update app file for space for summarisation from russian text

app.py

@@ -1,79 +1,26 @@
 import gradio as gr
-import tensorflow as tf
-import copy
-import numpy as np
-
-def generate_from_saved(seed_text):
-    # add the start generation of the lukashenko speech from the simple seed
-    # seed_text = 'я не глядя поддержу'
-    weights_path = 'results/weights_lukash.h5'
-    model_path = 'results/Lukashenko_tarakan'
-
-    model = tf.keras.models.load_model(model_path)
-    model.load_weights(weights_path)
-    # Show the Model summary
-    model.summary()
-
-    with open('data/source_text_lukash.txt', 'r') as source_text_file:
-        data = source_text_file.read().splitlines()
-
-    tmp_data = copy.deepcopy(data)
-    sent_length = 0
-    for idx, line in enumerate(data):
-        if len(line) < 5:
-            tmp_data.pop(idx)
-        else:
-            sent_length += len(line.split())
-    data = tmp_data
-    lstm_length = int(sent_length / len(data))
-
-    token = tf.keras.preprocessing.text.Tokenizer()
-    token.fit_on_texts(data)
-    encoded_text = token.texts_to_sequences(data)
-    # Vocabular size
-    vocab_size = len(token.word_counts) + 1
-
-    datalist = []
-    for d in encoded_text:
-        if len(d) > 1:
-            for i in range(2, len(d)):
-                datalist.append(d[:i])
-
-    max_length = 20
-    sequences = tf.keras.preprocessing.sequence.pad_sequences(datalist, maxlen=max_length, padding='pre')
-
-    # X - input data, y - target data
-    X = sequences[:, :-1]
-    y = sequences[:, -1]
-
-    y = tf.keras.utils.to_categorical(y, num_classes=vocab_size)
-    seq_length = X.shape[1]
-    print(f"Sequence length: {seq_length}")
-
-    generated_text = ''
-    number_lines = 3
-    for i in range(number_lines):
-        text_word_list = []
-        for _ in range(lstm_length * 2):
-            encoded = token.texts_to_sequences([seed_text])
-            encoded = tf.keras.preprocessing.sequence.pad_sequences(encoded, maxlen=seq_length, padding='pre')
-
-            y_pred = np.argmax(model.predict(encoded), axis=-1)
-
-            predicted_word = ""
-            for word, index in token.word_index.items():
-                if index == y_pred:
-                    predicted_word = word
-                    break
-
-            seed_text = seed_text + ' ' + predicted_word
-            text_word_list.append(predicted_word)
-
-        seed_text = text_word_list [-1]
-        generated_text = ' '.join(text_word_list)
-        generated_text += '\n'
-
-    return generated_text
-
-demo = gr.Interface(fn=generate_from_saved, inputs="text", outputs="text")
+import torch
+from transformers import GPT2Tokenizer, T5ForConditionalGeneration
+
+
+tokenizer = GPT2Tokenizer.from_pretrained('RussianNLP/FRED-T5-Summarizer', eos_token='</s>')
+model = T5ForConditionalGeneration.from_pretrained('RussianNLP/FRED-T5-Summarizer')
+device = 'cuda'
+model.to(device)
+
+input_text = "<LM> Сократи текст.\n "
+
+def make_summarization(user_text):
+    processing_text = input_text + user_text
+    input_ids = torch.tensor([tokenizer.encode(processing_text)]).to(device)
+    outputs = model.generate(input_ids, eos_token_id=tokenizer.eos_token_id,
+                             num_beams=3,
+                             min_new_tokens=17,
+                             max_new_tokens=200,
+                             do_sample=True,
+                             no_repeat_ngram_size=4,
+                             top_p=0.9)
+    return tokenizer.decode(outputs[0][1:])
+
+demo = gr.Interface(fn=make_summarization, inputs="text", outputs="text")
 demo.launch(share=True)