up

create_model.py

@@ -0,0 +1,33 @@
+import jax
+import jax.numpy as jnp
+from transformers import AutoFeatureExtractor, AutoTokenizer, FlaxSpeechEncoderDecoderModel
+
+encoder_id = "facebook/wav2vec2-large-lv60"
+decoder_id = "facebook/bart-large"
+
+model = FlaxSpeechEncoderDecoderModel.from_encoder_decoder_pretrained(encoder_id, decoder_id, encoder_add_adapter=True, decoder_from_pt=True)
+
+model.config.encoder.feat_proj_dropout = 0.0
+model.config.encoder.final_dropout = 0.0
+model.config.encoder.mask_time_prob = 0.1
+model.config.decoder_start_token_id = model.config.decoder.bos_token_id
+model.config.pad_token_id = model.config.decoder.pad_token_id
+model.config.eos_token_id = model.config.decoder.eos_token_id
+model.config.max_length = 40
+model.config.num_beams = 1
+model.config.encoder.layerdrop = 0.0
+model.config.use_cache = False
+model.config.processor_class = "Wav2Vec2Processor"
+
+# need to upcast bart-large weights from float16 to float32
+model.params = jax.tree_map(lambda x: x.astype(jnp.float32) if x.dtype != jnp.float32 else x, model.params)
+
+# check if generation works
+out = model.generate(jnp.ones((1, 2000)))
+
+model.save_pretrained("./")
+
+feature_extractor = AutoFeatureExtractor.from_pretrained(encoder_id)
+feature_extractor.save_pretrained("./")
+tokenizer = AutoTokenizer.from_pretrained(decoder_id)
+tokenizer.save_pretrained("./")

run_librispeech.sh

@@ -0,0 +1,34 @@
+#!/usr/bin/env bash
+python run_flax_speech_recognition_seq2seq.py \
+        --dataset_name="librispeech_asr" \
+        --model_name_or_path="./" \
+        --dataset_config_name="clean" \
+        --train_split_name="train.100" \
+        --eval_split_name="validation" \
+	--dataset_cache_dir="~/cache/huggingface/datasets" \
+        --output_dir="./" \
+        --preprocessing_num_workers="16" \
+        --length_column_name="input_length" \
+        --overwrite_output_dir \
+        --num_train_epochs="10" \
+        --per_device_train_batch_size="4" \
+        --per_device_eval_batch_size="4" \
+	--gradient_accumulation_steps="1" \
+        --logging_steps="25" \
+        --max_duration_in_seconds="10" \
+        --max_target_length="64" \
+        --generation_max_length="40" \
+        --generation_num_beams="1" \
+        --learning_rate="3e-4" \
+        --warmup_steps="500" \
+        --text_column_name="text" \
+        --save_total_limit="1" \
+        --freeze_feature_encoder \
+        --predict_with_generate \
+        --do_lower_case \
+        --do_eval \
+        --do_train \
+        --push_to_hub \
+        --use_auth_token \
+        --wandb_project="flax-wav2vec2-2-bart-large"	
+

wer-sweep.yaml

@@ -0,0 +1,104 @@
+command:
+- python3
+- ${program}
+- --overwrite_output_dir
+- --freeze_feature_encoder
+- --predict_with_generate
+- --do_lower_case
+- --do_train
+- --do_eval
+- ${args}
+method: random
+metric:
+  goal: minimize
+  name: eval/wer
+parameters:
+  activation_dropout:
+    distribution: log_uniform
+    max: -1.2
+    min: -3.4
+  dataset_cache_dir:
+    value: /home/sanchitgandhi/cache/huggingface/datasets
+  dataset_config_name:
+    value: clean
+  dataset_name:
+    value: librispeech_asr
+  decoder_activation_dropout:
+    distribution: log_uniform
+    max: -1.2
+    min: -3.4
+  decoder_attention_dropout:
+    distribution: log_uniform
+    max: -1.2
+    min: -3.4
+  decoder_dropout:
+    distribution: log_uniform
+    max: -1.2
+    min: -3.4
+  eval_split_name:
+    value: validation
+  eval_steps:
+    value: 500
+  feat_proj_dropout:
+    distribution: log_uniform
+    max: -1.2
+    min: -3.4
+  generation_max_length:
+    value: 40
+  generation_num_beams:
+    value: 1
+  gradient_accumulation_steps:
+    values:
+      - 2  
+      - 4
+      - 8
+  hidden_dropout:
+    distribution: log_uniform
+    max: -1.2
+    min: -3.4
+  layerdrop:
+    distribution: log_uniform
+    max: -1.2
+    min: -3.4
+  learning_rate:
+    distribution: log_uniform
+    max: -6.9
+    min: -9.2
+  length_column_name:
+    value: input_length
+  logging_steps:
+    value: 10
+  max_duration_in_seconds:
+    value: 10
+  max_grad_norm:
+    distribution: log_uniform
+    max: 0.0
+    min: -2.3
+  max_target_length:
+    value: 64
+  mixed_precision:
+    values:
+      - True
+      - False
+  model_name_or_path:
+    value: ./
+  num_train_epochs:
+    value: 10
+  output_dir:
+    value: ./output_dir
+  per_device_eval_batch_size:
+    value: 2
+  per_device_train_batch_size:
+    values:
+      - 1
+      - 2
+  preprocessing_num_workers:
+    value: 16
+  text_column_name:
+    value: text
+  train_split_name:
+    value: train.100
+  warmup_steps:
+    value: 500
+program: run_flax_speech_recognition_seq2seq.py
+project: flax-wav2vec2-2-bart-large