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( ) # Sending telemetry. Tracking the example usage helps us better allocate resources to maintain them. The # information sent is the one passed as arguments along with your Python/PyTorch versions. send_example_telemetry("run_t5_mlm", model_args, data_args, framework="flax") if (os.path.exists(training_args.output_dir) and os.listdir(training_args.output_dir) and training_args.do_train and not training_args.overwrite_output_dir): raise ValueError( f"Output directory ({training_args.output_dir}) already exists and is not empty." "Use --overwrite_output_dir to overcome.") # Setup logging logging.basicConfig( format="%(asctime)s - %(levelname)s - %(name)s - %(message)s", level=logging.INFO, datefmt="[%X]", ) # Log on each process the small summary: logger = logging.getLogger(__name__) # Set the verbosity to info of the Transformers logger (on main process only): logger.info(f"Training/evaluation parameters {training_args}") # Set seed before initializing model. set_seed(training_args.seed) # Handle the repository creation if training_args.push_to_hub: if training_args.hub_model_id is None: repo_name = get_full_repo_name(Path( training_args.output_dir).absolute().name, token=training_args.hub_token) else: repo_name = training_args.hub_model_id repo = Repository(training_args.output_dir, clone_from=repo_name) # 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). 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, cache_dir=model_args.cache_dir, use_auth_token=True if model_args.use_auth_token else None, ) if "validation" not in datasets.keys(): datasets["validation"] = load_dataset( data_args.dataset_name, data_args.dataset_config_name, split=f"train[:{data_args.validation_split_percentage}%]", cache_dir=model_args.cache_dir, use_auth_token=True if model_args.use_auth_token else None, ) datasets["train"] = load_dataset( data_args.dataset_name, data_args.dataset_config_name, split=f"train[{data_args.validation_split_percentage}%:]", cache_dir=model_args.cache_dir, use_auth_token=True if model_args.use_auth_token else None, ) 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 extension = data_args.train_file.split(".")[-1] if extension == "txt": extension = "text" datasets = load_dataset( extension, data_files=data_files, cache_dir=model_args.cache_dir, use_auth_token=True if model_args.use_auth_token else None, ) if "validation" not in datasets.keys(): datasets["validation"] = load_dataset( extension, data_files=data_files, split=f"train[:{data_args.validation_split_percentage}%]", cache_dir=model_args.cache_dir, use_auth_token=True if model_args.use_auth_token else None, ) datasets["train"] = load_dataset( extension, data_files=data_files, split=f"train[{data_args.validation_split_percentage}%:]", cache_dir=model_args.cache_dir, use_auth_token=True if model_args.use_auth_token else None, ) # 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. # Load pretrained model and tokenizer if model_args.tokenizer_name: tokenizer = AutoTokenizer.from_pretrained( model_args.tokenizer_name, cache_dir=model_args.cache_dir, use_fast=model_args.use_fast_tokenizer, use_auth_token=True if model_args.use_auth_token else None, ) elif model_args.model_name_or_path: tokenizer = AutoTokenizer.from_pretrained( model_args.model_name_or_path, cache_dir=model_args.cache_dir, use_fast=model_args.use_fast_tokenizer, use_auth_token=True if model_args.use_auth_token else None, ) else: raise ValueError( "You are instantiating a new tokenizer from scratch. This is not supported by this script." "You can do it from another script, save it, and load it from here, using --tokenizer_name." ) if model_args.config_name: config = T5Config.from_pretrained( model_args.config_name, cache_dir=model_args.cache_dir, vocab_size=len(tokenizer), use_auth_token=True if model_args.use_auth_token else None, ) elif model_args.model_name_or_path: config = T5Config.from_pretrained( model_args.model_name_or_path, cache_dir=model_args.cache_dir, use_auth_token=True if model_args.use_auth_token else None, ) else: config = CONFIG_MAPPING[model_args.model_type]() logger.warning( "You are instantiating a new config instance from scratch.") # Preprocessing the datasets. # First we tokenize all the texts. if training_args.do_train: column_names = datasets["train"].column_names else: column_names = datasets["validation"].column_names text_column_name = "text" if "text" in column_names else column_names[0] max_seq_length = min(data_args.max_seq_length, tokenizer.model_max_length) # Otherwise, we tokenize every text, then concatenate them together before splitting them in smaller parts. # Since we make sure that all sequences are of the same length, no attention_mask is needed. def tokenize_function(examples): return tokenizer(examples[text_column_name], return_attention_mask=False) tokenized_datasets = datasets.map( tokenize_function, batched=True, num_proc=data_args.preprocessing_num_workers, remove_columns=column_names, load_from_cache_file=not data_args.overwrite_cache, ) # T5-like span masked language modeling will fuse consecutively masked tokens to a single sentinel token. # To ensure that the input length is `max_seq_length`, we need to increase the maximum length # according to `mlm_probability` and `mean_noise_span_length`. We can also define the label length accordingly. expanded_inputs_length, targets_length = compute_input_and_target_lengths( inputs_length=max_seq_length, noise_density=data_args.mlm_probability, mean_noise_span_length=data_args.mean_noise_span_length, ) # Main data processing function that will concatenate all texts from our dataset and generate chunks of expanded_inputs_length. def group_texts(examples): # Concatenate all texts. concatenated_examples = { k: list(chain(*examples[k])) for k in examples.keys() } total_length = len(concatenated_examples[list(examples.keys())[0]]) # We drop the small remainder, we could add padding if the model supported it instead of this drop, you can # customize this part to your needs. if total_length >= expanded_inputs_length: total_length = (total_length // expanded_inputs_length) * expanded_inputs_length # Split by chunks of max_len. result = { k: [ t[i:i + expanded_inputs_length] for i in range(0, total_length, expanded_inputs_length) ] for k, t in concatenated_examples.items() } return result # Note that with `batched=True`, this map processes 1,000 texts together, so group_texts throws away a # remainder for each of those groups of 1,000 texts. You can adjust that batch_size here but a higher value # might be slower to preprocess. # # To speed up this part, we use multiprocessing. See the documentation of the map method for more information: # https://huggingface.co/docs/datasets/package_reference/main_classes.html#datasets.Dataset.map tokenized_datasets = tokenized_datasets.map( group_texts, batched=True, num_proc=data_args.preprocessing_num_workers, load_from_cache_file=not data_args.overwrite_cache, ) # Enable tensorboard only on the master node has_tensorboard = is_tensorboard_available() if has_tensorboard and jax.process_index() == 0: try: from flax.metrics.tensorboard import SummaryWriter summary_writer = SummaryWriter( log_dir=Path(training_args.output_dir)) except ImportError as ie: has_tensorboard = False logger.warning( f"Unable to display metrics through TensorBoard because some package are not installed: {ie}" ) else: logger.warning( "Unable to display metrics through TensorBoard because the package is not installed: " "Please run pip install tensorboard to enable.") # Initialize our training rng = jax.random.PRNGKey(training_args.seed) dropout_rngs = jax.random.split(rng, jax.local_device_count()) if model_args.model_name_or_path: model = FlaxT5ForConditionalGeneration.from_pretrained( model_args.model_name_or_path, config=config, seed=training_args.seed, dtype=getattr(jnp, model_args.dtype), use_auth_token=True if model_args.use_auth_token else None, ) else: config.vocab_size = len(tokenizer) model = FlaxT5ForConditionalGeneration( config, seed=training_args.seed, dtype=getattr(jnp, model_args.dtype), ) # Data collator # This one will take care of randomly masking the tokens. data_collator = FlaxDataCollatorForT5MLM( tokenizer=tokenizer, noise_density=data_args.mlm_probability, mean_noise_span_length=data_args.mean_noise_span_length, input_length=max_seq_length, target_length=targets_length, pad_token_id=model.config.pad_token_id, decoder_start_token_id=model.config.decoder_start_token_id, ) # Store some constant num_epochs = int(training_args.num_train_epochs) train_batch_size = int( training_args.per_device_train_batch_size) * jax.device_count() per_device_eval_batch_size = int(training_args.per_device_eval_batch_size) eval_batch_size = per_device_eval_batch_size * jax.device_count() num_train_steps = len( tokenized_datasets["train"]) // train_batch_size * num_epochs num_of_hosts = jax.process_count() current_host_idx = jax.process_index() # Create learning rate schedule warmup_fn = optax.linear_schedule( init_value=0.0, end_value=training_args.learning_rate, transition_steps=training_args.warmup_steps) decay_fn = optax.linear_schedule( init_value=training_args.learning_rate, end_value=0, transition_steps=num_train_steps - training_args.warmup_steps, ) linear_decay_lr_schedule_fn = optax.join_schedules( schedules=[warmup_fn, decay_fn], boundaries=[training_args.warmup_steps]) # We use Optax's "masking" functionality to not apply weight decay # to bias and LayerNorm scale parameters. decay_mask_fn returns a # mask boolean with the same structure as the parameters. # The mask is True for parameters that should be decayed. def decay_mask_fn(params): flat_params = traverse_util.flatten_dict(params) # find out all LayerNorm parameters layer_norm_candidates = ["layernorm", "layer_norm", "ln"] layer_norm_named_params = set([ layer[-2:] for layer_norm_name in layer_norm_candidates for layer in flat_params.keys() if layer_norm_name in "".join(layer).lower() ]) flat_mask = { path: (path[-1] != "bias" and path[-2:] not in layer_norm_named_params) for path in flat_params } return traverse_util.unflatten_dict(flat_mask) # create adam optimizer if training_args.adafactor: # We use the default parameters here to initialize adafactor, # For more details about the parameters please check https://github.com/deepmind/optax/blob/ed02befef9bf81cbbf236be3d2b0e032e9ed4a40/optax/_src/alias.py#L74 optimizer = optax.adafactor( learning_rate=linear_decay_lr_schedule_fn, ) else: optimizer = optax.adamw( learning_rate=linear_decay_lr_schedule_fn, b1=training_args.adam_beta1, b2=training_args.adam_beta2, weight_decay=training_args.weight_decay, mask=decay_mask_fn, ) # Setup train state state = train_state.TrainState.create(apply_fn=model.__call__, params=model.params, tx=optimizer) # Define gradient update step fn def train_step(state, batch, dropout_rng): dropout_rng, new_dropout_rng = jax.random.split(dropout_rng) def loss_fn(params): labels = batch.pop("labels") logits = state.apply_fn(**batch, params=params, dropout_rng=dropout_rng, train=True)[0] # compute loss loss = optax.softmax_cross_entropy( logits, onehot(labels, logits.shape[-1])).mean() return loss grad_fn = jax.value_and_grad(loss_fn) loss, grad = grad_fn(state.params) grad = jax.lax.pmean(grad, "batch") new_state = state.apply_gradients(grads=grad) metrics = jax.lax.pmean( { "loss": loss, "learning_rate": linear_decay_lr_schedule_fn(state.step) }, axis_name="batch") return new_state, metrics, new_dropout_rng # Create parallel version of the train step p_train_step = jax.pmap(train_step, "batch", donate_argnums=(0, )) # Define eval fn def eval_step(params, batch): labels = batch.pop("labels") logits = model(**batch, params=params, train=False)[0] # compute loss loss = optax.softmax_cross_entropy(logits, onehot(labels, logits.shape[-1])) # compute accuracy accuracy = jnp.equal(jnp.argmax(logits, axis=-1), labels) # summarize metrics metrics = {"loss": loss.mean(), "accuracy": accuracy.mean()} metrics = jax.lax.pmean(metrics, axis_name="batch") return metrics p_eval_step = jax.pmap(eval_step, "batch", donate_argnums=(0, )) # Replicate the train state on each device state = jax_utils.replicate(state) train_time = 0 epochs = tqdm(range(num_epochs), desc="Epoch ... ", position=0) for epoch in epochs: # ======================== Training ================================ train_start = time.time() train_metrics = [] # Create sampling rng rng, input_rng = jax.random.split(rng) # Generate an epoch by shuffling sampling indices from the train dataset num_train_samples = len(tokenized_datasets["train"]) # Avoid using jax.numpy here in case of TPU training train_samples_idx = np.random.permutation(np.arange(num_train_samples)) train_batch_idx = generate_batch_splits(train_samples_idx, train_batch_size) # Gather the indexes for creating the batch and do a training step for step, batch_idx in enumerate( tqdm(train_batch_idx, desc="Training...", position=1)): samples = [ tokenized_datasets["train"][int(idx)] for idx in batch_idx ] model_inputs = data_collator(samples) local_host_model_inputs = { key: np.split(model_inputs.data[key], num_of_hosts, axis=0)[current_host_idx] for key, value in model_inputs.data.items() } # Model forward model_inputs = shard(local_host_model_inputs) state, train_metric, dropout_rngs = p_train_step( state, model_inputs, dropout_rngs) train_metrics.append(train_metric) cur_step = epoch * (num_train_samples // train_batch_size) + step if cur_step % training_args.logging_steps == 0 and cur_step > 0: # Save metrics train_metric = jax_utils.unreplicate(train_metric) train_time += time.time() - train_start if has_tensorboard and jax.process_index() == 0: write_train_metric(summary_writer, train_metrics, train_time, cur_step) epochs.write( f"Step... ({cur_step} | Loss: {train_metric['loss'].mean()}, Learning Rate:" f" {train_metric['learning_rate'].mean()})") train_metrics = [] if cur_step % training_args.eval_steps == 0 and cur_step > 0: # ======================== Evaluating ============================== num_eval_samples = len(tokenized_datasets["validation"]) # Avoid using jax.numpy here in case of TPU training eval_samples_idx = np.arange(num_eval_samples) eval_batch_idx = generate_batch_splits(eval_samples_idx, eval_batch_size, drop_last=False) eval_metrics = [] for i, batch_idx in enumerate( tqdm(eval_batch_idx, desc="Evaluating ...", position=2)): samples = [ tokenized_datasets["validation"][int(idx)] for idx in batch_idx ] model_inputs = data_collator(samples) # Model forward metrics = pad_shard_unpad(p_eval_step, static_return=True)( state.params, model_inputs.data, min_device_batch=per_device_eval_batch_size) eval_metrics.append(metrics) # get eval metrics eval_metrics = get_metrics(eval_metrics) eval_metrics = jax.tree_map(jnp.mean, eval_metrics) # Update progress bar epochs.write( f"Step... ({cur_step} | Loss: {eval_metrics['loss']}, Acc: {eval_metrics['accuracy']})" ) # Save metrics if has_tensorboard and jax.process_index() == 0: write_eval_metric(summary_writer, eval_metrics, cur_step) if cur_step % training_args.save_steps == 0 and cur_step > 0: # save checkpoint after each epoch and push checkpoint to the hub if jax.process_index() == 0: params = jax.device_get( jax.tree_map(lambda x: x[0], state.params)) model.save_pretrained(training_args.output_dir, params=params) tokenizer.save_pretrained(training_args.output_dir) if training_args.push_to_hub: repo.push_to_hub( commit_message= f"Saving weights and logs of step {cur_step}", blocking=False) # Eval after training if training_args.do_eval: num_eval_samples = len(tokenized_datasets["validation"]) # Avoid using jax.numpy here in case of TPU training eval_samples_idx = np.arange(num_eval_samples) eval_batch_idx = generate_batch_splits(eval_samples_idx, eval_batch_size, drop_last=False) eval_metrics = [] for i, batch_idx in enumerate( tqdm(eval_batch_idx, desc="Evaluating ...", position=2)): samples = [ tokenized_datasets["validation"][int(idx)] for idx in batch_idx ] model_inputs = data_collator(samples) # Model forward metrics = pad_shard_unpad(p_eval_step, static_return=True)( state.params, model_inputs.data, min_device_batch=per_device_eval_batch_size) eval_metrics.append(metrics) # get eval metrics eval_metrics = get_metrics(eval_metrics) eval_metrics = jax.tree_map(lambda metric: jnp.mean(metric).item(), eval_metrics) if jax.process_index() == 0: eval_metrics = { f"eval_{metric_name}": value for metric_name, value in eval_metrics.items() } path = os.path.join(training_args.output_dir, "eval_results.json") with open(path, "w") as f: json.dump(eval_metrics, f, indent=4, sort_keys=True)
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( ) if (os.path.exists(training_args.output_dir) and os.listdir(training_args.output_dir) and training_args.do_train and not training_args.overwrite_output_dir): raise ValueError( f"Output directory ({training_args.output_dir}) already exists and is not empty." "Use --overwrite_output_dir to overcome.") # Make one log on every process with the configuration for debugging. logging.basicConfig( format="%(asctime)s - %(levelname)s - %(name)s - %(message)s", datefmt="%m/%d/%Y %H:%M:%S", level=logging.INFO, ) # Setup logging, we only want one process per machine to log things on the screen. logger.setLevel(logging.INFO if jax.process_index() == 0 else logging.ERROR) if jax.process_index() == 0: datasets.utils.logging.set_verbosity_warning() transformers.utils.logging.set_verbosity_info() else: datasets.utils.logging.set_verbosity_error() transformers.utils.logging.set_verbosity_error() # Set the verbosity to info of the Transformers logger (on main process only): logger.info(f"Training/evaluation parameters {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 guarantees 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. dataset = load_dataset(data_args.dataset_name, data_args.dataset_config_name, cache_dir=model_args.cache_dir, keep_in_memory=False) if "validation" not in dataset.keys(): dataset["validation"] = load_dataset( data_args.dataset_name, data_args.dataset_config_name, split=f"train[:{data_args.validation_split_percentage}%]", cache_dir=model_args.cache_dir, ) dataset["train"] = load_dataset( data_args.dataset_name, data_args.dataset_config_name, split=f"train[{data_args.validation_split_percentage}%:]", cache_dir=model_args.cache_dir, ) 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 extension = data_args.train_file.split(".")[-1] if extension == "txt": extension = "text" dataset = load_dataset(extension, data_files=data_files, cache_dir=model_args.cache_dir) if "validation" not in dataset.keys(): dataset["validation"] = load_dataset( extension, data_files=data_files, split=f"train[:{data_args.validation_split_percentage}%]", cache_dir=model_args.cache_dir, ) dataset["train"] = load_dataset( extension, data_files=data_files, split=f"train[{data_args.validation_split_percentage}%:]", cache_dir=model_args.cache_dir, ) # 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. # Load pretrained model and tokenizer # Distributed training: # The .from_pretrained methods guarantee that only one local process can concurrently # download model & vocab. if model_args.config_name: config = AutoConfig.from_pretrained(model_args.config_name, cache_dir=model_args.cache_dir) elif model_args.model_name_or_path: config = AutoConfig.from_pretrained(model_args.model_name_or_path, cache_dir=model_args.cache_dir) else: config = CONFIG_MAPPING[model_args.model_type]() logger.warning( "You are instantiating a new config instance from scratch.") if model_args.tokenizer_name: tokenizer = AutoTokenizer.from_pretrained( model_args.tokenizer_name, cache_dir=model_args.cache_dir, use_fast=model_args.use_fast_tokenizer) elif model_args.model_name_or_path: tokenizer = AutoTokenizer.from_pretrained( model_args.model_name_or_path, cache_dir=model_args.cache_dir, use_fast=model_args.use_fast_tokenizer) else: raise ValueError( "You are instantiating a new tokenizer from scratch. This is not supported by this script." "You can do it from another script, save it, and load it from here, using --tokenizer_name." ) if model_args.model_name_or_path: model = FlaxAutoModelForCausalLM.from_pretrained( model_args.model_name_or_path, config=config, seed=training_args.seed, dtype=getattr(jnp, model_args.dtype)) else: model = FlaxAutoModelForCausalLM.from_config(config, seed=training_args.seed, dtype=getattr( jnp, model_args.dtype)) # Preprocessing the datasets. # First we tokenize all the texts. if training_args.do_train: column_names = dataset["train"].column_names else: column_names = dataset["validation"].column_names text_column_name = "text" if "text" in column_names else column_names[0] # since this will be pickled to avoid _LazyModule error in Hasher force logger loading before tokenize_function tok_logger = transformers.utils.logging.get_logger( "transformers.tokenization_utils_base") def tokenize_function(examples): with CaptureLogger(tok_logger) as cl: output = tokenizer(examples[text_column_name]) # clm input could be much much longer than block_size if "Token indices sequence length is longer than the" in cl.out: tok_logger.warning( "^^^^^^^^^^^^^^^^ Please ignore the warning above - this long input will be chunked into smaller bits before being passed to the model." ) return output tokenized_datasets = dataset.map( tokenize_function, batched=True, num_proc=data_args.preprocessing_num_workers, remove_columns=column_names, load_from_cache_file=not data_args.overwrite_cache, ) if data_args.block_size is None: block_size = tokenizer.model_max_length if block_size > config.max_position_embeddings: logger.warning( f"The tokenizer picked seems to have a very large `model_max_length` ({tokenizer.model_max_length}). " "Picking 1024 instead. You can change that default value by passing --block_size xxx." ) block_size = 1024 else: if data_args.block_size > tokenizer.model_max_length: logger.warning( f"The block_size passed ({data_args.block_size}) is larger than the maximum length for the model" f"({tokenizer.model_max_length}). Using block_size={tokenizer.model_max_length}." ) block_size = min(data_args.block_size, tokenizer.model_max_length) # Main data processing function that will concatenate all texts from our dataset and generate chunks of block_size. def group_texts(examples): # Concatenate all texts. concatenated_examples = { k: sum(examples[k], []) for k in examples.keys() } total_length = len(concatenated_examples[list(examples.keys())[0]]) # We drop the small remainder, we could add padding if the model supported it instead of this drop, you can # customize this part to your needs. if total_length >= block_size: total_length = (total_length // block_size) * block_size # Split by chunks of max_len. result = { k: [t[i:i + block_size] for i in range(0, total_length, block_size)] for k, t in concatenated_examples.items() } result["labels"] = result["input_ids"].copy() return result # Note that with `batched=True`, this map processes 1,000 texts together, so group_texts throws away a remainder # for each of those groups of 1,000 texts. You can adjust that batch_size here but a higher value might be slower # to preprocess. # # To speed up this part, we use multiprocessing. See the documentation of the map method for more information: # https://huggingface.co/docs/datasets/package_reference/main_classes.html#datasets.Dataset.map lm_datasets = tokenized_datasets.map( group_texts, batched=True, num_proc=data_args.preprocessing_num_workers, load_from_cache_file=not data_args.overwrite_cache, ) if training_args.do_train: if "train" not in tokenized_datasets: raise ValueError("--do_train requires a train dataset") train_dataset = lm_datasets["train"] if data_args.max_train_samples is not None: train_dataset = train_dataset.select( range(data_args.max_train_samples)) if training_args.do_eval: if "validation" not in tokenized_datasets: raise ValueError("--do_eval requires a validation dataset") eval_dataset = lm_datasets["validation"] if data_args.max_eval_samples is not None: eval_dataset = eval_dataset.select( range(data_args.max_eval_samples)) # Enable tensorboard only on the master node has_tensorboard = is_tensorboard_available() if has_tensorboard and jax.process_index() == 0: try: from flax.metrics.tensorboard import SummaryWriter summary_writer = SummaryWriter( log_dir=Path(training_args.output_dir)) except ImportError as ie: has_tensorboard = False logger.warning( f"Unable to display metrics through TensorBoard because some package are not installed: {ie}" ) else: logger.warning( "Unable to display metrics through TensorBoard because the package is not installed: " "Please run pip install tensorboard to enable.") # Initialize our training rng = jax.random.PRNGKey(training_args.seed) rng, dropout_rng = jax.random.split(rng) # Store some constant num_epochs = int(training_args.num_train_epochs) train_batch_size = int( training_args.per_device_train_batch_size) * jax.device_count() eval_batch_size = int( training_args.per_device_eval_batch_size) * jax.device_count() steps_per_epoch = len(train_dataset) // train_batch_size total_train_steps = steps_per_epoch * num_epochs # Create learning rate schedule linear_decay_lr_schedule_fn = create_learning_rate_fn( len(train_dataset), train_batch_size, training_args.num_train_epochs, training_args.warmup_steps, training_args.learning_rate, ) # We use Optax's "masking" functionality to not apply weight decay # to bias and LayerNorm scale parameters. decay_mask_fn returns a # mask boolean with the same structure as the parameters. # The mask is True for parameters that should be decayed. # Note that this mask is specifically adapted for FlaxGPT2. # For other models, one should correct the layer norm parameter naming # accordingly. def decay_mask_fn(params): flat_params = traverse_util.flatten_dict(params) flat_mask = { path: (path[-1] != "bias" and path[-2:] not in [("ln_1", "scale"), ("ln_2", "scale"), ("ln_f", "scale")]) for path in flat_params } return traverse_util.unflatten_dict(flat_mask) # create adam optimizer if training_args.adafactor: # We use the default parameters here to initialize adafactor, # For more details about the parameters please check https://github.com/deepmind/optax/blob/ed02befef9bf81cbbf236be3d2b0e032e9ed4a40/optax/_src/alias.py#L74 optimizer = optax.adafactor( learning_rate=linear_decay_lr_schedule_fn, ) else: optimizer = optax.adamw( learning_rate=linear_decay_lr_schedule_fn, b1=training_args.adam_beta1, b2=training_args.adam_beta2, eps=training_args.adam_epsilon, weight_decay=training_args.weight_decay, mask=decay_mask_fn, ) # Setup train state state = TrainState.create(apply_fn=model.__call__, params=model.params, tx=optimizer, dropout_rng=dropout_rng) def loss_fn(logits, labels): shift_logits = logits[..., :-1, :] shift_labels = labels[..., 1:] loss = optax.softmax_cross_entropy( shift_logits, onehot(shift_labels, shift_logits.shape[-1])) return loss.mean() # Define gradient update step fn def train_step(state, batch): dropout_rng, new_dropout_rng = jax.random.split(state.dropout_rng) def compute_loss(params): labels = batch.pop("labels") logits = state.apply_fn(**batch, params=params, dropout_rng=dropout_rng, train=True)[0] loss = loss_fn(logits, labels) return loss grad_fn = jax.value_and_grad(compute_loss) loss, grad = grad_fn(state.params) grad = jax.lax.pmean(grad, "batch") new_state = state.apply_gradients(grads=grad, dropout_rng=new_dropout_rng) metrics = { "loss": loss, "learning_rate": linear_decay_lr_schedule_fn(state.step) } metrics = jax.lax.pmean(metrics, axis_name="batch") return new_state, metrics # Define eval fn def eval_step(params, batch): labels = batch.pop("labels") logits = model(**batch, params=params, train=False)[0] loss = loss_fn(logits, labels) # summarize metrics metrics = {"loss": loss} metrics = jax.lax.pmean(metrics, axis_name="batch") return metrics # Create parallel version of the train and eval step p_train_step = jax.pmap(train_step, "batch", donate_argnums=(0, )) p_eval_step = jax.pmap(eval_step, "batch") # Replicate the train state on each device state = state.replicate() logger.info("***** Running training *****") logger.info(f" Num examples = {len(train_dataset)}") logger.info(f" Num Epochs = {num_epochs}") logger.info( f" Instantaneous batch size per device = {training_args.per_device_train_batch_size}" ) logger.info( f" Total train batch size (w. parallel & distributed) = {train_batch_size}" ) logger.info(f" Total optimization steps = {total_train_steps}") train_time = 0 train_metrics = [] epochs = tqdm(range(num_epochs), desc="Epoch ... ", position=0) for epoch in epochs: # ======================== Training ================================ train_start = time.time() # Create sampling rng rng, input_rng = jax.random.split(rng) # Generate an epoch by shuffling sampling indices from the train dataset train_loader = data_loader(input_rng, train_dataset, train_batch_size, shuffle=True) steps_per_epoch = len(train_dataset) // train_batch_size # train for step in tqdm(range(steps_per_epoch), desc="Training...", position=1, leave=False): batch = next(train_loader) batch = shard(batch) state, train_metric = p_train_step(state, batch) train_metrics.append(train_metric) cur_step = epoch * (len(train_dataset) // train_batch_size) + step if cur_step % training_args.logging_steps == 0 and cur_step > 0: # Save metrics train_metric = unreplicate(train_metric) train_time += time.time() - train_start if has_tensorboard and jax.process_index() == 0: write_train_metric(summary_writer, train_metrics, train_time, cur_step) epochs.write( f"Step... ({cur_step} | Loss: {train_metric['loss'].mean()}, Learning Rate: {train_metric['learning_rate'].mean()})" ) train_metrics = [] if cur_step % training_args.eval_steps == 0 and cur_step > 0: # ======================== Evaluating ============================== eval_metrics = [] eval_loader = data_loader(input_rng, eval_dataset, eval_batch_size) eval_steps = len(eval_dataset) // eval_batch_size for _ in tqdm(range(eval_steps), desc="Evaluating...", position=2, leave=False): # Model forward batch = next(eval_loader) batch = shard(batch) metrics = p_eval_step(state.params, batch) eval_metrics.append(metrics) # normalize eval metrics eval_metrics = get_metrics(eval_metrics) eval_metrics = jax.tree_map(jnp.mean, eval_metrics) try: eval_metrics["perplexity"] = math.exp(eval_metrics["loss"]) except OverflowError: eval_metrics["perplexity"] = float("inf") # Print metrics and update progress bar desc = f"Step... ({cur_step} | Eval Loss: {eval_metrics['loss']} | Eval Perplexity: {eval_metrics['perplexity']})" epochs.write(desc) epochs.desc = desc # Save metrics if has_tensorboard and jax.process_index() == 0: write_eval_metric(summary_writer, eval_metrics, cur_step) if cur_step % training_args.save_steps == 0 and cur_step > 0: # save checkpoint after each epoch and push checkpoint to the hub if jax.process_index() == 0: params = jax.device_get(unreplicate(state.params)) model.save_pretrained( training_args.output_dir, params=params, push_to_hub=training_args.push_to_hub, commit_message= f"Saving weights and logs of step {cur_step}", )
# mask boolean with the same structure as the parameters. # The mask is True for parameters that should be decayed. # Note that this mask is specifically adapted for FlaxBERT-like models. # For other models, one should correct the layer norm parameter naming # accordingly. def decay_mask_fn(params): flat_params = traverse_util.flatten_dict(params) flat_mask = {path: (path[-1] != "bias" and path[-2:] != ("LayerNorm", "scale")) for path in flat_params} return traverse_util.unflatten_dict(flat_mask) # create adam optimizer if training_args.adafactor: # We use the default parameters here to initialize adafactor, # For more details about the parameters please check https://github.com/deepmind/optax/blob/ed02befef9bf81cbbf236be3d2b0e032e9ed4a40/optax/_src/alias.py#L74 optimizer = optax.adafactor( learning_rate=linear_decay_lr_schedule_fn, ) else: optimizer = optax.adamw( learning_rate=linear_decay_lr_schedule_fn, b1=training_args.adam_beta1, b2=training_args.adam_beta2, eps=training_args.adam_epsilon, weight_decay=training_args.weight_decay, mask=decay_mask_fn, ) # Setup train state state = train_state.TrainState.create(apply_fn=model.__call__, params=model.params, tx=optimizer) # Define gradient update step fn
def update( self, gradient: Weights, state: GenericGradientState, parameters: Optional[Weights] ) -> Tuple[Weights, GenericGradientState]: return GenericGradientState.wrap(*adafactor( **asdict(self)).update(gradient, state.data, parameters))
def init(self, parameters: Weights) -> GenericGradientState: return GenericGradientState(adafactor(**asdict(self)).init(parameters))
def vb_gauss_lowrank(key, logjoint_fn, data, nfeatures, initial_mean=None, initial_std=0.1, initial_scale=1., nsamples=20, niters=200, optimizer=optax.adafactor(1e-3), threshold=2500, window_size=None): ''' Parameters ---------- key : jax.random.PRNGKey logjoint_fn : Callable Log joint function data : Tuple The data to which the model is fitted, specified as a table or matrix. nfeatures : Number of features initial_mean : initial_std : Standard deviation of normal distribution for initialization initial_scale : float The constant factor to scale the initial values. num_samples : int Monte Carlo samples to estimate the lower bound niters : int Maximum number of iterations optimizer : optax.optimizers threshold : float Gradient clipping threshold window_size : int Rolling window size to smooth the lower bound. Default value of window size is None, which indicates that lower bounds won't be smoothed. Returns ------- Tuple: Consists of 1. mu : Estimation of variational mean 2. b : The vector factor loading vector component of the variational covariance matrix 3. c : The diagonal matrix component of the variational covariance matrix Array : Estimation of the lower bound over iterations ''' if initial_mean is None: mu_key, key = split(key, 2) mu = initial_std * normal(mu_key, shape=(nfeatures, 1)) else: mu = initial_mean b_key, key = split(key, 2) b = initial_std * normal(b_key, shape=(nfeatures, 1)) c = initial_scale * jnp.ones((nfeatures, 1)) # Variational parameters vector variational_params = (mu, b, c) # Initial state of the optimizer opt_state = optimizer.init(variational_params) def sample_fn(variational_params, U_normal): mu, b, c = variational_params # Parameters in Normal distribution epsilon1 = U_normal[0] epsilon2 = U_normal[1:].reshape((-1, 1)) theta = mu + b * epsilon1 + c * epsilon2 h_theta, grad_h_theta = value_and_grad(logjoint_fn)(theta, data) # Gradient of log variational distribution grad_log_q = grad_log_q_function(b, c, theta, mu) # Gradient of h(theta) and lower bound grad_theta = grad_h_theta - grad_log_q return grad_theta, epsilon1 * grad_theta, epsilon2 * grad_theta, h_theta def iter_fn(all_params, key): # Main VB iteration variational_params, opt_state = all_params mu, b, c = variational_params samples = normal(key, shape=(nsamples, nfeatures + 1)) *grad_lb_iter, lb_first_term = vmap(sample_fn, in_axes=(None, 0))(variational_params, samples) # Estimation of lowerbound logdet = jnp.log(jnp.linalg.det(1 + (b / c**2).T @ b)) + jnp.sum( jnp.log(c**2)) # Mean of log-q -> mean(log q(theta)) lb_log_q = -0.5 * nfeatures * jnp.log( 2 * jnp.pi) - 0.5 * logdet - nfeatures / 2 lower_bound = jnp.mean(lb_first_term) - lb_log_q # Gradient of log variational distribution grad_lb = tree_map(lambda x: x.mean(axis=0), grad_lb_iter) grads = compute_natural_gradients(b, c, grad_lb) # Gradient clipping grads = clip(grads, threshold=threshold) updates, opt_state = optimizer.update(grads, opt_state, variational_params) variational_params = optax.apply_updates(variational_params, updates) return (variational_params, opt_state), (variational_params, lower_bound) keys = split(key, niters) (best_params, _), (variational_params, lower_bounds) = lax.scan( iter_fn, (variational_params, opt_state), keys) if window_size is not None: def simple_moving_average(cur_sum, i): diff = (lower_bounds[i] - lower_bounds[i - window_size]) / window_size cur_sum += diff return cur_sum, cur_sum indices = jnp.arange(window_size, niters) cur_sum = jnp.sum(lower_bounds[:window_size]) / window_size _, lower_bounds = lax.scan(simple_moving_average, cur_sum, indices) lower_bounds = jnp.append(jnp.array([cur_sum]), lower_bounds) i = jnp.argmax(lower_bounds) + window_size - 1 best_params = tree_map(lambda x: x[i], variational_params) return best_params, lower_bounds
niters=800) w_ffvb = w_ffvb['params']['Dense_0']['kernel'].squeeze() lower_triangular = lower_triangular['params']['Dense_0']['kernel'] cov_ffvb = lower_triangular @ lower_triangular.T # Variational Bayes Low Rank Approximation (w_lowrank, b, c), lower_bounds = nagvac.vb_gauss_lowrank( key, partial_logjoint, (Phi, y), nfeatures, nsamples=20, niters=800, initial_std=0.1, initial_scale=0.3, initial_mean=-0.8 + 0.1 * random.normal(key, (nfeatures, 1)), optimizer=optax.adafactor(1e-4)) w_lowrank = w_lowrank.squeeze() cov_lowrank = b @ b.T + jnp.diag(c**2) # *** Ploting surface predictive distribution *** colors = ["black" if el else "white" for el in y] key = random.PRNGKey(31415) nsamples = 5000 # FFVB surface predictive distribution ffvb_samples = random.multivariate_normal(key, w_ffvb, cov_ffvb, (nsamples, )) Z_ffvb = nn.sigmoid(jnp.einsum("mij,sm->sij", Phispace, ffvb_samples)) Z_ffvb = Z_ffvb.mean(axis=0) # Variational Bayes Low Rank surface predictive distribution
return schedule if __name__ == '__main__': # Load the data url = 'https://raw.githubusercontent.com/probml/probml-data/main/data/vb_data_labour_force.csv' response = requests.get(url) rawdata = BytesIO(response.content) df = pd.read_csv(rawdata) data = df.to_numpy() X, y = jnp.array(data[:, :-1]), jnp.array(data[:, -1]) learning_rate, momentum = 0.001, 0.9 learning_rate_fn = learning_rate_schedule(learning_rate, 2500) optimizer = optax.adafactor(learning_rate_fn, momentum=momentum) # prior sigma for mu std_init = 0.01 # Shape of mu, model params nfeatures = X.shape[-1] # initial scale init_scale = 0.1 niters = 20 window_size = 50 num_samples = 200 threshold = 200