TestingMedgemma / app.py
throgletworld's picture
Upload 3 files
69ae058 verified
import gradio as gr
import torch
import numpy as np
import os
import traceback
from datetime import datetime
from transformers import WavLMModel
import torch.nn as nn
from transformers import AutoTokenizer, AutoModelForCausalLM
from huggingface_hub import login
# Log in using the Space's secret token if available
if "HF_TOKEN" in os.environ:
print("HF_TOKEN found, logging in...")
login(token=os.environ["HF_TOKEN"])
else:
print("WARNING: HF_TOKEN not found in environment variables. Access to gated models may fail.")
print(f"APP STARTUP: {datetime.now()}")
# =============================================================================
# WHY SIGMOID INSTEAD OF SOFTMAX? - A DETAILED EXPLANATION
# =============================================================================
"""
MULTI-LABEL vs MULTI-CLASS CLASSIFICATION
==========================================
Our stutter detection is a MULTI-LABEL problem:
- A single 3-second audio chunk can have MULTIPLE stutters simultaneously
- Example: Someone might have a "Block" AND a "SoundRep" in the same chunk
- Each of the 5 stutter types is INDEPENDENT of the others
SOFTMAX (❌ NOT suitable for us):
---------------------------------
- Used for MULTI-CLASS problems where classes are MUTUALLY EXCLUSIVE
- Example: "Is this image a Cat OR a Dog?" (can't be both)
- Formula: softmax(x_i) = exp(x_i) / sum(exp(x_j)) for all j
- All probabilities MUST sum to 1.0
- Problem: If we used softmax and got [0.7, 0.1, 0.1, 0.05, 0.05]:
- It would say "70% Prolongation" but FORCE other classes to be low
- We couldn't detect multiple stutters in one chunk!
SIGMOID (βœ… CORRECT for us):
----------------------------
- Used for MULTI-LABEL problems where classes are INDEPENDENT
- Each class gets its own independent probability (0 to 1)
- Formula: sigmoid(x) = 1 / (1 + exp(-x))
- Probabilities DON'T need to sum to 1
- Example output: [0.8, 0.7, 0.2, 0.1, 0.05]
- 80% chance of Prolongation
- 70% chance of Block
- Both can be detected simultaneously!
THE TRAINING & INFERENCE FLOW:
==============================
TRAINING:
---------
1. Model outputs: LOGITS (raw scores from -∞ to +∞)
Example: [2.5, -3.0, 0.1, -1.5, -2.0]
2. Loss Function: BCEWithLogitsLoss
- "WithLogits" means it applies Sigmoid INTERNALLY
- More numerically stable than separate Sigmoid + BCELoss
- Compares each prediction to each ground truth label independently
INFERENCE (this file):
----------------------
1. Model outputs: LOGITS (same as training)
Example: [2.5, -3.0, 0.1, -1.5, -2.0]
2. We manually apply Sigmoid to convert to probabilities:
probs = torch.sigmoid(logits)
Result: [0.92, 0.05, 0.52, 0.18, 0.12]
3. Apply threshold (e.g., 0.5) to each probability:
- 0.92 > 0.5 β†’ Prolongation DETECTED
- 0.05 < 0.5 β†’ Block NOT detected
- 0.52 > 0.5 β†’ SoundRep DETECTED
- etc.
4. If NO stutters detected (all below threshold):
β†’ Label the chunk as "Fluent"
THRESHOLD EXPLAINED:
====================
- Default: 0.5 (theoretically neutral, since sigmoid(0) = 0.5)
- Lower threshold (0.3-0.4): More SENSITIVE, catches more stutters, but more false positives
- Higher threshold (0.6-0.7): More STRICT, fewer false positives, but might miss subtle stutters
- The slider in the UI lets users adjust this based on their needs
- SAME threshold is applied to ALL 5 classes (simplest approach)
"""
class WaveLmStutterClassification(nn.Module):
def __init__(self, num_labels=5):
super().__init__()
self.wavlm = WavLMModel.from_pretrained("microsoft/wavlm-base")
self.hidden_size = self.wavlm.config.hidden_size
for param in self.wavlm.parameters():
param.requires_grad = False
self.classifier = nn.Linear(self.hidden_size, num_labels)
self.num_labels = num_labels
def forward(self, input_values, attention_mask=None):
outputs = self.wavlm(input_values, attention_mask=attention_mask)
hidden_states = outputs.last_hidden_state
pooled = hidden_states.mean(dim=1)
logits = self.classifier(pooled)
return logits
STUTTER_LABELS = ['Prolongation', 'Block', 'SoundRep', 'WordRep', 'Interjection']
STUTTER_DEFINITIONS = {
'Prolongation': 'Sound stretched longer than normal',
'Block': 'Complete stoppage of airflow/sound',
'SoundRep': 'Sound/syllable repetition',
'WordRep': 'Whole word repetition',
'Interjection': 'Filler words like um, uh'
}
device = "cuda" if torch.cuda.is_available() else "cpu"
print(f"Device: {device}")
wavlm_model = None
whisper_model = None
medgemma_model = None
medgemma_tokenizer = None
models_loaded = False
def load_models():
global wavlm_model, whisper_model, models_loaded, medgemma_model, medgemma_tokenizer
if models_loaded:
return True
try:
print("Loading WavLM...")
wavlm_model = WaveLmStutterClassification(num_labels=5)
checkpoint_path = "wavlm_stutter_classification_best.pth"
if os.path.exists(checkpoint_path):
checkpoint = torch.load(checkpoint_path, map_location=device, weights_only=False)
if isinstance(checkpoint, dict) and 'model_state_dict' in checkpoint:
wavlm_model.load_state_dict(checkpoint['model_state_dict'])
else:
wavlm_model.load_state_dict(checkpoint)
print("Checkpoint loaded!")
wavlm_model.to(device)
wavlm_model.eval()
print("Loading Whisper...")
import whisper
whisper_model = whisper.load_model("base", device=device)
# NOTE: We lazy load MedGemma only when requested to save startup time/VRAM
# or load it here if we have enough memory.
# For this demo, let's lazy load it in the generate function.
models_loaded = True
print("Models loaded!")
return True
except Exception as e:
print(f"Model loading error: {e}")
traceback.print_exc()
return False
def load_audio(audio_path):
print(f"Loading: {audio_path}")
try:
import librosa
waveform, sr = librosa.load(audio_path, sr=16000, mono=True)
return torch.from_numpy(waveform).float(), 16000
except Exception as e:
print(f"librosa error: {e}")
try:
import soundfile as sf
waveform, sr = sf.read(audio_path, dtype='float32')
if len(waveform.shape) > 1:
waveform = waveform.mean(axis=1)
waveform = torch.from_numpy(waveform).float()
if sr != 16000:
import torchaudio
waveform = torchaudio.transforms.Resample(sr, 16000)(waveform.unsqueeze(0)).squeeze(0)
return waveform, 16000
except Exception as e:
print(f"soundfile error: {e}")
raise Exception("Could not load audio")
# ============================================================================
# MEDGEMMA LOGIC
# ============================================================================
def load_medgemma_model():
global medgemma_model, medgemma_tokenizer
if medgemma_model is not None:
return True
print("Loading TxGemma 9B...")
try:
model_id = "google/txgemma-9b-predict"
# Use 4-bit quantization if CUDA is available to save VRAM
if device == "cuda":
from transformers import BitsAndBytesConfig
bnb_config = BitsAndBytesConfig(
load_in_4bit=True,
bnb_4bit_compute_dtype=torch.float16,
bnb_4bit_use_double_quant=True,
)
medgemma_model = AutoModelForCausalLM.from_pretrained(
model_id,
quantization_config=bnb_config,
device_map="auto"
)
else:
# CPU or MPS (load normally)
medgemma_model = AutoModelForCausalLM.from_pretrained(
model_id,
torch_dtype=torch.float32,
device_map="auto"
)
medgemma_tokenizer = AutoTokenizer.from_pretrained(model_id)
print("MedGemma Loaded!")
return True
except Exception as e:
print(f"Error loading MedGemma: {e}")
return False
def generate_medgemma_report(analysis_data, progress=gr.Progress()):
if not analysis_data:
return "⚠️ Please analyze an audio file first."
progress(0.1, desc="πŸ₯ Loading MedGemma...")
success = load_medgemma_model()
if not success:
return "❌ Failed to load MedGemma model. Please check logs."
progress(0.3, desc="πŸ“ Preparing clinical data...")
# Construct prompt
prompt = f"""You are an expert Speech-Language Pathologist (SLP) assistant.
Based on the following automated stuttering analysis data, generate a professional clinical report.
## PATIENT INFORMATION
- Audio Duration: {analysis_data['duration']:.2f} seconds
- Total Words (Est): {analysis_data['word_count']}
- Speaking Rate: {analysis_data['speaking_rate']:.1f} words/min
## TRANSCRIPTION
"{analysis_data['transcription']}"
## STUTTERING ANALYSIS RESULTS
- Total Stutter Events: {analysis_data['total_stutters']}
- Stuttering Frequency: {analysis_data['frequency']:.1f}% of chunks affected
## STUTTER TYPE DISTRIBUTION
{analysis_data['distribution_str']}
---
Based on this data, please generate:
1. **CLINICAL SUMMARY** (2-3 sentences): Overview of fluency patterns.
2. **DETAILED FINDINGS**: Elaborate on types observed (Blocks, Prolongations, Repetitions).
3. **RECOMMENDATIONS** (3 bullets): Evidence-based therapy suggestions.
Write in a professional, empathetic clinical tone suitable for patient records."""
messages = [
{"role": "system", "content": "You are an expert SLP assistant."},
{"role": "user", "content": prompt}
]
progress(0.5, desc="🧠 Generating clinical narrative...")
try:
inputs = medgemma_tokenizer.apply_chat_template(
messages,
add_generation_prompt=True,
tokenize=True,
return_dict=True,
return_tensors="pt"
).to(medgemma_model.device)
with torch.no_grad():
outputs = medgemma_model.generate(
**inputs,
max_new_tokens=800,
do_sample=True,
temperature=0.7
)
generated_text = medgemma_tokenizer.decode(outputs[0][inputs.input_ids.shape[1]:], skip_special_tokens=True)
return generated_text
except Exception as e:
return f"Error gathering report: {str(e)}"
def analyze_chunk(chunk_tensor, threshold=0.5):
with torch.no_grad():
logits = wavlm_model(chunk_tensor.unsqueeze(0).to(device))
probs = torch.sigmoid(logits).cpu().numpy()[0]
detected = [STUTTER_LABELS[i] for i, p in enumerate(probs) if p > threshold]
return detected, dict(zip(STUTTER_LABELS, probs.tolist()))
def analyze_audio(audio_input, threshold, progress=gr.Progress()):
print(f"\n=== ANALYZE CLICKED ===")
print(f"Input: {audio_input}, Type: {type(audio_input)}, Threshold: {threshold}")
progress(0, desc="πŸ”„ Starting analysis...")
if audio_input is None:
return "⚠️ Please upload an audio file first!", "", "", ""
audio_path = audio_input
if isinstance(audio_input, tuple):
import tempfile, soundfile as sf
sr, data = audio_input
f = tempfile.NamedTemporaryFile(suffix=".wav", delete=False)
sf.write(f.name, data, sr)
audio_path = f.name
if not os.path.exists(audio_path):
return f"File not found: {audio_path}", "", "", ""
print(f"File: {audio_path}, Size: {os.path.getsize(audio_path)}")
try:
progress(0.1, desc="πŸ”„ Loading models...")
if not models_loaded and not load_models():
return "❌ Failed to load models", "", "", ""
progress(0.2, desc="🎡 Loading audio file...")
waveform, sr = load_audio(audio_path)
duration = len(waveform) / sr
print(f"Duration: {duration:.1f}s")
progress(0.3, desc="βœ‚οΈ Splitting audio into chunks...")
chunk_samples = int(3.0 * sr)
stutter_counts = {l: 0 for l in STUTTER_LABELS}
timeline = []
total_chunks = (len(waveform) + chunk_samples - 1) // chunk_samples
for i, start in enumerate(range(0, len(waveform), chunk_samples)):
progress(0.3 + (0.4 * i / total_chunks), desc=f"πŸ” Analyzing chunk {i+1}/{total_chunks}...")
end = min(start + chunk_samples, len(waveform))
chunk = waveform[start:end]
if len(chunk) < chunk_samples:
chunk = torch.nn.functional.pad(chunk, (0, chunk_samples - len(chunk)))
detected, _ = analyze_chunk(chunk, threshold)
for l in detected:
stutter_counts[l] += 1
timeline.append({"time": f"{start/sr:.1f}-{end/sr:.1f}s", "detected": detected or ["Fluent"]})
progress(0.75, desc="πŸ—£οΈ Transcribing with Whisper...")
print("Running Whisper...")
transcription = whisper_model.transcribe(audio_path).get('text', '')
progress(0.9, desc="πŸ“Š Generating report...")
total = sum(stutter_counts.values())
summary = f"## βœ… Analysis Complete!\n\n**Duration:** {duration:.1f}s\n**Total Stutters Detected:** {total}\n\n### Stutter Counts:\n"
for l, c in stutter_counts.items():
emoji = "πŸ”΄" if c > 0 else "βšͺ"
summary += f"- {emoji} **{l}**: {c}\n"
timeline_md = "| Time | Detected |\n|---|---|\n"
for t in timeline[:15]:
timeline_md += f"| {t['time']} | {', '.join(t['detected'])} |\n"
if len(timeline) > 15:
timeline_md += f"\n*...and {len(timeline) - 15} more chunks*"
defs = "## πŸ“– Stutter Type Definitions\n\n"
defs += "\n".join([f"**{k}:** {v}" for k, v in STUTTER_DEFINITIONS.items()])
# Create analysis data for MedGemma
analysis_data = {
'duration': duration,
'word_count': len(transcription.split()),
'speaking_rate': (len(transcription.split())/duration) * 60 if duration > 0 else 0,
'transcription': transcription,
'total_stutters': total,
'frequency': (sum(1 for t in timeline if "Fluent" not in t['detected']) / total_chunks) * 100 if total_chunks > 0 else 0,
'distribution_str': "\n".join([f"- {k}: {v} occurrences" for k, v in stutter_counts.items() if v > 0])
}
progress(1.0, desc="βœ… Done!")
print("Done!")
return summary, transcription, timeline_md, defs, analysis_data
except Exception as e:
print(f"Error: {e}")
traceback.print_exc()
return f"Error: {e}\n\n{traceback.format_exc()}", "", "", "", None
print("Building UI...")
with gr.Blocks(title="Stutter Analysis", css="""
.loading-text {
font-size: 1.2em;
color: #666;
padding: 20px;
text-align: center;
}
""") as demo:
gr.Markdown("""
# πŸŽ™οΈ Speech Fluency Analysis System
Upload an audio file to analyze stuttering patterns using AI (WavLM + Whisper).
**Supported formats:** WAV, MP3, M4A, FLAC, OGG
""")
# Store analysis data for MedGemma
analysis_state = gr.State()
with gr.Row():
with gr.Column(scale=1):
audio = gr.Audio(label="🎀 Upload Audio", type="filepath")
threshold = gr.Slider(
minimum=0.3,
maximum=0.7,
value=0.5,
step=0.05,
label="Detection Threshold",
info="Lower = more sensitive, Higher = more strict"
)
btn = gr.Button("πŸ” Analyze Speech", variant="primary", size="lg")
gr.Markdown("*Analysis takes 30-60 seconds depending on audio length*")
with gr.Column(scale=2):
summary = gr.Markdown(value="### πŸ‘† Upload audio and click Analyze to start")
with gr.Tabs():
with gr.TabItem("πŸ“ Transcription"):
trans = gr.Markdown()
with gr.TabItem("πŸ“ˆ Timeline"):
timeline = gr.Markdown()
with gr.TabItem("πŸ“– Definitions"):
defs = gr.Markdown()
with gr.TabItem("πŸ₯ Clinical Report (MedGemma)"):
gr.Markdown("### Automatic Clinical Narrative Generation")
gr.Markdown("*Powered by Google MedGemma (HAI-DEF)*")
gen_btn = gr.Button("✨ Generate Professional Report", variant="secondary")
report_out = gr.Markdown("⚠️ Please run analysis first to generate report data.")
gr.Markdown("""
---
**Note:** The spinner will appear while processing. Please wait for analysis to complete.
""")
# The show_progress parameter shows a spinner during processing
btn.click(
fn=analyze_audio,
inputs=[audio, threshold],
outputs=[summary, trans, timeline, defs, analysis_state],
show_progress="full" # Shows loading spinner
)
gen_btn.click(
fn=generate_medgemma_report,
inputs=[analysis_state],
outputs=[report_out]
)
print("Loading models...")
load_models()
print("Launching...")
demo.queue()
demo.launch(ssr_mode=False)