The help for the evaluation

[Wangdanchunbufuz]

Hi, thank you for sharing this great work.I'm trying to test the effects of training，But I found that there were some missing code about the TABLE 7,such as 'Object Attr. Relation Color Count Size CLIP',So how do I get it？

[davidnvq]

Thanks for asking!
May you find the evaluation code/notebook from M2 Trasnformer repo?
I remember that I took the code from there.

[Wangdanchunbufuz]

Thanks for asking!
May you find the evaluation code/notebook from M2 Trasnformer repo?
I remember that I took the code from there.

Sorry，I browsed the code about M2 Trasnformer,but I can not find what I need.it is only have 'blue cider, meteor,rouge'

[davidnvq]

Sorry for my bad memory. Luckily, I've just checked the old code and found that the evaluation based on this repo, not M2 transformer: https://github.com/salaniz/pycocoevalcap.

To get 'Object Attr. Relation Color Count Size', you need to modify the code of https://github.com/salaniz/pycocoevalcap/blob/master/spice/spice.py as follows:

from __future__ import division
import os
import sys
import subprocess
import threading
import json
import numpy as np
import ast
import tempfile
import math

# Assumes spice.jar is in the same directory as spice.py.  Change as needed.
SPICE_JAR = 'spice-1.0.jar'
TEMP_DIR = 'tmp'
CACHE_DIR = 'cache'


class Spice:
    """
    Main Class to compute the SPICE metric 
    """

    def float_convert(self, obj):
        try:
            return float(obj)
        except:
            return np.nan

    def compute_score(self, gts, res):
        assert (sorted(gts.keys()) == sorted(res.keys()))
        imgIds = sorted(gts.keys())

        # Prepare temp input file for the SPICE scorer
        input_data = []
        for id in imgIds:
            hypo = res[id]
            ref = gts[id]

            # Sanity check.
            assert (type(hypo) is list)
            assert (len(hypo) == 1)
            assert (type(ref) is list)
            assert (len(ref) >= 1)

            input_data.append({"image_id": id, "test": hypo[0], "refs": ref})

        cwd = os.path.dirname(os.path.abspath(__file__))
        temp_dir = os.path.join(cwd, TEMP_DIR)
        if not os.path.exists(temp_dir):
            os.makedirs(temp_dir)
        in_file = tempfile.NamedTemporaryFile(mode='w+', delete=False, dir=temp_dir)
        json.dump(input_data, in_file, indent=2)
        in_file.close()

        # Start job
        out_file = tempfile.NamedTemporaryFile(mode='w+', delete=False, dir=temp_dir)
        out_file.close()
        cache_dir = os.path.join(cwd, CACHE_DIR)
        if not os.path.exists(cache_dir):
            os.makedirs(cache_dir)
        spice_cmd = [
            'java', '-jar', '-Xmx8G', SPICE_JAR, in_file.name, '-cache', cache_dir, '-out', out_file.name, '-subset',
            '-silent'
        ]
        subprocess.check_call(spice_cmd, cwd=os.path.dirname(os.path.abspath(__file__)))

        # Read and process results
        with open(out_file.name) as data_file:
            results = json.load(data_file)
        os.remove(in_file.name)
        os.remove(out_file.name)

        imgId_to_scores = {}
        spice_scores = []
        keys = ['Relation', 'Cardinality', 'Color', 'Attribute', 'Object', 'Size']
        other_scores = {key: [] for key in keys}
        for item in results:
            imgId_to_scores[item['image_id']] = item['scores']

            spice_scores.append(self.float_convert(item['scores']['All']['f']))
            for key in keys:
                value = self.float_convert(item['scores'][key]['f'])
                if not math.isnan(value):
                    other_scores[key].append(value)

        for key in keys:
            score = np.mean(np.array(other_scores[key]))
            print(f"SPICE key: {key} = {score}")

        average_score = np.mean(np.array(spice_scores))
        scores = []
        for image_id in imgIds:
            # Convert none to NaN before saving scores over subcategories
            score_set = {}
            for category, score_tuple in imgId_to_scores[image_id].items():
                score_set[category] = {k: self.float_convert(v) for k, v in score_tuple.items()}
            scores.append(score_set)

        print(f"SPICE Score: avg = {average_score}")

        return average_score, scores

    def method(self):
        return "SPICE"

[davidnvq]

I can't find my old code for CLIP score. However, you can easily follow this repo to compute:
https://github.com/jmhessel/clipscore