Chatbot

In [ ]:

! nvidia-smi

! nvidia-smi

In [ ]:

from google.colab import drive
drive.mount('/content/drive')

from google.colab import drive drive.mount('/content/drive')

In [ ]:

! pip install transformers datasets accelerate peft

! pip install transformers datasets accelerate peft

In [ ]:

import pandas as pd
from datasets import Dataset
from transformers import AutoTokenizer, AutoModelForSeq2SeqLM
from peft import LoraConfig, get_peft_model, TaskType
from transformers import Seq2SeqTrainer, Seq2SeqTrainingArguments, DataCollatorForSeq2Seq

import pandas as pd from datasets import Dataset from transformers import AutoTokenizer, AutoModelForSeq2SeqLM from peft import LoraConfig, get_peft_model, TaskType from transformers import Seq2SeqTrainer, Seq2SeqTrainingArguments, DataCollatorForSeq2Seq

In [ ]:

train_df = pd.read_parquet("train.parquet")
test_df = pd.read_parquet("test.parquet")
train_data = Dataset.from_pandas(train_df)
test_data = Dataset.from_pandas(test_df)

train_df = pd.read_parquet("train.parquet") test_df = pd.read_parquet("test.parquet") train_data = Dataset.from_pandas(train_df) test_data = Dataset.from_pandas(test_df)

In [ ]:

model_id="google/flan-t5-large"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForSeq2SeqLM.from_pretrained(model_id)

model_id="google/flan-t5-large" tokenizer = AutoTokenizer.from_pretrained(model_id) model = AutoModelForSeq2SeqLM.from_pretrained(model_id)

In [ ]:

def preprocess_function(sample,padding="max_length"):
    model_inputs = tokenizer(sample["Human"], max_length=256, padding=padding, truncation=True)
    labels = tokenizer(sample["Assistant"], max_length=256, padding=padding, truncation=True)
    if padding == "max_length":
        labels["input_ids"] = [
            [(l if l != tokenizer.pad_token_id else -100) for l in label] for label in labels["input_ids"]
        ]
    model_inputs["labels"] = labels["input_ids"]
    return model_inputs

def preprocess_function(sample,padding="max_length"): model_inputs = tokenizer(sample["Human"], max_length=256, padding=padding, truncation=True) labels = tokenizer(sample["Assistant"], max_length=256, padding=padding, truncation=True) if padding == "max_length": labels["input_ids"] = [ [(l if l != tokenizer.pad_token_id else -100) for l in label] for label in labels["input_ids"] ] model_inputs["labels"] = labels["input_ids"] return model_inputs

In [ ]:

train_tokenized_dataset = train_data.map(preprocess_function, batched=True, remove_columns=train_data.column_names)
test_tokenized_dataset = test_data.map(preprocess_function, batched=True, remove_columns=test_data.column_names)
print(f"Keys of tokenized dataset: {list(train_tokenized_dataset.features)}")

train_tokenized_dataset = train_data.map(preprocess_function, batched=True, remove_columns=train_data.column_names) test_tokenized_dataset = test_data.map(preprocess_function, batched=True, remove_columns=test_data.column_names) print(f"Keys of tokenized dataset: {list(train_tokenized_dataset.features)}")

In [ ]:

lora_config = LoraConfig(
 r=16,
 lora_alpha=32,
 target_modules=["q", "v"],
 lora_dropout=0.1,
 bias="none",
 task_type=TaskType.SEQ_2_SEQ_LM
)

lora_config = LoraConfig( r=16, lora_alpha=32, target_modules=["q", "v"], lora_dropout=0.1, bias="none", task_type=TaskType.SEQ_2_SEQ_LM )

In [ ]:

model = get_peft_model(model, lora_config)
model.print_trainable_parameters()

model = get_peft_model(model, lora_config) model.print_trainable_parameters()

In [ ]:

label_pad_token_id = -100
data_collator = DataCollatorForSeq2Seq(
    tokenizer,
    model=model,
    label_pad_token_id=label_pad_token_id,
    pad_to_multiple_of=8
)

label_pad_token_id = -100 data_collator = DataCollatorForSeq2Seq( tokenizer, model=model, label_pad_token_id=label_pad_token_id, pad_to_multiple_of=8 )

In [ ]:

from huggingface_hub import notebook_login
notebook_login()

from huggingface_hub import notebook_login notebook_login()

In [ ]:

output_dir="lora-flan-t5-large-chat"
training_args = Seq2SeqTrainingArguments(
    output_dir=output_dir,
    per_device_train_batch_size=4,
    learning_rate=1e-3,
    num_train_epochs=1,
    logging_dir=f"{output_dir}/logs",
    logging_strategy="epoch",
    save_strategy="epoch",
    report_to="tensorboard",
    push_to_hub = True
)

trainer = Seq2SeqTrainer(
    model=model,
    args=training_args,
    data_collator=data_collator,
    train_dataset=train_tokenized_dataset,
)
model.config.use_cache = False

trainer.train()
peft_save_model_id="lora-flan-t5-large-chat"
trainer.model.save_pretrained(peft_save_model_id, push_to_hub=True)
tokenizer.save_pretrained(peft_save_model_id, push_to_hub=True)
trainer.model.base_model.save_pretrained(peft_save_model_id, push_to_hub=True)

output_dir="lora-flan-t5-large-chat" training_args = Seq2SeqTrainingArguments( output_dir=output_dir, per_device_train_batch_size=4, learning_rate=1e-3, num_train_epochs=1, logging_dir=f"{output_dir}/logs", logging_strategy="epoch", save_strategy="epoch", report_to="tensorboard", push_to_hub = True ) trainer = Seq2SeqTrainer( model=model, args=training_args, data_collator=data_collator, train_dataset=train_tokenized_dataset, ) model.config.use_cache = False trainer.train() peft_save_model_id="lora-flan-t5-large-chat" trainer.model.save_pretrained(peft_save_model_id, push_to_hub=True) tokenizer.save_pretrained(peft_save_model_id, push_to_hub=True) trainer.model.base_model.save_pretrained(peft_save_model_id, push_to_hub=True)

In [ ]:

! cp -r /content/lora-flan-t5-large-chat/ /content/drive/MyDrive/Chatbot/

! cp -r /content/lora-flan-t5-large-chat/ /content/drive/MyDrive/Chatbot/

In [ ]:

import torch
from peft import PeftModel, PeftConfig
from transformers import AutoModelForSeq2SeqLM, AutoTokenizer

# Load peft config for pre-trained checkpoint etc.
peft_model_id = "lora-flan-t5-large-chat"
config = PeftConfig.from_pretrained(peft_model_id)

model = AutoModelForSeq2SeqLM.from_pretrained(config.base_model_name_or_path)
tokenizer = AutoTokenizer.from_pretrained(config.base_model_name_or_path)

model = PeftModel.from_pretrained(model, peft_model_id, device_map={"":0}).cuda()
model.eval()

sample = "Human: \nExplain me about the working of Artificial Intelligence. \nAssistant: "
input_ids = tokenizer(sample, return_tensors="pt", truncation=True, max_length=256).input_ids.cuda()
outputs = model.generate(input_ids=input_ids, do_sample=True, top_p=0.9, max_length=256)
print(f"{sample}")

print(tokenizer.batch_decode(outputs.detach().cpu().numpy(), skip_special_tokens=True)[0])

import torch from peft import PeftModel, PeftConfig from transformers import AutoModelForSeq2SeqLM, AutoTokenizer # Load peft config for pre-trained checkpoint etc. peft_model_id = "lora-flan-t5-large-chat" config = PeftConfig.from_pretrained(peft_model_id) model = AutoModelForSeq2SeqLM.from_pretrained(config.base_model_name_or_path) tokenizer = AutoTokenizer.from_pretrained(config.base_model_name_or_path) model = PeftModel.from_pretrained(model, peft_model_id, device_map={"":0}).cuda() model.eval() sample = "Human: \nExplain me about the working of Artificial Intelligence. \nAssistant: " input_ids = tokenizer(sample, return_tensors="pt", truncation=True, max_length=256).input_ids.cuda() outputs = model.generate(input_ids=input_ids, do_sample=True, top_p=0.9, max_length=256) print(f"{sample}") print(tokenizer.batch_decode(outputs.detach().cpu().numpy(), skip_special_tokens=True)[0])

In [ ]: