first deuncaser submit

README.md

@@ -1 +1 @@
-Placeholder for a model to come...
+Private sample code for running categorisation on the mT5X

finetune_deuncaser_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": 256, "targets": 64}
+INITIAL_CHECKPOINT_PATH = %gin.REQUIRED
+TRAIN_STEPS = %gin.REQUIRED  # 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 +1,207 @@
+# /home/perk/mymodel/categorisation-mt5x/tasks.py
+
+import functools
+import seqio
+import tensorflow_datasets as tfds
+from t5.evaluation import metrics
+import my_metrics
+from t5.data import preprocessors
+import t5
+import tensorflow.compat.v1 as tf
+
+tsv_path = {
+        "train": "gs://eu-jav-t5x/corpus/labeled/datasetA_train_3categories.tsv",
+        "validation": "gs://eu-jav-t5x/corpus/labeled/datasetA_dev_3categories.tsv",
+        "test": "gs://eu-jav-t5x/corpus/labeled/ datasetA_test_3categories.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)
+
+
+def categorise_fulltext_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 fulltext(t):
+    if t=="0":
+        t="il testo è favorevole alla vaccinazione"
+    elif t=="1":
+        t="il testo è neutro rispetto alla vaccinazione"
+    elif t=="2":
+        t="is testo è sfavorevole alla vaccinazione"
+    return t
+
+  def to_inputs_and_targets(ex):
+    """Map {"source": ..., "source": ...}->{"target": ..., "target": ...}."""
+    return {
+        "inputs":
+             tf.strings.join(
+                 [normalize_text(ex["source"])]),
+        "targets": 
+	    tf.strings.join(
+                 [fulltext(normalize_text(ex["target"]))]),
+    }
+  return ds.map(to_inputs_and_targets, 
+                num_parallel_calls=tf.data.experimental.AUTOTUNE)
+
+
+def categorise_fulltext_word_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 fulltext(t):
+    if t=="0":
+        t="promozionale"
+    elif t=="1":
+        t="neutro"
+    elif t=="2":
+        t="scoraggiante"
+    return t
+
+  def to_inputs_and_targets(ex):
+    """Map {"source": ..., "source": ...}->{"target": ..., "target": ...}."""
+    return {
+        "inputs":
+             tf.strings.join(
+                 [normalize_text(ex["source"])]),
+        "targets": 
+	    tf.strings.join(
+                 [fulltext(normalize_text(ex["target"]))]),
+    }
+  return ds.map(to_inputs_and_targets, 
+                num_parallel_calls=tf.data.experimental.AUTOTUNE)
+
+
+
+def categorise_binary_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 fulltext(t):
+    if t=="0":
+        t="1"
+    elif t=="1":
+        t="1"
+    elif t=="2":
+        t="2"
+    return t
+
+  def to_inputs_and_targets(ex):
+    """Map {"source": ..., "source": ...}->{"target": ..., "target": ...}."""
+    return {
+        "inputs":
+             tf.strings.join(
+                 [normalize_text(ex["source"])]),
+        "targets": 
+	    tf.strings.join(
+                 [fulltext(normalize_text(ex["target"]))]),
+    }
+  return ds.map(to_inputs_and_targets, 
+                num_parallel_calls=tf.data.experimental.AUTOTUNE)
+
+
+
+seqio.TaskRegistry.add(
+    "classify_tweets",
+    source=seqio.TextLineDataSource(
+        split_to_filepattern=tsv_path,
+        #num_input_examples=num_nq_examples
+        ),
+    preprocessors=[
+      functools.partial(
+          t5.data.preprocessors.parse_tsv,
+          field_names=["annotator1","annotator2","annotator3","target","source","id"]),
+      categorise_preprocessor,
+      seqio.preprocessors.tokenize_and_append_eos,
+    ],
+    metric_fns=[metrics.accuracy,my_metrics.f1_macro],
+    output_features=DEFAULT_OUTPUT_FEATURES,
+)
+
+seqio.TaskRegistry.add(
+    "classify_tweets_fulltext",
+    source=seqio.TextLineDataSource(
+        split_to_filepattern=tsv_path,
+        #num_input_examples=num_nq_examples
+        ),
+    preprocessors=[
+      functools.partial(
+          t5.data.preprocessors.parse_tsv,
+          field_names=["annotator1","annotator2","annotator3","target","source","id"]),
+      categorise_fulltext_preprocessor,
+      seqio.preprocessors.tokenize_and_append_eos,
+    ],
+    metric_fns=[metrics.accuracy,my_metrics.f1_macro],
+    output_features=DEFAULT_OUTPUT_FEATURES,
+)
+
+seqio.TaskRegistry.add(
+    "classify_tweets_binary",
+    source=seqio.TextLineDataSource(
+        split_to_filepattern=tsv_path,
+        #num_input_examples=num_nq_examples
+        ),
+    preprocessors=[
+      functools.partial(
+          t5.data.preprocessors.parse_tsv,
+          field_names=["annotator1","annotator2","annotator3","target","source","id"]),
+      categorise_binary_preprocessor,
+      seqio.preprocessors.tokenize_and_append_eos,
+    ],
+    metric_fns=[metrics.accuracy,my_metrics.f1_macro],
+    output_features=DEFAULT_OUTPUT_FEATURES,
+)
+
+seqio.TaskRegistry.add(
+    "classify_tweets_fulltext_word",
+    source=seqio.TextLineDataSource(
+        split_to_filepattern=tsv_path,
+        #num_input_examples=num_nq_examples
+        ),
+    preprocessors=[
+      functools.partial(
+          t5.data.preprocessors.parse_tsv,
+          field_names=["annotator1","annotator2","annotator3","target","source","id"]),
+      categorise_fulltext_word_preprocessor,
+      seqio.preprocessors.tokenize_and_append_eos,
+    ],
+    metric_fns=[metrics.accuracy,my_metrics.f1_macro],
+    output_features=DEFAULT_OUTPUT_FEATURES,
+)
+