changes

neuromodulation · Oct 2, 2024 · 7716b56 · 7716b56
1 parent 32f83b2
commit 7716b56
Show file tree

Hide file tree

Showing 3 changed files with 17 additions and 11 deletions.
diff --git a/requirements-dev.lock b/requirements-dev.lock
@@ -120,6 +120,7 @@ packaging==24.1
     # via matplotlib
     # via mne
     # via pooch
+    # via tensorboard
     # via transformers
 pandas==2.2.3
     # via datasets
@@ -170,7 +171,7 @@ soundfile==0.12.1
     # via repo-sophie-1
 sympy==1.13.3
     # via torch
-tensorboard==2.17.0
+tensorboard==2.18.0
     # via repo-sophie-1
 tensorboard-data-server==0.7.2
     # via tensorboard
@@ -203,6 +204,8 @@ xxhash==3.5.0
     # via evaluate
 yarl==1.13.1
     # via aiohttp
+pip==24.2
+    # via repo-sophie-1
 setuptools==75.1.0
     # via tensorboard
     # via torch
diff --git a/requirements.lock b/requirements.lock
@@ -120,6 +120,7 @@ packaging==24.1
     # via matplotlib
     # via mne
     # via pooch
+    # via tensorboard
     # via transformers
 pandas==2.2.3
     # via datasets
@@ -170,7 +171,7 @@ soundfile==0.12.1
     # via repo-sophie-1
 sympy==1.13.3
     # via torch
-tensorboard==2.17.0
+tensorboard==2.18.0
     # via repo-sophie-1
 tensorboard-data-server==0.7.2
     # via tensorboard
@@ -203,6 +204,8 @@ xxhash==3.5.0
     # via evaluate
 yarl==1.13.1
     # via aiohttp
+pip==24.2
+    # via repo-sophie-1
 setuptools==75.1.0
     # via tensorboard
     # via torch
diff --git a/speech_training_demo.py b/speech_training_demo.py
@@ -113,23 +113,23 @@ def parse_args():
     parser.add_argument(
         "--dataset_name",
         type=str,
-        default="MLCommons/peoples_speech",
+        default="hf-internal-testing/librispeech_asr_dummy", #  MLCommons/peoples_speech
         help="The name of the dataset to use (via the datasets library).",
     )
     parser.add_argument(
         "--dataset_config_names",
         nargs="+",
         type=str,
         required=False,
-        default=["default"],
+        default=["clean"],
         help="The configuration names of the dataset to use (via the datasets library).",
     )
     parser.add_argument(
         "--dataset_split_names",
         nargs="+",
         type=str,
         required=False,
-        default=["validation[2%:]", "test[:2%]"], #each about 600h (30k in total)
+        default=["validation", "test"], #each about 600h (30k in total)
         help="The names of the training data set splits to use (via the datasets library).",
     )
     parser.add_argument(
@@ -164,7 +164,7 @@ def parse_args():
     parser.add_argument(
         "--validation_split_percentage",
         type=int,
-        default=1,
+        default=10,
         help="Percentage of training data that should be used for validation if no validation is present in dataset.",
     )
     parser.add_argument(
@@ -176,7 +176,7 @@ def parse_args():
     parser.add_argument(
         "--saving_steps",
         type=int,
-        default=10000,
+        default=10,
         help="Number of steps between each logging",
     )
     parser.add_argument(
@@ -233,7 +233,7 @@ def parse_args():
     parser.add_argument(
         "--max_train_steps",
         type=int,
-        default=20000,
+        default=200,
         help="Total number of training steps to perform. If provided, overrides num_train_epochs.",
     )
     parser.add_argument(
@@ -520,7 +520,7 @@ def main():
             trust_remote_code=True, #################################################################################################################################
         )
         datasets_splits.append(dataset_split)
-        datasets_splits.to(device) ###############################################################################################################################
+        datasets_splits ###############################################################################################################################
 
     # Next, we concatenate all configurations and splits into a single training dataset
     raw_datasets = DatasetDict()
@@ -704,7 +704,7 @@ def prepare_dataset(batch):
             percent_masked = num_losses / sub_attention_mask.sum()
 
             # forward
-            outputs = model(**batch).to(device)
+            outputs = model(**batch) # .to(device)
 
             # divide loss by gradient accumulation steps since gradients
             # are accumulated for multiple backward passes in PyTorch
@@ -759,7 +759,7 @@ def prepare_dataset(batch):
                         writer.add_scalar("loss/train", float((loss * args.gradient_accumulation_steps) / num_losses), step)
                         writer.flush()
                         print("yay its logging")
-                        
+
                 progress_bar.update(1)
                 completed_steps += 1