first submit

README.md

@@ -1 +1 @@
-Training Files for the EU-JAV Tweet Categorisation. The training files are not stored here. The model will be placed here when done.
+Private sample code for running categorisation on the mT5X

finetuning_categorisation.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 = "categorise"
+TASK_FEATURE_LENGTHS = {"inputs": 256, "targets": 256}
+TRAIN_STEPS = 1_010_000  # 1000000 pre-trained steps + 10000 fine-tuning steps.
+USE_CACHED_TASKS = False
+DROPOUT_RATE = 0.0
+RANDOM_SEED = 0
+
+# 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
+
+INITIAL_CHECKPOINT_PATH = "gs://t5-data/pretrained_models/t5x/mt5_base/checkpoint_1000000"
+
+#train_script.train:
+#  eval_period = 500
+#  partitioner = @partitioning.ModelBasedPjitPartitioner()
+# partitioning.ModelBasedPjitPartitioner.num_partitions = 2
+
+# `num_decodes` is equivalent to a beam size in a beam search decoding.
+models.EncoderDecoderModel.predict_batch_with_aux.num_decodes = 4
+
+#mesh_transformer.learning_rate_schedules.constant_learning_rate.learning_rate = 0.0005
+#run.learning_rate_schedule = @learning_rate_schedules.constant_learning_rate
+

finetuning_categorisation_large.gin

@@ -0,0 +1,40 @@
+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/large.gin"
+include "t5x/configs/runs/finetune.gin"
+
+MIXTURE_OR_TASK_NAME = "categorise"
+TASK_FEATURE_LENGTHS = {"inputs": 256, "targets": 256}
+TRAIN_STEPS = 1_010_000  # 1000000 pre-trained steps + 10000 fine-tuning steps.
+USE_CACHED_TASKS = False
+DROPOUT_RATE = 0.0
+RANDOM_SEED = 0
+
+# 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
+
+INITIAL_CHECKPOINT_PATH = "gs://t5-data/pretrained_models/t5x/mt5_large/checkpoint_1000000"
+# BATCH_SIZE = 64
+
+#train_script.train:
+#  eval_period = 500
+#  partitioner = @partitioning.ModelBasedPjitPartitioner()
+# partitioning.ModelBasedPjitPartitioner.num_partitions = 2
+
+# `num_decodes` is equivalent to a beam size in a beam search decoding.
+models.EncoderDecoderModel.predict_batch_with_aux.num_decodes = 4
+
+#mesh_transformer.learning_rate_schedules.constant_learning_rate.learning_rate = 0.0005
+#run.learning_rate_schedule = @learning_rate_schedules.constant_learning_rate
+

interference.sh

@@ -0,0 +1,16 @@
+INFER_OUTPUT_DIR="output2"  # directory to write infer output
+T5X_DIR="../../t5x"  # directory where the t5x is cloned, e.g., ${HOME}"/t5x".
+TFDS_DATA_DIR="gs://nb-t5x/corpus_multi_sentencefix_mt5/"
+CHECKPOINT_PATH="gs://nb-t5x-us-central2/model_mT5X_large_16_e/checkpoint_1140000"
+PROJECT_DIR=${HOME}"/models/multi-sentencefix-mt5"
+export PYTHONPATH=${PROJECT_DIR}
+
+python3 ${T5X_DIR}/t5x/infer.py \
+  --gin_search_paths=${PROJECT_DIR} \
+  --gin_file="large_wmt_infer.gin" \
+  --gin.CHECKPOINT_PATH=\"${CHECKPOINT_PATH}\" \
+  --gin.INFER_OUTPUT_DIR=\"${INFER_OUTPUT_DIR}\" \
+  --tfds_data_dir=${TFDS_DATA_DIR}
+
+
+

tasks.py

@@ -0,0 +1,65 @@
+# /home/perk/mymodel/categorisation-mt5x/tasks.py
+
+import functools
+import seqio
+import tensorflow_datasets as tfds
+from t5.evaluation import metrics
+from t5.data import preprocessors
+import t5
+import tensorflow.compat.v1 as tf
+
+tsv_path = {
+        "train": "gs://peregilcloud/italian_tweets/train.tsv",
+        "validation": "gs://peregilcloud/italian_tweets/dev.tsv",
+        "test": "gs://peregilcloud/italian_tweets/test.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(
+    "categorise",
+    source=seqio.TextLineDataSource(
+        split_to_filepattern=tsv_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.bleu],
+    output_features=DEFAULT_OUTPUT_FEATURES,
+)
+

train_base.sh

@@ -0,0 +1,11 @@
+PROJECT_DIR=${HOME}"/models/categorisation-mt5x"
+T5X_DIR="../../t5x"  # directory where the t5x is cloned.
+#Needs to be updated when moving to tpu-v4  it should then be in another zone
+MODEL_DIR="gs://nb-t5x/categorisation"
+export PYTHONPATH=${PROJECT_DIR}
+
+python3 ${T5X_DIR}/t5x/train.py \
+  --gin_search_paths=${PROJECT_DIR} \
+  --gin_file="finetuning_categorisation.gin" \
+  --gin.MODEL_DIR="'${MODEL_DIR}'" 
+

train_large.sh

@@ -0,0 +1,11 @@
+PROJECT_DIR=${HOME}"/models/categorisation-mt5x"
+T5X_DIR="../../t5x"  # directory where the t5x is cloned.
+#Needs to be updated when moving to tpu-v4  it should then be in another zone
+MODEL_DIR="gs://nb-t5x/categorisation_large"
+export PYTHONPATH=${PROJECT_DIR}
+
+python3 ${T5X_DIR}/t5x/train.py \
+  --gin_search_paths=${PROJECT_DIR} \
+  --gin_file="finetuning_categorisation_large.gin" \
+  --gin.MODEL_DIR="'${MODEL_DIR}'" 
+