Training in progress, step 1000

config.json

@@ -34,7 +34,7 @@
   "num_mel_bins": 80,
   "pad_token_id": 50257,
   "scale_embedding": false,
-  "torch_dtype": "float32",
+  "torch_dtype": "float16",
   "transformers_version": "4.26.0.dev0",
   "use_cache": true,
   "vocab_size": 51865

pytorch_model.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:986bef39a80f7f6f66fbf34b6c0b8147ebdbfc17d957ffb9d202ac9a8a6c0234
+size 3086784957

runs/Dec17_11-49-41_193-122-155-205/1671277819.4203665/events.out.tfevents.1671277819.193-122-155-205.97469.1

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8f400b63f9c9eda8b5b38b5cae4abc7db4e25d599938e2140dbd33aad975b1ff
+size 5881

runs/Dec17_11-49-41_193-122-155-205/events.out.tfevents.1671277819.193-122-155-205.97469.0

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0de2ba200c4c5c99ccb703e53542723ea75f970ee12d321cdd6fbfff34929f62
+size 10867

start_train.sh

@@ -13,12 +13,12 @@ deepspeed run_speech_recognition_seq2seq_streaming.py \
 	--model_index_name="Whisper Large V2 Finnish" \
 	--max_steps="30000" \
 	--output_dir="./" \
-	--per_device_train_batch_size="64" \
+	--per_device_train_batch_size="32" \
 	--per_device_eval_batch_size="16" \
-    --gradient_accumulation_steps="2" \
+    --gradient_accumulation_steps="4" \
 	--logging_steps="25" \
 	--learning_rate="3e-6" \
-	--warmup_steps="500" \
+	--warmup_steps="1000" \
 	--evaluation_strategy="steps" \
 	--eval_steps="1000" \
 	--save_strategy="steps" \

training_args.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:57637a1aea2d4e060ece70ed0646f8fe72e7dc317260229ab32b7dc0ab1823bf
+size 4731