File size: 8,321 Bytes
094135a
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
import gradio as gr
import torch
import numpy as np
import pandas as pd
from tqdm.auto import tqdm
import matplotlib.pyplot as plt
import matplotlib
from IPython.display import display, HTML
from transformers import AutoTokenizer
from DecompX.src.globenc_utils import GlobencConfig
from DecompX.src.modeling_bert import BertForSequenceClassification
from DecompX.src.modeling_roberta import RobertaForSequenceClassification

plt.style.use("ggplot")
MODELS = ["WillHeld/roberta-base-sst2"]

def plot_clf(tokens, logits, label_names, title="", file_name=None):
    print(tokens)
    plt.figure(figsize=(4.5, 5))
    colors = ["#019875" if l else "#B8293D" for l in (logits >= 0)]
    plt.barh(range(len(tokens)), logits, color=colors)
    plt.axvline(0, color='black', ls='-', lw=2, alpha=0.2)
    plt.gca().invert_yaxis()
    
    max_limit = np.max(np.abs(logits)) + 0.2
    min_limit = -0.01 if np.min(logits) > 0 else -max_limit
    plt.xlim(min_limit, max_limit)
    plt.gca().set_xticks([min_limit, max_limit])
    plt.gca().set_xticklabels(label_names, fontsize=14, fontweight="bold")
    plt.gca().set_yticks(range(len(tokens)))
    plt.gca().set_yticklabels(tokens)

    plt.gca().yaxis.tick_right()
    for xtick, color in zip(plt.gca().get_yticklabels(), colors):
        xtick.set_color(color)
        xtick.set_fontweight("bold")
        xtick.set_verticalalignment("center")

    for xtick, color in zip(plt.gca().get_xticklabels(), ["#B8293D", "#019875"]):
        xtick.set_color(color)
    # plt.title(title, fontsize=14, fontweight="bold")
    plt.title(title)
    plt.tight_layout()

def print_importance(importance, tokenized_text, discrete=False, prefix="", no_cls_sep=False):
    """
    importance: (sent_len)
    """
    if no_cls_sep:
        importance = importance[1:-1]
        tokenized_text = tokenized_text[1:-1]
    importance = importance / np.abs(importance).max() / 1.5  # Normalize
    if discrete:
        importance = np.argsort(np.argsort(importance)) / len(importance) / 1.6
    
    html = "<pre style='color:black; padding: 3px;'>"+prefix
    for i in range(len(tokenized_text)):
        if importance[i] >= 0:
            rgba = matplotlib.colormaps.get_cmap('Greens')(importance[i])   # Wistia
        else:
            rgba = matplotlib.colormaps.get_cmap('Reds')(np.abs(importance[i]))   # Wistia
        text_color = "color: rgba(255, 255, 255, 1.0); " if np.abs(importance[i]) > 0.9 else ""
        color = f"background-color: rgba({rgba[0]*255}, {rgba[1]*255}, {rgba[2]*255}, {rgba[3]}); " + text_color
        html += (f"<span style='"
                 f"{color}"
                 f"color:black; border-radius: 5px; padding: 3px;"
                 f"font-weight: {int(800)};"
                 "'>")
        html += tokenized_text[i].replace('<', "[").replace(">", "]")
        html += "</span> "
    html += "</pre>"
    # display(HTML(html))
    return html

def print_preview(decompx_outputs_df, idx=0, discrete=False):
    html = ""
    NO_CLS_SEP = False
    df = decompx_outputs_df
    for col in ["importance_last_layer_aggregated", "importance_last_layer_classifier"]:
        if col in df and df[col][idx] is not None:
            if "aggregated" in col:
                sentence_importance = df[col].iloc[idx][0, :]
            if "classifier" in col:
                for label in range(df[col].iloc[idx].shape[-1]):
                    sentence_importance = df[col].iloc[idx][:, label]
                    html += print_importance(
                        sentence_importance,
                        df["tokens"].iloc[idx], 
                        prefix=f"{col.split('_')[-1]} Label{label}:".ljust(20),
                        no_cls_sep=NO_CLS_SEP,
                        discrete=False
                    )
                break
                sentence_importance = df[col].iloc[idx][:, df["label"].iloc[idx]]
            html += print_importance(
                sentence_importance,
                df["tokens"].iloc[idx], 
                prefix=f"{col.split('_')[-1]}:".ljust(20),
                no_cls_sep=NO_CLS_SEP,
                discrete=discrete
            )
    return "<div style='overflow:auto; background-color:white; padding: 10px;'>" + html

def run_decompx(text, model):
    """
    Provide DecompX Token Explanation of Model on Text
    """
    SENTENCES = [text, "nothing"]
    CONFIGS = {
        "DecompX":
            GlobencConfig(
                include_biases=True,
                bias_decomp_type="absdot",
                include_LN1=True,
                include_FFN=True,
                FFN_approx_type="GeLU_ZO",
                include_LN2=True,
                aggregation="vector",
                include_classifier_w_pooler=True,
                tanh_approx_type="ZO",
                output_all_layers=True,
                output_attention=None,
                output_res1=None,
                output_LN1=None,
                output_FFN=None,
                output_res2=None,
                output_encoder=None,
                output_aggregated="norm",
                output_pooler="norm",
                output_classifier=True,
            ),
    }
    MODEL = model
    # LOAD MODEL AND TOKENIZER
    tokenizer = AutoTokenizer.from_pretrained(MODEL)
    tokenized_sentence = tokenizer(SENTENCES, return_tensors="pt", padding=True)
    batch_lengths = tokenized_sentence['attention_mask'].sum(dim=-1)
    if "roberta" in MODEL:
        model = RobertaForSequenceClassification.from_pretrained(MODEL)
    elif "bert" in MODEL:
        model = BertForSequenceClassification.from_pretrained(MODEL)
    else:
        raise Exception(f"Not implemented model: {MODEL}")
    # RUN DECOMPX
    with torch.no_grad():
        model.eval()
        logits, hidden_states, globenc_last_layer_outputs, globenc_all_layers_outputs = model(
            **tokenized_sentence, 
            output_attentions=False, 
            return_dict=False, 
            output_hidden_states=True, 
            globenc_config=CONFIGS["DecompX"]
        )
    decompx_outputs = {
        "tokens": [tokenizer.convert_ids_to_tokens(tokenized_sentence["input_ids"][i][:batch_lengths[i]]) for i in range(len(SENTENCES))],
        "logits": logits.cpu().detach().numpy().tolist(),  # (batch, classes)
        "cls": hidden_states[-1][:, 0, :].cpu().detach().numpy().tolist()# Last layer & only CLS -> (batch, emb_dim)
    }

    ### globenc_last_layer_outputs.classifier ~ (8, 55, 2) ###
    importance = np.array([g.squeeze().cpu().detach().numpy() for g in globenc_last_layer_outputs.classifier]).squeeze()  # (batch, seq_len, classes)
    importance = [importance[j][:batch_lengths[j], :] for j in range(len(importance))]
    decompx_outputs["importance_last_layer_classifier"] = importance

    ### globenc_all_layers_outputs.aggregated ~ (12, 8, 55, 55) ###
    importance = np.array([g.squeeze().cpu().detach().numpy() for g in globenc_all_layers_outputs.aggregated])  # (layers, batch, seq_len, seq_len)
    importance = np.einsum('lbij->blij', importance)  # (batch, layers, seq_len, seq_len)
    importance = [importance[j][:, :batch_lengths[j], :batch_lengths[j]] for j in range(len(importance))]
    decompx_outputs["importance_all_layers_aggregated"] = importance

    decompx_outputs_df = pd.DataFrame(decompx_outputs)
    idx = 0
    pred_label = np.argmax(decompx_outputs_df.iloc[idx]["logits"], axis=-1)
    label = decompx_outputs_df.iloc[idx]["importance_last_layer_classifier"][:, pred_label]
    tokens = decompx_outputs_df.iloc[idx]["tokens"][1:-1]
    label = label[1:-1]
    label = label / np.max(np.abs(label))
    plot_clf(tokens, label, ['-','+'], title=f"DecompX for Predicted Label: {pred_label}", file_name="example_sst2_our_method")
    return plt, print_preview(decompx_outputs_df)

demo = gr.Interface(
    fn=run_decompx,
    inputs=[
        gr.components.Textbox(label="Text"),
        gr.components.Dropdown(label="Model", choices=MODELS),
    ],
    outputs=["plot", "html"],
    examples=[["Building a translation demo with Gradio is so easy!", "WillHeld/roberta-base-sst2"]],
    cache_examples=False,
    title="DecompX Demo",
    description="This demo is a simplified version of the original [NLLB-Translator](https://huggingface.co/spaces/Narrativaai/NLLB-Translator) space"
)

demo.launch()