From 5c28ad1905f63838afbc7ee6c5ec31ff1310e3a1 Mon Sep 17 00:00:00 2001
From: Gustaf Rydholm <gustaf.rydholm@gmail.com>
Date: Thu, 9 Jun 2022 22:36:04 +0200
Subject: Move conformer config

---
 training/conf/network/conformer.yaml           | 36 ++++++++++++++++++++++++++
 training/conf/network/conformer/conformer.yaml | 27 -------------------
 2 files changed, 36 insertions(+), 27 deletions(-)
 create mode 100644 training/conf/network/conformer.yaml
 delete mode 100644 training/conf/network/conformer/conformer.yaml

(limited to 'training/conf/network')

diff --git a/training/conf/network/conformer.yaml b/training/conf/network/conformer.yaml
new file mode 100644
index 0000000..1d72dd5
--- /dev/null
+++ b/training/conf/network/conformer.yaml
@@ -0,0 +1,36 @@
+_target_: text_recognizer.networks.conformer.Conformer
+depth: 8
+num_classes: 57
+dim: &dim 144
+dim_gru: 144
+block:
+  _target_: text_recognizer.networks.conformer.ConformerBlock
+  dim: *dim
+  attn:
+    _target_: text_recognizer.networks.conformer.Attention
+    dim: *dim
+    heads: 8
+    dim_head: 64
+    mult: 4
+  ff:
+    _target_: text_recognizer.networks.conformer.Feedforward
+    dim: *dim
+    expansion_factor: 4
+    dropout: 0.1
+  conv:
+    _target_: text_recognizer.networks.conformer.ConformerConv
+    dim: *dim
+    expansion_factor: 2
+    kernel_size: 31
+    dropout: 0.1
+subsampler:
+  _target_: text_recognizer.networks.conformer.Subsampler
+  pixel_pos_embedding:
+    _target_: text_recognizer.networks.transformer.AxialPositionalEmbedding
+    dim: 64
+    shape: [6, 127]
+  channels: 64
+  height: 6
+  dim: *dim
+  depth: 3
+  dropout: 0.1
diff --git a/training/conf/network/conformer/conformer.yaml b/training/conf/network/conformer/conformer.yaml
deleted file mode 100644
index 2bb423f..0000000
--- a/training/conf/network/conformer/conformer.yaml
+++ /dev/null
@@ -1,27 +0,0 @@
----
-_target_: text_recognizer.networks.conformer.Conformer
-depth: 16
-block:
-  _target_: text_recognizer.networks.conformer.ConformerBlock
-  dim: &dim 128
-  attn:
-    _target_: text_recognizer.networks.transformer.Attention
-    dim: *dim
-    num_heads: 8
-    dim_head: 64
-    dropout_rate: 0.1
-    causal: false
-    rotary_embedding:
-      _target_: text_recognizer.networks.transformer.RotaryEmbedding
-      dim: 64
-  ff:
-    _target_: text_recognizer.networks.conformer.Feedforward
-    dim: *dim
-    expansion_factor: 4
-    dropout: 0.1
-  conv:
-    _target_: text_recognizer.networks.conformer.ConformerConv
-    dim: *dim
-    expansion_factor: 2
-    kernel_size: 31
-    dropout: 0.1
-- 
cgit v1.2.3-70-g09d2