    
    Model & Performance Reference
  

💻 1 · Computational Cost

This model is a single-head Transformer trained for binary text classification. Its computational footprint is intentionally lightweight:

PARAMETERS

~1–5 M

typical small Transformer

MEMORY (RAM)

~50–200 MB

model weights + TF runtime

HARDWARE

CPU only

GPU not required at inference

MAX SEQ LEN

80 tokens

O(n²) attention applies

      ℹ️  Self-attention complexity scales as O(n²·d) where n = sequence length and d = embedding dimension.
      At n=80 this is negligible — the dominant cost is the first inference call due to TF graph compilation
      (TensorFlow's XLA warm-up). Subsequent calls are significantly faster.
    

🌐 2 · Latency

Latency is the total wall-clock time from receiving user input to returning the result — it includes tokenization, padding, model forward-pass, and HTML rendering.

1st REQUEST (cold)

500 ms – 3 s

TF graph compilation

WARM REQUESTS

20 – 80 ms

CPU inference, typical

WITH GPU

< 10 ms

if CUDA is available

      ℹ️  Network latency (Gradio UI ↔ Python backend) adds ~5–15 ms on localhost.
      On Hugging Face Spaces or remote servers, add 50–150 ms round-trip depending on geography.
    

⚡ 3 · Inference Time

Inference time measures only the model.predict() call — the neural network forward pass itself. This is the pure compute cost and excludes tokenization, Gradio overhead, and network round-trips. Each run shows the live measured value above in the result card.

BATCH SIZE

1 sample

single-text inference

TYPICAL RANGE

5 – 50 ms

warm, CPU, seq=80

VARIABILITY

±10–30%

OS scheduling, cache state

      ℹ️  First-call inference is slower because TensorFlow traces and compiles the computation graph
      (XLA/JIT). After the first call the compiled graph is cached in memory, making all
      subsequent inferences significantly faster and more consistent.
    

⚠️ 4 · When Predictions May Vary or Be Unreliable

🔀

MIXED SENTIMENT

Text containing both positive and negative signals (e.g., "পণ্যটি সুন্দর কিন্তু দাম বেশি") confuses the binary classifier. Confidence will hover near 50 % and may flip with small edits.

🔤

OOV / RARE VOCABULARY

Words not seen during training are mapped to an unknown token <UNK>. Heavy use of slang, dialect, transliteration, or technical jargon will degrade accuracy.

🌐

CODE-SWITCHING (BANGLISH)

Mixing Bangla script with English words or Roman-script Bangla causes many tokens to be OOV, making predictions unreliable. Use pure Unicode Bangla for best results.

📏

VERY SHORT OR VERY LONG TEXT

Single-word inputs lack context. Texts exceeding 80 tokens are silently truncated — sentiment expressed only in the trailing portion will be ignored entirely.

😏

SARCASM & IRONY

The model has no pragmatic understanding. Ironic phrases like "হ্যাঁ, অসাধারণ সার্ভিস!" (sarcastic) are likely classified as Positive because the surface tokens are positive.

🎭

DOMAIN SHIFT

If the training data consisted mainly of product reviews, the model may perform poorly on political commentary, news text, or social media posts with different linguistic patterns.

🎲

DROPOUT AT INFERENCE (if training=True)

Dropout layers are disabled during inference (training=False), so results are deterministic. If you call the model with training=True, outputs will differ randomly on each run.