up

config.json

@@ -1,6 +1,5 @@
 {
   "activation_dropout": 0.0,
-  "apply_spec_augment": true,
   "architectures": [
     "Wav2Vec2ForPreTraining"
   ],
@@ -73,7 +72,6 @@
   "num_hidden_layers": 12,
   "num_negatives": 100,
   "pad_token_id": 0,
-  "proj_codevector_dim": 256,
-  "transformers_version": "4.10.0.dev0",
-  "vocab_size": 32
+  "proj_codevector_dim": 768,
+  "transformers_version": "4.10.0.dev0"
 }

ds_config_wav2vec2_zero2.json

@@ -0,0 +1,51 @@
+{
+    "fp16": {
+        "enabled": "auto",
+        "loss_scale": 0,
+        "loss_scale_window": 1000,
+        "initial_scale_power": 16,
+        "hysteresis": 2,
+        "min_loss_scale": 1
+    },
+
+    "optimizer": {
+        "type": "AdamW",
+        "params": {
+            "lr": "auto",
+            "betas": "auto",
+            "eps": "auto",
+            "weight_decay": "auto"
+        }
+    },
+
+    "scheduler": {
+        "type": "WarmupLR",
+        "params": {
+            "warmup_min_lr": "auto",
+            "warmup_max_lr": "auto",
+            "warmup_num_steps": "auto"
+        }
+    },
+
+    "zero_optimization": {
+        "stage": 2,
+        "offload_optimizer": {
+            "device": "cpu",
+            "pin_memory": true
+        },
+        "find_unused_parameters": true,
+        "allgather_partitions": true,
+        "allgather_bucket_size": 2e8,
+        "overlap_comm": true,
+        "reduce_scatter": true,
+        "reduce_bucket_size": 2e8,
+        "contiguous_gradients": true
+    },
+
+    "gradient_accumulation_steps": "auto",
+    "gradient_clipping": "auto",
+    "steps_per_print": 2000,
+    "train_batch_size": "auto",
+    "train_micro_batch_size_per_gpu": "auto",
+    "wall_clock_breakdown": false
+}

run_main.sh

@@ -1,15 +1,23 @@
 #!/usr/bin/env bash
-./run_wav2vec2_pretrain_flax.py \
-    --output_dir="./" \
-    --num_train_epochs="2" \
-		--dataset_name="patrickvonplaten/librispeech_local" \
-		--dataset_config_name="clean" \
-		--train_split_name="train" \
-    --per_device_train_batch_size="16" \
-    --per_device_eval_batch_size="16" \
-    --learning_rate="5e-3" \
-    --weight_decay="0.01" \
-    --warmup_steps="1000" \
-    --model_name_or_path="./" \
-    --pad_to_multiple_of="16384" \
-		--logging_steps="2" \
+PYTHONPATH=/home/patrick/hugging_face/transformers/src deepspeed --num_gpus 4
+./run_pretrain.py \
+--output_dir="./test" \
+--num_train_epochs="20" \
+--per_device_train_batch_size="32" \
+--per_device_eval_batch_size="32" \
+--gradient_accumulation_steps="2" \
+--save_total_limit="3" \
+--save_steps="500" \
+--logging_steps="10" \
+--learning_rate="5e-3" \
+--weight_decay="0.01" \
+--warmup_steps="3000" \
+--model_name_or_path="patrickvonplaten/wav2vec2-base-libri-100h" \
+--dataset_name="patrickvonplaten/librispeech_local" \
+--dataset_config_name="clean" \
+--train_split_name="train" \
+--preprocessing_num_workers="4" \
+--verbose_logging \
+--fp16 \
+--deepspeed ds_config_wav2vec2_zero2.json \
+# --group_by_length \

run_pretrain.py

@@ -0,0 +1 @@
+/home/patrick/hugging_face/transformers/examples/research_projects/wav2vec2/run_pretrain.py