updating requirements (#4860)

* updating requirements

* lint

projects/roscoe/README.md

@@ -83,18 +83,16 @@ bash projects/roscoe/roscoe_data/download_annotated.sh
 
 ### Baseline scoring
 One-time setup: 
-It is higly recommended to run baseline scoring from a separate conda environment.
-```bash
-conda create --name roscoe_baselines python=3.8
-conda activate roscoe_baselines
-```
+Follow BLEURT [installation quidelines](https://github.com/google-research/bleurt#installation).
+Clone [BartScore repo](https://github.com/neulab/BARTScore) and update path in projects/roscoe/baselines/score.py. Install Bart_score requirements.
+Upload fine-tuned [BART model](https://dl.fbaipublicfiles.com/parlai/projects/roscoe/fine_tuned_bartscore.pth).
+Download PRISM [installation quidelines](https://github.com/thompsonb/prism) and download the model. Do not install requirements.
 Install requirements to run baselines:
 ```bash
+python -c "import nltk; nltk.download('punkt')"
+python -c "import nltk; nltk.download('stopwords')"
 pip install -r projects/roscoe/baselines/requirements.txt
-pip install -r projects/roscoe/baselines/bart_requirements.txt
 ```
-Follow BLEURT [installation quidelines](https://github.com/google-research/bleurt#installation)
-Upload fine-tuned [BART model](https://dl.fbaipublicfiles.com/parlai/projects/roscoe/fine_tuned_bartscore.pth)
 
 Then you can run baselines on all datasets, scores, and use of reference (when possible) with the following:
 ```bash

projects/roscoe/baselines/requirements.txt

@@ -1,8 +1,6 @@
-rouge-score
-bert-score
-ctc_score
+rouge-score>=0.1.2
+bert-score>=0.3.9
 sentencepiece>=0.1.86
 fairseq==0.9.0
 sacrebleu>=1.4.8
-torch>=1.4.0
 ctc_score

projects/roscoe/baselines/run.py

@@ -235,7 +235,7 @@ def save_scores_map(path, metric_to_score):
         type=str,
         choices=[x.value for x in UseRef],
         nargs="+",
-        default=[x.value for x in UseRef],
+        default=UseRef.NO.value,
         help='do we want to generate reference-based or reference-free scores',
     )
     parser.add_argument(

projects/roscoe/baselines/scores.py

@@ -26,8 +26,8 @@
 
 from importlib.machinery import SourceFileLoader
 
-BART_SCORE_REPO = "/path_to/BARTScore/"
-PRISM_SCORE_REPO = "/path_to/SUM"
+BART_SCORE_REPO = "/path_to/BARTScore"
+PRISM_SCORE_REPO = "/path_to/prism"
 BLEURT_SCORE_REPO = "/path_to/bleurt"
 
 ######### Base functionality
@@ -114,7 +114,9 @@ def get_scores(self, score_me):
 @register_scorer([BLEURT])
 class BleurtBaselineScorer(BaselineScorer):
     def __init__(self):
-        self.scorer = bleurt_score.BleurtScorer(BLEURT_SCORE_REPO + "/test_checkpoint")
+        self.scorer = bleurt_score.BleurtScorer(
+            BLEURT_SCORE_REPO + "/bleurt/test_checkpoint"
+        )
 
     def get_scores(self, score_me):
         scores = self.scorer.score(
@@ -143,8 +145,6 @@ def get_scores(self, score_me):
 
 
 ######### BartScore (and its variants)
-# Note: You might have to load some of its dependencies yourself.
-# HuggingFace is the major one
 # Second argument here should be path to `bart_score.py` of the BARTScore repo
 try:
     bart_score = SourceFileLoader(
@@ -190,8 +190,12 @@ def __init__(self):
         self.scorer = BARTScorer(
             device=DEFAULT_DEVICE, checkpoint='facebook/bart-large-cnn'
         )
-        # Path here should be to fine tuned BART model from https://github.com/neulab/BARTScore#direct-use
-        self.scorer.load(BART_SCORE_REPO + "/bart_score_para_finetuned.pth")
+        try:
+            self.scorer.load(BART_SCORE_REPO + "/bart_score_para_finetuned.pth")
+        except FileNotFoundError:
+            raise FileNotFoundError(
+                f"Path here should be to fine tuned BART model from https://github.com/neulab/BARTScore#direct-use"
+            )
         self.score_type = BARTSCORE_CNN_PARA_F
 
 
@@ -218,24 +222,25 @@ def load(self, path=None):
             device=DEFAULT_DEVICE, checkpoint='facebook/bart-large-cnn'
         )
         # Path here to fine-tuend BART Model
-        self.scorer.load(BART_SCORE_REPO + "/train/reproduce/trained/bart_6000.pth")
+        try:
+            self.scorer.load(BART_SCORE_REPO + "/train/reproduce/trained/bart_6000.pth")
+        except FileNotFoundError:
+            raise FileNotFoundError(
+                f"Path here should be to fine tuned BART model from"
+                + "https://dl.fbaipublicfiles.com/parlai/projects/roscoe/fine_tuned_bartscore.pth"
+            )
         self.score_type = BARTSCORE_FINETUNED_F
 
 
 ######### Prism
-# Prism deps (minimal set)
-# sentencepiece>=0.1.86
-# fairseq==0.9.0
-# sacrebleu>=1.4.8#
-# torch>=1.4.0
 prism = SourceFileLoader("prism", PRISM_SCORE_REPO + "/prism.py").load_module()
 
 
 @register_scorer([PRISM_AVG])
 class PrismBaselineScorer(BaselineScorer):
     def __init__(self):
         self.scorer = prism.Prism(
-            model_dir=PRISM_SCORE_REPO + '/models/m39v1/',
+            model_dir=PRISM_SCORE_REPO + '/m39v1/',
             lang='en',
         )