data/create_enriched_annotated_speechcommands.py

#!/usr/bin/env python
# coding: utf-8

# # Create embeddings with the transformer library
# 
# We use the Huggingface transformers library to create an embedding for a an audio dataset
# 

# ## tldr; Play as callable functions



import datasets
from transformers import AutoFeatureExtractor, AutoModel, ASTForAudioClassification
import torch
import pandas as pd
import umap
import numpy as np

_SPLIT = "train"
_MODELNAME = "MIT/ast-finetuned-speech-commands-v2"

def __set_device():
    device = "cuda" if torch.cuda.is_available() else "cpu"   
    if device == "cuda":
        torch.cuda.empty_cache()
    return device

def __permute_model2dataset_probabilities(model_probabilities):
    """
    Model and dataset int mapping is different. Therefore, we need to permute the model's 
    logits to match up to the dataset's order.
    Due to small size of vector do this on cpu
    """
    cpu_copy = model_probabilities.to("cpu").detach().numpy()
    
    # only select model2dataset_truncated values as they map model ints to existing dataset ints
    permuted_vector = np.empty(shape=(len(labels),))
    # Problem: model output has more output classes than dataset
    # Hence, only add those outputs that are mapped
    for key in model2dataset_truncated:
        dataset_int = model2dataset_truncated[key]
        permuted_vector[dataset_int] = cpu_copy[key]

    return permuted_vector



def batch_probabilities_and_embeddings(model, feature_extractor, classifier):
    device = model.device
    
    def processing(batch):
        audios = [element["array"] for element in batch["audio"]]
        inputs = feature_extractor(raw_speech=audios, return_tensors="pt", padding=True, sampling_rate=16000).to(device)
        outputs = model(**inputs)

        embeddings = classifier(**inputs).last_hidden_state[:, 0].cpu()
        return {"logits":torch.tensor([__permute_model2dataset_probabilities(logit) for logit in outputs.logits]), "embedding": embeddings}
    return processing
    

def annotate_probabilities_and_embeddings(dataset, modelname, batched=True, batch_size= 8):

    model = ASTForAudioClassification.from_pretrained(modelname)
    feature_extractor = AutoFeatureExtractor.from_pretrained(modelname, padding=True)
    classifier = AutoModel.from_pretrained(modelname, output_hidden_states=True)
    
    device = __set_device()
    calc_outputs = batch_probabilities_and_embeddings(model.to(device), feature_extractor, classifier.to(device))
    output_dataset = dataset.map(calc_outputs, batched = batched, batch_size = batch_size)
    
    return output_dataset






def annotate_batch(model, dataset):
    device = model.device
    
    def batch_annotation(batch):
        logits = [torch.tensor(element) for element in batch["logits"]]
        probabilities_per_class = [torch.nn.functional.softmax(logit, dim=-1) for logit in logits]
        predicted_class_ids = [torch.argmax(logit).item() for logit in logits]

        # logits are already permuted to match dataset ordering -> no additional work needed

        predicted_labels = [labels[predicted_class_id] for predicted_class_id in predicted_class_ids]
        annotated_labels = [labels[element] for element in batch["label"]]
        probabilities = []
        for index, prob_per_class in enumerate(probabilities_per_class):
            probabilities.append(prob_per_class[predicted_class_ids[index]].item())
        return {"label_string": annotated_labels, "probability": probabilities, 
                "probability_vector": probabilities_per_class,
                "prediction": predicted_class_ids,
                "prediction_string": predicted_labels}

    return batch_annotation

def annotate_dataset(dataset, modelname, batched=True, batch_size=8):
    model = ASTForAudioClassification.from_pretrained(modelname)
    device = __set_device()
    
    annotate = annotate_batch(model.to(device), dataset)
    annotated_dataset = dataset.map(annotate, batched=batched, batch_size=batch_size)
 
    return annotated_dataset
    




# ## Step-by-step example on speech-commands
# 
# ### Load speech-commands from Huggingface hub

# Use validation split to evaluate model's performance on unseen data



dataset = datasets.load_dataset('speech_commands', 'v0.01', split=_SPLIT)


labels = dataset.features["label"].names
# use dict comprehension to build a dict that maps the dataset's string label to the dataset's int label
label_dict = {label: i for label, i in zip(labels, range(len(labels)))}
# {key: value for key, value in zip(keys, values)}
model = ASTForAudioClassification.from_pretrained(_MODELNAME)
# look up label from model int in the label dict and translate that label into the dataset int
model2dataset_int_conversion = {i: label_dict[model.config.id2label[i]] if model.config.id2label[i] in label_dict.keys() 
                                 else -1 for i in range(model.config.num_labels)}

model2dataset_truncated = {key:value for key,value in model2dataset_int_conversion.items() if value != -1}



# Let's have a look at all of the labels that we want to predict

print(labels)


# ### Compute probabilities and embeddings and annotate dataset


# calculate logits and embedding for each sample and annotate
dataset_annotated = annotate_probabilities_and_embeddings(dataset, _MODELNAME)


# Now annotate labels and probabilities
dataset_enriched = annotate_dataset(dataset_annotated, _MODELNAME)


# ### Reduce embeddings for faster visualization

embeddings = np.stack(np.array(dataset_enriched['embedding']))
reducer = umap.UMAP()
reduced_embedding = reducer.fit_transform(embeddings)
dataset_enriched = dataset_enriched.add_column("embedding_reduced", list(reduced_embedding))


df = dataset_enriched.to_pandas()


df.to_parquet("data/dataset_audio_" + _SPLIT + ".parquet.gzip", compression='gzip')