1-create-data.sh

@@ -1,11 +0,0 @@
-#!/usr/bin/env bash
-
-set -e
-set -x
-
-# Create spm vocab
-# spm_train --input=train/in.tsv --model_prefix=vocab_spm_bpe --model_type=bpe --vocab_size=50000 --pad_id 1 --bos_id 2 --eos_id 3
-
-spm_encode --model vocab_spm_bpe.model < data/train/in.tsv > data/train.txt
-spm_encode --model vocab_spm_bpe.model < data/dev-0/in.tsv > data/valid.txt
-spm_encode --model vocab_spm_bpe.model < data/test-A/in.tsv > data/test.txt

2-preproc-classifier.sh

@@ -1,19 +0,0 @@
-#!/usr/bin/env bash
-
-set -e
-set -x
-
-TEXT=data/
-fairseq-preprocess \
-    --only-source --nwordssrc 50000 \
-    --trainpref $TEXT/train.txt \
-    --validpref $TEXT/valid.txt \
-    --destdir data-bin/classifier-spm-bpe/input0 \
-    --workers 8
-
-fairseq-preprocess \
-    --only-source \
-    --trainpref $TEXT/train/expected.tsv \
-    --validpref $TEXT/dev-0/expected.tsv \
-    --destdir data-bin/classifier-spm-bpe/label \
-    --workers 8

3-train.sh

@@ -1,35 +0,0 @@
-#!/usr/bin/env bash
-
-set -e
-set -x
-
-TOTAL_NUM_UPDATES=600_000  # Total number of training steps == 10 epoch (1 peoch = 60_000)
-WARMUP_UPDATES=24_000       # Warmup the learning rate over this many updates
-PEAK_LR=0.0001             # Peak learning rate, adjust as needed
-HEAD_NAME='he_she'         # Custom name for the classification head.
-TOKENS_PER_SAMPLE=256      # Max sequence length
-NUM_CLASSES=2              # Number of classes for the classification task.
-MAX_SENTENCES=50           # Batch size.
-UPDATE_FREQ=1              # Increase the batch size
-
-MODEL_PATH='checkpoints/lm_roberta_small/checkpoint_best.pt'
-DATA_DIR=data-bin/classifier-spm-bpe
-
-fairseq-train $DATA_DIR \
-    --restore-file "$MODEL_PATH" \
-    --fp16 --max-sentences $MAX_SENTENCES --max-positions $TOKENS_PER_SAMPLE  --update-freq $UPDATE_FREQ \
-    --max-tokens 32768 --save-dir checkpoints/lm_roberta_small_finetune \
-    --task sentence_prediction \
-    --reset-optimizer --reset-dataloader --reset-meters \
-    --required-batch-size-multiple 1 \
-    --init-token 0 --separator-token 2 \
-    --arch roberta \
-    --criterion sentence_prediction \
-    --num-classes $NUM_CLASSES \
-    --dropout 0.1 --attention-dropout 0.1 --encoder-layers 8 --encoder-embed-dim 512 --encoder-ffn-embed-dim 2048 --encoder-attention-heads 8 \
-    --weight-decay 0.1 --clip-norm 0.0 \
-    --optimizer adam --adam-betas "(0.9, 0.98)" --adam-eps 1e-06 \
-    --lr-scheduler polynomial_decay --lr $PEAK_LR --total-num-update $TOTAL_NUM_UPDATES --warmup-updates $WARMUP_UPDATES \
-    --max-epoch 10 --log-format tqdm --log-interval 1 --save-interval-updates 15000 --keep-interval-updates 5 --skip-invalid-size-inputs-valid-test \
-    --best-checkpoint-metric accuracy --maximize-best-checkpoint-metric \
-    --find-unused-parameters

4-eval.py

@@ -1,53 +0,0 @@
-#!/usr/bin/env python3
-# -*- coding: utf-8 -*-
-
-import torch
-from fairseq.models.roberta import RobertaModel
-from tqdm import tqdm
-
-if __name__ == '__main__':
-    for model_epoch in ['epoch10', 'epoch20']:
-        roberta = RobertaModel.from_pretrained(
-            model_name_or_path=f'checkpoints/classifier_roberta_small_{model_epoch}',
-            data_name_or_path='data-bin/classifier-spm-bpe',
-            sentencepiece_vocab='vocab_spm_bpe.model',
-            checkpoint_file='checkpoint_best.pt',
-            bpe='sentencepiece',
-        )
-
-        roberta.cuda()
-        roberta.eval()
-
-        max_seq = 256
-        batch_size = 15
-        pad_index = roberta.task.source_dictionary.pad()
-
-        for dir_test in ['dev-0', 'dev-1', 'test-A']:
-            lines = []
-            with open(f'data/{dir_test}/in.tsv', 'rt') as f:
-                for line in tqdm(f, desc=f'Reading {dir_test}'):
-                    line = roberta.encode(line.rstrip('\n'))[:max_seq]
-                    lines.append(line)
-
-            predictions = []
-            for i in tqdm(range(0, len(lines), batch_size), desc='Processing'):
-                batch_text = lines[i: i + batch_size]
-                # Get max length of batch
-                max_len = max([tokens.size(0) for tokens in batch_text])
-
-                # Create empty tensor with padding index
-                input_tensor = torch.LongTensor(len(batch_text), max_len).fill_(pad_index)
-                # Fill tensor with tokens
-                for i, tokens in enumerate(batch_text):
-                    input_tensor[i][:tokens.size(0)] = tokens
-
-                with torch.no_grad():
-                    raw_prediction = roberta.predict('sentence_classification_head', input_tensor)
-                    # Get probability for second class (M class)
-                    out_tensor = torch.exp(raw_prediction[:, 1])
-                    for line_prediction in out_tensor:
-                        # Get probability for first class
-                        predictions.append(line_prediction.item())
-
-            with open(f'data/{dir_test}/out-epoch={model_epoch}.tsv', 'wt') as fw:
-                fw.write('\n'.join([f'{p:.8f}' for p in predictions]))