web_app/debug_model.py

#!/usr/bin/env python
import sys
sys.path.append('/home/fwl/src/textmining')

from web_app.model import CEFRModel
import torch

print("Loading model...")
model = CEFRModel(model_path='runs/metric-proto-k3/metric_proto.pt')

# Test simple sentence
sentences = ["Jag är bra."]
print(f"\nTesting: {sentences}")

# Tokenize
encoded = model.tokenize(sentences)
input_ids = encoded["input_ids"].to(model.device)
attention_mask = encoded["attention_mask"].to(model.device)

print(f"Input shape: {input_ids.shape}")
print(f"Device: {model.device}")

# Predict
with torch.no_grad():
    logits = model.model(input_ids, attention_mask)["logits"]
    print(f"Logits shape: {logits.shape}")
    print(f"Logits: {logits}")

    probs = torch.softmax(logits, dim=1)
    print(f"Probs shape: {probs.shape}")
    print(f"Probs: {probs}")

    predictions = torch.argmax(logits, dim=1)
    print(f"Predictions: {predictions}")

    # Test different ways to extract confidence
    cpu_probs = probs.cpu()
    for i, pred in enumerate(predictions.cpu().numpy()):
        print(f"\nSentence {i}: '{sentences[i]}'")
        print(f"  Predicted class: {pred}")
        print(f"  Predicted level: {model.id_to_label[pred]}")
        print(f"  Method 1 - probs[i][pred]: {probs[i][pred].item()}")
        print(f"  Method 2 - cpu_probs[i][pred]: {cpu_probs[i][pred].item()}")
        print(f"  Method 3 - float(cpu_probs[i][pred].item()): {float(cpu_probs[i][pred].item())}")

# Test using predict_batch
print("\n" + "="*60)
print("Using predict_batch method:")
results = model.predict_batch(sentences)
for sent, (level, conf) in zip(sentences, results):
    print(f"  {level} ({conf*100:.1f}%): {sent}")

# Test using predict_sentence
print("\n" + "="*60)
print("Using predict_sentence method:")
level, conf = model.predict_sentence(sentences[0])
print(f"  {level} ({conf*100:.1f}%): {sentences[0]}")