gwelliannau/improvements & Common Voice 8

.dockerignore

@@ -0,0 +1,2 @@
+models
+homedir

README.md

@@ -1,6 +1,8 @@
 
+
 [![DOI](https://zenodo.org/badge/DOI/10.5281/zenodo.5270295.svg)](https://doi.org/10.5281/zenodo.5270295)
 
+
 # docker-wav2vec2-xlsr-ft-cy
 
 [(click here to read the README in English)](README_en.md)

inference/.dockerignore

@@ -1,2 +1,3 @@
 models
-recordings
+recordings
+data

inference/.gitignore

@@ -1,2 +1,3 @@
+data
 models
-recordings
+recordings

inference/Makefile

@@ -1,32 +1,54 @@
 default: build
 
+$(eval DEVICE = cpu)
+#$(eval DEVICE = gpu)
 
 config:
+	# to use a local model, provide the full /models/.... path for WAV2VEC2_MODEL_NAME and 
+	# leave the MODEL_VERSION blank empty string.
 	$(eval WAV2VEC2_MODEL_NAME = techiaith/wav2vec2-xlsr-ft-cy)
 	$(eval MODEL_VERSION = 21.08)
+	mkdir -p ${PWD}/data/
 
 
 build: config
-	docker build --rm -t techiaith/wav2vec2-xlsr-ft-cy \
+	docker build --rm -t techiaith/wav2vec2-xlsr-ft-cy-${USER} \
 		--build-arg WAV2VEC2_MODEL_NAME=${WAV2VEC2_MODEL_NAME} \
 		--build-arg MODEL_VERSION=${MODEL_VERSION} \
 		.
 
+run: config run-${DEVICE}
 
-run: config
-	mkdir -p ${PWD}/recordings/
-	docker run --name techiaith-wav2vec2-xlsr-ft-cy \
+run-gpu:
+	docker run --gpus all --name techiaith-wav2vec2-xlsr-ft-cy-${USER} \
 		--restart=always \
 		-it \
 		-v ${PWD}/models/:/models \
-		-v ${PWD}/recordings/:/recordings \
-		techiaith/wav2vec2-xlsr-ft-cy
+		-v ${PWD}/data/:/data \
+		techiaith/wav2vec2-xlsr-ft-cy-${USER}
 
+run-cpu:
+	docker run --name techiaith-wav2vec2-xlsr-ft-cy-${USER} \
+		--restart=always \
+		-it \
+		-v ${PWD}/models/:/models \
+		-v ${PWD}/data/:/data \
+		techiaith/wav2vec2-xlsr-ft-cy-${USER}
+
+
+fetch-test:
+	if [ ! -d "data/corpws-profi-adnabod-lleferydd" ]; then \
+	    mkdir -p data; \
+	    cd data && git clone -b fersiwn2 --single-branch https://git.techiaith.bangor.ac.uk/data-porth-technolegau-iaith/corpws-profi-adnabod-lleferydd.git; \
+	fi
 
 stop: config
-	-docker stop techiaith-wav2vec2-xlsr-ft-cy
-	-docker rm techiaith-wav2vec2-xlsr-ft-cy
+	-docker stop techiaith-wav2vec2-xlsr-ft-cy-${USER}
+	-docker rm techiaith-wav2vec2-xlsr-ft-cy-${USER}
 
 
 clean: config stop
-	-docker rmi techiaith/wav2vec2-xlsr-ft-cy
+	-docker rmi techiaith/wav2vec2-xlsr-ft-cy-${USER}
+
+purge: clean
+	sudo rm -rf models

inference/python/models.py

@@ -1,11 +1,14 @@
 import os
+import yaml
 import tarfile
 import urllib.request
 from urllib.parse import urlparse
 
 from pathlib import Path
 from tqdm import tqdm
 
+from transformers import Wav2Vec2ForCTC, Wav2Vec2Processor
+from ctcdecode import CTCBeamDecoder
 
 
 class DownloadProgressBar(tqdm):
@@ -14,40 +17,90 @@ def update_to(self, b=1, bsize=1, tsize=None):
             self.total = tsize
         self.update(b * bsize - self.n)
 
-
-def default_root_dir():
-    return os.path.join("/", "models")
-
+
+def create(model_path, revision):
 
-def download_file(models_root_dir, model_name, version, file_name):
-    # expecting model name as HuggingFace Model name e.g. techiaith/wav2vec2-xlsr-ft-cy
-    # expecting file name within the HuggingFace model git repository e.g. kenlm.tar.gz
+    cache_dir=model_path
 
-    model_file = os.path.join(models_root_dir, file_name)
-
-    if Path(model_file).is_file():
-        print ("Model file {} already downloaded.".format(model_file))
+    # initialize acoustic model...
+    #
+    if Path(model_path).is_dir():
+        # from a local directory containing our own trained model
+        print("Initiaising wav2vec2 model from local directory: %s" % model_path)
+        processor = Wav2Vec2Processor.from_pretrained(model_path)
+        model = Wav2Vec2ForCTC.from_pretrained(model_path)
     else:
-        print ("Downloading {} version {}".format(file_name, version))
-        Path(models_root_dir).mkdir(parents=True, exist_ok=True)
-
-        file_url = os.path.join("https://huggingface.co", model_name, "resolve", version, file_name)
-        download_and_extract(file_url, model_file)
+        # from the HuggingFace models repository. 
+        print("Initialising wav2vec2 model \"%s\" from HuggingFace model repository" % model_path)
+        cache_dir = os.path.join('/models', model_path)
+        processor = Wav2Vec2Processor.from_pretrained(model_path, cache_dir=cache_dir, revision=revision)
+        model = Wav2Vec2ForCTC.from_pretrained(model_path, cache_dir=cache_dir, revision=revision)
+
+    # initialize language model...
+    #
+    targz_file_path=os.path.join(cache_dir, "kenlm.tar.gz")
+    if not Path(targz_file_path).is_file():
+        print ("Downloading kenlm language model version {}".format(revision))
+        try:
+            file_url = os.path.join("https://huggingface.co", model_path, "resolve", revision, 'kenlm.tar.gz')
+            download(file_url, os.path.join(cache_dir, targz_file_path))
+        except Exception as e:
+            print (e)
+
+    if not Path(os.path.join(cache_dir, "config_ctc.yaml")).is_file():
+        if Path(targz_file_path).is_file():
+            extract(targz_file_path)
 
-    return models_root_dir
+    if Path(os.path.join(cache_dir, "config_ctc.yaml")).is_file():
+        with open(os.path.join(cache_dir, "config_ctc.yaml"), 'r') as config_file:
+            ctc_lm_params=yaml.load(config_file, Loader=yaml.FullLoader)
+
+    #
+    vocab=processor.tokenizer.convert_ids_to_tokens(range(0, processor.tokenizer.vocab_size))
+    space_ix = vocab.index('|')
+    vocab[space_ix]=' '
 
+    ctcdecoder = CTCBeamDecoder(vocab, 
+        model_path='', 
+        alpha=0,
+        beta=0,
+        cutoff_top_n=40,
+        cutoff_prob=1.0,
+        beam_width=100,
+        num_processes=4,
+        blank_id=processor.tokenizer.pad_token_id,
+        log_probs_input=True
+        )
 
-def download_and_extract(file_url, output_file_path):
+    kenlm_ctcdecoder=None
+    if Path(os.path.join(cache_dir, "lm.binary")).is_file():        
+        if ctc_lm_params:
+            print ("Initializing KenLM language model...")
+            kenlm_ctcdecoder = CTCBeamDecoder(vocab,
+                model_path=os.path.join(cache_dir, "lm.binary"),
+                alpha=ctc_lm_params['alpha'],
+                beta=ctc_lm_params['beta'],
+                cutoff_top_n=40,
+                cutoff_prob=1.0,
+                beam_width=100,
+                num_processes=4,
+                blank_id=processor.tokenizer.pad_token_id,
+                log_probs_input=True
+            )
 
-    #
+    return processor, model, vocab, ctcdecoder, kenlm_ctcdecoder
+
+
+def download(file_url, output_file_path):
     with DownloadProgressBar(unit='B', unit_scale=True, miniters=1, desc=file_url.split('/')[-1]) as t:
         urllib.request.urlretrieve(file_url, filename=output_file_path, reporthook=t.update_to)
 
+def extract(targz_file_path):
     # extract.
-    if output_file_path.endswith(".tar.gz"):
+    if targz_file_path.endswith(".tar.gz"):
         print ("Extracting...")
-        model_dir = Path(output_file_path).parent.absolute()
-        tar = tarfile.open(output_file_path, "r:gz")
+        model_dir = Path(targz_file_path).parent.absolute()
+        tar = tarfile.open(targz_file_path, "r:gz")
         tar.extractall(model_dir)
         tar.close()
 

inference/python/requirements.txt

@@ -3,11 +3,16 @@ numpy
 wave
 jiwer
 webrtcvad
-transformers
+transformers==4.9.2
 tqdm==4.61.0
 pyyaml==5.4.1
 torchaudio==0.7.2
 torch==1.7.1
+datasets
 librosa
 srt
-praatio<5
+praatio<5
+pydub
+pandas
+python_speech_features
+scipy