first commit

finetune_base.sh

@@ -0,0 +1,10 @@
+PROJECT_DIR=${HOME}"/models/t5-parliament-categorisation"
+export PYTHONPATH=${PROJECT_DIR}
+
+python3 ../../t5x/t5x/train.py \
+  --gin_search_paths="./" \
+  --gin_file="finetune_translate_base.gin" \
+  --gin.MIXTURE_OR_TASK_NAME=\"translate\" \
+  --gin.MODEL_DIR=\"gs://nb-t5x-us-central2/finetuned/v1_nynorsk_translate_mT5_base\" \
+
+

finetune_translate_base.gin

@@ -0,0 +1,39 @@
+from __gin__ import dynamic_registration
+import tasks
+
+import __main__ as train_script
+from t5.data import mixtures
+from t5x import models
+from t5x import partitioning
+from t5x import utils
+
+include "t5x/examples/t5/mt5/base.gin"
+include "t5x/configs/runs/finetune.gin"
+
+MIXTURE_OR_TASK_NAME = %gin.REQUIRED
+TASK_FEATURE_LENGTHS = {"inputs": 512, "targets": 512}
+INITIAL_CHECKPOINT_PATH = "gs://t5-data/pretrained_models/t5x/mt5_base/checkpoint_1000000"
+TRAIN_STEPS = 1_010_000  # 1000000 pre-trained steps + 10000 fine-tuning steps.
+USE_CACHED_TASKS = False
+DROPOUT_RATE = 0.1
+RANDOM_SEED = 0
+
+#Fixing a small error
+infer_eval/utils.DatasetConfig:
+  task_feature_lengths = %TASK_FEATURE_LENGTHS
+
+#Saving every 1000 steps
+utils.SaveCheckpointConfig:
+  period = 1000
+
+
+# Pere: Only necessary if we load a t5 model. We can start with an t5x model here
+# `LOSS_NORMALIZING_FACTOR`: When fine-tuning a model that was pre-trained
+# using Mesh Tensorflow (e.g. the public T5 / mT5 / ByT5 models), this should be
+# set to `pretraining batch_size` * `target_token_length`. For T5 and T5.1.1:
+# `2048 * 114`. For mT5: `1024 * 229`. For ByT5: `1024 * 189`.
+# LOSS_NORMALIZING_FACTOR = 234496
+
+# Might have to ba changed based on architecture
+# partitioning.PjitPartitioner.num_partitions = 1
+

my_metrics.py

@@ -0,0 +1,7 @@
+import sklearn.metrics
+import numpy as np
+
+def f1_macro(targets, predictions):
+      targets, predictions = np.asarray(targets).astype(str), np.asarray(predictions).astype(str)
+      return {"f1_macro": 100*sklearn.metrics.f1_score(targets, predictions, average='macro')}
+

tasks.py

@@ -0,0 +1,183 @@
+# /home/perk/mymodel/categorisation-mt5x/tasks.py
+
+import functools
+import seqio
+import my_metrics
+import tensorflow_datasets as tfds
+from t5.evaluation import metrics
+from t5.data import preprocessors
+#import my_preprocessors
+import t5
+import tensorflow.compat.v1 as tf
+
+
+
+tsv_parliament_path = {
+        "train": "gs://notram-public/finetune_datasets/parliament_speeches_1998_2016_frp_or_sv/train.tsv",
+        "validation": "gs://notram-public/finetune_datasets/parliament_speeches_1998_2016_frp_or_sv/dev.tsv",
+        "test": "gs://notram-public/finetune_datasets/parliament_speeches_1998_2016_frp_or_sv/test.tsv"
+}
+
+tsv_translate_path = {
+        "train": "gs://nb-t5x-us-central2/corpus_bokmal_nynorsk/train.tsv",
+        "validation": "gs://nb-t5x-us-central2/corpus_bokmal_nynorsk/dev.tsv",
+        "test": "gs://nb-t5x-us-central2/corpus_bokmal_nynorsk/test.tsv"
+}
+
+tsv_sentiment_path = {
+        "train": "gs://notram-public/finetune_datasets/norec_sentiment/train.tsv",
+        "validation": "gs://notram-public/finetune_datasets/norec_sentiment/dev.tsv",
+        "test": "gs://notram-public/finetune_datasets/norec_sentiment/test.tsv"
+}
+
+json_angry_tweets_path = {
+        "train": "gs://notram-public/finetune_datasets/angry_tweets/train.jsonl",
+        "validation": "gs://notram-public/finetune_datasets/angry_tweets/test.jsonl",
+        "test": "gs://notram-public/finetune_datasets/angry_tweets/test.jsonl"
+}
+
+tsv_angry_tweets_path = {
+        "train": "gs://notram-public/finetune_datasets/angry_tweets/train.tsv",
+        "validation": "gs://notram-public/finetune_datasets/angry_tweets/test.tsv",
+        "test": "gs://notram-public/finetune_datasets/angry_tweets/test.tsv"
+}
+
+
+tsv_dane_path = {
+        "train": "gs://notram-public/finetune_datasets/dane/train.tsv",
+        "validation": "gs://notram-public/finetune_datasets/dane/test.tsv",
+        "test": "gs://notram-public/finetune_datasets/dane/test.tsv"
+}
+
+tsv_dane_tokens_path = {
+        "train": "gs://notram-public/finetune_datasets/dane/train_tokens.tsv",
+        "validation": "gs://notram-public/finetune_datasets/dane/test_tokens.tsv",
+        "test": "gs://notram-public/finetune_datasets/dane/test_tokens.tsv"
+}
+
+
+tsv_dane_long_tokens_path = {
+        "train": "gs://notram-public/finetune_datasets/dane/train_long_tokens.tsv",
+        "validation": "gs://notram-public/finetune_datasets/dane/test_long_tokens.tsv",
+        "test": "gs://notram-public/finetune_datasets/dane/test_long_tokens.tsv"
+}
+
+
+vocabulary = seqio.SentencePieceVocabulary(
+                'gs://t5-data/vocabs/mc4.250000.100extra/sentencepiece.model', extra_ids=0)
+
+DEFAULT_OUTPUT_FEATURES = {
+    "inputs":
+        seqio.Feature(
+            vocabulary=vocabulary, add_eos=True),
+    "targets":
+        seqio.Feature(
+            vocabulary=vocabulary, add_eos=True)
+}
+
+def categorise_preprocessor(ds):
+  def normalize_text(text):
+    """Lowercase and remove quotes from a TensorFlow string."""
+    #text = tf.strings.regex_replace(text,"'(.*)'", r"\1")
+    ...
+    return text
+
+  def to_inputs_and_targets(ex):
+    """Map {"source": ..., "source": ...}->{"target": ..., "target": ...}."""
+    return {
+        "inputs":
+             tf.strings.join(
+                 [normalize_text(ex["source"])]),
+        "targets": 
+	    tf.strings.join(
+                 [normalize_text(ex["target"])]),
+    }
+  return ds.map(to_inputs_and_targets, 
+                num_parallel_calls=tf.data.experimental.AUTOTUNE)
+
+
+seqio.TaskRegistry.add(
+    "parliament",
+    source=seqio.TextLineDataSource(
+        split_to_filepattern=tsv_parliament_path,
+        #num_input_examples=num_nq_examples
+        ),
+    preprocessors=[
+      functools.partial(
+          t5.data.preprocessors.parse_tsv,
+          field_names=["target","source"]),
+      categorise_preprocessor,
+      seqio.preprocessors.tokenize_and_append_eos,
+    ],
+    metric_fns=[metrics.accuracy,my_metrics.f1_macro],
+    output_features=DEFAULT_OUTPUT_FEATURES,
+)   
+
+seqio.TaskRegistry.add(
+    "sentiment",
+    source=seqio.TextLineDataSource(
+        split_to_filepattern=tsv_sentiment_path,
+        #num_input_examples=num_nq_examples
+        ),
+    preprocessors=[
+      functools.partial(
+          t5.data.preprocessors.parse_tsv,
+          field_names=["target","source"]),
+      categorise_preprocessor,
+      seqio.preprocessors.tokenize_and_append_eos,
+    ],
+    metric_fns=[metrics.accuracy,my_metrics.f1_macro],
+    output_features=DEFAULT_OUTPUT_FEATURES,
+)   
+
+seqio.TaskRegistry.add(
+    "angry_tweets",
+    source=seqio.TextLineDataSource(
+        split_to_filepattern=tsv_angry_tweets_path,
+        #num_input_examples=num_nq_examples
+        ),
+    preprocessors=[
+      functools.partial(
+          t5.data.preprocessors.parse_tsv,
+          field_names=["target","source"]),
+      categorise_preprocessor,
+      seqio.preprocessors.tokenize_and_append_eos,
+    ],
+    metric_fns=[metrics.accuracy,my_metrics.f1_macro],
+    output_features=DEFAULT_OUTPUT_FEATURES,
+)   
+
+seqio.TaskRegistry.add(
+    "dane",
+    source=seqio.TextLineDataSource(
+        split_to_filepattern=tsv_dane_long_tokens_path,
+        #num_input_examples=num_nq_examples
+        ),
+    preprocessors=[
+      functools.partial(
+          t5.data.preprocessors.parse_tsv,
+          field_names=["placeholder1","placeholder2","placeholder3","target","source"]),
+      categorise_preprocessor,
+      seqio.preprocessors.tokenize_and_append_eos,
+    ],
+    metric_fns=[metrics.accuracy,my_metrics.f1_macro],
+    output_features=DEFAULT_OUTPUT_FEATURES,
+)   
+
+seqio.TaskRegistry.add(
+    "translate",
+    source=seqio.TextLineDataSource(
+        split_to_filepattern=tsv_translate_path,
+        #num_input_examples=num_nq_examples
+        ),
+    preprocessors=[
+      functools.partial(
+          t5.data.preprocessors.parse_tsv,
+          field_names=["source","target"]),
+      categorise_preprocessor,
+      seqio.preprocessors.tokenize_and_append_eos,
+    ],
+    metric_fns=[metrics.accuracy,my_metrics.f1_macro],
+    output_features=DEFAULT_OUTPUT_FEATURES,
+)   
+