Remove promptsource and use t-zero

bigscience/eval-spreadsheet/interactive_results_parser.py

@@ -3,7 +3,7 @@
 import time
 
 from parse_eai_results import process_eai_results
-from parse_promptsource import process_ps_results
+from parse_t0 import process_t0_results
 
 if __name__ == "__main__":
     previous_clipboard = clipboard.paste()

bigscience/eval-spreadsheet/parse_promptsource.py → bigscience/eval-spreadsheet/parse_t0.py

@@ -85,13 +85,13 @@ def process_task_prompt(task_prompt: str) -> tuple[str, str]:
 
 if __name__ == "__main__":
     parser = argparse.ArgumentParser(
-        description="Get formatted promptsource tasks & results for c/p in spreadsheet"
+        description="Get formatted t0 tasks & results for c/p in spreadsheet"
     )
     parser.add_argument(
         "-i",
         "--ps-results-file",
         type=str,
-        help="Path to a promptsource .json result file",
+        help="Path to a t0 .json result file",
     )
     parser.add_argument(
         "-t",

bigscience/gins/c_dec_t0_adapt.gin

@@ -3,7 +3,7 @@ from __gin__ import dynamic_registration
 from t5x import models
 from t5x import utils
 import seqio
-from promptsource import seqio_tasks
+from t0 import seqio_tasks
 
 include "bigscience/gins/c_dec_xxl.gin"
 include "t5x/configs/runs/finetune.gin"

bigscience/gins/enc_dec_t0_adapt.gin

@@ -4,7 +4,7 @@ from t5x import models
 from t5x import trainer
 from t5x import utils
 import seqio
-from promptsource import seqio_tasks
+from t0 import seqio_tasks
 
 include "bigscience/gins/enc_dec_xxl.gin"
 include "t5x/configs/runs/finetune.gin"

bigscience/gins/nc_dec_t0_adapt.gin

@@ -3,7 +3,7 @@ from __gin__ import dynamic_registration
 from t5x import models
 from t5x import utils
 import seqio
-from promptsource import seqio_tasks
+from t0 import seqio_tasks
 
 include "bigscience/gins/nc_dec_xxl.gin"
 include "t5x/configs/runs/finetune.gin"

bigscience/gins/task.py

@@ -2,7 +2,7 @@
 
 import seqio
 
-from promptsource import seqio_tasks
+from t0 import seqio_tasks
 
 # --- Seqio ---
 seqio.add_global_cache_dirs([

bigscience/scripts/plot_eai_eval.py

@@ -8,8 +8,6 @@
 import numpy as np
 from matplotlib import pyplot as plt
 
-# gcloud alpha compute tpus tpu-vm scp thomas-dev-tpu:~/arch_objective_exps_v2 .  --zone us-central2-b --recurse
-
 def get_args():
     parser = ArgumentParser()
     parser.add_argument('--all', action="store_true", help="Plot all results in a single plot")

bigscience/scripts/run_on_all_vms.sh

@@ -12,4 +12,4 @@ gcloud alpha compute tpus tpu-vm ssh ${TPU_NAME} --zone ${ZONE} --worker=all --c
 # Example to run t5_c4_span_corruption
 #  - run setup vms: sh bigscience/scripts/run_on_all_vms.sh enc_dec_c4_span_corruption "$(cat bigscience/scripts/setup_vm.sh)"
 #  - run t5_c4_span_corruption: sh bigscience/scripts/run_on_all_vms.sh enc_dec_c4_span_corruption "cd code/t5x; git pull; sh bigscience/scripts/launch_command_in_tmux.sh \"bash bigscience/scripts/pretrain.sh enc_dec_c4_span_corruption\""
-#  - kill zombie process: sh bigscience/scripts/run_on_all_vms.sh enc_dec_c4_span_corruption "killall -u thomas"
+#  - kill zombie process: sh bigscience/scripts/run_on_all_vms.sh enc_dec_c4_span_corruption "killall -u {USER}"

bigscience/scripts/setup_vm.sh

@@ -5,39 +5,25 @@ rm libtpu_tpuv4-0.1.dev*
 gsutil cp gs://cloud-tpu-tpuvm-v4-artifacts/wheels/libtpu/latest/libtpu_tpuv4-0.1.dev* .
 pip3 install libtpu_tpuv4-0.1.dev*
 
-### ...
-##pip3 install tensorflow==2.7.0
-#rm tensorflow-2.7.0-cp38-cp38-linux_x86_64.whl
-#gsutil cp gs://cloud-tpu-tpuvm-v4-artifacts/wheels/tensorflow/tf-2-7-0/tensorflow-2.7.0-cp38-cp38-linux_x86_64.whl .
-#pip3 install tensorflow-2.7.0-cp38-cp38-linux_x86_64.whl tensorflow-text==2.7.0
-
 mkdir -p ~/code
 cd ~/code
 
-# Install t5 first
+# Install t5 master version
 git clone https://github.com/google-research/text-to-text-transfer-transformer.git
 pushd text-to-text-transfer-transformer
 pip3 install -e .
 popd
-#rm -rf text-to-text-transfer-transformer
-#git clone https://github.com/thomasw21/text-to-text-transfer-transformer.git
-#pushd text-to-text-transfer-transformer
-#git checkout fix_prefix_lm_obj
-#pip3 install -e .
-#popd
 
-git clone https://github.com/bigscience-workshop/promptsource.git
-pushd promptsource
-git reset e65186c2b8a544de1eb5c283b11b235033b01514 --hard
-pip3 install black==21.12b0 # conflicts with streamlit
-pip3 install -r requirements.txt
-pip3 install --ignore-requires-python -e . #needed because `promptsource` forces the use of python 3.7
+git clone https://github.com/bigscience-workshop/t-zero.git
+pushd t-zero
+# TODO: remove once https://github.com/bigscience-workshop/t-zero/pull/24 is merged
+git checkout thomas/update_promptsource_dependency
+pip3 install -e ".[seqio_tasks]"
 popd
 
-#rm -rf t5x
 git clone https://github.com/bigscience-workshop/t5x.git
 pushd t5x
-pip3 install -e .
+pip3 install -e ".[bigscience]"
 popd
 
 git clone https://github.com/EleutherAI/lm-evaluation-harness.git

setup.py

@@ -55,14 +55,16 @@
         't5',
         'tensorflow',
         'tensorstore==0.1.13',
-        'promptsource'
     ],
     extras_require={
         'gcp': [
             'gevent', 'google-api-python-client', 'google-compute-engine',
             'google-cloud-storage', 'oauth2client'
         ],
         'test': ['pytest'],
+        'bigscience': [
+            "t0[seqio_tasks]"
+        ]
     },
     classifiers=[
         'Development Status :: 4 - Beta',