Checkpoint 0.6.1 restore problem

DeepSpeech
#1

I try to fine tune with the last version and two variable are missing :

(0) Not found: Key cond_1/beta1_power not found in checkpoint
[[{{node save/RestoreV2}}]]
(1) Not found: Key cond_1/beta1_power not found in checkpoint
[[{{node save/RestoreV2}}]]
[[save/RestoreV2/_1]]

Have you change tensorflow version ?

#2

No, this is still based on r1.14. How are you running the fine-tuning ?

#3 
if [ ! -f "./datacorpus/models/output_graph.pb" ]; then
EARLY_STOP_FLAG="--early_stop"
if [ "${EARLY_STOP}" = "0" ]; then
	EARLY_STOP_FLAG="--noearly_stop"
fi;


python3 -u DeepSpeech.py \
	--show_progressbar True \
	--alphabet_config_path ./datacorpus/models/alphabet.txt \
	--use_cudnn_rnn True \
	--automatic_mixed_precision True \
	--lm_binary_path ./datacorpus/lm/lm.binary \
	--lm_trie_path ./datacorpus/lm/trie \
	--feature_cache ./datacorpus/sources/feature_cache \
	--train_files ${all_train_csv} \
	--dev_files ${all_dev_csv} \
	--test_files ${all_test_csv} \
	--train_batch_size ${BATCH_SIZE} \
	--dev_batch_size ${BATCH_SIZE} \
	--test_batch_size ${BATCH_SIZE} \
	--n_hidden ${N_HIDDEN} \
	--epochs ${EPOCHS} \
	--learning_rate ${LEARNING_RATE} \
	--dropout_rate ${DROPOUT} \
	--lm_alpha ${LM_ALPHA} \
	--lm_beta ${LM_BETA} \
	${EARLY_STOP_FLAG} \
	--checkpoint_dir ./datacorpus/checkpoints/ \
	--export_dir ./datacorpus/models/ \
	--export_language "fra"

fi;
all_train_csv="$(find ./datacorpus/extracted/data/cv-fr/ -type f -name ‘*train.csv’ -printf ‘%p,’ | sed -e ‘s/,$//g’)"
all_dev_csv="$(find ./datacorpus/extracted/data/cv-fr/ -type f -name ‘*dev.csv’ -printf ‘%p,’ | sed -e ‘s/,$//g’)"
all_test_csv="$(find ./datacorpus/extracted/data/cv-fr/ -type f -name ‘*test.csv’ -printf ‘%p,’ | sed -e ‘s/,$//g’)"
EARLY_STOP_FLAG=“0”
BATCH_SIZE=“96”
N_HIDDEN=“2048”
EPOCHS=“75”
LEARNING_RATE=“0.0001”
DROPOUT=“0.20”
LM_ALPHA=“0.75”
LM_BETA=“1.85”

#4

Try and remove that flag ?

#5

Thanks lissyx it’s better.

1 Like