def test_inference_masked_lm(self):
tokenizer = PerceiverTokenizer.from_pretrained("deepmind/language-perceiver")
model = PerceiverForMaskedLM.from_pretrained("deepmind/language-perceiver")
model.to(torch_device)
# prepare inputs
text = "This is an incomplete sentence where some words are missing."
encoding = tokenizer(text, padding="max_length", return_tensors="pt")
# mask " missing.".
encoding.input_ids[0, 52:61] = tokenizer.mask_token_id
inputs, input_mask = encoding.input_ids.to(torch_device), encoding.attention_mask.to(torch_device)
# forward pass
with torch.no_grad():
outputs = model(inputs=inputs, attention_mask=input_mask)
logits = outputs.logits
# verify logits
expected_shape = torch.Size((1, tokenizer.model_max_length, tokenizer.vocab_size))
self.assertEqual(logits.shape, expected_shape)
expected_slice = torch.tensor(
[[-10.8609, -10.7651, -10.9187], [-12.1689, -11.9389, -12.1479], [-12.1518, -11.9707, -12.2073]],
device=torch_device,
)
self.assertTrue(torch.allclose(logits[0, :3, :3], expected_slice, atol=1e-4))
expected_greedy_predictions = [38, 115, 111, 121, 121, 111, 116, 109, 52]
masked_tokens_predictions = logits[0, 52:61].argmax(dim=-1).tolist()
self.assertListEqual(expected_greedy_predictions, masked_tokens_predictions)
def convert_perceiver_checkpoint(pickle_file,
pytorch_dump_folder_path,
architecture="MLM"):
"""
Copy/paste/tweak model's weights to our Perceiver structure.
"""
# load parameters as FlatMapping data structure
with open(pickle_file, "rb") as f:
checkpoint = pickle.loads(f.read())
state = None
if isinstance(checkpoint, dict) and architecture in [
"image_classification",
"image_classification_fourier",
"image_classification_conv",
]:
# the image classification_conv checkpoint also has batchnorm states (running_mean and running_var)
params = checkpoint["params"]
state = checkpoint["state"]
else:
params = checkpoint
# turn into initial state dict
state_dict = dict()
for scope_name, parameters in hk.data_structures.to_mutable_dict(
params).items():
for param_name, param in parameters.items():
state_dict[scope_name + "/" + param_name] = param
if state is not None:
# add state variables
for scope_name, parameters in hk.data_structures.to_mutable_dict(
state).items():
for param_name, param in parameters.items():
state_dict[scope_name + "/" + param_name] = param
# rename keys
rename_keys(state_dict, architecture=architecture)
# load HuggingFace model
config = PerceiverConfig()
subsampling = None
repo_id = "datasets/huggingface/label-files"
if architecture == "MLM":
config.qk_channels = 8 * 32
config.v_channels = 1280
model = PerceiverForMaskedLM(config)
elif "image_classification" in architecture:
config.num_latents = 512
config.d_latents = 1024
config.d_model = 512
config.num_blocks = 8
config.num_self_attends_per_block = 6
config.num_cross_attention_heads = 1
config.num_self_attention_heads = 8
config.qk_channels = None
config.v_channels = None
# set labels
config.num_labels = 1000
filename = "imagenet-1k-id2label.json"
id2label = json.load(
open(cached_download(hf_hub_url(repo_id, filename)), "r"))
id2label = {int(k): v for k, v in id2label.items()}
config.id2label = id2label
config.label2id = {v: k for k, v in id2label.items()}
if architecture == "image_classification":
config.image_size = 224
model = PerceiverForImageClassificationLearned(config)
elif architecture == "image_classification_fourier":
config.d_model = 261
model = PerceiverForImageClassificationFourier(config)
elif architecture == "image_classification_conv":
config.d_model = 322
model = PerceiverForImageClassificationConvProcessing(config)
else:
raise ValueError(f"Architecture {architecture} not supported")
elif architecture == "optical_flow":
config.num_latents = 2048
config.d_latents = 512
config.d_model = 322
config.num_blocks = 1
config.num_self_attends_per_block = 24
config.num_self_attention_heads = 16
config.num_cross_attention_heads = 1
model = PerceiverForOpticalFlow(config)
elif architecture == "multimodal_autoencoding":
config.num_latents = 28 * 28 * 1
config.d_latents = 512
config.d_model = 704
config.num_blocks = 1
config.num_self_attends_per_block = 8
config.num_self_attention_heads = 8
config.num_cross_attention_heads = 1
config.num_labels = 700
# define dummy inputs + subsampling (as each forward pass is only on a chunk of image + audio data)
images = torch.randn((1, 16, 3, 224, 224))
audio = torch.randn((1, 30720, 1))
nchunks = 128
image_chunk_size = np.prod((16, 224, 224)) // nchunks
audio_chunk_size = audio.shape[1] // config.samples_per_patch // nchunks
# process the first chunk
chunk_idx = 0
subsampling = {
"image":
torch.arange(image_chunk_size * chunk_idx,
image_chunk_size * (chunk_idx + 1)),
"audio":
torch.arange(audio_chunk_size * chunk_idx,
audio_chunk_size * (chunk_idx + 1)),
"label":
None,
}
model = PerceiverForMultimodalAutoencoding(config)
# set labels
filename = "kinetics700-id2label.json"
id2label = json.load(
open(cached_download(hf_hub_url(repo_id, filename)), "r"))
id2label = {int(k): v for k, v in id2label.items()}
config.id2label = id2label
config.label2id = {v: k for k, v in id2label.items()}
else:
raise ValueError(f"Architecture {architecture} not supported")
model.eval()
# load weights
model.load_state_dict(state_dict)
# prepare dummy input
input_mask = None
if architecture == "MLM":
tokenizer = PerceiverTokenizer.from_pretrained(
"/Users/NielsRogge/Documents/Perceiver/Tokenizer files")
text = "This is an incomplete sentence where some words are missing."
encoding = tokenizer(text, padding="max_length", return_tensors="pt")
# mask " missing.". Note that the model performs much better if the masked chunk starts with a space.
encoding.input_ids[0, 51:60] = tokenizer.mask_token_id
inputs = encoding.input_ids
input_mask = encoding.attention_mask
elif architecture in [
"image_classification", "image_classification_fourier",
"image_classification_conv"
]:
feature_extractor = PerceiverFeatureExtractor()
image = prepare_img()
encoding = feature_extractor(image, return_tensors="pt")
inputs = encoding.pixel_values
elif architecture == "optical_flow":
inputs = torch.randn(1, 2, 27, 368, 496)
elif architecture == "multimodal_autoencoding":
images = torch.randn((1, 16, 3, 224, 224))
audio = torch.randn((1, 30720, 1))
inputs = dict(image=images,
audio=audio,
label=torch.zeros((images.shape[0], 700)))
# forward pass
if architecture == "multimodal_autoencoding":
outputs = model(inputs=inputs,
attention_mask=input_mask,
subsampled_output_points=subsampling)
else:
outputs = model(inputs=inputs, attention_mask=input_mask)
logits = outputs.logits
# verify logits
if not isinstance(logits, dict):
print("Shape of logits:", logits.shape)
else:
for k, v in logits.items():
print(f"Shape of logits of modality {k}", v.shape)
if architecture == "MLM":
expected_slice = torch.tensor([[-11.8336, -11.6850, -11.8483],
[-12.8149, -12.5863, -12.7904],
[-12.8440, -12.6410, -12.8646]])
assert torch.allclose(logits[0, :3, :3], expected_slice)
masked_tokens_predictions = logits[0, 51:60].argmax(dim=-1).tolist()
expected_list = [38, 115, 111, 121, 121, 111, 116, 109, 52]
assert masked_tokens_predictions == expected_list
print("Greedy predictions:")
print(masked_tokens_predictions)
print()
print("Predicted string:")
print(tokenizer.decode(masked_tokens_predictions))
elif architecture in [
"image_classification", "image_classification_fourier",
"image_classification_conv"
]:
print("Predicted class:",
model.config.id2label[logits.argmax(-1).item()])
# Finally, save files
Path(pytorch_dump_folder_path).mkdir(exist_ok=True)
print(f"Saving model to {pytorch_dump_folder_path}")
model.save_pretrained(pytorch_dump_folder_path)
def perceiver_tokenizer(self):
return PerceiverTokenizer.from_pretrained(
"deepmind/language-perceiver")
def setUp(self):
super().setUp()
tokenizer = PerceiverTokenizer()
tokenizer.save_pretrained(self.tmpdirname)
def main():
# See all possible arguments in src/transformers/training_args.py
# or by passing the --help flag to this script.
# We now keep distinct sets of args, for a cleaner separation of concerns.
parser = HfArgumentParser(
(ModelArguments, DataTrainingArguments, TrainingArguments))
if len(sys.argv) == 2 and sys.argv[1].endswith(".json"):
# If we pass only one argument to the script and it's the path to a json file,
# let's parse it to get our arguments.
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(
)
# Detecting last checkpoint.
last_checkpoint = None
if (os.path.isdir(training_args.output_dir) and training_args.do_train
and not training_args.overwrite_output_dir):
last_checkpoint = get_last_checkpoint(training_args.output_dir)
if last_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 last_checkpoint is not None:
logger.info(
f"Checkpoint detected, resuming training at {last_checkpoint}. To avoid this behavior, change "
"the `--output_dir` or add `--overwrite_output_dir` to train from scratch."
)
# Setup logging
logging.basicConfig(
format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
datefmt="%m/%d/%Y %H:%M:%S",
handlers=[logging.StreamHandler(sys.stdout)],
)
logger.setLevel(logging.INFO if is_main_process(training_args.local_rank
) else logging.WARN)
# Log on each process the small summary:
logger.warning(
f"Process rank: {training_args.local_rank}, device: {training_args.device}, n_gpu: {training_args.n_gpu}"
+
f"distributed training: {bool(training_args.local_rank != -1)}, 16-bits training: {training_args.fp16}"
)
# Set the verbosity to info of the Transformers logger (on main process only):
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("Training/evaluation parameters %s", training_args)
# Set seed before initializing model.
set_seed(training_args.seed)
# Get the datasets: you can either provide your own CSV/JSON/TXT training and evaluation files (see below)
# or just provide the name of one of the public datasets available on the hub at https://huggingface.co/datasets/
# (the dataset will be downloaded automatically from the datasets Hub).
#
# For CSV/JSON files, this script will use the column called 'text' or the first column if no column called
# 'text' is found. You can easily tweak this behavior (see below).
#
# In distributed training, the load_dataset function guarantee that only one local process can concurrently
# download the dataset.
if data_args.dataset_name is not None:
# Downloading and loading a dataset from the hub.
datasets = load_dataset(data_args.dataset_name,
data_args.dataset_config_name)
else:
data_files = {}
if data_args.train_file is not None:
data_files["train"] = data_args.train_file
if data_args.validation_file is not None:
data_files["validation"] = data_args.validation_file
if data_args.test_file is not None:
data_files["test"] = data_args.test_file
extension = data_args.train_file.split(".")[-1]
datasets = load_dataset(extension, data_files=data_files)
# See more about loading any type of standard or custom dataset (from files, python dict, pandas DataFrame, etc) at
# https://huggingface.co/docs/datasets/loading_datasets.html.
if training_args.do_train:
column_names = datasets["train"].column_names
features = datasets["train"].features
else:
column_names = datasets["validation"].column_names
features = datasets["validation"].features
text_column_name = "tokens" if "tokens" in column_names else column_names[0]
label_column_name = (f"{data_args.task_name}_tags"
if f"{data_args.task_name}_tags" in column_names else
column_names[1])
# In the event the labels are not a `Sequence[ClassLabel]`, we will need to go through the dataset to get the
# unique labels.
def get_label_list(labels):
unique_labels = set()
for label in labels:
unique_labels = unique_labels | set(label)
label_list = list(unique_labels)
label_list.sort()
return label_list
if isinstance(features[label_column_name].feature, ClassLabel):
label_list = features[label_column_name].feature.names
# No need to convert the labels since they are already ints.
label_to_id = {i: i for i in range(len(label_list))}
else:
label_list = get_label_list(datasets["train"][label_column_name])
label_to_id = {l: i for i, l in enumerate(label_list)}
num_labels = len(label_list)
config = PerceiverConfig(
num_labels=num_labels,
finetuning_task=data_args.task_name,
cache_dir=model_args.cache_dir,
d_latents=model_args.d_latents,
d_model=model_args.d_model,
max_position_embeddings=model_args.max_position_embeddings,
)
tokenizer = PerceiverTokenizer(
cache_dir=model_args.cache_dir,
model_max_length=model_args.model_max_length)
model = PerceiverForTokenClassification(config=config)
model.main_input_name = "input_ids"
# Preprocessing the dataset
# Padding strategy
padding = "max_length" if data_args.pad_to_max_length else False
# Tokenize all texts and align the labels with them.
def tokenize_and_align_labels(examples):
tokenized_inputs = tokenizer(
examples[text_column_name],
padding=padding,
truncation=True,
# We use this argument because the texts in our dataset are lists of words (with a label for each word).
is_split_into_words=True,
# return_offsets_mapping=True,
)
labels = []
"""
There is no PerceiverTokenizerFast, follow code works for conll2003 datasets.
words: ['EU', 'rejects', 'German', 'call', 'to', 'boycott', 'British', 'lamb', '.']
words_labels: [3, 0, 7, 0, 0, 0, 7, 0, 0]
tokens_labels: [-100, 3, 3, 0, 0, 0, 0, 0, 0, 0, 7, 7, 7, 7, 7, 7, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 7, 7, 7, 7, 7, 7, 7, 0, 0, 0, 0, 0, -100]
"""
for i in range(len(tokenized_inputs["input_ids"])):
# -2 special token [BOS] [SEP]
id_length = len(tokenized_inputs["input_ids"][i]) - 2
char_length = len("".join(examples[text_column_name][i]))
assert id_length == char_length
label_ids = [-100]
words = examples[text_column_name][i]
words_labels = examples[label_column_name][i]
for w, w_label in zip(words, words_labels):
current_label = label_to_id[w_label]
label_ids.extend([current_label] * len(w))
label_ids.append(-100)
assert len(label_ids) == id_length + 2
labels.append(label_ids)
tokenized_inputs["labels"] = labels
return tokenized_inputs
if training_args.do_train:
if "train" not in datasets:
raise ValueError("--do_train requires a train dataset")
train_dataset = datasets["train"]
if data_args.max_train_samples is not None:
train_dataset = train_dataset.select(
range(data_args.max_train_samples))
train_dataset = train_dataset.map(
tokenize_and_align_labels,
batched=True,
num_proc=data_args.preprocessing_num_workers,
load_from_cache_file=not data_args.overwrite_cache,
)
if training_args.do_eval:
if "validation" not in datasets:
raise ValueError("--do_eval requires a validation dataset")
eval_dataset = datasets["validation"]
if data_args.max_val_samples is not None:
eval_dataset = eval_dataset.select(range(
data_args.max_val_samples))
eval_dataset = eval_dataset.map(
tokenize_and_align_labels,
batched=True,
num_proc=data_args.preprocessing_num_workers,
load_from_cache_file=not data_args.overwrite_cache,
)
if training_args.do_predict:
if "test" not in datasets:
raise ValueError("--do_predict requires a test dataset")
test_dataset = datasets["test"]
if data_args.max_test_samples is not None:
test_dataset = test_dataset.select(
range(data_args.max_test_samples))
test_dataset = test_dataset.map(
tokenize_and_align_labels,
batched=True,
num_proc=data_args.preprocessing_num_workers,
load_from_cache_file=not data_args.overwrite_cache,
)
# Data collator
data_collator = DataCollatorForTokenClassification(
tokenizer, pad_to_multiple_of=8 if training_args.fp16 else None)
# Metrics
metric = load_metric("seqeval")
def compute_metrics(p):
predictions, labels = p
predictions = np.argmax(predictions, axis=2)
# Remove ignored index (special tokens)
true_predictions = [[
label_list[p] for (p, l) in zip(prediction, label) if l != -100
] for prediction, label in zip(predictions, labels)]
true_labels = [[
label_list[l] for (p, l) in zip(prediction, label) if l != -100
] for prediction, label in zip(predictions, labels)]
results = metric.compute(predictions=true_predictions,
references=true_labels)
if data_args.return_entity_level_metrics:
# Unpack nested dictionaries
final_results = {}
for key, value in results.items():
if isinstance(value, dict):
for n, v in value.items():
final_results[f"{key}_{n}"] = v
else:
final_results[key] = value
return final_results
else:
return {
"precision": results["overall_precision"],
"recall": results["overall_recall"],
"f1": results["overall_f1"],
"accuracy": results["overall_accuracy"],
}
# Initialize our Trainer
trainer = Trainer(
model=model,
args=training_args,
train_dataset=train_dataset if training_args.do_train else None,
eval_dataset=eval_dataset if training_args.do_eval else None,
tokenizer=tokenizer,
data_collator=data_collator,
compute_metrics=compute_metrics,
)
# Training
if training_args.do_train:
if last_checkpoint is not None:
checkpoint = last_checkpoint
elif os.path.isdir(model_args.model_name_or_path):
checkpoint = model_args.model_name_or_path
else:
checkpoint = None
train_result = trainer.train(resume_from_checkpoint=checkpoint)
metrics = train_result.metrics
trainer.save_model() # Saves the tokenizer too for easy upload
max_train_samples = (data_args.max_train_samples
if data_args.max_train_samples is not None else
len(train_dataset))
metrics["train_samples"] = min(max_train_samples, len(train_dataset))
trainer.log_metrics("train", metrics)
trainer.save_metrics("train", metrics)
trainer.save_state()
# Evaluation
if training_args.do_eval:
logger.info("*** Evaluate ***")
metrics = trainer.evaluate()
max_val_samples = (data_args.max_val_samples
if data_args.max_val_samples is not None else
len(eval_dataset))
metrics["eval_samples"] = min(max_val_samples, len(eval_dataset))
trainer.log_metrics("eval", metrics)
trainer.save_metrics("eval", metrics)
# Predict
if training_args.do_predict:
logger.info("*** Predict ***")
predictions, labels, metrics = trainer.predict(test_dataset)
predictions = np.argmax(predictions, axis=2)
# Remove ignored index (special tokens)
true_predictions = [[
label_list[p] for (p, l) in zip(prediction, label) if l != -100
] for prediction, label in zip(predictions, labels)]
trainer.log_metrics("test", metrics)
trainer.save_metrics("test", metrics)
# Save predictions
output_test_predictions_file = os.path.join(training_args.output_dir,
"test_predictions.txt")
if trainer.is_world_process_zero():
with open(output_test_predictions_file, "w") as writer:
for prediction in true_predictions:
writer.write(" ".join(prediction) + "\n")