from scratch RoBERTa MLM + classfier

1-create-data.sh

@@ -0,0 +1,11 @@
+#!/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

@@ -0,0 +1,19 @@
+#!/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

2-preproc-mlm.sh

@@ -0,0 +1,13 @@
+#!/usr/bin/env bash
+
+set -e
+set -x
+
+TEXT=data
+fairseq-preprocess \
+    --only-source --nwordssrc 50000 \
+    --trainpref $TEXT/train.txt \
+    --validpref $TEXT/valid.txt \
+    --testpref $TEXT/test.txt \
+    --destdir data-bin/lm-spm-bpe \
+    --workers 8

3-train-mlm.sh

@@ -0,0 +1,23 @@
+#!/usr/bin/env bash
+
+set -e
+set -x
+
+TOTAL_UPDATES=200_000     # Total number of training steps == 10 epoch (1 peoch = 20_000)
+WARMUP_UPDATES=8_000     # Warmup the learning rate over this many updates
+PEAK_LR=0.0001            # Peak learning rate, adjust as needed
+TOKENS_PER_SAMPLE=256     # Max sequence length
+MAX_POSITIONS=256         # Num. positional embeddings (usually same as above)
+MAX_SENTENCES=50          # Number of sequences per batch (batch size)
+UPDATE_FREQ=1             # Increase the batch size
+
+DATA_DIR=data-bin/lm-spm-bpe
+
+fairseq-train $DATA_DIR \
+    --fp16 --task masked_lm --criterion masked_lm  --save-dir checkpoints/lm_roberta_small \
+    --arch roberta_base --sample-break-mode complete --tokens-per-sample $TOKENS_PER_SAMPLE \
+    --optimizer adam --adam-betas '(0.9,0.98)' --adam-eps 1e-6 --clip-norm 0.0 \
+    --lr-scheduler polynomial_decay --lr $PEAK_LR --warmup-updates $WARMUP_UPDATES --total-num-update $TOTAL_UPDATES \
+    --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 0.01 \
+    --max-sentences $MAX_SENTENCES --max-positions $MAX_POSITIONS --update-freq $UPDATE_FREQ \
+    --max-epoch 10 --log-format tqdm --log-interval 1 --save-interval-updates 5000 --keep-interval-updates 5 --skip-invalid-size-inputs-valid-test

4-finetune.sh

@@ -0,0 +1,35 @@
+#!/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

5-eval.py

@@ -0,0 +1,52 @@
+#!/usr/bin/env python3
+# -*- coding: utf-8 -*-
+
+import torch
+from fairseq.models.roberta import RobertaModel
+from tqdm import tqdm
+
+if __name__ == '__main__':
+    roberta = RobertaModel.from_pretrained(
+        model_name_or_path='checkpoints/lm_roberta_small_finetune',
+        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.tsv', 'wt') as fw:
+            fw.write('\n'.join([f'{p:.8f}' for p in predictions]))