1.3',
+ version='0.1.4',
description="A transformer-based framework with finetuning as the first class citizen.",
long_description=Path("README.md").read_text(),
long_description_content_type="text/markdown",
--
GitLab