File size: 5,760 Bytes
54dc7b4
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
47121bf
 
 
 
0c7ca41
 
 
 
 
 
47121bf
54dc7b4
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
from .utils import get_transformed_image

import streamlit as st
import numpy as np
import pandas as pd
import os
import matplotlib.pyplot as plt
from mtranslate import translate
from .utils import (
    read_markdown,
    tokenizer,
    language_mapping,
    code_to_name
)
import requests
from PIL import Image
from .model.flax_clip_vision_mbart.modeling_clip_vision_mbart import (
    FlaxCLIPVisionMBartForConditionalGeneration,
)
from streamlit import caching


def app(state):
    mic_state = state
    with st.beta_expander("Usage"):
        st.write(read_markdown("usage.md"))
    st.write("\n")
    st.write(read_markdown("intro.md"))

    # st.sidebar.title("Generation Parameters")
    max_length = 64

    with st.sidebar.beta_expander('Generation Parameters'):
        do_sample = st.checkbox("Sample", value=False, help="Sample from the model instead of using beam search.")
        top_k = st.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.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.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.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.button("Clear All Cache"):
            caching.clear_cache()


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

    @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 = mic_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

    mic_checkpoints = ["flax-community/clip-vit-base-patch32_mbart-large-50"]  # TODO: Maybe add more checkpoints?
    dummy_data = pd.read_csv("reference.tsv", sep="\t")

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

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

    if mic_state.model is None:
        # Display Top-5 Predictions
        with st.spinner("Loading model..."):
            mic_state.model = load_model(mic_checkpoints[0])

    query1 = st.text_input(
        "Enter a URL to an image",
        value="http://images.cocodataset.org/val2017/000000039769.jpg",
    )

    col1, col2, col3 = st.beta_columns([2,1, 2])
    if col1.button(
        "Get a random example",
        help="Get a random example from the 100 `seeded` image-text pairs.",
    ):
        sample = dummy_data.sample(1).reset_index()
        mic_state.image_file = sample.loc[0, "image_file"]
        mic_state.caption = sample.loc[0, "caption"].strip("- ")
        mic_state.lang_id = sample.loc[0, "lang_id"]

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

    col2.write("OR")

    if col3.button("Use above URL"):
        image_data = requests.get(query1, stream=True).raw
        image = np.asarray(Image.open(image_data))
        mic_state.image = image

    transformed_image = get_transformed_image(mic_state.image)

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

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

    # Select Language
    options = list(code_to_name.keys())
    lang_id = new_col2.selectbox(
        "Language",
        index=options.index(mic_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])
        )




# 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"))