roberta_base

0-get-models.sh

@@ -0,0 +1,8 @@
+#!/usr/bin/env bash
+
+set -e
+set -x
+
+wget https://github.com/sdadas/polish-roberta/releases/download/models/roberta_base_fairseq.zip
+
+unzip roberta_base_fairseq.zip -d roberta_base_fairseq

1-create-data.sh

@@ -0,0 +1,10 @@
+#!/usr/bin/env bash
+
+set -e
+set -x
+
+spm_encode --model=roberta_base_fairseq/sentencepiece.bpe.model < data/train/in.tsv > data/train.input0.spm
+spm_encode --model=roberta_base_fairseq/sentencepiece.bpe.model < data/dev-0/in.tsv > data/dev.input.spm
+
+cp data/dev-0/expected.tsv data/dev.label 
+cp data/train/expected.tsv data/train.label

2-preproc.sh

@@ -0,0 +1,18 @@
+#!/usr/bin/env bash
+
+set -e
+set -x
+
+fairseq-preprocess \
+    --only-source \
+    --trainpref "data/train.input0.spm" \
+    --validpref "data/dev.input0.spm" \
+    --destdir "data-bin/input0" \
+    --workers 4 --srcdict roberta_base_fairseq/dict.txt
+
+fairseq-preprocess \
+    --only-source \
+    --trainpref "data/train.label" \
+    --validpref "data/dev.label" \
+    --destdir "data-bin/label" \
+    --workers 4

3-train.py

@@ -0,0 +1,31 @@
+#!/usr/bin/env bash
+
+TOTAL_NUM_UPDATES=1000000000000000  # 10 epochs through IMDB for bsz 32
+WARMUP_UPDATES=216085   # 6 percent of the number of updates
+LR=1e-05                # Peak LR for polynomial LR scheduler.
+HEAD_NAME=hesaid        # Custom name for the classification head.
+NUM_CLASSES=2           # Number of classes for the classification task.
+MAX_SENTENCES=35        # Batch size.
+ROBERTA_PATH="roberta_base_fairseq/model.pt"
+
+fairseq-train data-bin/ \
+    --restore-file $ROBERTA_PATH \
+    --max-positions 512 \
+    --max-sentences $MAX_SENTENCES \
+    --max-tokens 8192 \
+    --task sentence_prediction \
+    --reset-optimizer --reset-dataloader --reset-meters \
+    --required-batch-size-multiple 2 \
+    --init-token 0 --separator-token 2 \
+    --arch roberta_base \
+    --criterion sentence_prediction \
+    --classification-head-name $HEAD_NAME \
+    --num-classes $NUM_CLASSES \
+    --dropout 0.1 --attention-dropout 0.1 \
+    --weight-decay 0.1 --optimizer adam --adam-betas "(0.9, 0.98)" --adam-eps 1e-06 \
+    --clip-norm 0.0 \
+    --lr-scheduler polynomial_decay --lr $LR --total-num-update $TOTAL_NUM_UPDATES --warmup-updates $WARMUP_UPDATES \
+    --max-epoch 5 --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 \
+    --update-freq 1

4-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',
+        data_name_or_path='data-bin',
+        sentencepiece_vocab='roberta_base_fairseq/sentencepiece.bpe.model',
+        checkpoint_file='checkpoint_best.pt',
+        bpe='sentencepiece',
+    )
+
+    roberta.cuda()
+    roberta.eval()
+
+    max_seq = 512
+    batch_size = 5
+    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('hesaid', 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]))