Merge pull request #66 from EleutherAI/uychire-new-harness-1

Add naive eval script for StoryCloze

LICENSE.md

+MIT License
+
+Copyright (c) 2020 EleutherAI
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.

batch_eval/main.py

+import csv
+import os
+
+import click
+import torch
+from transformers import AutoConfig, AutoModelForCausalLM, AutoTokenizer
+
+
+@click.command()
+@click.argument("datadir", required=True)
+def main(datadir):
+    model = AutoModelForCausalLM.from_pretrained(
+        # 117M
+        pretrained_model_name_or_path="gpt2",
+        config=AutoConfig.from_pretrained(
+            "gpt2",
+            # <|endoftext|>
+            pad_token_id=50256,
+        ),
+    ).to("cuda")
+    model = model.eval()
+
+    tokenizer = AutoTokenizer.from_pretrained("gpt2")
+
+    prompt = "The quick brown fox jumps over"
+    encoded_prompt = tokenizer.encode(
+        prompt, add_special_tokens=False, return_tensors="pt"
+    ).to("cuda")
+
+    # Sanity check the model
+    [output_token_ids] = model.generate(
+        input_ids=encoded_prompt,
+        max_length=100,
+        tempareture=0,
+        do_sample=False,
+        num_return_sequences=1,
+    )
+    decoded_output = tokenizer.decode(output_token_ids.tolist())
+    # Next word should be "the" ("The quick brown fox jumps over *the*...")
+    print(decoded_output[len(prompt + " ") :][:10])
+    assert decoded_output[len(prompt + " ") :].startswith("the")
+
+    with open(
+        os.path.join(datadir, "cloze_test_test__spring2016 - cloze_test_ALL_test.csv")
+    ) as f:
+        storycloze_test_examples = list(csv.DictReader(f))
+
+    example_evaluations = [
+        evaluate_example(model, tokenizer, example)
+        for example in storycloze_test_examples
+    ]
+    fraction_correct = len(
+        [
+            evaluation
+            for evaluation in example_evaluations
+            if evaluation["was_model_correct"]
+        ]
+    ) / float(len(example_evaluations))
+    print(f"Fraction correct: {fraction_correct}")
+
+
+def evaluate_example(model, tokenizer, example):
+    storycloze_prompt = "{} {} {} {}".format(
+        example["InputSentence1"],
+        example["InputSentence2"],
+        example["InputSentence3"],
+        example["InputSentence4"],
+    )
+
+    # Calculate *per-token* likelihoods, as the paper did
+    per_token_logit_for_sentence1 = compute_per_token_logit_for_completion(
+        model, tokenizer, storycloze_prompt, example["RandomFifthSentenceQuiz1"]
+    )
+    per_token_logit_for_sentence2 = compute_per_token_logit_for_completion(
+        model, tokenizer, storycloze_prompt, example["RandomFifthSentenceQuiz2"]
+    )
+
+    if per_token_logit_for_sentence1 > per_token_logit_for_sentence2:
+        model_answer = example["RandomFifthSentenceQuiz1"]
+        model_answer_code = "1"
+    else:
+        model_answer = example["RandomFifthSentenceQuiz2"]
+        model_answer_code = "2"
+
+    return {
+        "model_answer": model_answer,
+        "was_model_correct": model_answer_code == example["AnswerRightEnding"],
+    }
+
+
+def compute_per_token_logit_for_completion(model, tokenizer, prompt, completion):
+    prompt_token_count = (
+        tokenizer.encode(prompt, add_special_tokens=False, return_tensors="pt")
+        .to("cuda")
+        .shape[1]
+    )
+    encoded_prompt_with_completion = tokenizer.encode(
+        prompt + " " + completion,
+        add_special_tokens=False,
+        return_tensors="pt",
+    ).to("cuda")
+    output_logits = model(encoded_prompt_with_completion).logits
+
+    # Align the output logits to the input tokens.
+    # The last logit needs to be dropped, because it's predicting the "next token", and it doesn't correspond to any input token
+    logits_for_input_positions = output_logits[0, :-1, :]
+    # The model does not predict the first input token, so it needs to be dropped as well.
+    input_tokens_at_positions_with_logits = encoded_prompt_with_completion[0, 1:]
+    # At each position, the model outputs ~50k logits, one for every possible token.
+    # To get the logits of the tokens that were actually provided, we need to select the right logit at each position.
+    logits_for_provided_tokens = torch.gather(
+        logits_for_input_positions,
+        1,
+        input_tokens_at_positions_with_logits.unsqueeze(1),
+    ).squeeze(1)
+
+    return (
+        logits_for_provided_tokens[
+            prompt_token_count
+            # Again, the model does not predict the first input token, so we need
+            - 1 :
+        ]
+        .mean()
+        .item()
+    )
+
+
+if __name__ == "__main__":
+    main()
\ No newline at end of file

requirements.txt

+black==20.8b1
+click>=7.1
+torch>=1.7
+transformers>=4.1
\ No newline at end of file