Update app.py

app.py

@@ -1,8 +1,86 @@
+rom transformers import AutoTokenizer, AutoModelForSequenceClassification, AutoConfig
 import gradio as gr
-
-model = "huggingface/yseop/distilbert-base-financial-relation-extraction"
-description = "Please enter input to predict relation"
-title = "FReE Demo"
-examples = [["An A-B trust is a joint trust created by a married couple for the purpose of minimizing estate taxes."]]
-
-gr.Interface.load(model, description=description, title=title, examples=examples).launch()
+from torch.nn import functional as F
+import seaborn
+import matplotlib
+import platform
+from transformers.file_utils import ModelOutput
+if platform.system() == "Darwin":
+    print("MacOS")
+    matplotlib.use('Agg')
+import matplotlib.pyplot as plt
+import io
+from PIL import Image
+import matplotlib.font_manager as fm
+import util
+# global var
+MODEL_NAME = 'yseop/distilbert-base-financial-relation-extraction'
+tokenizer = AutoTokenizer.from_pretrained(MODEL_NAME)
+model = AutoModelForSequenceClassification.from_pretrained(MODEL_NAME)
+config = AutoConfig.from_pretrained(MODEL_NAME)
+MODEL_BUF = {
+    "name": MODEL_NAME,
+    "tokenizer": tokenizer,
+    "model": model,
+    "config": config
+}
+font_dir = ['./']
+for font in fm.findSystemFonts(font_dir):
+    print(font)
+    fm.fontManager.addfont(font)
+plt.rcParams["font.family"] = 'NanumGothicCoding'
+def visualize_attention(sent, attention_matrix, n_words=10):
+    def draw(data, x, y, ax):
+        seaborn.heatmap(data, 
+                        xticklabels=x, square=True, yticklabels=y, vmin=0.0, vmax=1.0, 
+                        cbar=False, ax=ax)
+        
+    # make plt figure with 1x6 subplots
+    fig = plt.figure(figsize=(16, 8))
+    # fig.subplots_adjust(hspace=0.7, wspace=0.2)
+    for i, layer in enumerate(range(1, 12, 2)):
+        ax = fig.add_subplot(2, 3, i+1)
+        ax.set_title("Layer {}".format(layer))
+        draw(attention_matrix[layer], sent if layer > 6 else [], sent if layer in [1,7] else [], ax=ax)
+ 
+    fig.tight_layout()
+    plt.close()
+    return fig
+def change_model_name(name):
+    MODEL_BUF["name"] = name
+    MODEL_BUF["tokenizer"] = AutoTokenizer.from_pretrained(name)
+    MODEL_BUF["model"] = AutoModelForSequenceClassification.from_pretrained(name)
+    MODEL_BUF["config"] = AutoConfig.from_pretrained(name)
+def predict(model_name, text):
+    if model_name != MODEL_NAME:
+        change_model_name(model_name)
+    
+    tokenizer = MODEL_BUF["tokenizer"]
+    model = MODEL_BUF["model"]
+    config = MODEL_BUF["config"]
+    tokenized_text = tokenizer([text], return_tensors='pt')
+    input_tokens = tokenizer.convert_ids_to_tokens(tokenized_text.input_ids[0])
+    input_tokens = util.bytetokens_to_unicdode(input_tokens) if config.model_type in ['roberta', 'gpt', 'gpt2'] else input_tokens
+    model.eval()
+    output, attention = model(**tokenized_text, output_attentions=True, return_dict=False)
+    output = F.softmax(output, dim=-1)
+    result = {}
+    
+    for idx, label in enumerate(output[0].detach().numpy()):
+        result[config.id2label[idx]] = float(label)
+    fig = visualize_attention(input_tokens, attention[0][0].detach().numpy())
+    return result, fig#.logits.detach()#.numpy()#, output.attentions.detach().numpy()
+if __name__ == '__main__':
+    text = 'An A-B trust is a joint trust created by a married couple for the purpose of minimizing estate taxes.'
+    model_name_list = [
+        'yseop/distilbert-base-financial-relation-extraction'
+    ]
+    #Create a gradio app with a button that calls predict()
+    app = gr.Interface(
+        fn=predict,
+        inputs=[gr.inputs.Dropdown(model_name_list, label="Model Name"), 'text'], outputs=['label', 'plot'], 
+        examples = [[MODEL_BUF["name"], text]],
+        title="FReE",
+        description="Financial relations classifier"
+        )
+    app.launch(inline=False)