Upload assessment

.gitattributes

@@ -33,3 +33,4 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 *.zip filter=lfs diff=lfs merge=lfs -text
 *.zst filter=lfs diff=lfs merge=lfs -text
 *tfevents* filter=lfs diff=lfs merge=lfs -text
+*.pdf filter=lfs diff=lfs merge=lfs -text

Article 11 Hidden Technical Debt in Machine Learning Systems.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1a67da09a8bd5ba9a3577176e30aa2fbd88534e6baf0bc31522b4999f643d2a1
+size 165614

Article 5 A Comprehensive Survey on Applications of Transformers for Deep Learning Tasks.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6d8187b33070ee9d6ccdca29bbcd9494875099927ff8c66d45b210dc612eae7a
+size 3080026

Article 7 Efficient Estimation of Word Representations in Vector Space.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a44d7e22d2005752271c9cc1929c6462d4c8270916b063977992a883e3a54362
+size 228716

app.py

@@ -1,7 +1,130 @@
+# https://huggingface.co/spaces/manuel-calzolari/assessment3_part2
+
+# Import modules
+import re
+import fitz
+import scipy
+import torch
+from transformers import pipeline
 import gradio as gr
 
-def greet(name):
-    return "Hello " + name + "!!"
+# Path of the audio file to save
+TTS_AUDIO_PATH = "tts.wav"
+
+
+def extract_text(pdf_path):
+    """
+    Function to extract text from PDF.
+    """
+    with fitz.open(pdf_path) as pdf:
+        text = ""
+        for i, page in enumerate(pdf, start=1):
+            text += page.get_text()
+    return text
+
+
+def get_abstract(text):
+    """
+    Function to get the abstract:
+    - Remove the text before the abstract
+    - Remove the text after the abstract (after the next title)
+    - Remove new lines
+
+    This works for some tested PDFs but obviously may not work with every
+    possible layout.
+    """
+    abstract = re.split(r"\n(?:Abstract|ABSTRACT)\n", text)[1]
+    abstract = re.split(r"\n[1A-Z∗][a-zA-Z @:,.{}]{,96}\n", abstract)[0]
+    abstract = abstract.replace("-\n", "").replace("\n", " ")
+    return abstract
+
+
+def generate_summary(abstract):
+    """
+    Function to generate the summary.
+    Use a model trained to generate one-line summary based on abstract of papers
+    See: https://huggingface.co/snrspeaks/t5-one-line-summary
+    """
+    summarizer = pipeline(
+        "summarization",
+        model="snrspeaks/t5-one-line-summary",
+        device="cuda:0" if torch.cuda.is_available() else "cpu",
+    )
+
+    summary = summarizer(
+        abstract,
+        max_length=64,  # Maximum length of the summary
+    )[0]["summary_text"]
+
+    # If it's not already there, add a full stop at the end of the summary
+    summary = summary if summary.endswith(".") else summary + "."
+
+    return summary
+
+
+def generate_speech(summary):
+    """
+    Function to generate the speech (TTS model).
+    See: https://huggingface.co/suno/bark-small
+    Note 1: I get some PyTorch warnings but it seems to work.
+    Note 2: Sometimes (not always) this TTS model adds spurious sounds or words
+    at the end (or more rarely at the beginning) of the speech related to the
+    text being provided.
+    """
+    synthesiser = pipeline(
+        "text-to-speech",
+        model="suno/bark-small",
+        device="cuda:0" if torch.cuda.is_available() else "cpu",
+    )
+
+    speech = synthesiser(
+        summary,
+        forward_params={"do_sample": True},  # From the bark-small usage example
+    )
+
+    return speech
+
+
+def save_speech(speech, audio_path):
+    """
+    Function to save the speech to a WAV file (from the bark-small usage example)
+    """
+    scipy.io.wavfile.write(
+        audio_path,
+        rate=speech["sampling_rate"],
+        data=speech["audio"].T,  # Transpose to get shape (n_samples, n_channels)
+    )
+
+
+def synthesis(pdf_path):
+    try:
+        # Extract text from PDF
+        text = extract_text(pdf_path)
+        # Get the abstract
+        abstract = get_abstract(text)
+    except:
+        return "ERROR: ABSTRACT NOT FOUND!!!", None
+    # Generate the summary
+    summary = generate_summary(abstract)
+    # Generate the speech of the summary
+    speech = generate_speech(summary)
+    # Save the speech to a file
+    save_speech(speech, TTS_AUDIO_PATH)
+
+    return summary, TTS_AUDIO_PATH
+
 
-iface = gr.Interface(fn=greet, inputs="text", outputs="text")
-iface.launch()
+# Build and launch the app
+summary_tts = gr.Interface(
+    fn=synthesis,
+    inputs=gr.File(type="filepath", file_types=[".pdf"], label="Upload PDF with abstract"),
+    outputs=[gr.Text(label="Summary"), gr.Audio(label="Summary TTS")],
+    title="PDF voice abstract summarization",
+    description="Upload a PDF with an abstract and this app will summarize the abstract in one sentence and read the summary aloud.",
+    examples=[
+        "Article 11 Hidden Technical Debt in Machine Learning Systems.pdf",
+        "Article 7 Efficient Estimation of Word Representations in Vector Space.pdf",
+        "Article 5 A Comprehensive Survey on Applications of Transformers for Deep Learning Tasks.pdf",
+    ],
+  )
+summary_tts.launch()

requirements.txt

@@ -0,0 +1,5 @@
+PyMuPDF
+transformers
+scipy
+torch
+# gradio is automatically made available in Hugging Face Spaces