File size: 6,184 Bytes
324f080
 
 
 
 
 
42dfeae
324f080
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
888419c
d74c94a
324f080
d74c94a
324f080
d74c94a
324f080
 
 
 
 
 
 
78f11a8
324f080
 
 
 
 
 
0802e6e
324f080
 
 
 
 
 
 
950c460
d74c94a
 
 
324f080
c1274fe
 
611eaf4
 
0802e6e
 
 
 
324f080
 
 
950c460
 
 
d825f6f
ff13355
cbc5727
ff13355
950c460
 
 
ff13355
950c460
 
 
 
611eaf4
 
 
950c460
611eaf4
324f080
 
 
 
 
 
 
 
 
 
611eaf4
324f080
611eaf4
652973c
324f080
 
 
 
 
 
 
 
 
 
 
 
950c460
324f080
652973c
 
 
 
 
324f080
 
 
950c460
324f080
 
 
 
950c460
324f080
950c460
324f080
950c460
324f080
d74c94a
324f080
 
 
652973c
324f080
 
 
652973c
324f080
 
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
from io import BytesIO
import streamlit as st
import pandas as pd
import json
import os
import numpy as np
from streamlit import caching
from PIL import Image
from model.flax_clip_vision_mbart.modeling_clip_vision_mbart import (
    FlaxCLIPVisionMBartForConditionalGeneration,
)
from transformers import MBart50TokenizerFast
from utils import (
    get_transformed_image,
)
import matplotlib.pyplot as plt
from mtranslate import translate


from session import _get_state

state = _get_state()


@st.cache
def load_model(ckpt):
    return FlaxCLIPVisionMBartForConditionalGeneration.from_pretrained(ckpt)


tokenizer = MBart50TokenizerFast.from_pretrained("facebook/mbart-large-50")

language_mapping = {
    "en": "en_XX",
    "de": "de_DE",
    "fr": "fr_XX",
    "es": "es_XX"
}

code_to_name = {
    "en": "English",
    "fr": "French",
    "de": "German",
    "es": "Spanish",
}

@st.cache
def generate_sequence(pixel_values, lang_code, num_beams, temperature, top_p, do_sample, top_k, max_length):
    lang_code = language_mapping[lang_code]
    output_ids = state.model.generate(input_ids=pixel_values, forced_bos_token_id=tokenizer.lang_code_to_id[lang_code], max_length=max_length, num_beams=num_beams, temperature=temperature, top_p = top_p, top_k=top_k, do_sample=do_sample)
    print(output_ids)
    output_sequence = tokenizer.batch_decode(output_ids[0], skip_special_tokens=True, max_length=max_length)
    return output_sequence

def read_markdown(path, parent="./sections/"):
    with open(os.path.join(parent, path)) as f:
        return f.read()


checkpoints = ["./ckpt/ckpt-49499"]  # TODO: Maybe add more checkpoints?
dummy_data = pd.read_csv("reference.tsv", sep="\t")

st.set_page_config(
    page_title="Multilingual Image Captioning",
    layout="wide",
    initial_sidebar_state="collapsed",
    page_icon="./misc/mic-logo.png",
)

st.title("Multilingual Image Captioning")
st.write(
    "[Bhavitvya Malik](https://huggingface.co/bhavitvyamalik), [Gunjan Chhablani](https://huggingface.co/gchhablani)"
)

st.sidebar.title("Generation Parameters")
max_length = st.sidebar.number_input("Max Length", min_value=16, max_value=128, value=64, step=1, help="The maximum length of sequence to be generated.")
do_sample = st.sidebar.checkbox("Sample", value=False, help="Sample from the model instead of using beam search.")
top_k = st.sidebar.number_input("Top K", min_value=10, max_value=200, value=50, step=1, help="The number of highest probability vocabulary tokens to keep for top-k-filtering.")
num_beams = st.sidebar.number_input(label="Number of Beams", min_value=2, max_value=10, value=4, step=1, help="Number of beams to be used in beam search.")
temperature = st.sidebar.select_slider(label="Temperature", options = list(np.arange(0.0,1.1, step=0.1)), value=1.0, help ="The value used to module the next token probabilities.", format_func=lambda x: f"{x:.2f}")
top_p = st.sidebar.select_slider(label = "Top-P", options = list(np.arange(0.0,1.1, step=0.1)),value=1.0, help="Nucleus Sampling : If set to float < 1, only the most probable tokens with probabilities that add up to :obj:`top_p` or higher are kept for generation.", format_func=lambda x: f"{x:.2f}")
if st.sidebar.button("Clear All Cache"):
    caching.clear_cache()
image_col, intro_col = st.beta_columns([3, 8])
image_col.image("./misc/mic-logo.png", use_column_width="always")
intro_col.write(read_markdown("intro.md"))

with st.beta_expander("Usage"):
    st.markdown(read_markdown("usage.md"))

with st.beta_expander("Article"):
    st.write(read_markdown("abstract.md"))
    st.write(read_markdown("caveats.md"))
    st.write("## Methodology")
    st.image(
        "./misc/Multilingual-IC.png"
    )
    st.markdown(read_markdown("pretraining.md"))
    st.write(read_markdown("challenges.md"))
    st.write(read_markdown("social_impact.md"))
    st.write(read_markdown("future_scope.md"))
    st.write(read_markdown("references.md"))
    # st.write(read_markdown("checkpoints.md"))
    st.write(read_markdown("acknowledgements.md"))

if state.model is None:
    with st.spinner("Loading model..."):
        state.model = load_model(checkpoints[0])

first_index = 25
# Init Session State
if state.image_file is None:
    state.image_file = dummy_data.loc[first_index, "image_file"]
    state.caption = dummy_data.loc[first_index, "caption"].strip("- ")
    state.lang_id = dummy_data.loc[first_index, "lang_id"]

    image_path = os.path.join("images", state.image_file)
    image = plt.imread(image_path)
    state.image = image

new_col1, new_col2 = st.beta_columns([5,5])


if new_col1.button("Get a random example", help="Get a random example from one of the seeded examples."):
    sample = dummy_data.sample(1).reset_index()
    state.image_file = sample.loc[0, "image_file"]
    state.caption = sample.loc[0, "caption"].strip("- ")
    state.lang_id = sample.loc[0, "lang_id"]

    image_path = os.path.join("images", state.image_file)
    image = plt.imread(image_path)
    state.image = image

transformed_image = get_transformed_image(state.image)

# Display Image
new_col1.image(state.image, use_column_width="always")
# Display Reference Caption
with new_col1.beta_expander("Reference Caption"):
    st.write("**Reference Caption**: " + state.caption)
    st.markdown(
        f"""**English Translation**: {state.caption if state.lang_id == "en" else translate(state.caption, 'en')}"""
    )

# Select Language
options = list(code_to_name.keys())
lang_id = new_col2.selectbox(
    "Language",
    index=options.index(state.lang_id),
    options=options,
    format_func=lambda x: code_to_name[x],
    help="The language in which caption is to be generated."
)

sequence = ['']
if new_col2.button("Generate Caption", help="Generate a caption in the specified language."):
    with st.spinner("Generating Sequence..."):
        sequence = generate_sequence(transformed_image, lang_id, num_beams, temperature, top_p, do_sample, top_k, max_length)
# print(sequence)

if sequence!=['']:
    new_col2.write(
        "**Generated Caption**: "+sequence[0]
    )

    new_col2.write(
        "**English Translation**: "+ sequence[0] if lang_id=="en" else translate(sequence[0])
    )