up

all_results.json

@@ -1,14 +1,14 @@
 {
-    "epoch": 30.0,
+    "epoch": 5.0,
     "eval_loss": 23.11814308166504,
     "eval_runtime": 563.3925,
     "eval_samples": 2642,
     "eval_samples_per_second": 4.689,
     "eval_steps_per_second": 0.147,
     "eval_wer": 1.0,
-    "train_loss": 4.37666116677247,
-    "train_runtime": 28536.6396,
+    "train_loss": 1.844849973621323,
+    "train_runtime": 5604.6492,
     "train_samples": 28538,
-    "train_samples_per_second": 30.001,
-    "train_steps_per_second": 0.117
+    "train_samples_per_second": 25.459,
+    "train_steps_per_second": 0.398
 }

config.json

@@ -173,7 +173,7 @@
     "feat_extract_norm": "layer",
     "feat_proj_dropout": 0.0,
     "feat_quantizer_dropout": 0.0,
-    "final_dropout": 0.1,
+    "final_dropout": 0.0,
     "finetuning_task": null,
     "forced_bos_token_id": null,
     "forced_eos_token_id": null,
@@ -202,7 +202,7 @@
     "mask_feature_prob": 0.0,
     "mask_time_length": 10,
     "mask_time_min_masks": 2,
-    "mask_time_prob": 0.0,
+    "mask_time_prob": 0.1,
     "max_length": 20,
     "min_length": 0,
     "model_type": "wav2vec2",

create_model.py

@@ -8,7 +8,8 @@ decoder_id = "facebook/bart-large"
 
 model = SpeechEncoderDecoderModel.from_encoder_decoder_pretrained(encoder_id, decoder_id, encoder_add_adapter=True)
 model.config.encoder.feat_proj_dropout = 0.0
-model.config.encoder.mask_time_prob = 0.0
+model.config.encoder.final_dropout = 0.0
+model.config.encoder.mask_time_prob = 0.1
 model.config.decoder_start_token_id = model.decoder.config.bos_token_id
 model.config.pad_token_id = model.decoder.config.pad_token_id
 model.config.eos_token_id = model.decoder.config.eos_token_id

emissions.csv

@@ -1,2 +1,3 @@
 timestamp,experiment_id,project_name,duration,emissions,energy_consumed,country_name,country_iso_code,region,on_cloud,cloud_provider,cloud_region
 2021-12-23T04:13:13,7231458a-1d3e-45ac-82fc-95d20ece134c,codecarbon,28532.068876504898,4.1430815744945955,11.264495852350722,USA,USA,nan,Y,aws,us-east-1
+2021-12-23T12:31:06,beb31896-3df5-43c8-9163-a3d02456a623,codecarbon,5600.265459775925,0.8146651355314377,2.2149677420648115,USA,USA,nan,Y,aws,us-east-1

pytorch_model.bin

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:dba48df68903e549126e4427d6762ddae08942558823a18be5af8842e0974dd3
-size 2278376895
+oid sha256:c3128180dc6f3faf1bbf77470f8f5f25284d19d333976e4b0bc855ef517990f5
+size 2278381245

run_librispeech.sh

@@ -11,10 +11,10 @@ python -m torch.distributed.launch \
 	--preprocessing_num_workers="16" \
 	--length_column_name="input_length" \
 	--overwrite_output_dir \
-	--num_train_epochs="30" \
+	--num_train_epochs="5" \
 	--per_device_train_batch_size="4" \
 	--per_device_eval_batch_size="4" \
-	--gradient_accumulation_steps="8" \
+	--gradient_accumulation_steps="2" \
 	--generation_max_length="40" \
 	--generation_num_beams="1" \
 	--learning_rate="3e-4" \
@@ -30,4 +30,5 @@ python -m torch.distributed.launch \
 	--fp16 \
 	--group_by_length \
 	--predict_with_generate \
+	--do_lower_case \
 	--do_eval --do_train

runs/Dec23_10-49-11_ip-172-31-1-149/1640256589.3304777/events.out.tfevents.1640256589.ip-172-31-1-149.645478.1

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:fa39857b3345487112d203460efc1683d1d5c47ad9c17ad4c726875a77c8eab1
+size 4957

runs/Dec23_10-49-11_ip-172-31-1-149/events.out.tfevents.1640256589.ip-172-31-1-149.645478.0

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6cf1064c7262f6db0ec0a26ffb6db7f278d6e43449c768502b4d968cafabbe0c
+size 11608

runs/Dec23_10-57-06_ip-172-31-1-149/1640257063.6087983/events.out.tfevents.1640257063.ip-172-31-1-149.714399.1

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f9a8004194c841a35c30ccfa2b691f69fce2812780667915857670bd098314d3
+size 4957

runs/Dec23_10-57-06_ip-172-31-1-149/events.out.tfevents.1640257063.ip-172-31-1-149.714399.0

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9414a120df6d371de49275ec6a034d69e8937014b66a4302c635687686668c28
+size 360963

train_results.json

@@ -1,8 +1,8 @@
 {
-    "epoch": 30.0,
-    "train_loss": 4.37666116677247,
-    "train_runtime": 28536.6396,
+    "epoch": 5.0,
+    "train_loss": 1.844849973621323,
+    "train_runtime": 5604.6492,
     "train_samples": 28538,
-    "train_samples_per_second": 30.001,
-    "train_steps_per_second": 0.117
+    "train_samples_per_second": 25.459,
+    "train_steps_per_second": 0.398
 }

training_args.bin

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:f66d5b0ed0dac64f5d9da2e863d02120b6ad92f92ffa73b63bf2509640e268fd
+oid sha256:0ebcadf8a557cc2a21ac9306cd8a4038db6cdcad1cd7a661d41c3de1913899af
 size 3055