update sweep params

audio_classification_scripts/run_sweep.yaml

@@ -3,11 +3,13 @@ command:
   - ${program}
   - --load_best_model_at_end
   - --fp16
+  - --fp16_full_eval
   - --do_train
   - --do_eval
   - --trust_remote_code
   - --overwrite_output_dir
   - --ignore_mismatched_sizes
+  - --gradient_checkpointing
   - ${args}
 method: grid
 metric:
@@ -18,7 +20,6 @@ parameters:
     values:
       - facebook/mms-lid-126
       - openai/whisper-large-v3
-      - facebook/w2v-bert-2.0
   train_dataset_name:
     value: sanchit-gandhi/vctk+facebook/voxpopuli+sanchit-gandhi/edacc
   train_dataset_config_name:
@@ -44,7 +45,7 @@ parameters:
   lr_scheduler_type:
     value: constant_with_warmup
   max_length_seconds:
-    value: 20
+    value: 10  # give some data diversity for longer audio samples
   min_length_seconds:
     value: 5
   attention_mask:
@@ -52,15 +53,15 @@ parameters:
   warmup_steps:
     value: 50
   max_steps:
-    value: 1000
+    value: 2000
   per_device_train_batch_size:
     value: 32
   per_device_eval_batch_size:
-    value: 32
+    value: 16
   preprocessing_num_workers:
     value: 16
   dataloader_num_workers:
-    value: 8
+    value: 4
   logging_strategy:
     value: steps
   logging_steps:
@@ -68,17 +69,15 @@ parameters:
   evaluation_strategy:
     value: steps
   eval_steps:
-    value: 1000
+    value: 2000
   save_strategy:
     value: steps
   save_steps:
-    value: 1000
+    value: 2000
   metric_for_best_model:
     value: accuracy
   freeze_base_model:
-    values:
-      - true
-      - false
+    value: false
   push_to_hub:
     value: false
 program: run_audio_classification.py