transformer/eval.py at master · machine-intelligence/transformer · GitHub

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
# -*- coding: utf-8 -*-
# /usr/bin/python2
'''
June 2017 by kyubyong park.
kbpark.linguist@gmail.com.
https://www.github.com/kyubyong/transformer
'''

from __future__ import print_function
import codecs
import os
import typing

import tensorflow as tf
import numpy as np
import numpy.linalg

from hyperparams import Hyperparams as hp
from data_load import load_test_data, load_de_vocab, load_en_vocab
from train import Graph
from nltk.translate.bleu_score import corpus_bleu
from PIL import Image, ImageDraw, ImageFont

batches_to_visualize = 1

def getRelevantWordIndicesAndScores(wordEmbeddings, wordVectors) -> typing.Iterable[typing.Tuple[int, float]]:
    # Normalize the lengths of all the embedding vectors, so that the similarity scores are in [0, 1]:
    normalizedWordEmbeddings = wordEmbeddings / np.linalg.norm(wordEmbeddings, axis=-1)[:, np.newaxis]
    normalizedWordVectors = wordVectors / np.linalg.norm(wordVectors, axis=-1)[:, np.newaxis]

    similarity_scores = np.dot(normalizedWordVectors, np.transpose(normalizedWordEmbeddings))
    maximum_indices = similarity_scores.argmax(axis=-1)
    return [(maximum_index, max(0, similarity_scores[query_index, maximum_index])) for query_index, maximum_index in enumerate(maximum_indices)]

def visualizeEncoderAttention(sources, idx2en, tensors_of_interest, batch_index: int):
    x_step_size = 100
    y_step_size = 100
    colors = ((0, 0, 255), (0, 255, 0), (255, 0, 0), (255, 0, 255), (0, 180, 255), (200, 200, 200), (255, 100, 0), (100, 0, 255))
    assert len(colors) == hp.num_heads
    attention_strength_threshold = 0.3

    for sentence_index, sentence in enumerate(sources):
        print("Source sentence:", sentence)

        im = Image.new('RGB', (x_step_size * (hp.maxlen + 1), y_step_size * hp.num_blocks), color=(255, 255, 255, 255))
        draw = ImageDraw.Draw(im, 'RGBA')

        # If you're not on paperspace, you get to have nicer text with the following:
        # noinspection PyBroadException
        try:
            font = ImageFont.truetype("arial.ttf", 12)
            draw.font = font
        except:
            pass

        for layer in range(hp.num_blocks):
            attn_signal_strength = tensors_of_interest['Attention-Signal-Strength%s' % layer]
            residual_signal_strength = tensors_of_interest['Residual-Signal-Strength%s' % layer]
            activation = tensors_of_interest['Activation%s' % layer]
            english_word_embedding = tensors_of_interest['English-Word-Embedding']
            resulting_signal = tensors_of_interest['Output%s' % layer]

            for head_index in range(hp.num_heads):
                for q_index in range(hp.maxlen):
                    for k_index in range(hp.maxlen):
                        color = colors[head_index]
                        attention_strength = activation[head_index * hp.batch_size + sentence_index][q_index][k_index]
                        if attention_strength > attention_strength_threshold:
                            draw.line(
                                (((q_index + 1) * x_step_size, im.size[1] - (layer + 1) * y_step_size),
                                 ((k_index + 1) * x_step_size, im.size[1] - layer * y_step_size)),
                                fill=color + (int(255 * attention_strength),),
                                width=3)

                    residual_variance = residual_signal_strength[sentence_index][q_index] ** 2
                    attention_variance = attn_signal_strength[sentence_index][q_index] ** 2
                    # print("Variances:", residual_variance, attention_variance)
                    strength_ratio = residual_variance / (residual_variance + attention_variance)
                    scale = strength_ratio * 10.0

                    dot_x = (q_index + 1) * x_step_size
                    dot_y = im.size[1] - (layer + 1) * y_step_size
                    draw.ellipse((dot_x - scale, dot_y - scale, dot_x + scale, dot_y + scale), fill=(0, 0, 0, 127))

            relevant_words_and_scores = getRelevantWordIndicesAndScores(english_word_embedding, resulting_signal[sentence_index])
            for q_index, (relevant_word_index, relevance_score) in enumerate(relevant_words_and_scores):
                relevant_word = idx2en[relevant_word_index]
                text_size_x, text_size_y = draw.textsize(relevant_word)
                text_x = (q_index + 1) * x_step_size - text_size_x / 2.0
                text_y = im.size[1] - (layer + 1) * y_step_size + 15

                assert 0.0 <= relevance_score <= 1.0
                draw.rectangle(((text_x, text_y), (text_x + text_size_x, text_y + text_size_y)), fill=(255, 255, 255, int(180 * relevance_score)))
                text_opacity = int(200 * relevance_score + 55)
                draw.text((text_x, text_y), text=relevant_word, fill=(255 - text_opacity, 255 - text_opacity, 255 - text_opacity, text_opacity))

        for i, word in enumerate(sentence.split()):
            text_size_x, text_size_y = draw.textsize(word)
            text_x = (i + 1) * x_step_size - text_size_x / 2.0
            text_y = im.size[1] - 20

            draw.rectangle(((text_x, text_y), (text_x + text_size_x, text_y + text_size_y)), fill=(255, 255, 255, 127))
            draw.text((text_x, text_y), text=word, fill=(0, 0, 0, 255))

        del draw
        im.save("fig/Activation-Batch-{}-Sentence-{}.png".format(batch_index, sentence_index), "PNG")

def eval():
    # Load graph
    g = Graph(is_training=False)
    print("Graph loaded")

    # Load data
    X, Sources, Targets = load_test_data()
    en2idx, idx2en = load_en_vocab()
    de2idx, idx2de = load_de_vocab()

    # Start session
    with g.graph.as_default():
        sv = tf.train.Supervisor()
        with sv.managed_session(config=tf.ConfigProto(allow_soft_placement=True)) as sess:
            # Restore parameters
            sv.saver.restore(sess, tf.train.latest_checkpoint(hp.logdir))
            print("Restored!")

            # Get model name
            mname = open(hp.logdir + '/checkpoint', 'r').read().split('"')[1]  # model name

            # Inference
            if not os.path.exists('results'):
                os.mkdir('results')
            with codecs.open("results/" + mname, "w", "utf-8") as fout:
                list_of_refs, hypotheses = [], []
                for i in range(len(X) // hp.batch_size):

                    # Get mini-batches
                    x = X[i * hp.batch_size: (i + 1) * hp.batch_size]
                    sources = Sources[i * hp.batch_size: (i + 1) * hp.batch_size]
                    targets = Targets[i * hp.batch_size: (i + 1) * hp.batch_size]

                    # Autoregressive inference
                    preds = np.zeros((hp.batch_size, hp.maxlen), np.int32)
                    for j in range(hp.maxlen):
                        tensors = [g.preds] + list(g.tensors_of_interest.values())
                        tensors_out = sess.run(tensors, {g.x: x, g.y: preds})
                        _preds = tensors_out[0]
                        preds[:, j] = _preds[:, j]

                        print([idx2de[idx] for idx in preds[0]])

                        # For the first few batches, we save figures giving the attention structure in the encoder.
                        if j == 0 and i < batches_to_visualize:
                            tensor_keys = [None] + list(g.tensors_of_interest.keys())  # Add a null key at the start so it lines up with the tensors_out list
                            visualizeEncoderAttention(sources=sources, idx2en=idx2en, tensors_of_interest={key: value for key, value in zip(tensor_keys, tensors_out)}, batch_index=i)

                    # Write to file
                    for source, target, pred in zip(sources, targets, preds):  # sentence-wise
                        got = " ".join(idx2de[idx] for idx in pred).split("</S>")[0].strip()
                        fout.write("- source: " + source + "\n")
                        fout.write("- expected: " + target + "\n")
                        fout.write("- got: " + got + "\n\n")
                        fout.flush()

                        # bleu score
                        ref = target.split()
                        hypothesis = got.split()
                        if len(ref) > 3 and len(hypothesis) > 3:
                            list_of_refs.append([ref])
                            hypotheses.append(hypothesis)

                # Calculate bleu score
                score = corpus_bleu(list_of_refs, hypotheses)
                fout.write("Bleu Score = " + str(100 * score))

if __name__ == '__main__':
    eval()
    print("Done")