larger models added

finetuning_categorisation_xl.gin

@@ -0,0 +1,41 @@
+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/xl.gin"
+include "t5x/configs/runs/finetune.gin"
+
+MIXTURE_OR_TASK_NAME = "categorise"
+TASK_FEATURE_LENGTHS = {"inputs": 96, "targets": 2}
+TRAIN_STEPS = 1_010_000  # 1000000 pre-trained steps + 10000 fine-tuning steps.
+USE_CACHED_TASKS = False
+DROPOUT_RATE = 0.0
+RANDOM_SEED = 0
+BATCH_SIZE = 8
+
+
+# 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_xl/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_xxl.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/xxl.gin"
+include "t5x/configs/runs/finetune.gin"
+
+MIXTURE_OR_TASK_NAME = "categorise"
+TASK_FEATURE_LENGTHS = {"inputs": 96, "targets": 2}
+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_xxl/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
+

tasks.py

@@ -9,9 +9,9 @@ import t5
 import tensorflow.compat.v1 as tf
 
 tsv_path = {
-        "train": "gs://peregilcloud/italian_tweets/train2.tsv",
-        "validation": "gs://peregilcloud/italian_tweets/dev2.tsv",
-        "test": "gs://peregilcloud/italian_tweets/test2.tsv"
+        "train": "gs://peregilcloud/italian_tweets/train3.tsv",
+        "validation": "gs://peregilcloud/italian_tweets/dev.tsv",
+        "test": "gs://peregilcloud/italian_tweets/test.tsv"
 }
 
 vocabulary = seqio.SentencePieceVocabulary(

train_large.sh

@@ -1,7 +1,7 @@
 PROJECT_DIR=${HOME}"/models/eu-jav-categorisation"
 T5X_DIR="../../t5x"  # directory where the t5x is cloned.
 #Needs to be updated when moving to tpu-v4  it should then be in another zone
-MODEL_DIR="gs://nb-t5x/eujav_large2"
+MODEL_DIR="gs://nb-t5x/eujav_large3"
 export PYTHONPATH=${PROJECT_DIR}
 
 python3 ${T5X_DIR}/t5x/train.py \

train_xl.sh

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

train_xxl.sh

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