works

analysis/README.md

@@ -3,6 +3,25 @@
 We're going to add visualizations for both the eval. data and results here.
 So far, we have the following tools:
 
+### Per token uterrance reward
+This returns the reward per-token to show how the reward evolves over a piece of text.
+```
+python analysis/per_token_reward.py --model=OpenAssistant/reward-model-deberta-v3-large-v2 --chat_template=raw --text="I love to walk the dog, what do you like?"
+```
+E.g. with OpenAssistant/reward-model-deberta-v3-large-v2
+Reward: -0.544 | Substring: I
+Reward: -0.556 | Substring: I love
+Reward: -0.566 | Substring: I love to
+Reward: 0.099 | Substring: I love to walk
+Reward: 0.096 | Substring: I love to walk the
+Reward: 0.092 | Substring: I love to walk the dog
+Reward: 0.09 | Substring: I love to walk the dog,
+Reward: 0.087 | Substring: I love to walk the dog, what
+Reward: 0.085 | Substring: I love to walk the dog, what do
+Reward: 0.089 | Substring: I love to walk the dog, what do you
+Reward: 0.09 | Substring: I love to walk the dog, what do you like
+Reward: 0.093 | Substring: I love to walk the dog, what do you like?
+
 ### Model usage within eval. dataset
 To run this, execute:
 ```

analysis/per_token_reward.py

@@ -0,0 +1,237 @@
+# Copyright 2023 AllenAI. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+# Script to output the per-token reward across a piece of text given a reward model
+
+import argparse
+import logging
+import sys
+
+import torch
+import transformers
+from accelerate import Accelerator
+from accelerate.logging import get_logger
+from datasets import Dataset
+from tqdm import tqdm
+from transformers import (
+    AutoModelForSequenceClassification,
+    AutoTokenizer,
+    T5ForConditionalGeneration,
+    pipeline,
+)
+
+
+def get_args():
+    """
+    Parse arguments strings model and chat_template
+    """
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--model", type=str, default="natolambert/gpt2-dummy-rm", help="path to model")
+    parser.add_argument(
+        "--tokenizer", type=str, default=None, help="path to non-matching tokenizer, requires --direct_load"
+    )
+    parser.add_argument("--chat_template", type=str, default="tulu", help="path to chat template")
+    parser.add_argument(
+        "--batch_size", type=int, default=64, help="batch size for inference (if above number of tokens)"
+    )
+    parser.add_argument("--text", type=str, default="I love to drink coffee at work.", help="text to evaluate")
+    args = parser.parse_args()
+
+    if "PairRM" in args.model or "PairRM" in args.chat_template or "SHP" in args.model or "SHP" in args.chat_template:
+        # Note: SHP could be used in single-output mode, but the code is not yet added
+        raise ValueError("PairRM and SHP require pairwise inputs, not supported")
+    return args
+
+
+def main():
+    args = get_args()
+    quantized = True  # only Starling isn't quantized for now
+    custom_dialogue = False
+    # some models need custom code to be run
+    if "oasst" in args.model or "oasst" in args.chat_template:
+        from herm.models import openassistant  # noqa
+
+        model_builder = AutoModelForSequenceClassification.from_pretrained
+        pipeline_builder = pipeline
+    elif "Starling" in args.model or "Starling" in args.chat_template:
+        from herm.models.starling import StarlingPipeline, build_starling_rm
+
+        model_builder = build_starling_rm
+        pipeline_builder = StarlingPipeline
+        quantized = False
+    elif "openbmb" in args.model or "openbmb" in args.chat_template:
+        from herm.models.openbmb import LlamaRewardModel, OpenBMBPipeline
+
+        model_builder = LlamaRewardModel.from_pretrained
+        pipeline_builder = OpenBMBPipeline
+    elif "PairRM" in args.model or "PairRM" in args.chat_template:
+        from herm.models.pairrm import DebertaV2PairRM, PairRMPipeline
+
+        custom_dialogue = True
+        model_builder = DebertaV2PairRM.from_pretrained
+        pipeline_builder = PairRMPipeline
+    elif "SHP" in args.model or "SHP" in args.chat_template:
+        from herm.models.shp import SHPPipeline
+
+        custom_dialogue = True
+        model_builder = T5ForConditionalGeneration.from_pretrained
+        pipeline_builder = SHPPipeline
+    else:
+        model_builder = AutoModelForSequenceClassification.from_pretrained
+        pipeline_builder = pipeline
+
+    if custom_dialogue:
+        raise ValueError("Custom dialogue formatting not yet supported in this script")
+
+    ###############
+    # Setup logging
+    ###############
+    accelerator = Accelerator()
+    current_device = accelerator.process_index
+
+    logger = get_logger(__name__)
+    logging.basicConfig(
+        format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
+        datefmt="%Y-%m-%d %H:%M:%S",
+        handlers=[logging.StreamHandler(sys.stdout)],
+    )
+    log_level = logging.INFO
+    logger.setLevel(log_level)
+    transformers.utils.logging.set_verbosity(log_level)
+    transformers.utils.logging.enable_default_handler()
+    transformers.utils.logging.enable_explicit_format()
+
+    logger.info(f"Running reward model on {args.model} with chat template {args.chat_template}")
+
+    ############################
+    # Load reward model pipeline
+    ############################
+    tokenizer_path = args.tokenizer if args.tokenizer else args.model
+    tokenizer = AutoTokenizer.from_pretrained(tokenizer_path)
+
+    BATCH_SIZE = args.batch_size
+    logger.info("*** Load reward model ***")
+    reward_pipeline_kwargs = {
+        "batch_size": BATCH_SIZE,  # eval_args.inference_batch_size,
+        "truncation": True,
+        "padding": True,
+        "max_length": 2048,
+        "function_to_apply": "none",  # Compute raw logits
+        "return_token_type_ids": False,
+    }
+    if quantized:
+        model_kwargs = {
+            "load_in_8bit": True,
+            "device_map": {"": current_device},
+            "torch_dtype": torch.float16 if torch.cuda.is_available() else None,
+        }
+    else:
+        model_kwargs = {"device_map": {"": current_device}}
+    # TODO remove direct load logic
+    # if pipeline_builder is pipeline, use built in pipeline, else custom
+    if not pipeline == pipeline_builder:
+        model = model_builder(args.model, **model_kwargs)
+        tokenizer = AutoTokenizer.from_pretrained(tokenizer_path)
+        reward_pipe = pipeline_builder(
+            "text-classification",
+            model=model,
+            tokenizer=tokenizer,
+        )
+    else:
+        reward_pipe = pipeline(
+            "text-classification",
+            model=args.model,
+            tokenizer=tokenizer,
+            revision="main",
+            model_kwargs=model_kwargs,
+        )
+
+    ############################
+    # Tokenization settings & dataset preparation
+    ############################
+    # set pad token to eos token if not set
+    if reward_pipe.tokenizer.pad_token_id is None:
+        reward_pipe.model.config.pad_token_id = reward_pipe.tokenizer.eos_token_id
+        reward_pipe.tokenizer.pad_token_id = reward_pipe.tokenizer.eos_token_id
+
+    def tokenify_string(string, tokenizer):
+        # Tokenize the entire text
+        tokens = tokenizer.tokenize(string)
+
+        cumulative_texts = []
+        # Iterate over each token
+        for i, _ in enumerate(tokens):
+            # Append the current cumulative text to the list
+            cumulative_texts.append(tokenizer.convert_tokens_to_string(tokens[: i + 1]))
+
+        return cumulative_texts
+
+    substrings = tokenify_string(args.text, tokenizer)
+    # create dataset from list of strings substrings with huggingface
+    dataset = [{"text": substring} for substring in substrings]
+    dataset = Dataset.from_list(dataset)
+
+    ############################
+    # Run inference [1/2]" built in transformers
+    ############################
+    # if using HF pipeline, can pass entire dataset and get results
+    # first, handle custom pipelines that we must batch normally
+    if not pipeline_builder == pipeline:
+        logger.info("*** Running forward pass via built in pipeline abstraction ***")
+        # this setup can be optimized slightly with one pipeline call
+        # prepare for inference
+        reward_pipe = accelerator.prepare(reward_pipe)
+
+        rewards = reward_pipe(dataset["text"], **reward_pipeline_kwargs)
+
+    ############################
+    # Run inference [2/2] custom pipelines
+    ############################
+    else:
+        logger.info("*** Running dataloader to collect results ***")
+
+        dataloader = torch.utils.data.DataLoader(
+            dataset,
+            batch_size=BATCH_SIZE,
+            collate_fn=None,
+            shuffle=False,
+            drop_last=False,
+        )
+
+        dataloader, model = accelerator.prepare(dataloader, reward_pipe.model)
+        reward_pipe.model = model
+
+        results = []
+        for step, batch in enumerate(tqdm(dataloader, desc="RM batch steps")):
+            logger.info(f"RM inference step {step}/{len(dataloader)}")
+            rewards = reward_pipe(batch["text"], **reward_pipeline_kwargs)
+
+            # for each item in batch, record 1 if chosen > rejected
+            # extra score from dict within batched results (e.g. logits)
+            # [{'label': 'LABEL_1', 'score': 0.6826171875},... ]
+            if isinstance(rewards[0], dict):
+                scores = [result["score"] for result in rewards]
+            # for classes that directly output scores (custom code)
+            else:
+                scores = rewards.cpu().numpy().tolist()
+
+            results.extend(scores)
+
+    # print the results
+    for i, substring in enumerate(substrings):
+        print(f"Reward: {round(results[i], 3)} | Substring: {substring}")
+
+
+if __name__ == "__main__":
+    main()

scripts/run_rm.py

@@ -161,6 +161,8 @@ def main():
         }
     else:
         model_kwargs = {"device_map": {"": current_device}}
+    # TODO remove direct load logic
+    # if pipeline_builder is pipeline, use built in pipeline, else custom
     if args.direct_load:
         model = model_builder(args.model, **model_kwargs)
         tokenizer = AutoTokenizer.from_pretrained(tokenizer_path)