up

make_graph.py

@@ -0,0 +1,20 @@
+#!/usr/bin/env python3
+import sys
+import matplotlib.pyplot as plt
+
+file_path = sys.argv[1]
+
+with open(file_path, "r") as f: 
+    lines = f.readlines()
+
+loss = []
+
+key_word = "constrast_loss: "
+for line in lines:
+    if key_word in line:
+        loss.append(line.split(key_word)[-1].split("|")[0])
+
+
+X = range(len(loss))
+plt.plot(X, loss)
+plt.savefig(file_path + "_plot.png")

run_main.sh

@@ -3,11 +3,11 @@
 accelerate launch  --config_file ./default_config.yaml ./run_pretrain_no_trainer.py \
 --output_dir="./test" \
 --max_train_steps="200000" \
---num_warmup_steps="100" \
+--num_warmup_steps="100000" \
 --gradient_accumulation_steps="4" \
---learning_rate="0.0005" \
+--learning_rate="0.0001" \
 --weight_decay="0.01" \
---max_duration_in_seconds="10.0" \
+--max_duration_in_seconds="8.0" \
 --model_name_or_path="./" \
 --dataset_name="patrickvonplaten/librispeech_local" \
 --manual_data_dir="/home/patrick/wav2vec2_reproduce" \

run_pretrain_no_trainer.py

@@ -378,8 +378,8 @@ def main():
             split="train",
         )
 
-    raw_datasets["train"] = raw_datasets["train"].select(range(128))
-    raw_datasets["validation"] = raw_datasets["validation"].select(range(16))
+#    raw_datasets["train"] = raw_datasets["train"].select(range(128))
+#    raw_datasets["validation"] = raw_datasets["validation"].select(range(16))
 
     # only normalized-inputs-training is supported
     feature_extractor = Wav2Vec2FeatureExtractor.from_pretrained(