fulltext

batch_finetune_eu_jav_base_exp_english.sh

@@ -1,6 +1,6 @@
 PROJECT_DIR=${HOME}"/models/eu-jav-categorisation"
 export PYTHONPATH=${PROJECT_DIR}
-INITIAL_CHECKPOINT_PATH=\"gs://t5-data/pretrained_models/t5x/t5_1_1_basecheckpoint_1000000\"
+INITIAL_CHECKPOINT_PATH=\"gs://t5-data/pretrained_models/t5x/t5_1_1_base/checkpoint_1000000\"
 TRAIN_STEPS=1005000
 
 python3 ../../t5x/t5x/train.py --gin_search_paths="./" --gin.TRAIN_STEPS=${TRAIN_STEPS} --gin_file="finetune_classification_base_exp_english.gin" --gin.INITIAL_CHECKPOINT_PATH=${INITIAL_CHECKPOINT_PATH}  --gin.MIXTURE_OR_TASK_NAME=\"classify_tweets\" --gin.MODEL_DIR=\"gs://eu-jav-t5x/finetuned/italian_tweets/english_classify_tweets_base_v1\" &&

batch_finetune_eu_jav_base_exp_fulltext.sh

@@ -0,0 +1,11 @@
+PROJECT_DIR=${HOME}"/models/eu-jav-categorisation"
+export PYTHONPATH=${PROJECT_DIR}
+INITIAL_CHECKPOINT_PATH=\"gs://t5-data/pretrained_models/t5x/mt5_base/checkpoint_1000000\"
+TRAIN_STEPS=1005000
+
+python3 ../../t5x/t5x/train.py --gin_search_paths="./" --gin.TRAIN_STEPS=${TRAIN_STEPS} --gin_file="finetune_classification_base.gin" --gin.INITIAL_CHECKPOINT_PATH=${INITIAL_CHECKPOINT_PATH}  --gin.MIXTURE_OR_TASK_NAME=\"classify_tweets_fulltext\" --gin.MODEL_DIR=\"gs://eu-jav-t5x/finetuned/italian_tweets/fulltext_classify_tweets_fulltext_base_v1\" &&
+python3 ../../t5x/t5x/train.py --gin_search_paths="./" --gin.TRAIN_STEPS=${TRAIN_STEPS} --gin_file="finetune_classification_base.gin" --gin.INITIAL_CHECKPOINT_PATH=${INITIAL_CHECKPOINT_PATH}  --gin.MIXTURE_OR_TASK_NAME=\"classify_tweets_fulltext\" --gin.MODEL_DIR=\"gs://eu-jav-t5x/finetuned/italian_tweets/fulltext_classify_tweets_fulltext_base_v2\" &&
+python3 ../../t5x/t5x/train.py --gin_search_paths="./" --gin.TRAIN_STEPS=${TRAIN_STEPS} --gin_file="finetune_classification_base.gin" --gin.INITIAL_CHECKPOINT_PATH=${INITIAL_CHECKPOINT_PATH}  --gin.MIXTURE_OR_TASK_NAME=\"classify_tweets_fulltext\" --gin.MODEL_DIR=\"gs://eu-jav-t5x/finetuned/italian_tweets/fulltext_classify_tweets_fulltext_base_v3\" &&
+python3 ../../t5x/t5x/train.py --gin_search_paths="./" --gin.TRAIN_STEPS=${TRAIN_STEPS} --gin_file="finetune_classification_base.gin" --gin.INITIAL_CHECKPOINT_PATH=${INITIAL_CHECKPOINT_PATH}  --gin.MIXTURE_OR_TASK_NAME=\"classify_tweets_fulltext\" --gin.MODEL_DIR=\"gs://eu-jav-t5x/finetuned/italian_tweets/fulltext_classify_tweets_fulltext_base_v4\" &&
+python3 ../../t5x/t5x/train.py --gin_search_paths="./" --gin.TRAIN_STEPS=${TRAIN_STEPS} --gin_file="finetune_classification_base.gin" --gin.INITIAL_CHECKPOINT_PATH=${INITIAL_CHECKPOINT_PATH}  --gin.MIXTURE_OR_TASK_NAME=\"classify_tweets_fulltext\" --gin.MODEL_DIR=\"gs://eu-jav-t5x/finetuned/italian_tweets/fulltext_classify_tweets_fulltext_base_v5\"
+

finetune_classification_base.gin

@@ -11,7 +11,7 @@ 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": 2}
+TASK_FEATURE_LENGTHS = {"inputs": 256, "targets": 32}
 INITIAL_CHECKPOINT_PATH = %gin.REQUIRED
 TRAIN_STEPS = %gin.REQUIRED  # 1000000 pre-trained steps + 10000 fine-tuning steps.
 USE_CACHED_TASKS = False

finetune_classification_base_exp_english.gin

@@ -7,7 +7,7 @@ from t5x import models
 from t5x import partitioning
 from t5x import utils
 
-include "t5x/examples/t5/mt5/base.gin"
+include "t5x/examples/t5/t5_1_1/base.gin"
 include "t5x/configs/runs/finetune.gin"
 
 MIXTURE_OR_TASK_NAME = %gin.REQUIRED
@@ -24,7 +24,7 @@ infer_eval/utils.DatasetConfig:
 
 #Saving every 1000 steps
 utils.SaveCheckpointConfig:
-  period = 100
+  period = 500
 
 
 # Pere: Only necessary if we load a t5 model. We can start with an t5x model here

tasks.py

@@ -43,6 +43,36 @@ def categorise_preprocessor(ds):
 	    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=="3"
+        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)
 
@@ -65,7 +95,7 @@ seqio.TaskRegistry.add(
 )
 
 seqio.TaskRegistry.add(
-    "classify_tweets_a1",
+    "classify_tweetsi_fulltext",
     source=seqio.TextLineDataSource(
         split_to_filepattern=tsv_path,
         #num_input_examples=num_nq_examples
@@ -73,44 +103,11 @@ seqio.TaskRegistry.add(
     preprocessors=[
       functools.partial(
           t5.data.preprocessors.parse_tsv,
-          field_names=["target","annotator2","annotator3","placeholder","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_a2",
-    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","target","annotator3","placeholder","source","id"]),
-      categorise_preprocessor,
+          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_a3",
-    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","target","placeholder","source","id"]),
-      categorise_preprocessor,
-      seqio.preprocessors.tokenize_and_append_eos,
-    ],
-    metric_fns=[metrics.accuracy,my_metrics.f1_macro],
-    output_features=DEFAULT_OUTPUT_FEATURES,
-)

tasks_exp_english.py

@@ -16,7 +16,7 @@ tsv_path = {
 }
 
 vocabulary = seqio.SentencePieceVocabulary(
-                'gs://t5-data/vocabs/cc_en.32000/sentencepiece.model', extra_ids=0)
+                'gs://t5-data/vocabs/cc_all.32000.100extra/sentencepiece.model', extra_ids=0)
 
 DEFAULT_OUTPUT_FEATURES = {
     "inputs":
@@ -63,54 +63,3 @@ seqio.TaskRegistry.add(
     metric_fns=[metrics.accuracy,my_metrics.f1_macro],
     output_features=DEFAULT_OUTPUT_FEATURES,
 )
-
-seqio.TaskRegistry.add(
-    "classify_tweets_a1",
-    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","annotator2","annotator3","placeholder","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_a2",
-    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","target","annotator3","placeholder","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_a3",
-    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","target","placeholder","source","id"]),
-      categorise_preprocessor,
-      seqio.preprocessors.tokenize_and_append_eos,
-    ],
-    metric_fns=[metrics.accuracy,my_metrics.f1_macro],
-    output_features=DEFAULT_OUTPUT_FEATURES,
-)