bertology.py

#!/usr/bin/env python3
import os
import argparse
import logging
from datetime import timedelta, datetime
from tqdm import tqdm

import numpy as np

import torch
from torch.utils.data import DataLoader, SequentialSampler, TensorDataset, Subset
from torch.utils.data.distributed import DistributedSampler
from torch.nn import CrossEntropyLoss, MSELoss

from pytorch_pretrained_bert import BertForSequenceClassification, BertTokenizer

from run_classifier_dataset_utils import processors, output_modes, convert_examples_to_features, compute_metrics


logger = logging.getLogger(__name__)


def entropy(p):
    plogp = p * torch.log(p)
    plogp[p == 0] = 0
    return -plogp.sum(dim=-1)

def print_1d_tensor(tensor, prefix=""):
    if tensor.dtype != torch.long:
        logger.info(prefix + "\t".join(f"{x:.5f}" for x in tensor.cpu().data))
    else:
        logger.info(prefix + "\t".join(f"{x:d}" for x in tensor.cpu().data))

def print_2d_tensor(tensor):
    logger.info("lv, h >\t" + "\t".join(f"{x + 1}" for x in range(len(tensor))))
    for row in range(len(tensor)):
        print_1d_tensor(tensor[row], prefix=f"layer {row + 1}:\t")

def compute_heads_importance(args, model, eval_dataloader, compute_entropy=True, compute_importance=True, head_mask=None):
    """ Example on how to use model outputs to compute:
        - head attention entropy (activated by setting output_attentions=True when we created the model
        - head importance scores according to http://arxiv.org/abs/1905.10650
            (activated by setting keep_multihead_output=True when we created the model)
    """
    # Prepare our tensors
    n_layers, n_heads = model.bert.config.num_hidden_layers, model.bert.config.num_attention_heads
    head_importance = torch.zeros(n_layers, n_heads).to(args.device)
    attn_entropy = torch.zeros(n_layers, n_heads).to(args.device)
    preds = None
    labels = None
    tot_tokens = 0.0

    for step, batch in enumerate(tqdm(eval_dataloader, desc="Iteration", disable=args.local_rank not in [-1, 0])):
        batch = tuple(t.to(args.device) for t in batch)
        input_ids, input_mask, segment_ids, label_ids = batch

        # Do a forward pass (not in torch.no_grad() since we need gradients for importance score - see below)
        all_attentions, logits = model(input_ids, token_type_ids=segment_ids, attention_mask=input_mask, head_mask=head_mask)

        if compute_entropy:
            # Update head attention entropy
            for layer, attn in enumerate(all_attentions):
                masked_entropy = entropy(attn.detach()) * input_mask.float().unsqueeze(1)
                attn_entropy[layer] += masked_entropy.sum(-1).sum(0).detach()

        if compute_importance:
            # Update head importance scores with regards to our loss
            # First, backpropagate to populate the gradients
            if args.output_mode == "classification":
                loss_fct = CrossEntropyLoss()
                loss = loss_fct(logits.view(-1, args.num_labels), label_ids.view(-1))
            elif args.output_mode == "regression":
                loss_fct = MSELoss()
                loss = loss_fct(logits.view(-1), label_ids.view(-1))
            loss.backward()
            # Second, compute importance scores according to http://arxiv.org/abs/1905.10650
            multihead_outputs = model.bert.get_multihead_outputs()
            for layer, mh_layer_output in enumerate(multihead_outputs):
                dot = torch.einsum("bhli,bhli->bhl", [mh_layer_output.grad, mh_layer_output])
                head_importance[layer] += dot.abs().sum(-1).sum(0).detach()

        # Also store our logits/labels if we want to compute metrics afterwards
        if preds is None:
            preds = logits.detach().cpu().numpy()
            labels = label_ids.detach().cpu().numpy()
        else:
            preds = np.append(preds, logits.detach().cpu().numpy(), axis=0)
            labels = np.append(labels, label_ids.detach().cpu().numpy(), axis=0)

        tot_tokens += input_mask.float().detach().sum().data

    # Normalize
    attn_entropy /= tot_tokens
    head_importance /= tot_tokens
    # Layerwise importance normalization
    if not args.dont_normalize_importance_by_layer:
        exponent = 2
        norm_by_layer = torch.pow(torch.pow(head_importance, exponent).sum(-1), 1/exponent)
        head_importance /= norm_by_layer.unsqueeze(-1) + 1e-20

    if not args.dont_normalize_global_importance:
        head_importance = (head_importance - head_importance.min()) / (head_importance.max() - head_importance.min())

    return attn_entropy, head_importance, preds, labels

def run_model():
    parser = argparse.ArgumentParser()
    parser.add_argument('--model_name_or_path', type=str, default='bert-base-cased-finetuned-mrpc', help='pretrained model name or path to local checkpoint')
    parser.add_argument("--task_name", type=str, default='mrpc', help="The name of the task to train.")
    parser.add_argument("--data_dir", type=str, required=True, help="The input data dir. Should contain the .tsv files (or other data files) for the task.")
    parser.add_argument("--output_dir", type=str, required=True, help="The output directory where the model predictions and checkpoints will be written.")
    parser.add_argument("--data_subset", type=int, default=-1, help="If > 0: limit the data to a subset of data_subset instances.")
    parser.add_argument("--overwrite_output_dir", action='store_true', help="Whether to overwrite data in output directory")

    parser.add_argument("--dont_normalize_importance_by_layer", action='store_true', help="Don't normalize importance score by layers")
    parser.add_argument("--dont_normalize_global_importance", action='store_true', help="Don't normalize all importance scores between 0 and 1")

    parser.add_argument("--try_masking", action='store_true', help="Whether to try to mask head until a threshold of accuracy.")
    parser.add_argument("--masking_threshold", default=0.9, type=float, help="masking threshold in term of metrics"
                                                                             "(stop masking when metric < threshold * original metric value).")
    parser.add_argument("--masking_amount", default=0.1, type=float, help="Amount to heads to masking at each masking step.")
    parser.add_argument("--metric_name", default="acc", type=str, help="Metric to use for head masking.")

    parser.add_argument("--max_seq_length", default=128, type=int, help="The maximum total input sequence length after WordPiece tokenization. \n"
                             "Sequences longer than this will be truncated, and sequences shorter \n"
                             "than this will be padded.")
    parser.add_argument("--batch_size", default=1, type=int, help="Batch size.")

    parser.add_argument("--seed", type=int, default=42)
    parser.add_argument("--local_rank", type=int, default=-1, help="local_rank for distributed training on gpus")
    parser.add_argument("--no_cuda", action='store_true', help="Whether not to use CUDA when available")
    parser.add_argument('--server_ip', type=str, default='', help="Can be used for distant debugging.")
    parser.add_argument('--server_port', type=str, default='', help="Can be used for distant debugging.")
    args = parser.parse_args()

    if args.server_ip and args.server_port:
        # Distant debugging - see https://code.visualstudio.com/docs/python/debugging#_attach-to-a-local-script
        import ptvsd
        print("Waiting for debugger attach")
        ptvsd.enable_attach(address=(args.server_ip, args.server_port), redirect_output=True)
        ptvsd.wait_for_attach()

    # Setup devices and distributed training
    if args.local_rank == -1 or args.no_cuda:
        args.device = torch.device("cuda" if torch.cuda.is_available() and not args.no_cuda else "cpu")
        n_gpu = torch.cuda.device_count()
    else:
        torch.cuda.set_device(args.local_rank)
        args.device = torch.device("cuda", args.local_rank)
        n_gpu = 1
        torch.distributed.init_process_group(backend='nccl')  # Initializes the distributed backend

    # Setup logging
    logging.basicConfig(level = logging.INFO if args.local_rank in [-1, 0] else logging.WARN)
    logger.info("device: {} n_gpu: {}, distributed: {}".format(args.device, n_gpu, bool(args.local_rank != -1)))

    # Set seeds
    np.random.seed(args.seed)
    torch.random.manual_seed(args.seed)
    if n_gpu > 0:
        torch.cuda.manual_seed(args.seed)

    # Prepare GLUE task
    task_name = args.task_name.lower()
    processor = processors[task_name]()
    label_list = processor.get_labels()
    args.output_mode = output_modes[task_name]
    args.num_labels = len(label_list)

    # Prepare output directory
    if os.path.exists(args.output_dir) and os.listdir(args.output_dir) and not args.overwrite_output_dir:
        raise ValueError("Output directory ({}) already exists and is not empty.".format(args.output_dir))
    if not os.path.exists(args.output_dir) and args.local_rank in [-1, 0]:
        os.makedirs(args.output_dir)

    # Load model & tokenizer
    if args.local_rank not in [-1, 0]:
        torch.distributed.barrier()  # Make sure only one distributed process download model & vocab
    tokenizer = BertTokenizer.from_pretrained(args.model_name_or_path)

    # Load a model with all BERTology options on:
    #   output_attentions => will output attention weights
    #   keep_multihead_output => will store gradient of attention head outputs for head importance computation
    #       see: http://arxiv.org/abs/1905.10650
    model = BertForSequenceClassification.from_pretrained(args.model_name_or_path,
                                                          num_labels=args.num_labels,
                                                          output_attentions=True,
                                                          keep_multihead_output=True)
    if args.local_rank == 0:
        torch.distributed.barrier()  # Make sure only one distributed process download model & vocab
    model.to(args.device)
    if args.local_rank != -1:
        model = torch.nn.parallel.DistributedDataParallel(model, device_ids=[args.local_rank], output_device=args.local_rank, find_unused_parameters=True)
    model.eval()

    # Prepare dataset for the GLUE task
    eval_examples = processor.get_dev_examples(args.data_dir)
    cached_eval_features_file = os.path.join(args.data_dir, 'dev_{0}_{1}_{2}'.format(
        list(filter(None, args.model_name_or_path.split('/'))).pop(), str(args.max_seq_length), str(task_name)))
    try:
        eval_features = torch.load(cached_eval_features_file)
    except:
        eval_features = convert_examples_to_features(eval_examples, label_list, args.max_seq_length, tokenizer, args.output_mode)
        if args.local_rank in [-1, 0]:
            logger.info("Saving eval features to cache file %s", cached_eval_features_file)
            torch.save(eval_features, cached_eval_features_file)

    all_input_ids = torch.tensor([f.input_ids for f in eval_features], dtype=torch.long)
    all_input_mask = torch.tensor([f.input_mask for f in eval_features], dtype=torch.long)
    all_segment_ids = torch.tensor([f.segment_ids for f in eval_features], dtype=torch.long)
    all_label_ids = torch.tensor([f.label_id for f in eval_features], dtype=torch.long if args.output_mode == "classification" else torch.float)
    eval_data = TensorDataset(all_input_ids, all_input_mask, all_segment_ids, all_label_ids)

    if args.data_subset > 0:
        eval_data = Subset(eval_data, list(range(args.data_subset)))

    eval_sampler = SequentialSampler(eval_data) if args.local_rank == -1 else DistributedSampler(eval_data)
    eval_dataloader = DataLoader(eval_data, sampler=eval_sampler, batch_size=args.batch_size)

    # Print/save training arguments
    print(args)
    torch.save(args, os.path.join(args.output_dir, 'run_args.bin'))

    # Compute head entropy and importance score
    attn_entropy, head_importance, _, _ = compute_heads_importance(args, model, eval_dataloader)

    # Print/save matrices
    np.save(os.path.join(args.output_dir, 'attn_entropy.npy'), attn_entropy.detach().cpu().numpy())
    np.save(os.path.join(args.output_dir, 'head_importance.npy'), head_importance.detach().cpu().numpy())

    logger.info("Attention entropies")
    print_2d_tensor(attn_entropy)
    logger.info("Head importance scores")
    print_2d_tensor(head_importance)
    logger.info("Head ranked by importance scores")
    head_ranks = torch.zeros(head_importance.numel(), dtype=torch.long, device=args.device)
    head_ranks[head_importance.view(-1).sort(descending=True)[1]] = torch.arange(head_importance.numel(), device=args.device)
    head_ranks = head_ranks.view_as(head_importance)
    print_2d_tensor(head_ranks)

    # Do masking if we want to
    if args.try_masking and args.masking_threshold > 0.0 and args.masking_threshold < 1.0:
        _, head_importance, preds, labels = compute_heads_importance(args, model, eval_dataloader, compute_entropy=False)
        preds = np.argmax(preds, axis=1) if args.output_mode == "classification" else np.squeeze(preds)
        original_score = compute_metrics(task_name, preds, labels)[args.metric_name]
        logger.info("Pruning: original score: %f, threshold: %f", original_score, original_score * args.masking_threshold)

        new_head_mask = torch.ones_like(head_importance)
        num_to_mask = int(new_head_mask.numel() * args.masking_amount)

        current_score = original_score
        while current_score >= original_score * args.masking_threshold:
            head_mask = new_head_mask  # save current head mask
            # heads from most important to least - keep only not-masked heads
            head_importance = head_importance.view(-1)[head_mask.view(-1).nonzero()][:, 0]
            current_heads_to_mask = head_importance.sort()[1]

            if len(current_heads_to_mask) <= num_to_mask:
                break

            # mask heads
            current_heads_to_mask = current_heads_to_mask[:num_to_mask]
            logger.info("Heads to mask: %s", str(current_heads_to_mask.tolist()))
            new_head_mask = head_mask.view(-1)
            new_head_mask[current_heads_to_mask] = 0.0
            new_head_mask = new_head_mask.view_as(head_mask)
            print_2d_tensor(new_head_mask)

            # Compute metric and head importance again
            _, head_importance, preds, labels = compute_heads_importance(args, model, eval_dataloader, compute_entropy=False, head_mask=new_head_mask)
            preds = np.argmax(preds, axis=1) if args.output_mode == "classification" else np.squeeze(preds)
            current_score = compute_metrics(task_name, preds, labels)[args.metric_name]
            logger.info("Masking: current score: %f, remaning heads %d (%.1f percents)", current_score, new_head_mask.sum(), new_head_mask.sum()/new_head_mask.numel() * 100)

        # Try pruning and test time speedup
        # Pruning is like masking but we actually remove the masked weights
        before_time = datetime.now()
        _, _, preds, labels = compute_heads_importance(args, model, eval_dataloader,
                                                       compute_entropy=False, compute_importance=False, head_mask=head_mask)
        preds = np.argmax(preds, axis=1) if args.output_mode == "classification" else np.squeeze(preds)
        score_masking = compute_metrics(task_name, preds, labels)[args.metric_name]
        original_time = datetime.now() - before_time

        heads_to_prune = dict((layer, (1 - head_mask[layer].long()).nonzero().tolist()) for layer in range(len(head_mask)))
        assert sum(len(h) for h in heads_to_prune.values()) == (1 - head_mask.long()).sum().item()
        model.bert.prune_heads(heads_to_prune)

        before_time = datetime.now()
        _, _, preds, labels = compute_heads_importance(args, model, eval_dataloader,
                                                       compute_entropy=False, compute_importance=False, head_mask=None)
        preds = np.argmax(preds, axis=1) if args.output_mode == "classification" else np.squeeze(preds)
        score_pruning = compute_metrics(task_name, preds, labels)[args.metric_name]
        new_time = datetime.now() - before_time

        logger.info("Pruning: score with masking: %f score with pruning: %f", score_masking, score_pruning)
        logger.info("Pruning: speed ratio (new timing / original timing): %f percents", original_time/new_time * 100)

if __name__ == '__main__':
    run_model()