002

run.sh

@@ -12,6 +12,7 @@ deepspeed run_speech_recognition_seq2seq_streaming-farsipal.py \
 	--torch_compile_mode="reduce-overhead" \
 	--torch_compile_mode="max-autotune" \
 	--logging_steps="25" \
+	--learning_rate="1e-5" \
 	--max_steps="5000" \
 	--output_dir="./" \
 	--per_device_train_batch_size="32" \
@@ -38,7 +39,7 @@ deepspeed run_speech_recognition_seq2seq_streaming-farsipal.py \
 	--do_eval \
 	--predict_with_generate \
 	--do_normalize_eval \
-	--streaming="False" \
+	--streaming="True" \
 	--use_auth_token \
 	--push_to_hub
 

run.sh.mlcommons

@@ -12,6 +12,7 @@ deepspeed run_speech_recognition_seq2seq_streaming-farsipal.py \
 	--torch_compile_mode="reduce-overhead" \
 	--torch_compile_mode="max-autotune" \
 	--logging_steps="25" \
+	--learning_rate="1e-5" \
 	--max_steps="5000" \
 	--output_dir="./" \
 	--per_device_train_batch_size="32" \

run.sh.mlcommons-streaming

@@ -0,0 +1,45 @@
+deepspeed run_speech_recognition_seq2seq_streaming-farsipal.py \
+	--deepspeed="ds_config.json" \
+	--model_name_or_path="emilios/whisper-medium-el" \
+	--dataset_name="mozilla-foundation/common_voice_11_0,google/fleurs,MLCommons/ml_spoken_words" \
+	--dataset_config_name="el,el_gr,el_wav" \
+	--language="greek" \
+	--train_split_name="train+validation,train+validation,train+validation" \
+	--eval_split_name="test,-,-" \
+	--model_index_name="Whisper medium Greek El Greco" \
+	--text_column_name="sentence,transcription,keyword" \
+	--torch_compile="True" \
+	--torch_compile_mode="reduce-overhead" \
+	--torch_compile_mode="max-autotune" \
+	--logging_steps="25" \
+	--learning_rate="1e-5" \
+	--max_steps="5000" \
+	--output_dir="./" \
+	--per_device_train_batch_size="32" \
+	--gradient_accumulation_steps="1" \
+	--per_device_eval_batch_size="16" \
+	--seed="42" \
+	--warmup_steps="500" \
+	--evaluation_strategy="steps" \
+	--eval_steps="1000" \
+	--save_strategy="steps" \
+	--save_steps="1000" \
+	--generation_max_length="225" \
+	--length_column_name="input_length" \
+	--max_duration_in_seconds="30" \
+	--freeze_feature_encoder="False" \
+	--report_to="tensorboard" \
+	--metric_for_best_model="wer" \
+	--greater_is_better="False" \
+	--load_best_model_at_end \
+	--gradient_checkpointing \
+	--fp16 \
+	--overwrite_output_dir \
+	--do_train \
+	--do_eval \
+	--predict_with_generate \
+	--do_normalize_eval \
+	--streaming="True" \
+	--use_auth_token \
+	--push_to_hub
+

runs/Dec21_11-16-44_129-146-176-120/events.out.tfevents.1671627320.129-146-176-120.3027155.0

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:0d20614f4043f3402001ae757a598a2ff189f2582afcd74ebb519420939924fb
-size 10883
+oid sha256:76c4eaee524692bbaaf5721be54975bfe624865adb19636f7a9b60cfec04955c
+size 11668