Saumya3007/en-hi-transformer-tuned

English to Hindi Transformer (Optuna Optimized)

Custom PyTorch Transformer tuned with Optuna + SuccessiveHalving (ASHA).

Results

Model	Epochs	Loss	BLEU
Baseline	100	0.1484	0.5123
Best Tuned	20	0.3500	0.5260

Best Hyperparameters

{
  "d_model": 512,
  "num_heads": 8,
  "num_enc_layers": 4,
  "num_dec_layers": 3,
  "d_ff": 2048,
  "dropout": 0.05109771787083142,
  "lr": 0.00020147922100273766,
  "batch_size": 128
}

Usage

import torch, pickle, json
from huggingface_hub import hf_hub_download

weights = hf_hub_download("Saumya3007/en-hi-transformer-tuned", "rollno_ass_4_best_model.pth")
cfg     = json.load(open(hf_hub_download("Saumya3007/en-hi-transformer-tuned", "best_config.json")))
model   = Transformer(src_vocab, tgt_vocab, **{k: cfg[k] for k in
          ['d_model','num_heads','num_enc_layers','num_dec_layers','d_ff','dropout']})
model.load_state_dict(torch.load(weights, map_location='cpu'))
model.eval()