up

create_model.py

@@ -0,0 +1,32 @@
+import jax.numpy as jnp
+from transformers import AutoFeatureExtractor, AutoTokenizer
+from models.modeling_flax_speech_encoder_decoder import FlaxSpeechEncoderDecoderModel
+
+encoder_id = "hf-internal-testing/tiny-random-wav2vec2"
+decoder_id = "hf-internal-testing/tiny-random-bart"
+
+model = FlaxSpeechEncoderDecoderModel.from_encoder_decoder_pretrained(
+    encoder_id, decoder_id, encoder_from_pt=True, decoder_from_pt=True, encoder_add_adapter=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 = 20
+model.config.num_beams = 1
+model.config.encoder.layerdrop = 0.0
+model.config.use_cache = False
+model.config.processor_class = "Wav2Vec2Processor"
+
+# 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("./")

create_scan_model.py

@@ -0,0 +1,61 @@
+import jax.numpy as jnp
+from transformers import AutoFeatureExtractor, AutoTokenizer
+from models.modeling_flax_speech_encoder_decoder import FlaxSpeechEncoderDecoderModel
+from flax.traverse_util import flatten_dict, unflatten_dict
+
+encoder_id = "hf-internal-testing/tiny-random-wav2vec2"
+decoder_id = "hf-internal-testing/tiny-random-bart"
+
+unrolled_model = FlaxSpeechEncoderDecoderModel.from_encoder_decoder_pretrained(encoder_id, decoder_id, encoder_add_adapter=True, encoder_from_pt=True, decoder_from_pt=True, encoder_use_scan=False, decoder_use_scan=False)
+model = FlaxSpeechEncoderDecoderModel.from_encoder_decoder_pretrained(encoder_id, decoder_id, encoder_add_adapter=True, encoder_from_pt=True, decoder_from_pt=True, encoder_use_scan=True, decoder_use_scan=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"
+
+def unrolled_to_scanned(model):
+    params = model.params
+    new_enc_params = {}
+    # get the key of a scanned module
+    for k in flatten_dict(params['encoder']['encoder']['layers']['0']):
+        # stack the weights for each layer of the scanned module into one matrix
+        new_enc_params[k] = jnp.stack([flatten_dict(params['encoder']['encoder']['layers'][str(i)])[k] for i in range(model.config.encoder.num_hidden_layers)])
+    # append the correct prefix to the scanned modules' keys
+    new_enc_params = unflatten_dict({('encoder', 'layers', 'FlaxWav2Vec2EncoderLayers'): unflatten_dict(new_enc_params)})
+
+    # repeat for the decoder (note that the key 'layers' appears one index to the right than in the encoder, thus we'll treat the encoder and decoder independently for now)
+    new_dec_params = {}
+    for k in flatten_dict(params['decoder']['model']['decoder']['layers']['0']):
+        new_dec_params[k] = jnp.stack([flatten_dict(params['decoder']['model']['decoder']['layers'][str(i)])[k] for i in range(model.config.decoder.decoder_layers)])
+    new_dec_params = unflatten_dict({('model', 'decoder', 'layers', 'FlaxBartDecoderLayers'): unflatten_dict(new_dec_params)})
+
+    # combine the encoder and decoder parameters
+    new_params = {'encoder': new_enc_params, 'decoder': new_dec_params}
+    new_params = flatten_dict(new_params)
+
+    # append parameters for non-scanned modules (i.e. all modules that do not contain the key 'layers')
+    for k in flatten_dict(params):
+        if 'layers' not in k or 'adapter' in k:
+            new_params[k] = flatten_dict(params)[k]
+
+    return unflatten_dict(new_params)
+
+model.params = unrolled_to_scanned(unrolled_model)
+
+# 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 +1,36 @@
+#!/usr/bin/env bash
+python run_flax_speech_recognition_seq2seq.py \
+	--dataset_name="hf-internal-testing/librispeech_asr_dummy" \
+	--model_name_or_path="./" \
+	--dataset_config_name="clean" \
+	--train_split_name="validation" \
+	--eval_split_name="validation" \
+	--test_split_name="validation[:90%]+validation[:95%]" \
+	--output_dir="./" \
+	--dataset_cache_dir="/home/sanchitgandhi/cache/huggingface/datasets" \
+	--preprocessing_num_workers="1" \
+	--length_column_name="input_length" \
+	--overwrite_output_dir \
+	--max_steps="15" \
+	--eval_steps="5" \
+	--save_steps="5" \
+	--per_device_train_batch_size="2" \
+	--per_device_eval_batch_size="2" \
+	--logging_steps="1" \
+	--max_duration_in_seconds="15" \
+	--max_target_length="64" \
+	--generation_max_length="40" \
+	--generation_num_beams="1" \
+	--final_generation_max_length="50" \
+	--final_generation_num_beams="2" \
+	--learning_rate="3e-4" \
+	--warmup_steps="500" \
+	--text_column_name="text" \
+	--save_total_limit="1" \
+        --wandb_project="flax-wav2vec2-2-bart-dummy" \
+	--freeze_feature_encoder \
+	--predict_with_generate \
+	--do_lower_case \
+       	--do_train \
+	--do_eval \
+	--do_predict