{ "cells": [ { "cell_type": "markdown", "metadata": {}, "source": [ "Восстановим константы, словарь и модели из прошлого нотубка" ] }, { "cell_type": "code", "execution_count": 1, "metadata": {}, "outputs": [], "source": [ "import os\n", "import pandas as pd\n", "import torch\n", "import torch.nn as nn\n", "import torch.nn.functional as F\n", "from torch.utils.data import DataLoader, TensorDataset\n", "\n", "from src.models.models import TransformerClassifier, LSTMClassifier, CustomMambaClassifier, SimpleMambaBlock\n", "from src.data_utils.config import DatasetConfig\n", "from src.data_utils.dataset_params import DatasetName\n", "from src.data_utils.dataset_generator import DatasetGenerator\n", "\n", "MAX_SEQ_LEN = 300\n", "EMBEDDING_DIM = 128\n", "BATCH_SIZE = 32 \n", "NUM_CLASSES = 2\n", "SAVE_DIR = \"../pretrained_comparison\" \n", "DATA_DIR = \"../datasets\" \n", "DEVICE = torch.device(\"cuda\" if torch.cuda.is_available() else \"cpu\")\n", "\n", "config = DatasetConfig(load_from_disk=True, path_to_data=DATA_DIR)\n", "generator = DatasetGenerator(DatasetName.IMDB, config=config)\n", "\n", "_, _, _ = generator.generate_dataset() \n", "vocab = generator.vocab\n", "VOCAB_SIZE = len(vocab)\n", "text_processor = generator.get_text_processor()" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "Возьмем всопомгательную функцию из пролшло нотубка" ] }, { "cell_type": "code", "execution_count": 2, "metadata": {}, "outputs": [], "source": [ "from sklearn.metrics import accuracy_score, precision_recall_fscore_support\n", "\n", "def evaluate_on_test(model, test_loader, device, criterion):\n", " model.eval()\n", " total_test_loss = 0\n", " all_preds = []\n", " all_labels = []\n", "\n", " with torch.no_grad():\n", " for batch_X, batch_y in test_loader:\n", " batch_X, batch_y = batch_X.to(device), batch_y.to(device)\n", " outputs = model(batch_X)\n", " loss = criterion(outputs, batch_y)\n", " total_test_loss += loss.item()\n", " \n", " _, predicted = torch.max(outputs.data, 1)\n", " all_preds.extend(predicted.cpu().numpy())\n", " all_labels.extend(batch_y.cpu().numpy())\n", " \n", " avg_test_loss = total_test_loss / len(test_loader)\n", " \n", " accuracy = accuracy_score(all_labels, all_preds)\n", " precision, recall, f1, _ = precision_recall_fscore_support(all_labels, all_preds, average='binary')\n", " \n", " return {'loss': avg_test_loss, 'accuracy': accuracy, 'precision': precision, 'recall': recall, 'f1_score': f1}" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "Создадим генератор датасета и передадим в него уже готовый текстовый процессор, заберем датасет из другого распределения" ] }, { "cell_type": "code", "execution_count": 3, "metadata": {}, "outputs": [], "source": [ "\n", "def create_dataloader(X, y, batch_size, shuffle=True):\n", " X_tensor = torch.as_tensor(X, dtype=torch.long)\n", " y_tensor = torch.as_tensor(y, dtype=torch.long)\n", " dataset = TensorDataset(X_tensor, y_tensor)\n", " return DataLoader(dataset, batch_size=batch_size, shuffle=shuffle)\n", "\n", "text_processor = generator.get_text_processor()\n", "config_polarity = DatasetConfig(\n", " load_from_disk=True,\n", " path_to_data=\"../datasets\",\n", " train_size=25000, # взяли весь датасет\n", " val_size=12500,\n", " test_size=12500,\n", " build_vocab=False\n", ")\n", "generator_polarity = DatasetGenerator(DatasetName.POLARITY, config=config_polarity)\n", "generator_polarity.vocab = generator.vocab\n", "generator_polarity.id2word = generator.id2word\n", "generator_polarity.text_processor = text_processor\n", "(X_train, y_train), (X_val, y_val), (X_test, y_test) = generator_polarity.generate_dataset()\n", "\n", "\n", "test_loader = create_dataloader(X_test, y_test, BATCH_SIZE, shuffle=False)" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "Восстановим конфигурации конфигов моделей из прошлого нотубка" ] }, { "cell_type": "code", "execution_count": 4, "metadata": {}, "outputs": [], "source": [ "model_configs = {\n", " \"CustomMamba\": {\n", " \"class\": CustomMambaClassifier,\n", " \"params\": {'vocab_size': VOCAB_SIZE, 'd_model': EMBEDDING_DIM, 'd_state': 8, \n", " 'd_conv': 4, 'num_layers': 2, 'num_classes': NUM_CLASSES},\n", " },\n", " \"Lib_LSTM\": {\n", " \"class\": LSTMClassifier,\n", " \"params\": {'vocab_size': VOCAB_SIZE, 'embed_dim': EMBEDDING_DIM, 'hidden_dim': 128, \n", " 'num_layers': 2, 'num_classes': NUM_CLASSES, 'dropout': 0.5},\n", " },\n", " \"Lib_Transformer\": {\n", " \"class\": TransformerClassifier,\n", " \"params\": {'vocab_size': VOCAB_SIZE, 'embed_dim': EMBEDDING_DIM, 'num_heads': 4, \n", " 'num_layers': 2, 'num_classes': NUM_CLASSES, 'max_seq_len': MAX_SEQ_LEN},\n", " },\n", "}" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "Теперь посмотрим на результаты" ] }, { "cell_type": "code", "execution_count": 5, "metadata": {}, "outputs": [ { "name": "stderr", "output_type": "stream", "text": [ "/home/gab1k/.cache/pypoetry/virtualenvs/monkey-coding-dl-project-F4QJzkF_-py3.12/lib/python3.12/site-packages/torch/nn/modules/transformer.py:505: UserWarning: The PyTorch API of nested tensors is in prototype stage and will change in the near future. We recommend specifying layout=torch.jagged when constructing a nested tensor, as this layout receives active development, has better operator coverage, and works with torch.compile. (Triggered internally at /pytorch/aten/src/ATen/NestedTensorImpl.cpp:178.)\n", " output = torch._nested_tensor_from_mask(\n" ] }, { "name": "stdout", "output_type": "stream", "text": [ "\n", "\n", "--- Итоговая таблица сравнения моделей на тестовых данных ---\n", " loss accuracy precision recall f1_score\n", "CustomMamba 0.583675 0.70344 0.653410 0.871734 0.746945\n", "Lib_LSTM 0.675894 0.59520 0.574803 0.744423 0.648709\n", "Lib_Transformer 0.618924 0.66432 0.612190 0.904238 0.730091\n" ] } ], "source": [ "results = {}\n", "for model_name, config in model_configs.items(): \n", " model_path = os.path.join(SAVE_DIR, f\"best_model_{model_name.lower()}.pth\") \n", " model = config['class'](**config['params']).to(DEVICE)\n", "\n", " model.load_state_dict(torch.load(model_path, map_location=DEVICE))\n", " criterion = nn.CrossEntropyLoss()\n", " test_metrics = evaluate_on_test(model, test_loader, DEVICE, criterion)\n", " results[model_name] = test_metrics\n", " \n", "results_df = pd.DataFrame(results).T\n", "print(\"\\n\\n--- Итоговая таблица сравнения моделей на тестовых данных ---\")\n", "print(results_df.to_string())\n" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "Снимали тут на \"игрушечных данных\". На даже на них видно, что:\n", " - accuracy выше всего на Mamba\n", " - Трансформер справился тоже неплохо\n", " - LSTM опять проиграл\n", "\n", "В следующем нотбуке обучим Mamba и Transformer на всем датасете и снимем качество на втором. Та модель, которая будет лучше, \"поедет в продакшн\" " ] } ], "metadata": { "kernelspec": { "display_name": "monkey-coding-dl-project-F4QJzkF_-py3.12", "language": "python", "name": "python3" }, "language_info": { "codemirror_mode": { "name": "ipython", "version": 3 }, "file_extension": ".py", "mimetype": "text/x-python", "name": "python", "nbconvert_exporter": "python", "pygments_lexer": "ipython3", "version": "3.12.11" } }, "nbformat": 4, "nbformat_minor": 2 }