File size: 7,997 Bytes
2018b94
 
 
 
f80827c
2018b94
f80827c
 
 
 
 
 
 
 
 
2018b94
 
 
 
 
 
 
9867f8a
2018b94
 
 
f80827c
2018b94
 
f80827c
2018b94
 
 
 
 
 
f80827c
2018b94
 
 
 
f80827c
 
 
 
 
2018b94
 
 
 
 
f80827c
 
2018b94
 
 
 
 
 
 
 
 
f80827c
2018b94
 
 
 
 
 
 
 
 
 
 
f80827c
2018b94
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
f80827c
2018b94
f80827c
2018b94
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
f80827c
2018b94
 
f80827c
 
2018b94
 
 
f80827c
 
2018b94
 
 
f80827c
2018b94
 
f80827c
 
2018b94
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
f80827c
2018b94
 
f80827c
2018b94
 
f80827c
2018b94
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
f80827c
 
2018b94
 
 
 
f80827c
2018b94
 
 
 
f80827c
 
2018b94
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
import gradio as gr
from datasets import disable_caching, load_dataset
from transformer_ranker import TransformerRanker

from demo.config import SAMPLE_SIZE, MAX_SAMPLE_SIZE, ALL_LMS, PRESELECTED_LMS
from demo.utils import (
    BANNER,
    FOOTER,
    CSS,
    UNSET,
    EmbeddingProgressTracker,
    compute_ratio,
    validate_dataset,
    preprocess_dataset,
    ensure_dataset_is_loaded,
)


disable_caching()

with gr.Blocks(css=CSS, theme=None) as demo:

    gr.Markdown(BANNER, elem_classes="banner")

    ##### 1. Load from datasets #####

    gr.Markdown("## 📚 Load Data")

    gr.Markdown(
        "Pick a dataset from the Hugging Face Hub (e.g. `trec`). This defines your downstream task."
    )

    with gr.Group():
        dataset = gr.State(None)

        dataset_id = gr.Textbox(
            label="Dataset identifier",
            placeholder="try: trec, conll2003, ag_news",
            max_lines=1,
        )

        load_dataset_button = gr.Button(
            value="Load data",
            variant="primary",
            interactive=True,
        )

        # enable loading if dataset exists on hub
        dataset_id.change(validate_dataset, inputs=dataset_id, outputs=load_dataset_button)

    gr.Markdown(
        "⚡️ Speed mode on: tweak the downsampling ratio in *Dataset Setup* for quicker runs. "
        "Unlock the full data via [framework](https://github.com/flairNLP/transformer-ranker)."
    )

    ##### data preprocessing #####

    with gr.Accordion("Dataset Setup", open=False) as dataset_config:
        with gr.Row() as dataset_details:
            dataset_id_label = gr.Label("", label="Dataset")
            num_samples = gr.State(0)
            num_samples_label = gr.Label("", label="Dataset size")
            num_samples.change(lambda x: str(x), inputs=[num_samples], outputs=[num_samples_label])

        with gr.Row():
            text_column = gr.Dropdown("", label="Text Column")
            text_pair_column = gr.Dropdown("", label="Text Pair")

        with gr.Row():
            label_column = gr.Dropdown("", label="Labels")
            task_category = gr.Dropdown("", label="Downstream Task")

        with gr.Group():
            downsample_ratio = gr.State(0.0)
            sampling_rate = gr.Slider(20, MAX_SAMPLE_SIZE, label="Sampling rate", value=SAMPLE_SIZE, step=1)
            downsample_ratio_label = gr.Label("", label="Sampling rate")
            downsample_ratio.change(
                lambda x: f"{x:.1%}",
                inputs=[downsample_ratio],
                outputs=[downsample_ratio_label],
            )

            sampling_rate.change(
                compute_ratio,
                inputs=[sampling_rate, num_samples],
                outputs=downsample_ratio,
            )
            num_samples.change(
                compute_ratio,
                inputs=[sampling_rate, num_samples],
                outputs=downsample_ratio,
            )

    def load_hf_dataset(dataset_id):
        try:
            dataset = load_dataset(dataset_id, trust_remote_code=True)
            dataset_details = preprocess_dataset(dataset)
        except ValueError as e:
            gr.Warning(f"Watch out — single datasets only. Cannot load dataset: {e}")

        return (gr.update(value="Loaded"), dataset_id, dataset, *dataset_details)

    load_dataset_button.click(
        load_hf_dataset,
        inputs=[dataset_id],
        outputs=[
            load_dataset_button,
            dataset_id_label,
            dataset,
            task_category,
            text_column,
            text_pair_column,
            label_column,
            num_samples,
        ],
        scroll_to_output=True,
    )

    ########## 2. Select LMs ##########

    gr.Markdown("## 🧠 Select Language Models")

    gr.Markdown(
        "Add two or more pretrained models to compare. "
        "Stick to smaller models here since the demo runs on CPU."
    )

    with gr.Group():
        model_options = [(model_handle.split("/")[-1], model_handle) for model_handle in ALL_LMS]
        models = gr.CheckboxGroup(choices=model_options, label="Model List", value=PRESELECTED_LMS)

    ########## 3. Run ranking ##########

    gr.Markdown("## 🏆 Rank Models")

    gr.Markdown(
        "Rank models by transferability to your task. "
        "More control? Tweak transferability metric and layer aggregation in *Settings*."
    )

    with gr.Group():

        submit_button = gr.Button("Run ranking", variant="primary", interactive=False)

        with gr.Accordion("Advanced Settings", open=False):
            with gr.Row():
                estimator = gr.Dropdown(
                    choices=["hscore", "logme", "knn"],
                    label="Transferability metric",
                    value="hscore",
                )
                layer_aggregator = gr.Dropdown(
                    choices=["lastlayer", "layermean", "bestlayer"],
                    label="Layer aggregation",
                    value="layermean",
                )

        # ranking button works after dataset loads
        dataset.change(
            ensure_dataset_is_loaded,
            inputs=[dataset, text_column, label_column, task_category],
            outputs=submit_button
        )

        label_column.change(
            ensure_dataset_is_loaded,
            inputs=[dataset, text_column, label_column, task_category],
            outputs=submit_button
        )

        text_column.change(
            ensure_dataset_is_loaded,
            inputs=[dataset, text_column, label_column, task_category],
            outputs=submit_button
        )

    def rank_models(
        dataset,
        downsample_ratio,
        selected_models,
        layer_aggregator,
        estimator,
        text_column,
        text_pair_column,
        label_column,
        task_category,
        progress=gr.Progress(),
    ):

        if text_column == UNSET:
            raise gr.Error("Text column is required.")

        if label_column == UNSET:
            raise gr.Error("Label column is required.")

        if task_category == UNSET:
            raise gr.Error("Task category is required.")

        if text_pair_column == UNSET:
            text_pair_column = None

        progress(0.0, "Starting")

        with EmbeddingProgressTracker(progress=progress, model_names=selected_models) as tracker:
            try:
                ranker = TransformerRanker(
                    dataset,
                    dataset_downsample=downsample_ratio,
                    text_column=text_column,
                    text_pair_column=text_pair_column,
                    label_column=label_column,
                    task_category=task_category,
                )

                results = ranker.run(
                    models=selected_models,
                    layer_aggregator=layer_aggregator,
                    estimator=estimator,
                    batch_size=64,
                    tracker=tracker,
                )

                sorted_results = sorted(results._results.items(), key=lambda item: item[1], reverse=True)
                return [(i + 1, model, score) for i, (model, score) in enumerate(sorted_results)]
            except Exception as e:
                gr.Warning(f"Ranking issue: {e}")
                return []

    gr.Markdown("**Leaderboard:** higher score → better downstream performance.")

    ranking_results = gr.Dataframe(
        headers=["Rank", "Model", "Score"],
        datatype=["number", "str", "number"],
        value=[["-", "-", "-"]],
        interactive=False
    )

    submit_button.click(
        rank_models,
        inputs=[
            dataset,
            downsample_ratio,
            models,
            layer_aggregator,
            estimator,
            text_column,
            text_pair_column,
            label_column,
            task_category,
        ],
        outputs=ranking_results,
        scroll_to_output=True,
    )

    gr.Markdown(FOOTER)