barbier_victor_2_bert_tuning_112022.ipynb

{"cells":[{"cell_type":"markdown","metadata":{"id":"9h3j6wQy5T8y"},"source":["# Set up"]},{"cell_type":"markdown","metadata":{"id":"G5jSfmKIQmjP"},"source":["Define work location :"]},{"cell_type":"code","execution_count":3,"metadata":{"executionInfo":{"elapsed":6,"status":"ok","timestamp":1669826360622,"user":{"displayName":"Victor Barbier","userId":"10827788304456120491"},"user_tz":-60},"id":"PCnjx3aDQpUG"},"outputs":[],"source":["# Define work location\n","work_loc = 'local'  # colab (google) or local (pc)"]},{"cell_type":"markdown","metadata":{"id":"sIFTGXGPw4Ic"},"source":["Mount Google Drive :"]},{"cell_type":"code","execution_count":4,"metadata":{"colab":{"base_uri":"https://localhost:8080/"},"executionInfo":{"elapsed":25919,"status":"ok","timestamp":1669826386536,"user":{"displayName":"Victor Barbier","userId":"10827788304456120491"},"user_tz":-60},"id":"jBIscIIgw1pF","outputId":"309cd372-3c3f-49fc-b9b5-e32eaf4c3e89"},"outputs":[],"source":["project_path = \"My Drive/OPENCLASSROOMS/Projets DATA/6. Classification biens consommation\"\n","\n","if (work_loc == 'colab'):\n","    # Mount Drive folder\n","    from google.colab import drive\n","    drive.mount('/content/drive')\n","\n","    # Google Drive path\n","    work_path = \"/content/drive/\"\n","\n","elif (work_loc == 'local'):\n","    # Local drive path\n","    work_path = \"/Users/victor/Library/CloudStorage/GoogleDrive-vicozer@gmail.com/\"\n","\n","else:\n","    raise ValueError(\"Choose between 'colab' or 'local' work!\")"]},{"cell_type":"markdown","metadata":{"id":"TiKzLZWRRufr"},"source":["Python librairies install :"]},{"cell_type":"code","execution_count":5,"metadata":{"colab":{"base_uri":"https://localhost:8080/"},"executionInfo":{"elapsed":56506,"status":"ok","timestamp":1669826443038,"user":{"displayName":"Victor Barbier","userId":"10827788304456120491"},"user_tz":-60},"id":"VQAahsxeRxaq","outputId":"521fadb1-66de-4e4a-9913-1ea97200729b"},"outputs":[],"source":["if work_loc == 'colab':\n","  %pip install transformers"]},{"cell_type":"markdown","metadata":{"id":"Ep7bFK5h5YDB"},"source":["Python librairies imports :"]},{"cell_type":"code","execution_count":6,"metadata":{"executionInfo":{"elapsed":13034,"status":"ok","timestamp":1669826456064,"user":{"displayName":"Victor Barbier","userId":"10827788304456120491"},"user_tz":-60},"id":"o7zNBmZWRW1m"},"outputs":[],"source":["import numpy as np\n","import pandas as pd\n","\n","import matplotlib.pyplot as plt\n","import seaborn as sns\n","\n","from sklearn.model_selection import train_test_split\n","\n","# import tensorflow\n","import torch\n","from transformers import TrainingArguments, Trainer\n","from transformers import BertForSequenceClassification, BertTokenizerFast\n","\n","import evaluate\n","\n","import warnings\n","warnings.simplefilter(action='ignore', category=FutureWarning)\n","\n","%matplotlib inline\n","sns.set_theme(palette=\"Set2\")"]},{"cell_type":"markdown","metadata":{"id":"U8ETewrVdoNK"},"source":["Load pycodestyle (PEP8) :"]},{"cell_type":"code","execution_count":7,"metadata":{"executionInfo":{"elapsed":10,"status":"ok","timestamp":1669826456065,"user":{"displayName":"Victor Barbier","userId":"10827788304456120491"},"user_tz":-60},"id":"WpsLq42YdoNL"},"outputs":[],"source":["if (work_loc == 'local'):\n","    # Pycodestyle (PEP8)\n","    %load_ext pycodestyle_magic\n","    %pycodestyle_on --max_line_length 125  # GitHub limit"]},{"cell_type":"markdown","metadata":{"id":"ttBI6Jfp-KFe"},"source":["Load data :"]},{"cell_type":"code","execution_count":58,"metadata":{"executionInfo":{"elapsed":2551,"status":"ok","timestamp":1669826458608,"user":{"displayName":"Victor Barbier","userId":"10827788304456120491"},"user_tz":-60},"id":"w06w3Eou-G7H"},"outputs":[{"name":"stderr","output_type":"stream","text":["6:13: W291 trailing whitespace\n","14:12: E225 missing whitespace around operator\n"]},{"name":"stdout","output_type":"stream","text":["['Specifications of Sivanna Baking Powder Rouge  Highlighter (Dark pink brown shimmer 10) General Traits Organic Type Mineral Quantity 12 Shade Dark pink brown shimmer 10 Ideal For Girls, Women Container Type Compact Case Highlighter Traits Multi-shaded Yes Texture Powder based In the Box Highlighter', 'Buy Recon SuperChef Kadhai for Rs.715 online. Recon SuperChef Kadhai at best prices with FREE shipping & cash on delivery. Only Genuine Products. 30 Day Replacement Guarantee.', 'Flipkart.com: Buy Vincent Valentine Paris Set of New Dark & Dark Fire Deodorants Combo Set online only for Rs. 310 from Flipkart.com. Only Genuine Products. 30 Day Replacement Guarantee. Free Shipping. Cash On Delivery!', 'Tatvaarts Elephant Mini ( Fine-Finish) Showpiece  -  8.89 cm (Brass, Gold)\\r\\n                         Price: Rs. 4,200\\r\\n\\t\\t\\t\\t\\r\\n\\t\\t\\tElephants are such animals that everybody loves them, be it in reality or in form of toys, paintings, on accessories or artifacts. This is an adorable elephant which is adorned with traditional and beautiful hand carved designs.Specialty: Made from brass (alloy of copper and zinc) Hand carved designs Traditional and natural feel Lustrous and gold like appearance Handmade by the craftsmen\\xa0\\r\\nElephants are such animals that everybody loves them, be it in reality or in form of toys, paintings, on accessories or artifacts. This is an adorable elephant which is adorned with traditional and beautiful hand carved designs.Specialty: Made from brass (alloy of copper and zinc) Hand carved designs Traditional and natural feel Lustrous and gold like appearance Handmade by the craftsmen\\xa0', 'Buy Anjalika Brass Laddu Gopal Showpiece  -  6 cm for Rs.900 online. Anjalika Brass Laddu Gopal Showpiece  -  6 cm at best prices with FREE shipping & cash on delivery. Only Genuine Products. 30 Day Replacement Guarantee.']\n","['Disney DW100405 Digital Watch  - For Boys - Buy Disney DW100405 Digital Watch  - For Boys  DW100405 Online at Rs.1345 in India Only at Flipkart.com. - Great Discounts, Only Genuine Products, 30 Day Replacement Guarantee, Free Shipping. Cash On Delivery!', 'Recrafto Buddha Showpiece  -  15.24 cm (Brass, Black, Green)\\r\\n                         Price: Rs. 3,500\\r\\n\\t\\t\\t\\t\\r\\n\\t\\t\\tRecrafto Brass Buddha Religious Idols\\r\\nRecrafto Brass Buddha Religious Idols', 'Buy Asus RP-N53 Dual-Band Wireless-N600 Range Extender only for Rs. 8000 from Flipkart.com. Only Genuine Products. 30 Day Replacement Guarantee. Free Shipping. Cash On Delivery!', 'Buy Crystle Crstl - Pink Crystle01 USB Led Light for Rs.170 online. Crystle Crstl - Pink Crystle01 USB Led Light at best prices with FREE shipping & cash on delivery. Only Genuine Products. 30 Day Replacement Guarantee.', 'Buy Crystle Crystle01 CRSTL - Bluecensor USB Led Light, USB Cable for Rs.199 online. Crystle Crystle01 CRSTL - Bluecensor USB Led Light, USB Cable at best prices with FREE shipping & cash on delivery. Only Genuine Products. 30 Day Replacement Guarantee.']\n"]}],"source":["# Load data\n","clean_df = pd.read_csv(\"clean_df.csv\")\n","train_df = pd.read_csv(\"train_df.csv\")\n","test_df = pd.read_csv(\"test_df.csv\")\n","\n","# Get corpus \n","corpus = clean_df['corpus'].to_list()\n","train_corpus = train_df['corpus'].to_list()\n","test_corpus = test_df['corpus'].to_list()\n","\n","# Get labels\n","labels = clean_df['labels'].to_list()\n","train_labels = train_df['labels'].to_list()\n","test_labels= test_df['labels'].to_list()\n","labels_encoded = clean_df['labels_encoded'].to_list()\n","train_labels_encoded = train_df['labels_encoded'].to_list()\n","test_labels_encoded = test_df['labels_encoded'].to_list()\n","\n","print(train_corpus[:5])\n","print(test_corpus[:5])"]},{"cell_type":"markdown","metadata":{"id":"NimfDEr5Qket"},"source":["# Functions"]},{"cell_type":"code","execution_count":10,"metadata":{"executionInfo":{"elapsed":11,"status":"ok","timestamp":1669826459256,"user":{"displayName":"Victor Barbier","userId":"10827788304456120491"},"user_tz":-60},"id":"pitsfZpYQket"},"outputs":[],"source":["# Display histplot\n","def display_histplot(data, x=None, y=None, title='Histplot', figsize=(12, 8), kde=False):\n","    plt.figure(figsize=figsize)\n","    plt.title(title)\n","    sns.histplot(data=data, x=x, y=y, kde=kde)\n","    plt.show()"]},{"cell_type":"markdown","metadata":{},"source":["# BERT"]},{"cell_type":"markdown","metadata":{},"source":["## Tokenize the corpus"]},{"cell_type":"markdown","metadata":{},"source":["I create **BERT** tokenizer."]},{"cell_type":"code","execution_count":59,"metadata":{},"outputs":[{"name":"stderr","output_type":"stream","text":["loading file vocab.txt from cache at /Users/victor/.cache/huggingface/hub/models--bert-base-uncased/snapshots/0a6aa9128b6194f4f3c4db429b6cb4891cdb421b/vocab.txt\n","loading file tokenizer.json from cache at /Users/victor/.cache/huggingface/hub/models--bert-base-uncased/snapshots/0a6aa9128b6194f4f3c4db429b6cb4891cdb421b/tokenizer.json\n","loading file added_tokens.json from cache at None\n","loading file special_tokens_map.json from cache at None\n","loading file tokenizer_config.json from cache at /Users/victor/.cache/huggingface/hub/models--bert-base-uncased/snapshots/0a6aa9128b6194f4f3c4db429b6cb4891cdb421b/tokenizer_config.json\n","loading configuration file config.json from cache at /Users/victor/.cache/huggingface/hub/models--bert-base-uncased/snapshots/0a6aa9128b6194f4f3c4db429b6cb4891cdb421b/config.json\n","Model config BertConfig {\n","  \"_name_or_path\": \"bert-base-uncased\",\n","  \"architectures\": [\n","    \"BertForMaskedLM\"\n","  ],\n","  \"attention_probs_dropout_prob\": 0.1,\n","  \"classifier_dropout\": null,\n","  \"gradient_checkpointing\": false,\n","  \"hidden_act\": \"gelu\",\n","  \"hidden_dropout_prob\": 0.1,\n","  \"hidden_size\": 768,\n","  \"initializer_range\": 0.02,\n","  \"intermediate_size\": 3072,\n","  \"layer_norm_eps\": 1e-12,\n","  \"max_position_embeddings\": 512,\n","  \"model_type\": \"bert\",\n","  \"num_attention_heads\": 12,\n","  \"num_hidden_layers\": 12,\n","  \"pad_token_id\": 0,\n","  \"position_embedding_type\": \"absolute\",\n","  \"transformers_version\": \"4.24.0\",\n","  \"type_vocab_size\": 2,\n","  \"use_cache\": true,\n","  \"vocab_size\": 30522\n","}\n","\n"]}],"source":["# Create bert tokenizer\n","bert_tokenizer = BertTokenizerFast.from_pretrained('bert-base-uncased')"]},{"cell_type":"markdown","metadata":{},"source":["I tokenize the full corpus with the **BERT** tokenizer."]},{"cell_type":"code","execution_count":60,"metadata":{},"outputs":[{"name":"stderr","output_type":"stream","text":["Token indices sequence length is longer than the specified maximum sequence length for this model (577 > 512). Running this sequence through the model will result in indexing errors\n"]}],"source":["# Tokenize corpus with BERT\n","corpus_encodings = bert_tokenizer(corpus)"]},{"cell_type":"markdown","metadata":{},"source":["I display the distribution of the number of tokens per description to find an optimal description max length."]},{"cell_type":"code","execution_count":61,"metadata":{},"outputs":[{"data":{"text/plain":["count    1050.000000\n","mean      108.136190\n","std        98.834329\n","min        19.000000\n","25%        47.000000\n","50%        70.000000\n","75%       129.750000\n","max       708.000000\n","dtype: float64"]},"metadata":{},"output_type":"display_data"},{"data":{"image/png":"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","text/plain":["<Figure size 1200x800 with 1 Axes>"]},"metadata":{},"output_type":"display_data"}],"source":["# Get number of tokens\n","corpus_tokens_number = pd.Series(corpus_encodings['input_ids']).apply(lambda x: len(x))\n","\n","display(corpus_tokens_number.describe())\n","display_histplot(corpus_tokens_number, kde=True, title=\"Token numbers distribution\")"]},{"cell_type":"markdown","metadata":{},"source":["I choose to use third quantile as max length."]},{"cell_type":"code","execution_count":62,"metadata":{},"outputs":[{"name":"stdout","output_type":"stream","text":["max length =  129\n"]}],"source":["# Q3 as max length\n","max_length = int(corpus_tokens_number.quantile(0.75))\n","print(\"max length = \", max_length)"]},{"cell_type":"markdown","metadata":{},"source":["I tokenize the training and testing corpus."]},{"cell_type":"code","execution_count":64,"metadata":{},"outputs":[],"source":["# Encode train and test texts\n","train_encodings = bert_tokenizer(train_corpus, max_length=max_length, truncation=True, padding=True)\n","test_encodings = bert_tokenizer(test_corpus, max_length=max_length, truncation=True, padding=True)"]},{"cell_type":"markdown","metadata":{},"source":["## Create the datasets"]},{"cell_type":"markdown","metadata":{},"source":["I define a torch Dataset class to create my datasets."]},{"cell_type":"code","execution_count":65,"metadata":{},"outputs":[],"source":["# Custom torch dataset class\n","class FlipkartDataset(torch.utils.data.Dataset):\n","    def __init__(self, encodings, labels):\n","        self.encodings = encodings\n","        self.labels = labels\n","\n","    def __getitem__(self, idx):\n","        item = {key: torch.tensor(val[idx]) for key, val in self.encodings.items()}\n","        item['labels'] = torch.tensor(self.labels[idx])\n","        return item\n","\n","    def __len__(self):\n","        return len(self.labels)"]},{"cell_type":"markdown","metadata":{},"source":["I create the torch training and testing sets."]},{"cell_type":"code","execution_count":66,"metadata":{},"outputs":[],"source":["# Create train and test dataset\n","train_dataset = FlipkartDataset(train_encodings, train_labels_encoded)\n","test_dataset = FlipkartDataset(test_encodings, test_labels_encoded)"]},{"cell_type":"markdown","metadata":{},"source":["I define the **BERT** model for classification."]},{"cell_type":"code","execution_count":67,"metadata":{},"outputs":[{"name":"stdout","output_type":"stream","text":["Number of labels = 19\n"]},{"name":"stderr","output_type":"stream","text":["loading configuration file config.json from cache at /Users/victor/.cache/huggingface/hub/models--bert-base-uncased/snapshots/0a6aa9128b6194f4f3c4db429b6cb4891cdb421b/config.json\n","Model config BertConfig {\n","  \"architectures\": [\n","    \"BertForMaskedLM\"\n","  ],\n","  \"attention_probs_dropout_prob\": 0.1,\n","  \"classifier_dropout\": null,\n","  \"gradient_checkpointing\": false,\n","  \"hidden_act\": \"gelu\",\n","  \"hidden_dropout_prob\": 0.1,\n","  \"hidden_size\": 768,\n","  \"id2label\": {\n","    \"0\": \"LABEL_0\",\n","    \"1\": \"LABEL_1\",\n","    \"2\": \"LABEL_2\",\n","    \"3\": \"LABEL_3\",\n","    \"4\": \"LABEL_4\",\n","    \"5\": \"LABEL_5\",\n","    \"6\": \"LABEL_6\",\n","    \"7\": \"LABEL_7\",\n","    \"8\": \"LABEL_8\",\n","    \"9\": \"LABEL_9\",\n","    \"10\": \"LABEL_10\",\n","    \"11\": \"LABEL_11\",\n","    \"12\": \"LABEL_12\",\n","    \"13\": \"LABEL_13\",\n","    \"14\": \"LABEL_14\",\n","    \"15\": \"LABEL_15\",\n","    \"16\": \"LABEL_16\",\n","    \"17\": \"LABEL_17\",\n","    \"18\": \"LABEL_18\"\n","  },\n","  \"initializer_range\": 0.02,\n","  \"intermediate_size\": 3072,\n","  \"label2id\": {\n","    \"LABEL_0\": 0,\n","    \"LABEL_1\": 1,\n","    \"LABEL_10\": 10,\n","    \"LABEL_11\": 11,\n","    \"LABEL_12\": 12,\n","    \"LABEL_13\": 13,\n","    \"LABEL_14\": 14,\n","    \"LABEL_15\": 15,\n","    \"LABEL_16\": 16,\n","    \"LABEL_17\": 17,\n","    \"LABEL_18\": 18,\n","    \"LABEL_2\": 2,\n","    \"LABEL_3\": 3,\n","    \"LABEL_4\": 4,\n","    \"LABEL_5\": 5,\n","    \"LABEL_6\": 6,\n","    \"LABEL_7\": 7,\n","    \"LABEL_8\": 8,\n","    \"LABEL_9\": 9\n","  },\n","  \"layer_norm_eps\": 1e-12,\n","  \"max_position_embeddings\": 512,\n","  \"model_type\": \"bert\",\n","  \"num_attention_heads\": 12,\n","  \"num_hidden_layers\": 12,\n","  \"pad_token_id\": 0,\n","  \"position_embedding_type\": \"absolute\",\n","  \"transformers_version\": \"4.24.0\",\n","  \"type_vocab_size\": 2,\n","  \"use_cache\": true,\n","  \"vocab_size\": 30522\n","}\n","\n","loading weights file pytorch_model.bin from cache at /Users/victor/.cache/huggingface/hub/models--bert-base-uncased/snapshots/0a6aa9128b6194f4f3c4db429b6cb4891cdb421b/pytorch_model.bin\n","Some weights of the model checkpoint at bert-base-uncased were not used when initializing BertForSequenceClassification: ['cls.seq_relationship.weight', 'cls.seq_relationship.bias', 'cls.predictions.transform.dense.weight', 'cls.predictions.bias', 'cls.predictions.transform.LayerNorm.bias', 'cls.predictions.transform.LayerNorm.weight', 'cls.predictions.decoder.weight', 'cls.predictions.transform.dense.bias']\n","- This IS expected if you are initializing BertForSequenceClassification from the checkpoint of a model trained on another task or with another architecture (e.g. initializing a BertForSequenceClassification model from a BertForPreTraining model).\n","- This IS NOT expected if you are initializing BertForSequenceClassification from the checkpoint of a model that you expect to be exactly identical (initializing a BertForSequenceClassification model from a BertForSequenceClassification model).\n","Some weights of BertForSequenceClassification were not initialized from the model checkpoint at bert-base-uncased and are newly initialized: ['classifier.weight', 'classifier.bias']\n","You should probably TRAIN this model on a down-stream task to be able to use it for predictions and inference.\n"]}],"source":["# Number of labels\n","n_labels = pd.Series(labels).nunique()\n","print(\"Number of labels =\", n_labels)\n","\n","# Load BertForSequenceClassification\n","bert_model = BertForSequenceClassification.from_pretrained(\"bert-base-uncased\", num_labels=n_labels)"]},{"cell_type":"markdown","metadata":{},"source":["I define an evaluation function for the model."]},{"cell_type":"code","execution_count":68,"metadata":{},"outputs":[],"source":["# Evaluation function\n","def bert_compute_metrics(eval_preds):\n","    metric = evaluate.load(\"accuracy\")\n","    logits, labels = eval_preds\n","    predictions = np.argmax(logits, axis=-1)\n","\n","    return metric.compute(predictions=predictions, references=labels)"]},{"cell_type":"markdown","metadata":{},"source":["## Train the model"]},{"cell_type":"markdown","metadata":{},"source":["I define the training arguments."]},{"cell_type":"code","execution_count":69,"metadata":{},"outputs":[{"name":"stderr","output_type":"stream","text":["PyTorch: setting up devices\n","The default value for the training argument `--report_to` will change in v5 (from all installed integrations to none). In v5, you will need to use `--report_to all` to get the same behavior as now. You should start updating your code and make this info disappear :-).\n"]}],"source":["# Training arguments\n","bert_training_args = TrainingArguments(\n","    output_dir='./bert_results',\n","    evaluation_strategy=\"epoch\",\n","    save_strategy=\"epoch\",\n","    learning_rate=2e-5,\n","    per_device_train_batch_size=16,\n","    per_device_eval_batch_size=16,\n","    num_train_epochs=3,\n","    load_best_model_at_end=True,\n","    metric_for_best_model=\"accuracy\"\n",")"]},{"cell_type":"markdown","metadata":{},"source":["I create the torch trainer."]},{"cell_type":"code","execution_count":70,"metadata":{},"outputs":[],"source":["# Torch trainer\n","bert_trainer = Trainer(\n","    model=bert_model,\n","    args=bert_training_args,\n","    train_dataset=train_dataset,\n","    eval_dataset=test_dataset,\n","    tokenizer=bert_tokenizer,\n","    compute_metrics=bert_compute_metrics\n",")"]},{"cell_type":"markdown","metadata":{},"source":["I train the **BERT** model."]},{"cell_type":"code","execution_count":71,"metadata":{},"outputs":[{"name":"stderr","output_type":"stream","text":["***** Running training *****\n","  Num examples = 735\n","  Num Epochs = 3\n","  Instantaneous batch size per device = 16\n","  Total train batch size (w. parallel, distributed & accumulation) = 16\n","  Gradient Accumulation steps = 1\n","  Total optimization steps = 138\n","  Number of trainable parameters = 109496851\n"]},{"data":{"application/vnd.jupyter.widget-view+json":{"model_id":"01b322a64d17432c86c22c973d1052f0","version_major":2,"version_minor":0},"text/plain":["  0%|          | 0/138 [00:00<?, ?it/s]"]},"metadata":{},"output_type":"display_data"},{"name":"stderr","output_type":"stream","text":["You're using a BertTokenizerFast tokenizer. Please note that with a fast tokenizer, using the `__call__` method is faster than using a method to encode the text followed by a call to the `pad` method to get a padded encoding.\n","***** Running Evaluation *****\n","  Num examples = 315\n","  Batch size = 16\n"]},{"data":{"application/vnd.jupyter.widget-view+json":{"model_id":"6947eb48a5a945d5956ff2556f79c9c2","version_major":2,"version_minor":0},"text/plain":["  0%|          | 0/20 [00:00<?, ?it/s]"]},"metadata":{},"output_type":"display_data"},{"name":"stderr","output_type":"stream","text":["Saving model checkpoint to ./bert_results/checkpoint-46\n","Configuration saved in ./bert_results/checkpoint-46/config.json\n"]},{"name":"stdout","output_type":"stream","text":["{'eval_loss': 2.4034199714660645, 'eval_accuracy': 0.5206349206349207, 'eval_runtime': 19.7932, 'eval_samples_per_second': 15.915, 'eval_steps_per_second': 1.01, 'epoch': 1.0}\n"]},{"name":"stderr","output_type":"stream","text":["Model weights saved in ./bert_results/checkpoint-46/pytorch_model.bin\n","tokenizer config file saved in ./bert_results/checkpoint-46/tokenizer_config.json\n","Special tokens file saved in ./bert_results/checkpoint-46/special_tokens_map.json\n","***** Running Evaluation *****\n","  Num examples = 315\n","  Batch size = 16\n"]},{"data":{"application/vnd.jupyter.widget-view+json":{"model_id":"6bd844ea63304063b0fcfa215758e542","version_major":2,"version_minor":0},"text/plain":["  0%|          | 0/20 [00:00<?, ?it/s]"]},"metadata":{},"output_type":"display_data"},{"name":"stderr","output_type":"stream","text":["Saving model checkpoint to ./bert_results/checkpoint-92\n","Configuration saved in ./bert_results/checkpoint-92/config.json\n"]},{"name":"stdout","output_type":"stream","text":["{'eval_loss': 1.9956859350204468, 'eval_accuracy': 0.638095238095238, 'eval_runtime': 19.832, 'eval_samples_per_second': 15.883, 'eval_steps_per_second': 1.008, 'epoch': 2.0}\n"]},{"name":"stderr","output_type":"stream","text":["Model weights saved in ./bert_results/checkpoint-92/pytorch_model.bin\n","tokenizer config file saved in ./bert_results/checkpoint-92/tokenizer_config.json\n","Special tokens file saved in ./bert_results/checkpoint-92/special_tokens_map.json\n","***** Running Evaluation *****\n","  Num examples = 315\n","  Batch size = 16\n"]},{"data":{"application/vnd.jupyter.widget-view+json":{"model_id":"9eb550b161db4fe49ade16cd64c3de9c","version_major":2,"version_minor":0},"text/plain":["  0%|          | 0/20 [00:00<?, ?it/s]"]},"metadata":{},"output_type":"display_data"},{"name":"stderr","output_type":"stream","text":["Saving model checkpoint to ./bert_results/checkpoint-138\n","Configuration saved in ./bert_results/checkpoint-138/config.json\n"]},{"name":"stdout","output_type":"stream","text":["{'eval_loss': 1.8318103551864624, 'eval_accuracy': 0.6666666666666666, 'eval_runtime': 19.7458, 'eval_samples_per_second': 15.953, 'eval_steps_per_second': 1.013, 'epoch': 3.0}\n"]},{"name":"stderr","output_type":"stream","text":["Model weights saved in ./bert_results/checkpoint-138/pytorch_model.bin\n","tokenizer config file saved in ./bert_results/checkpoint-138/tokenizer_config.json\n","Special tokens file saved in ./bert_results/checkpoint-138/special_tokens_map.json\n","\n","\n","Training completed. Do not forget to share your model on huggingface.co/models =)\n","\n","\n","Loading best model from ./bert_results/checkpoint-138 (score: 0.6666666666666666).\n"]},{"name":"stdout","output_type":"stream","text":["{'train_runtime': 585.6626, 'train_samples_per_second': 3.765, 'train_steps_per_second': 0.236, 'train_loss': 2.2912593233412593, 'epoch': 3.0}\n"]},{"data":{"text/plain":["TrainOutput(global_step=138, training_loss=2.2912593233412593, metrics={'train_runtime': 585.6626, 'train_samples_per_second': 3.765, 'train_steps_per_second': 0.236, 'train_loss': 2.2912593233412593, 'epoch': 3.0})"]},"execution_count":71,"metadata":{},"output_type":"execute_result"}],"source":["# Train the model\n","bert_trainer.train()"]},{"cell_type":"markdown","metadata":{},"source":["I evaluate the trained model."]},{"cell_type":"code","execution_count":73,"metadata":{},"outputs":[{"name":"stderr","output_type":"stream","text":["***** Running Evaluation *****\n","  Num examples = 315\n","  Batch size = 16\n"]},{"data":{"application/vnd.jupyter.widget-view+json":{"model_id":"f011e54e9d134d95afdf2b3440920c19","version_major":2,"version_minor":0},"text/plain":["  0%|          | 0/20 [00:00<?, ?it/s]"]},"metadata":{},"output_type":"display_data"},{"data":{"text/plain":["{'eval_loss': 1.8318103551864624,\n"," 'eval_accuracy': 0.6666666666666666,\n"," 'eval_runtime': 17.558,\n"," 'eval_samples_per_second': 17.941,\n"," 'eval_steps_per_second': 1.139,\n"," 'epoch': 3.0}"]},"execution_count":73,"metadata":{},"output_type":"execute_result"}],"source":["# Evaluate the model\n","bert_trainer.evaluate()"]},{"cell_type":"markdown","metadata":{},"source":["I save the trained model."]},{"cell_type":"code","execution_count":74,"metadata":{},"outputs":[{"name":"stderr","output_type":"stream","text":["Saving model checkpoint to ./bert_save/\n","Configuration saved in ./bert_save/config.json\n"]},{"name":"stdout","output_type":"stream","text":["Saving model to ./bert_save/\n"]},{"name":"stderr","output_type":"stream","text":["Model weights saved in ./bert_save/pytorch_model.bin\n","tokenizer config file saved in ./bert_save/tokenizer_config.json\n","Special tokens file saved in ./bert_save/special_tokens_map.json\n"]}],"source":["import os\n","\n","# Define output_directory\n","output_dir = './bert_save/'\n","\n","# Create output directory if needed\n","if not os.path.exists(output_dir):\n","    os.makedirs(output_dir)\n","\n","print(\"Saving model to %s\" % output_dir)\n","\n","# Save trained model\n","bert_trainer.save_model(output_dir)"]}],"metadata":{"accelerator":"GPU","colab":{"provenance":[],"toc_visible":true},"gpuClass":"standard","kernelspec":{"display_name":"Python 3.10.5 64-bit","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":"3.10.5"},"vscode":{"interpreter":{"hash":"aee8b7b246df8f9039afb4144a1f6fd8d2ca17a180786b69acc140d282b71a49"}}},"nbformat":4,"nbformat_minor":0}