adding train files

model_file.py

@@ -0,0 +1,226 @@
+import jax
+import jax.numpy as jnp
+
+import flax
+import flax.linen as nn
+from flax.core.frozen_dict import FrozenDict, unfreeze
+
+from typing import Any, Optional, Tuple
+
+from transformers import (
+    GPT2Config)
+
+from transformers.file_utils import add_start_docstrings, add_start_docstrings_to_model_forward
+from transformers.models.gpt2.modeling_flax_gpt2  import  FlaxGPT2BlockCollection
+from transformers.modeling_flax_outputs import FlaxBaseModelOutput
+from transformers.modeling_flax_utils import FlaxPreTrainedModel
+from transformers.models.gpt2.modeling_flax_gpt2 import FlaxGPT2Module
+
+from transformers import GPT2Tokenizer
+tokenizer = GPT2Tokenizer.from_pretrained("gpt2",pad_token='<|endoftext|>') 
+
+GPT2_START_DOCSTRING = r"""
+    This model inherits from :class:`~transformers.FlaxPreTrainedModel`. Check the superclass documentation for the
+    generic methods the library implements for all its model (such as downloading or saving, resizing the input
+    embeddings, pruning heads etc.)
+    This model is also a Flax Linen `flax.nn.Module
+    <https://flax.readthedocs.io/en/latest/_autosummary/flax.nn.module.html>`__ subclass. Use it as a regular Flax
+    Module and refer to the Flax documentation for all matter related to general usage and behavior.
+    Finally, this model supports inherent JAX features such as:
+    - `Just-In-Time (JIT) compilation <https://jax.readthedocs.io/en/latest/jax.html#just-in-time-compilation-jit>`__
+    - `Automatic Differentiation <https://jax.readthedocs.io/en/latest/jax.html#automatic-differentiation>`__
+    - `Vectorization <https://jax.readthedocs.io/en/latest/jax.html#vectorization-vmap>`__
+    - `Parallelization <https://jax.readthedocs.io/en/latest/jax.html#parallelization-pmap>`__
+    Parameters:
+        config (:class:`~transformers.GPT2Config`): Model configuration class with all the parameters of the model.
+            Initializing with a config file does not load the weights associated with the model, only the
+            configuration. Check out the :meth:`~transformers.FlaxPreTrainedModel.from_pretrained` method to load the
+            model weights.
+"""
+GPT2_INPUTS_DOCSTRING = r"""
+    Args:
+        input_ids (:obj:`numpy.ndarray` of shape :obj:`(batch_size,input_ids_length)`):
+            :obj:`input_ids_length` = ``sequence_length``. Indices of input sequence tokens in the vocabulary.
+            Indices can be obtained using :class:`~transformers.GPT2Tokenizer`. See
+            :meth:`transformers.PreTrainedTokenizer.encode` and :meth:`transformers.PreTrainedTokenizer.__call__` for
+            details.
+            `What are input IDs? <../glossary.html#input-ids>`__
+        attention_mask (:obj:`numpy.ndarray` of shape :obj:`(batch_size, sequence_length)`, `optional`):
+            Mask to avoid performing attention on padding token indices. Mask values selected in ``[0, 1]``:
+            - 1 for tokens that are **not masked**,
+            - 0 for tokens that are **masked**.
+            `What are attention masks? <../glossary.html#attention-mask>`__
+        position_ids (:obj:`numpy.ndarray` of shape :obj:`(batch_size, sequence_length)`, `optional`):
+            Indices of positions of each input sequence tokens in the position embeddings. Selected in the range ``[0,
+            config.max_position_embeddings - 1]``.
+        past_key_values (:obj:`Dict[str, np.ndarray]`, `optional`, returned by ``init_cache`` or when passing previous ``past_key_values``):
+            Dictionary of pre-computed hidden-states (key and values in the attention blocks) that can be used for fast
+            auto-regressive decoding. Pre-computed key and value hidden-states are of shape `[batch_size, max_length]`.
+        output_attentions (:obj:`bool`, `optional`):
+            Whether or not to return the attentions tensors of all attention layers. See ``attentions`` under returned
+            tensors for more detail.
+        output_hidden_states (:obj:`bool`, `optional`):
+            Whether or not to return the hidden states of all layers. See ``hidden_states`` under returned tensors for
+            more detail.
+        return_dict (:obj:`bool`, `optional`):
+            Whether or not to return a :class:`~transformers.file_utils.ModelOutput` instead of a plain tuple.
+"""
+
+class FlaxGPT2PreTrainedModel(FlaxPreTrainedModel):  #modify
+    """
+    An abstract class to handle weights initialization and a simple interface for downloading and loading pretrained
+    models.
+    """
+
+    config_class = GPT2Config
+    base_model_prefix = "transformer"
+    module_class: nn.Module = None
+
+    def __init__(
+        self,
+        config: GPT2Config,
+        input_shape: Tuple = (1, 1),
+        seed: int = 0,
+        dtype: jnp.dtype = jnp.float32,
+        **kwargs,
+    ):
+        module = self.module_class(config=config, dtype=dtype, **kwargs)
+        super().__init__(config, module, input_shape=input_shape, seed=seed, dtype=dtype)
+
+    def init_weights(self, rng: jax.random.PRNGKey, input_shape: Tuple) -> FrozenDict:
+        # init input tensors
+        input_ids = jnp.zeros(input_shape, dtype="i4")
+        attention_mask = jnp.ones_like(input_ids)
+        position_ids = jnp.broadcast_to(jnp.arange(jnp.atleast_2d(input_ids).shape[-1]), input_shape)
+        params_rng, dropout_rng = jax.random.split(rng)
+        rngs = {"params": params_rng, "dropout": dropout_rng}
+
+        return self.module.init(rngs, input_ids, attention_mask, position_ids, return_dict=False)["params"]
+
+    def init_cache(self, batch_size, max_length):
+        r"""
+        Args:
+            batch_size (:obj:`int`):
+                batch_size used for fast auto-regressive decoding. Defines the batch size of the initialized cache.
+            max_length (:obj:`int`):
+                maximum possible length for auto-regressive decoding. Defines the sequence length of the initialized
+                cache.
+        """
+        # init input variables to retrieve cache
+        input_ids = jnp.ones((batch_size, max_length))
+        attention_mask = jnp.ones_like(input_ids)
+        position_ids = jnp.broadcast_to(jnp.arange(jnp.atleast_2d(input_ids).shape[-1]), input_ids.shape)
+
+        init_variables = self.module.init(
+            jax.random.PRNGKey(0), input_ids, attention_mask, position_ids, return_dict=False, init_cache=True
+        )
+        return init_variables["cache"]
+
+    @add_start_docstrings_to_model_forward(GPT2_INPUTS_DOCSTRING)
+    def __call__(
+        self,
+        input_ids,
+        attention_mask=None,
+        position_ids=None,
+        params: dict = None,
+        past_key_values: dict = None,
+        dropout_rng: jax.random.PRNGKey = None,
+        train: bool = False,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+    ):
+        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+        output_hidden_states = (
+            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+        )
+        return_dict = return_dict if return_dict is not None else self.config.return_dict
+       
+
+        if position_ids is None:
+            if past_key_values is not None:
+                raise ValueError("Make sure to provide `position_ids` when passing `past_key_values`.")
+
+            position_ids = jnp.broadcast_to(jnp.arange(jnp.atleast_2d(input_ids).shape[-1]), input_ids.shape)
+
+        if attention_mask is None:
+            attention_mask = jnp.ones_like(input_ids)
+
+        # Handle any PRNG if needed
+        rngs = {}
+        if dropout_rng is not None:
+            rngs["dropout"] = dropout_rng
+
+        inputs = {"params": params or self.params}
+
+        # if past_key_values are passed then cache is already initialized a private flag init_cache has to be passed down to ensure cache is used. It has to be made sure that cache is marked as mutable so that it can be changed by FlaxGPT2Attention module
+        if past_key_values:
+            inputs["cache"] = past_key_values
+            mutable = ["cache"]
+        else:
+            mutable = False
+
+        outputs = self.module.apply(
+            inputs,
+            jnp.array(input_ids, dtype="i4"),
+            jnp.array(attention_mask, dtype="i4"),
+            jnp.array(position_ids, dtype="i4"),
+            not train,
+            False,
+            output_attentions,
+            output_hidden_states,
+            return_dict,
+            rngs=rngs,
+            mutable=mutable,
+        )
+
+        # add updated cache to model output
+        if past_key_values is not None and return_dict:
+            outputs, past_key_values = outputs
+            outputs["past_key_values"] = unfreeze(past_key_values["cache"])
+            return outputs
+        elif past_key_values is not None and not return_dict:
+            outputs, past_key_values = outputs
+            outputs = outputs[:1] + (unfreeze(past_key_values["cache"]),) + outputs[1:]
+
+        return outputs
+
+class FlaxGPT2ForMultipleChoiceModule(nn.Module):
+  config:GPT2Config
+  dtype: jnp.dtype = jnp.float32
+  def setup(self):
+    self.transformer = FlaxGPT2Module(config=self.config, dtype=self.dtype)
+    self.dropout = nn.Dropout(rate=0.2)
+    self.classifier = nn.Dense(4, dtype=self.dtype)
+
+  def __call__(self,input_ids,attention_mask,position_ids,return_dict=True,deterministic=True,*args):
+    batch_size = input_ids.shape[0]
+    rng=jax.random.PRNGKey(0)
+    _, dropout_rng = jax.random.split(rng)
+    input_ids=input_ids.reshape(4*batch_size,-1)
+    position_ids=position_ids.reshape(4*batch_size,-1)
+    attention_mask=attention_mask.reshape(4*batch_size,-1) 
+
+    outputs=self.transformer(input_ids, attention_mask,position_ids,return_dict=return_dict)
+    
+
+    hidden_states = outputs[0]
+    hidden_states= jnp.mean(hidden_states, axis=1)
+
+    
+    
+    hidden_states=hidden_states.reshape(batch_size,-1)         #(32,8,768)->(32,8*768)
+
+    dropout_output = self.dropout(hidden_states,deterministic=deterministic,rng=dropout_rng)
+
+    
+
+    logits = self.classifier(dropout_output)
+    reshaped_logits = logits.reshape(-1, 4)   
+                  #(32,4)
+    if not return_dict:
+      return (reshaped_logits,) + outputs[2:]
+    return reshaped_logits
+
+class FlaxGPT2ForMultipleChoice(FlaxGPT2PreTrainedModel):
+    module_class = FlaxGPT2ForMultipleChoiceModule

requirements.txt

@@ -0,0 +1,5 @@
+jax
+flax
+transformers
+Datasets
+tqdm

train.py

@@ -0,0 +1,237 @@
+import jax
+print(jax.local_device_count())
+import jax.numpy as jnp
+
+import flax
+import flax.linen as nn
+from flax.training.common_utils import get_metrics,onehot,shard,shard_prng_key
+from flax.training import train_state
+from flax.metrics.tensorboard import SummaryWriter
+from flax.training import checkpoints
+
+
+import logging
+import optax
+import math
+from tqdm import tqdm
+
+from pathlib import Path
+from typing import Callable
+from itertools import chain
+from flax.metrics import tensorboard
+
+from datasets import load_dataset,load_metric
+from transformers import GPT2Config,GPT2Tokenizer
+
+from model_file  import FlaxGPT2ForMultipleChoice
+
+logger = logging.getLogger()
+logger.setLevel(logging.INFO)
+
+def main():
+
+
+  tokenizer=GPT2Tokenizer.from_pretrained('gpt2',pad_token='<|endoftext|>')
+
+  dataset=load_dataset('cosmos_qa')
+
+  def preprocess(example):
+    example['context&question']=example['context']+example['question']
+    example['first_sentence']=[example['context&question'],example['context&question'],example['context&question'],example['context&question']]
+    example['second_sentence']=example['answer0'],example['answer1'],example['answer2'],example['answer3']
+    return example
+
+  train_dataset=dataset['train'].map(preprocess)
+  validation_dataset=dataset['validation'].map(preprocess)
+  test_dataset=dataset['test'].map(preprocess)
+
+  #Remove after experiment
+  len_train_dataset=64
+  len_validation_dataset=64
+  len_test_dataset=64
+
+  train_dataset=train_dataset.select(range(len_train_dataset))
+  test_dataset=test_dataset.select(range(len_validation_dataset))
+  validation_dataset=validation_dataset.select(range(len_test_dataset))
+
+  #remove_cols=train_dataset.column_names
+
+  def tokenize(examples):
+    a=tokenizer(examples['first_sentence'],examples['second_sentence'],padding='max_length',truncation=True,max_length=256,return_tensors='jax')
+    a['labels']=examples['label']
+    return a
+
+  train_dataset=train_dataset.map(tokenize)
+  validation_dataset=validation_dataset.map(tokenize)
+  test_dataset=test_dataset.map(tokenize)
+
+  remov_col=['id', 'context', 'question', 'answer0', 'answer1', 'answer2', 'answer3', 'labels', 'context&question', 'first_sentence', 'second_sentence']
+
+  train_dataset=train_dataset.remove_columns(remov_col)
+  validation_dataset=validation_dataset.remove_columns(remov_col)
+  test_dataset=test_dataset.remove_columns(remov_col)
+
+  per_device_batch_size=4
+  seed=0
+  num_train_epochs=1
+  learning_rate=2e-5
+  
+
+  total_batch_size = per_device_batch_size * jax.local_device_count()
+  print('The overall batch size (both for training and eval) is', total_batch_size)
+  num_train_steps = len(train_dataset) // total_batch_size * num_train_epochs
+  num_validation_steps=len(validation_dataset)//total_batch_size*num_train_epochs
+
+  learning_rate_function = optax.linear_schedule(init_value=learning_rate, end_value=0, transition_steps=num_train_steps)
+
+  class TrainState(train_state.TrainState):
+    logits_function:Callable=flax.struct.field(pytree_node=False)
+    loss_function:Callable=flax.struct.field(pytree_node=False)
+
+  def adamw(weight_decay):
+    return optax.adamw(learning_rate=learning_rate_function,b1=0.9,b2=0.99,eps=1e-6,weight_decay=weight_decay)
+
+  decay_path=lambda p:not any(x in p for x in ['bias','LayerNorm.weight'])
+
+  def traverse(function):
+    def mask(data):
+      flat=flax.traverse_util.flatten_dict(data)
+      return flax.traverse_util.unflatten_dict({k:function(k,v) for k,v in flat.items()})
+    return mask
+
+  gradient_transformation=optax.chain(
+      optax.masked(adamw(0.0),mask=traverse(lambda path,_:decay_path(path))),
+      optax.masked(adamw(0.01),mask=traverse(lambda path,_:not decay_path(path))))
+
+  def loss_function(logits,labels):
+    logits=flax.linen.log_softmax(logits)
+    xentropy=optax.softmax_cross_entropy(logits,onehot(labels,num_classes=4))
+    return jnp.mean(xentropy)
+
+  def eval_function(logits):
+    return logits.argmax(-1)
+
+  model = FlaxGPT2ForMultipleChoice.from_pretrained('gpt2',input_shape=(1,4,1))
+
+  state=TrainState.create(apply_fn=model.__call__,
+                          params=model.params,
+                          tx=gradient_transformation,
+                          logits_function=eval_function,
+                          loss_function=loss_function)
+                      
+  def train_step(state,batch,dropout_rng):
+      targets=batch.pop("label")
+      dropout_rng,new_dropout_rng=jax.random.split(dropout_rng)
+      def loss_function(params):
+          logits=state.apply_fn(**batch,params=params,dropout_rng=dropout_rng,train=True)[0]
+          loss=state.loss_function(logits,targets)
+          return loss
+      grad_function=jax.value_and_grad(loss_function)
+      loss,grad=grad_function(state.params)
+      grad=jax.lax.pmean(grad,"batch")
+      new_state=state.apply_gradients(grads=grad)
+        #Added.
+      logits=new_state.apply_fn(**batch,params=new_state.params,dropout_rng=dropout_rng,train=True)[0]
+      accuracy=jnp.equal(jnp.argmax(logits,axis=-1),targets)
+      metrics=jax.lax.pmean({"loss":loss,"learning_rate":learning_rate_function(state.step),'accuracy':accuracy},axis_name="batch")
+      return new_state,metrics,new_dropout_rng
+
+  parallel_train_step = jax.pmap(train_step, axis_name="batch", donate_argnums=(0,))
+
+  def eval_step(state, batch):
+      targets=batch.pop('label')
+      logits = state.apply_fn(**batch, params=state.params, train=False)
+      loss=state.loss_function(logits,targets)
+      predictions=state.logits_function(logits)
+      eval_accuracy=jnp.equal(predictions,targets)
+      #eval_acc=jnp.equal(predictions,targets)
+      metrics=jax.lax.pmean({"loss":loss,'accuracy':eval_accuracy},axis_name="batch")
+      #return state.logits_function(logits)  #(8,4)
+      return targets,predictions,metrics
+
+  parallel_eval_step = jax.pmap(eval_step, axis_name="batch")
+
+  def glue_train_data_loader(rng,dataset,batch_size):
+    steps_per_epoch=len_train_dataset//batch_size
+    perms=jax.random.permutation(rng,len(dataset))
+    perms=perms[:steps_per_epoch*batch_size]
+    perms=perms.reshape((steps_per_epoch,batch_size))
+    for perm in perms:
+      batch=dataset[perm]
+      batch={k:jnp.array(v) for k,v in batch.items()}
+      batch=shard(batch)
+      yield batch
+
+  rng=jax.random.PRNGKey(seed)
+  dropout_rngs=jax.random.split(rng,jax.local_device_count())
+
+  def glue_eval_data_loader(dataset, batch_size):
+      for i in range(len_validation_dataset // batch_size):
+          batch = dataset[i * batch_size : (i + 1) * batch_size]
+          batch = {k: jnp.array(v) for k, v in batch.items()}
+          batch = shard(batch)
+
+          yield batch
+
+  state = flax.jax_utils.replicate(state)
+  #metrics_list = list_metrics()
+
+  actual_task = "mnli"
+  metric = load_metric('glue', "mnli")
+  actual_taskmetric = load_metric('glue', actual_task)
+
+  workdir='./results_tensorboard'
+  summary_writer = tensorboard.SummaryWriter(workdir)
+  #summary_writer.hparams(dict(GPT2Config()))
+
+  logger.info(f"***** Running training *****")
+  logger.info(f"  Num examples = {len_train_dataset}")
+  logger.info(f"  Num Epochs = {1}")
+  logger.info(f"  Instantaneous batch size per device = {per_device_batch_size}")
+  logger.info(f"  Total train batch size  = {total_batch_size}")
+  logger.info(f"  Total optimization steps = {num_train_steps}")
+
+  for i, epoch in enumerate(tqdm(range(1, 2), desc=f"Epoch ...", position=0, leave=True)):
+    rng, input_rng = jax.random.split(rng)
+    train_acc_metrics=[]
+    train_loss_metrics=[]
+    eval_acc_metrics=[]
+    eval_loss_metrics=[]
+    # train
+    with tqdm(total=len_train_dataset // total_batch_size, desc="Training...", leave=False) as progress_bar_train:
+      for idx,batch in enumerate(glue_train_data_loader(input_rng, train_dataset, total_batch_size)):
+        state, train_metric, dropout_rngs = parallel_train_step(state, batch, dropout_rngs)
+        train_acc_metrics.append(jax.device_get(train_metric['accuracy']).mean().item())
+        train_loss_metrics.append(flax.jax_utils.unreplicate(train_metric)['loss'].item())
+        if idx%1==0:
+          summary_writer.scalar('train_loss',flax.jax_utils.unreplicate(train_metric)['loss'].item(),idx)
+          summary_writer.scalar('train_accuracy', jax.device_get(train_metric['accuracy']).mean().item(),idx)
+        if idx%1==0:
+            params = jax.device_get(jax.tree_map(lambda x: x[0], state.params))
+            model.save_pretrained(
+                '../gpt2-common-sense-reasoning/checkpoints',
+                params=params,
+                push_to_hub=True,
+                commit_message=f"Saving weights of epoch {epoch} at step {idx}",)
+        progress_bar_train.update(1)
+
+    # evaluate
+    with tqdm(total=len_validation_dataset // total_batch_size, desc="Evaluating...", leave=False) as progress_bar_eval:
+      for idx,batch in enumerate(glue_eval_data_loader(validation_dataset, total_batch_size)):
+          labels,predictions,eval_metric=parallel_eval_step(state, batch)
+          eval_acc_metrics.append(jax.device_get(eval_metric['accuracy']).mean().item())
+          eval_loss_metrics.append(flax.jax_utils.unreplicate(eval_metric)['loss'].item())
+          progress_bar_eval.update(1)
+          if idx%1==0:
+            logger.info(f"eval_step_loss{idx}:{flax.jax_utils.unreplicate(eval_metric)['loss'].item()} eval_step_acc{idx}:{jax.device_get(eval_metric['accuracy']).mean().item()}")
+            summary_writer.scalar('eval_loss',flax.jax_utils.unreplicate(eval_metric)['loss'].item(),idx)
+            summary_writer.scalar('eval_accuracy', jax.device_get(eval_metric['accuracy']).mean().item(),idx)
+          
+            #correct
+    logger.info(f"Epoch {epoch} done")
+    logger.info(f"Train loss:{jax.device_get(jnp.array(train_loss_metrics)).mean().item()} Train accuracy:{jax.device_get(jnp.array(train_acc_metrics)).mean().item()}")
+    logger.info(f"Eval loss:{jax.device_get(jnp.array(eval_loss_metrics)).mean().item()} Eval accuracy:{jax.device_get(jnp.array(eval_acc_metrics)).mean().item()}")
+  summary_writer.flush()
+
+if __name__ == "__main__":
+    main()