|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
""" |
|
Fine-tuning a 🤗 Transformers model for image classification. |
|
|
|
Here is the full list of checkpoints on the hub that can be fine-tuned by this script: |
|
https://huggingface.co/models?filter=image-classification |
|
""" |
|
|
|
import json |
|
import logging |
|
import os |
|
import sys |
|
from dataclasses import dataclass, field |
|
from typing import Optional |
|
|
|
import evaluate |
|
import numpy as np |
|
import tensorflow as tf |
|
from datasets import load_dataset |
|
from PIL import Image |
|
|
|
import transformers |
|
from transformers import ( |
|
TF_MODEL_FOR_IMAGE_CLASSIFICATION_MAPPING, |
|
AutoConfig, |
|
AutoImageProcessor, |
|
DefaultDataCollator, |
|
HfArgumentParser, |
|
PushToHubCallback, |
|
TFAutoModelForImageClassification, |
|
TFTrainingArguments, |
|
create_optimizer, |
|
set_seed, |
|
) |
|
from transformers.keras_callbacks import KerasMetricCallback |
|
from transformers.trainer_utils import get_last_checkpoint, is_main_process |
|
from transformers.utils import check_min_version, send_example_telemetry |
|
from transformers.utils.versions import require_version |
|
|
|
|
|
logger = logging.getLogger(__name__) |
|
|
|
|
|
check_min_version("4.32.0.dev0") |
|
|
|
require_version("datasets>=1.8.0", "To fix: pip install -r examples/pytorch/image-classification/requirements.txt") |
|
|
|
MODEL_CONFIG_CLASSES = list(TF_MODEL_FOR_IMAGE_CLASSIFICATION_MAPPING.keys()) |
|
MODEL_TYPES = tuple(conf.model_type for conf in MODEL_CONFIG_CLASSES) |
|
|
|
|
|
def pil_loader(path: str): |
|
with open(path, "rb") as f: |
|
im = Image.open(f) |
|
return im.convert("RGB") |
|
|
|
|
|
@dataclass |
|
class DataTrainingArguments: |
|
""" |
|
Arguments pertaining to what data we are going to input our model for training and eval. |
|
Using `HfArgumentParser` we can turn this class into argparse arguments to be able to specify |
|
them on the command line. |
|
""" |
|
|
|
dataset_name: Optional[str] = field( |
|
default=None, |
|
metadata={ |
|
"help": "Name of a dataset from the hub (could be your own, possibly private dataset hosted on the hub)." |
|
}, |
|
) |
|
dataset_config_name: Optional[str] = field( |
|
default=None, metadata={"help": "The configuration name of the dataset to use (via the datasets library)."} |
|
) |
|
train_dir: Optional[str] = field(default=None, metadata={"help": "A folder containing the training data."}) |
|
validation_dir: Optional[str] = field(default=None, metadata={"help": "A folder containing the validation data."}) |
|
train_val_split: Optional[float] = field( |
|
default=0.15, metadata={"help": "Percent to split off of train for validation."} |
|
) |
|
overwrite_cache: bool = field( |
|
default=False, metadata={"help": "Overwrite the cached training and evaluation sets"} |
|
) |
|
preprocessing_num_workers: Optional[int] = field( |
|
default=None, |
|
metadata={"help": "The number of processes to use for the preprocessing."}, |
|
) |
|
max_train_samples: Optional[int] = field( |
|
default=None, |
|
metadata={ |
|
"help": ( |
|
"For debugging purposes or quicker training, truncate the number of training examples to this " |
|
"value if set." |
|
) |
|
}, |
|
) |
|
max_eval_samples: Optional[int] = field( |
|
default=None, |
|
metadata={ |
|
"help": ( |
|
"For debugging purposes or quicker training, truncate the number of evaluation examples to this " |
|
"value if set." |
|
) |
|
}, |
|
) |
|
max_predict_samples: Optional[int] = field( |
|
default=None, |
|
metadata={ |
|
"help": ( |
|
"For debugging purposes or quicker training, truncate the number of prediction examples to this " |
|
"value if set." |
|
) |
|
}, |
|
) |
|
|
|
def __post_init__(self): |
|
if self.dataset_name is None and (self.train_dir is None and self.validation_dir is None): |
|
raise ValueError( |
|
"You must specify either a dataset name from the hub or a train and/or validation directory." |
|
) |
|
|
|
|
|
@dataclass |
|
class ModelArguments: |
|
""" |
|
Arguments pertaining to which model/config/tokenizer we are going to fine-tune from. |
|
""" |
|
|
|
model_name_or_path: str = field( |
|
default="google/vit-base-patch16-224-in21k", |
|
metadata={"help": "Path to pretrained model or model identifier from huggingface.co/models"}, |
|
) |
|
model_type: Optional[str] = field( |
|
default=None, |
|
metadata={"help": "If training from scratch, pass a model type from the list: " + ", ".join(MODEL_TYPES)}, |
|
) |
|
config_name: Optional[str] = field( |
|
default=None, metadata={"help": "Pretrained config name or path if not the same as model_name"} |
|
) |
|
cache_dir: Optional[str] = field( |
|
default=None, metadata={"help": "Where do you want to store the pretrained models downloaded from s3"} |
|
) |
|
model_revision: str = field( |
|
default="main", |
|
metadata={"help": "The specific model version to use (can be a branch name, tag name or commit id)."}, |
|
) |
|
image_processor_name: str = field(default=None, metadata={"help": "Name or path of preprocessor config."}) |
|
use_auth_token: bool = field( |
|
default=False, |
|
metadata={ |
|
"help": ( |
|
"Will use the token generated when running `huggingface-cli login` (necessary to use this script " |
|
"with private models)." |
|
) |
|
}, |
|
) |
|
ignore_mismatched_sizes: bool = field( |
|
default=False, |
|
metadata={"help": "Will enable to load a pretrained model whose head dimensions are different."}, |
|
) |
|
|
|
|
|
def center_crop(image, size): |
|
size = (size, size) if isinstance(size, int) else size |
|
orig_height, orig_width, _ = image.shape |
|
crop_height, crop_width = size |
|
top = (orig_height - orig_width) // 2 |
|
left = (orig_width - crop_width) // 2 |
|
image = tf.image.crop_to_bounding_box(image, top, left, crop_height, crop_width) |
|
return image |
|
|
|
|
|
|
|
|
|
def random_crop(image, scale=(0.08, 1.0), ratio=(3.0 / 4.0, 4.0 / 3.0)): |
|
height, width, _ = image.shape |
|
area = height * width |
|
log_ratio = np.log(ratio) |
|
for _ in range(10): |
|
target_area = np.random.uniform(*scale) * area |
|
aspect_ratio = np.exp(np.random.uniform(*log_ratio)) |
|
w = int(round(np.sqrt(target_area * aspect_ratio))) |
|
h = int(round(np.sqrt(target_area / aspect_ratio))) |
|
if 0 < w <= width and 0 < h <= height: |
|
i = np.random.randint(0, height - h + 1) |
|
j = np.random.randint(0, width - w + 1) |
|
return image[i : i + h, j : j + w, :] |
|
|
|
|
|
in_ratio = float(width) / float(height) |
|
w = width if in_ratio < min(ratio) else int(round(height * max(ratio))) |
|
h = height if in_ratio > max(ratio) else int(round(width / min(ratio))) |
|
i = (height - h) // 2 |
|
j = (width - w) // 2 |
|
return image[i : i + h, j : j + w, :] |
|
|
|
|
|
def random_resized_crop(image, size, scale=(0.08, 1.0), ratio=(3.0 / 4.0, 4.0 / 3.0)): |
|
size = (size, size) if isinstance(size, int) else size |
|
image = random_crop(image, scale, ratio) |
|
image = tf.image.resize(image, size) |
|
return image |
|
|
|
|
|
def main(): |
|
|
|
|
|
|
|
parser = HfArgumentParser((ModelArguments, DataTrainingArguments, TFTrainingArguments)) |
|
if len(sys.argv) == 2 and sys.argv[1].endswith(".json"): |
|
|
|
|
|
model_args, data_args, training_args = parser.parse_json_file(json_file=os.path.abspath(sys.argv[1])) |
|
else: |
|
model_args, data_args, training_args = parser.parse_args_into_dataclasses() |
|
|
|
if not (training_args.do_train or training_args.do_eval or training_args.do_predict): |
|
exit("Must specify at least one of --do_train, --do_eval or --do_predict!") |
|
|
|
|
|
|
|
send_example_telemetry("run_image_classification", model_args, data_args, framework="tensorflow") |
|
|
|
|
|
checkpoint = None |
|
if os.path.isdir(training_args.output_dir) and training_args.do_train and not training_args.overwrite_output_dir: |
|
checkpoint = get_last_checkpoint(training_args.output_dir) |
|
if checkpoint is None and len(os.listdir(training_args.output_dir)) > 0: |
|
raise ValueError( |
|
f"Output directory ({training_args.output_dir}) already exists and is not empty. " |
|
"Use --overwrite_output_dir to overcome." |
|
) |
|
elif checkpoint is not None and training_args.resume_from_checkpoint is None: |
|
logger.info( |
|
f"Checkpoint detected, resuming training at {checkpoint}. To avoid this behavior, change " |
|
"the `--output_dir` or add `--overwrite_output_dir` to train from scratch." |
|
) |
|
|
|
|
|
logging.basicConfig( |
|
format="%(asctime)s - %(levelname)s - %(name)s - %(message)s", |
|
datefmt="%m/%d/%Y %H:%M:%S", |
|
handlers=[logging.StreamHandler(sys.stdout)], |
|
) |
|
log_level = training_args.get_process_log_level() |
|
logger.setLevel(log_level) |
|
|
|
|
|
if is_main_process(training_args.local_rank): |
|
transformers.utils.logging.set_verbosity_info() |
|
transformers.utils.logging.enable_default_handler() |
|
transformers.utils.logging.enable_explicit_format() |
|
logger.info(f"Training/evaluation parameters {training_args}") |
|
|
|
|
|
|
|
set_seed(training_args.seed) |
|
|
|
|
|
if data_args.dataset_name is not None: |
|
dataset = load_dataset( |
|
data_args.dataset_name, |
|
data_args.dataset_config_name, |
|
cache_dir=model_args.cache_dir, |
|
task="image-classification", |
|
use_auth_token=True if model_args.use_auth_token else None, |
|
) |
|
else: |
|
data_files = {} |
|
if data_args.train_dir is not None: |
|
data_files["train"] = os.path.join(data_args.train_dir, "**") |
|
if data_args.validation_dir is not None: |
|
data_files["validation"] = os.path.join(data_args.validation_dir, "**") |
|
dataset = load_dataset( |
|
"imagefolder", |
|
data_files=data_files, |
|
cache_dir=model_args.cache_dir, |
|
task="image-classification", |
|
) |
|
|
|
|
|
|
|
|
|
|
|
labels = dataset["train"].features["labels"].names |
|
label2id, id2label = {}, {} |
|
for i, label in enumerate(labels): |
|
label2id[label] = str(i) |
|
id2label[str(i)] = label |
|
|
|
|
|
config = AutoConfig.from_pretrained( |
|
model_args.config_name or model_args.model_name_or_path, |
|
num_labels=len(labels), |
|
label2id=label2id, |
|
id2label=id2label, |
|
finetuning_task="image-classification", |
|
cache_dir=model_args.cache_dir, |
|
revision=model_args.model_revision, |
|
use_auth_token=True if model_args.use_auth_token else None, |
|
) |
|
image_processor = AutoImageProcessor.from_pretrained( |
|
model_args.image_processor_name or model_args.model_name_or_path, |
|
cache_dir=model_args.cache_dir, |
|
revision=model_args.model_revision, |
|
use_auth_token=True if model_args.use_auth_token else None, |
|
) |
|
|
|
|
|
data_args.train_val_split = None if "validation" in dataset.keys() else data_args.train_val_split |
|
if isinstance(data_args.train_val_split, float) and data_args.train_val_split > 0.0: |
|
split = dataset["train"].train_test_split(data_args.train_val_split) |
|
dataset["train"] = split["train"] |
|
dataset["validation"] = split["test"] |
|
|
|
|
|
|
|
if "shortest_edge" in image_processor.size: |
|
|
|
image_size = (image_processor.size["shortest_edge"], image_processor.size["shortest_edge"]) |
|
else: |
|
image_size = (image_processor.size["height"], image_processor.size["width"]) |
|
|
|
def _train_transforms(image): |
|
img_size = image_size |
|
image = tf.keras.utils.img_to_array(image) |
|
image = random_resized_crop(image, size=img_size) |
|
image = tf.image.random_flip_left_right(image) |
|
image /= 255.0 |
|
image = (image - image_processor.image_mean) / image_processor.image_std |
|
image = tf.transpose(image, perm=[2, 0, 1]) |
|
return image |
|
|
|
def _val_transforms(image): |
|
image = tf.keras.utils.img_to_array(image) |
|
image = tf.image.resize(image, size=image_size) |
|
|
|
image = center_crop(image, size=image_size) |
|
image /= 255.0 |
|
image = (image - image_processor.image_mean) / image_processor.image_std |
|
image = tf.transpose(image, perm=[2, 0, 1]) |
|
return image |
|
|
|
def train_transforms(example_batch): |
|
"""Apply _train_transforms across a batch.""" |
|
example_batch["pixel_values"] = [ |
|
_train_transforms(pil_img.convert("RGB")) for pil_img in example_batch["image"] |
|
] |
|
return example_batch |
|
|
|
def val_transforms(example_batch): |
|
"""Apply _val_transforms across a batch.""" |
|
example_batch["pixel_values"] = [_val_transforms(pil_img.convert("RGB")) for pil_img in example_batch["image"]] |
|
return example_batch |
|
|
|
train_dataset = None |
|
if training_args.do_train: |
|
if "train" not in dataset: |
|
raise ValueError("--do_train requires a train dataset") |
|
train_dataset = dataset["train"] |
|
if data_args.max_train_samples is not None: |
|
train_dataset = train_dataset.shuffle(seed=training_args.seed).select(range(data_args.max_train_samples)) |
|
train_dataset = train_dataset.map( |
|
train_transforms, |
|
batched=True, |
|
num_proc=data_args.preprocessing_num_workers, |
|
load_from_cache_file=not data_args.overwrite_cache, |
|
) |
|
|
|
eval_dataset = None |
|
if training_args.do_eval: |
|
if "validation" not in dataset: |
|
raise ValueError("--do_eval requires a validation dataset") |
|
eval_dataset = dataset["validation"] |
|
if data_args.max_eval_samples is not None: |
|
eval_dataset = eval_dataset.select(range(data_args.max_eval_samples)) |
|
|
|
eval_dataset = eval_dataset.map( |
|
val_transforms, |
|
batched=True, |
|
num_proc=data_args.preprocessing_num_workers, |
|
load_from_cache_file=not data_args.overwrite_cache, |
|
) |
|
|
|
predict_dataset = None |
|
if training_args.do_predict: |
|
if "test" not in dataset: |
|
raise ValueError("--do_predict requires a test dataset") |
|
predict_dataset = dataset["test"] |
|
if data_args.max_predict_samples is not None: |
|
predict_dataset = predict_dataset.select(range(data_args.max_predict_samples)) |
|
|
|
predict_dataset = predict_dataset.map( |
|
val_transforms, |
|
batched=True, |
|
num_proc=data_args.preprocessing_num_workers, |
|
load_from_cache_file=not data_args.overwrite_cache, |
|
) |
|
|
|
collate_fn = DefaultDataCollator(return_tensors="np") |
|
|
|
|
|
metric = evaluate.load("accuracy") |
|
|
|
|
|
|
|
def compute_metrics(p): |
|
"""Computes accuracy on a batch of predictions""" |
|
logits, label_ids = p |
|
predictions = np.argmax(logits, axis=-1) |
|
metrics = metric.compute(predictions=predictions, references=label_ids) |
|
return metrics |
|
|
|
with training_args.strategy.scope(): |
|
if checkpoint is None: |
|
model_path = model_args.model_name_or_path |
|
else: |
|
model_path = checkpoint |
|
|
|
model = TFAutoModelForImageClassification.from_pretrained( |
|
model_path, |
|
config=config, |
|
from_pt=bool(".bin" in model_path), |
|
cache_dir=model_args.cache_dir, |
|
revision=model_args.model_revision, |
|
use_auth_token=True if model_args.use_auth_token else None, |
|
ignore_mismatched_sizes=model_args.ignore_mismatched_sizes, |
|
) |
|
num_replicas = training_args.strategy.num_replicas_in_sync |
|
total_train_batch_size = training_args.per_device_train_batch_size * num_replicas |
|
total_eval_batch_size = training_args.per_device_eval_batch_size * num_replicas |
|
|
|
dataset_options = tf.data.Options() |
|
dataset_options.experimental_distribute.auto_shard_policy = tf.data.experimental.AutoShardPolicy.OFF |
|
|
|
if training_args.do_train: |
|
num_train_steps = int(len(train_dataset) * training_args.num_train_epochs) |
|
if training_args.warmup_steps > 0: |
|
num_warmpup_steps = int(training_args.warmup_steps) |
|
elif training_args.warmup_ratio > 0: |
|
num_warmpup_steps = int(training_args.warmup_ratio * num_train_steps) |
|
else: |
|
num_warmpup_steps = 0 |
|
|
|
optimizer, _ = create_optimizer( |
|
init_lr=training_args.learning_rate, |
|
num_train_steps=num_train_steps, |
|
num_warmup_steps=num_warmpup_steps, |
|
adam_beta1=training_args.adam_beta1, |
|
adam_beta2=training_args.adam_beta2, |
|
adam_epsilon=training_args.adam_epsilon, |
|
weight_decay_rate=training_args.weight_decay, |
|
adam_global_clipnorm=training_args.max_grad_norm, |
|
) |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
train_dataset = model.prepare_tf_dataset( |
|
train_dataset, |
|
shuffle=True, |
|
batch_size=total_train_batch_size, |
|
collate_fn=collate_fn, |
|
).with_options(dataset_options) |
|
else: |
|
optimizer = None |
|
|
|
if training_args.do_eval: |
|
eval_dataset = model.prepare_tf_dataset( |
|
eval_dataset, |
|
shuffle=False, |
|
batch_size=total_eval_batch_size, |
|
collate_fn=collate_fn, |
|
).with_options(dataset_options) |
|
|
|
if training_args.do_predict: |
|
predict_dataset = model.prepare_tf_dataset( |
|
predict_dataset, |
|
shuffle=False, |
|
batch_size=total_eval_batch_size, |
|
collate_fn=collate_fn, |
|
).with_options(dataset_options) |
|
|
|
|
|
|
|
model.compile(optimizer=optimizer, jit_compile=training_args.xla, metrics=["accuracy"]) |
|
|
|
push_to_hub_model_id = training_args.push_to_hub_model_id |
|
if not push_to_hub_model_id: |
|
model_name = model_args.model_name_or_path.split("/")[-1] |
|
push_to_hub_model_id = f"{model_name}-finetuned-image-classification" |
|
|
|
model_card_kwargs = { |
|
"finetuned_from": model_args.model_name_or_path, |
|
"tasks": "image-classification", |
|
"dataset": data_args.dataset_name, |
|
"tags": ["image-classification", "tensorflow", "vision"], |
|
} |
|
|
|
callbacks = [] |
|
if eval_dataset is not None: |
|
callbacks.append(KerasMetricCallback(metric_fn=compute_metrics, eval_dataset=eval_dataset)) |
|
if training_args.push_to_hub: |
|
callbacks.append( |
|
PushToHubCallback( |
|
output_dir=training_args.output_dir, |
|
hub_model_id=push_to_hub_model_id, |
|
hub_token=training_args.push_to_hub_token, |
|
tokenizer=image_processor, |
|
**model_card_kwargs, |
|
) |
|
) |
|
|
|
if training_args.do_train: |
|
model.fit( |
|
train_dataset, |
|
validation_data=eval_dataset, |
|
epochs=int(training_args.num_train_epochs), |
|
callbacks=callbacks, |
|
) |
|
|
|
if training_args.do_eval: |
|
n_eval_batches = len(eval_dataset) |
|
eval_predictions = model.predict(eval_dataset, steps=n_eval_batches) |
|
eval_labels = dataset["validation"]["labels"][: n_eval_batches * total_eval_batch_size] |
|
eval_metrics = compute_metrics((eval_predictions.logits, eval_labels)) |
|
logging.info("Eval metrics:") |
|
for metric_name, value in eval_metrics.items(): |
|
logging.info(f"{metric_name}: {value:.3f}") |
|
|
|
if training_args.output_dir is not None: |
|
os.makedirs(training_args.output_dir, exist_ok=True) |
|
with open(os.path.join(training_args.output_dir, "all_results.json"), "w") as f: |
|
f.write(json.dumps(eval_metrics)) |
|
|
|
if training_args.do_predict: |
|
n_predict_batches = len(predict_dataset) |
|
test_predictions = model.predict(predict_dataset, steps=n_predict_batches) |
|
test_labels = dataset["validation"]["labels"][: n_predict_batches * total_eval_batch_size] |
|
test_metrics = compute_metrics((test_predictions.logits, test_labels)) |
|
logging.info("Test metrics:") |
|
for metric_name, value in test_metrics.items(): |
|
logging.info(f"{metric_name}: {value:.3f}") |
|
|
|
if training_args.output_dir is not None and not training_args.push_to_hub: |
|
|
|
model.save_pretrained(training_args.output_dir) |
|
|
|
|
|
if __name__ == "__main__": |
|
main() |
|
|