b22ee075/en-hi-transformer-ass4

EN-HI Transformer (Assignment 4 — Best Tuned Model)

Baseline (en_to_hi_new.ipynb, Colab T4)

Metric	Value
Epochs	100
Training Time	~2 hr 1 min
Final Loss	0.0963
BLEU Score	50.13

Tuned Model

Metric	Value
Epochs	30
Training Time	53.4 min
Final Loss	0.3788
BLEU Score	73.70
Baseline matched at epoch	15

Best Hyperparameters Found by Optuna + ASHA

{
  "lr": 0.0001350650706032417,
  "batch_size": 32,
  "num_heads": 4,
  "d_ff": 4096,
  "dropout": 0.29935168747119506,
  "num_layers": 6,
  "num_epochs": 30
}

Usage

import torch, pickle
from huggingface_hub import hf_hub_download

# Download artifacts
model_path  = hf_hub_download(repo_id="b22ee075/en-hi-transformer", filename="b22ee075_ass_4_best_model.pth")
en_vocab_path = hf_hub_download(repo_id="b22ee075/en-hi-transformer", filename="en_vocab.pkl")
hi_vocab_path = hf_hub_download(repo_id="b22ee075/en-hi-transformer", filename="hi_vocab.pkl")

with open(en_vocab_path, 'rb') as f: en_vocab = pickle.load(f)
with open(hi_vocab_path, 'rb') as f: hi_vocab = pickle.load(f)

model = Transformer(src_vocab=len(en_vocab), tgt_vocab=len(hi_vocab), ...)
model.load_state_dict(torch.load(model_path, map_location='cpu'))