From 8212994501dfff2bd42281ec796ab64e8a22ffb1 Mon Sep 17 00:00:00 2001
From: Nick White <git@njw.name>
Date: Wed, 22 Jan 2020 14:18:08 +0000
Subject: Replace traintessv4 with traintessv5 script, which was used for
 fra-engbase training (very minor edits)

---
 traintessv4.sh | 100 ---------------------------------------------------------
 1 file changed, 100 deletions(-)
 delete mode 100755 traintessv4.sh

(limited to 'traintessv4.sh')

diff --git a/traintessv4.sh b/traintessv4.sh
deleted file mode 100755
index 92da4be..0000000
--- a/traintessv4.sh
+++ /dev/null
@@ -1,100 +0,0 @@
-#!/bin/sh
-usage="Usage: $0 gtdir gtevaldir oldtraineddata trainingname"
-
-test $# -ne 4 && echo "$usage" && exit 1
-
-## Settings ##
-# This retrains the top layers
-#extra="--append_index 5"
-#netspec="[Lfx512 O1c1]"
-# This fine-tunes the existing layers (copying the existing best/eng netspec)
-extra=""
-netspec="[1,36,0,1 Ct3,3,16 Mp3,3 Lfys64 Lfx96 Lrx96 Lfx512 O1c1]"
-iterations=10000
-
-oldtraining="$3"
-name="$4"
-
-mkdir -p "$name"
-
-printf 'gtdir: %s\ngtevaldir: %s\noldtraineddata: %s\ntrainingname: %s\nnetspec: %s\niterations: %s\nextra_args: %s\n' \
-	"$1" "$2" "$3" "$4" "$netspec" $iterations "$extra" > "$name/settings"
-
-
-echo "Copying training ground truth"
-mkdir -p "$name/gt"
-find "$1" -type f -name '*tif' -o -name '*png' -o -name '*txt' -o -name '*box' | while read i; do
-	n=`basename "$i" | sed 's/\.bin\.png$/.png/g; s/\.gt\.txt$/.txt/g'`
-	cp "$i" "$name/gt/$n"
-done
-
-echo "Copying eval ground truth"
-mkdir -p "$name/eval"
-find "$2" -type f -name '*tif' -o -name '*png' -o -name '*txt' -o -name '*box' | while read i; do
-	n=`basename "$i" | sed 's/\.bin\.png$/.png/g; s/\.gt\.txt$/.txt/g'`
-	cp "$i" "$name/eval/$n"
-done
-
-echo "Making box files"
-find "$name/gt" "$name/eval" -type f -name '*txt' | while read i; do
-	b=`basename "$i" .txt`
-	d=`dirname "$i"`
-	n=""
-	test -f "$d/$b.tif" && n="$b.tif"
-	test -f "$d/$b.png" && n="$b.png"
-	test -z "$n" && echo "Skipping $i as no corresponding image found" && continue
-	test -f "$d/$b.box" && echo "Skipping $i as box file already present" && continue
-	python ~/training/generate_line_box.py -i "$d/$n" -t "$i" > "$d/$b.box" || exit 1
-done
-
-echo "Making unicharset"
-unicharset_extractor --output_unicharset "$name/gt/unicharset" --norm_mode 2 "$name/gt/"*box || exit 1
-
-echo "Making lstmf files"
-find "$name/gt" "$name/eval" -type f -name '*box' | while read i; do
-	b=`basename "$i" .box`
-	d=`dirname "$i"`
-	test -f "$d/$b.tif" && n="$b.tif"
-	test -f "$d/$b.png" && n="$b.png"
-	echo "making lstm for $d/$n"
-	tesseract "$d/$n" "$d/$b" --psm 6 lstm.train || exit 1
-done
-
-echo "Listing lstmf files"
-find "$name/gt" -type f -name '*lstmf' > "$name/gt/list"
-find "$name/eval" -type f -name '*lstmf' > "$name/eval/list"
-
-echo "Unpacking old training"
-mkdir -p "$name/orig"
-cp "$oldtraining" "$name/orig/orig.traineddata"
-combine_tessdata -u "$name/orig/orig.traineddata" "$name/orig/orig" || exit 1
-
-echo "Making complete unicharset"
-merge_unicharsets "$name/gt/unicharset" "$name/orig/orig.lstm-unicharset" "$name/unicharset" || exit 1
-
-echo "Making starter training"
-mkdir -p "$name/starter"
-curl -L -f 'https://github.com/tesseract-ocr/langdata_lstm/raw/master/radical-stroke.txt' > "$name/starter/radical-stroke.txt" || exit 1
-combine_lang_model --input_unicharset "$name/unicharset" --script_dir "$name/starter" --output_dir "$name/starter" --lang "$name" || exit 1
-
-mkdir -p "$name/checkpoint"
-
-echo "Starting training"
-lstmtraining \
-	--traineddata "$name/starter/$name/$name.traineddata" \
-	--old_traineddata "$name/orig/orig.traineddata" \
-	--continue_from "$name/orig/orig.lstm" \
-	--net_spec "$netspec" \
-	--model_output "$name/checkpoint/$name" \
-	--learning_rate 20e-4 \
-	--train_listfile "$name/gt/list" \
-	--eval_listfile "$name/eval/list" \
-	--max_iterations $iterations \
-	$extra || exit 1
-
-echo "Saving training"
-lstmtraining \
-	--stop_training \
-	--continue_from "$name/checkpoint/${name}_checkpoint" \
-	--traineddata "$name/starter/$name/$name.traineddata" \
-	--model_output "$name/$name.traineddata"
-- 
cgit v1.2.1-24-ge1ad