Skip to content

GitLab

Commit f4d79f44 authored by lukovnikov's avatar lukovnikov
Browse files

Merge remote-tracking branch 'upstream/master'

parents bd91ae65 5cd8d7ad
Show whitespace changes
Inline Side-by-side
Showing with 351 additions and 143 deletions
README.md
View file @ f4d79f44
...@@ -93,26 +93,28 @@ A series of tests is included in the [tests folder](https://github.com/huggingfa ...@@ -93,26 +93,28 @@ A series of tests is included in the [tests folder](https://github.com/huggingfa
You can run the tests with the command: You can run the tests with the command:
```bash ```bash
pytest -sv ./tests/ python -m pytest -sv tests/
``` ```
## Training on large batches: gradient accumulation, multi-GPU and distributed training ## Training on large batches: gradient accumulation, multi-GPU and distributed training
BERT-base and BERT-large are respectively 110M and 340M parameters models and it can be difficult to fine-tune them on a single GPU with the recommended batch size for good performance (in most case a batch size of 32). BERT-base and BERT-large are respectively 110M and 340M parameters models and it can be difficult to fine-tune them on a single GPU with the recommended batch size for good performance (in most case a batch size of 32).
To help with fine-tuning these models, we have included three techniques that you can activate in the fine-tuning scripts `run_classifier.py` and `run_squad.py`: gradient-accumulation, multi-gpu and distributed training. For more details on how to use these techniques you can read [the tips on training large batches in PyTorch](https://medium.com/huggingface/training-larger-batches-practical-tips-on-1-gpu-multi-gpu-distributed-setups-ec88c3e51255) that I published earlier this month. To help with fine-tuning these models, we have included five techniques that you can activate in the fine-tuning scripts `run_classifier.py` and `run_squad.py`: gradient-accumulation, multi-gpu training, distributed training, optimize on CPU and 16-bits training . For more details on how to use these techniques you can read [the tips on training large batches in PyTorch](https://medium.com/huggingface/training-larger-batches-practical-tips-on-1-gpu-multi-gpu-distributed-setups-ec88c3e51255) that I published earlier this month.
Here is how to use these techniques in our scripts: Here is how to use these techniques in our scripts:
- **Gradient Accumulation**: Gradient accumulation can be used by supplying a integer greater than 1 to the `--gradient_accumulation_steps` argument. The batch at each step will be divided by this integer and gradient will be accumulated over `gradient_accumulation_steps` steps. - **Gradient Accumulation**: Gradient accumulation can be used by supplying a integer greater than 1 to the `--gradient_accumulation_steps` argument. The batch at each step will be divided by this integer and gradient will be accumulated over `gradient_accumulation_steps` steps.
- **Multi-GPU**: Multi-GPU is automatically activated when several GPUs are detected and the batches are splitted over the GPUs. - **Multi-GPU**: Multi-GPU is automatically activated when several GPUs are detected and the batches are splitted over the GPUs.
- **Distributed training**: Distributed training can be activated by suppying an integer greater or equal to 0 to the `--local_rank` argument. To use Distributed training, you will need to run one training script on each of your machines. This can be done for example by running the following command on each server (see the above blog post for more details): - **Distributed training**: Distributed training can be activated by supplying an integer greater or equal to 0 to the `--local_rank` argument (see below).
- **Optimize on CPU**: The Adam optimizer stores 2 moving average of the weights of the model. If you keep them on GPU 1 (typical behavior), your first GPU will have to store 3-times the size of the model. This is not optimal for large models like `BERT-large` and means your batch size is a lot lower than it could be. This option will perform the optimization and store the averages on the CPU/RAM to free more room on the GPU(s). As the most computational intensive operation is usually the backward pass, this doesn't have a significant impact on the training time. Activate this option with `--optimize_on_cpu` on the `run_squad.py` script.
- **16-bits training**: 16-bits training, also called mixed-precision training, can reduce the memory requirement of your model on the GPU by using half-precision training, basically allowing to double the batch size. If you have a recent GPU (starting from NVIDIA Volta architecture) you should see no decrease in speed. A good introduction to Mixed precision training can be found [here](https://devblogs.nvidia.com/mixed-precision-training-deep-neural-networks/) and a full documentation is [here](https://docs.nvidia.com/deeplearning/sdk/mixed-precision-training/index.html). In our scripts, this option can be activated by setting the `--fp16` flag and you can play with loss scaling using the `--loss_scaling` flag (see the previously linked documentation for details on loss scaling). If the loss scaling is too high (`Nan` in the gradients) it will be automatically scaled down until the value is acceptable. The default loss scaling is 128 which behaved nicely in our tests.
Note: To use *Distributed Training*, you will need to run one training script on each of your machines. This can be done for example by running the following command on each server (see [the above mentioned blog post]((https://medium.com/huggingface/training-larger-batches-practical-tips-on-1-gpu-multi-gpu-distributed-setups-ec88c3e51255)) for more details):
```bash ```bash
python -m torch.distributed.launch --nproc_per_node=4 --nnodes=2 --node_rank=$THIS_MACHINE_INDEX --master_addr="192.168.1.1" --master_port=1234 run_classifier.py (--arg1 --arg2 --arg3 and all other arguments of the run_classifier script) python -m torch.distributed.launch --nproc_per_node=4 --nnodes=2 --node_rank=$THIS_MACHINE_INDEX --master_addr="192.168.1.1" --master_port=1234 run_classifier.py (--arg1 --arg2 --arg3 and all other arguments of the run_classifier script)
``` ```
Where `$THIS_MACHINE_INDEX` is an sequential index assigned to each of your machine (0, 1, 2...) and the machine with rank 0 has an IP address `192.168.1.1` and an open port `1234`.
Where `$THIS_MACHINE_INDEX` is an sequential index assigned to each of your machine (0, 1, 2...) and the machine with rank 0 has an IP adress `192.168.1.1` and an open port `1234`.
## TPU support and pretraining scripts ## TPU support and pretraining scripts
...@@ -128,9 +130,9 @@ Since, pre-training BERT is a particularly expensive operation that basically re ...@@ -128,9 +130,9 @@ Since, pre-training BERT is a particularly expensive operation that basically re
We also include [two Jupyter Notebooks](https://github.com/huggingface/pytorch-pretrained-BERT/tree/master/notebooks) that can be used to check that the predictions of the PyTorch model are identical to the predictions of the original TensorFlow model. We also include [two Jupyter Notebooks](https://github.com/huggingface/pytorch-pretrained-BERT/tree/master/notebooks) that can be used to check that the predictions of the PyTorch model are identical to the predictions of the original TensorFlow model.
- The first NoteBook ([Comparing TF and PT models.ipynb](https://github.com/huggingface/pytorch-pretrained-BERT/blob/master/notebooks/Comparing%20TF%20and%20PT%20models.ipynb)) extracts the hidden states of a full sequence on each layers of the TensorFlow and the PyTorch models and computes the sandard deviation between them. In the given example, we get a standard deviation of 1.5e-7 to 9e-7 on the various hidden state of the models. - The first NoteBook ([Comparing TF and PT models.ipynb](https://github.com/huggingface/pytorch-pretrained-BERT/blob/master/notebooks/Comparing%20TF%20and%20PT%20models.ipynb)) extracts the hidden states of a full sequence on each layers of the TensorFlow and the PyTorch models and computes the standard deviation between them. In the given example, we get a standard deviation of 1.5e-7 to 9e-7 on the various hidden state of the models.
- The second NoteBook ([Comparing TF and PT models SQuAD predictions.ipynb](https://github.com/huggingface/pytorch-pretrained-BERT/blob/master/notebooks/Comparing%20TF%20and%20PT%20models%20SQuAD%20predictions.ipynb)) compares the loss computed by the TensorFlow and the PyTorch models for identical initialization of the fine-tuning layer of the `BertForQuestionAnswering` and computes the sandard deviation between them. In the given example, we get a standard deviation of 2.5e-7 between the models. - The second NoteBook ([Comparing TF and PT models SQuAD predictions.ipynb](https://github.com/huggingface/pytorch-pretrained-BERT/blob/master/notebooks/Comparing%20TF%20and%20PT%20models%20SQuAD%20predictions.ipynb)) compares the loss computed by the TensorFlow and the PyTorch models for identical initialization of the fine-tuning layer of the `BertForQuestionAnswering` and computes the standard deviation between them. In the given example, we get a standard deviation of 2.5e-7 between the models.
Please follow the instructions given in the notebooks to run and modify them. They can also be nice example on how to use the models in a simpler way than the full fine-tuning scripts we provide. Please follow the instructions given in the notebooks to run and modify them. They can also be nice example on how to use the models in a simpler way than the full fine-tuning scripts we provide.
...@@ -138,7 +140,7 @@ Please follow the instructions given in the notebooks to run and modify them. Th ...@@ -138,7 +140,7 @@ Please follow the instructions given in the notebooks to run and modify them. Th
We showcase the same examples as [the original implementation](https://github.com/google-research/bert/): fine-tuning a sequence-level classifier on the MRPC classification corpus and a token-level classifier on the question answering dataset SQuAD. We showcase the same examples as [the original implementation](https://github.com/google-research/bert/): fine-tuning a sequence-level classifier on the MRPC classification corpus and a token-level classifier on the question answering dataset SQuAD.
Before running theses examples you should download the Before running these examples you should download the
[GLUE data](https://gluebenchmark.com/tasks) by running [GLUE data](https://gluebenchmark.com/tasks) by running
[this script](https://gist.github.com/W4ngatang/60c2bdb54d156a41194446737ce03e2e) [this script](https://gist.github.com/W4ngatang/60c2bdb54d156a41194446737ce03e2e)
and unpack it to some directory `$GLUE_DIR`. Please also download the `BERT-Base` and unpack it to some directory `$GLUE_DIR`. Please also download the `BERT-Base`
...@@ -166,7 +168,7 @@ python run_classifier.py \ ...@@ -166,7 +168,7 @@ python run_classifier.py \
--output_dir /tmp/mrpc_output/ --output_dir /tmp/mrpc_output/
``` ```
Our test ran on a few seeds with [the original implementation hyper-parameters](https://github.com/google-research/bert#sentence-and-sentence-pair-classification-tasks) gave evaluation results between 82 and 87. Our test ran on a few seeds with [the original implementation hyper-parameters](https://github.com/google-research/bert#sentence-and-sentence-pair-classification-tasks) gave evaluation results between 84% and 88%.
The second example fine-tunes `BERT-Base` on the SQuAD question answering task. The second example fine-tunes `BERT-Base` on the SQuAD question answering task.
...@@ -184,15 +186,81 @@ python run_squad.py \ ...@@ -184,15 +186,81 @@ python run_squad.py \
--bert_config_file $BERT_BASE_DIR/bert_config.json \ --bert_config_file $BERT_BASE_DIR/bert_config.json \
--init_checkpoint $BERT_PYTORCH_DIR/pytorch_model.bin \ --init_checkpoint $BERT_PYTORCH_DIR/pytorch_model.bin \
--do_train \ --do_train \
--train_file $SQUAD_DIR/train-v1.1.json \
--do_predict \ --do_predict \
--do_lower_case \
--train_file $SQUAD_DIR/train-v1.1.json \
--predict_file $SQUAD_DIR/dev-v1.1.json \ --predict_file $SQUAD_DIR/dev-v1.1.json \
--train_batch_size 12 \ --train_batch_size 12 \
--learning_rate 5e-5 \ --learning_rate 3e-5 \
--num_train_epochs 2.0 \ --num_train_epochs 2.0 \
--max_seq_length 384 \ --max_seq_length 384 \
--doc_stride 128 \ --doc_stride 128 \
--output_dir ../debug_squad/ --output_dir ../debug_squad/
``` ```
There is currently a bug in the `run_squad.py` script that we are investigating. The reported numbers are very low (F1 of 41.8 and exact match of 21.7) even though the correct answer is usually in the n-best predictions. We are investigating that right now on the develop branch, follow [this issue](https://github.com/huggingface/pytorch-pretrained-BERT/issues/3) for more updates. Training with the previous hyper-parameters gave us the following results:
```bash
{"f1": 88.52381567990474, "exact_match": 81.22043519394512}
```
# Fine-tuning BERT-large on GPUs
The options we list above allow to fine-tune BERT-large rather easily on GPU(s) instead of the TPU used by the original implementation.
For example, fine-tuning BERT-large on SQuAD can be done on a server with 4 k-80 (these are pretty old now) in 18 hours. Our results are similar to the TensorFlow implementation results (actually slightly higher):
```bash
{"exact_match": 84.56953642384106, "f1": 91.04028647786927}
```
To get these results we used a combination of:
- multi-GPU training (automatically activated on a multi-GPU server),
- 2 steps of gradient accumulation and
- perform the optimization step on CPU to store Adam's averages in RAM.
Here is the full list of hyper-parameters for this run:
```bash
python ./run_squad.py \
--vocab_file $BERT_LARGE_DIR/vocab.txt \
--bert_config_file $BERT_LARGE_DIR/bert_config.json \
--init_checkpoint $BERT_LARGE_DIR/pytorch_model.bin \
--do_lower_case \
--do_train \
--do_predict \
--train_file $SQUAD_TRAIN \
--predict_file $SQUAD_EVAL \
--learning_rate 3e-5 \
--num_train_epochs 2 \
--max_seq_length 384 \
--doc_stride 128 \
--output_dir $OUTPUT_DIR \
--train_batch_size 24 \
--gradient_accumulation_steps 2 \
--optimize_on_cpu
```
If you have a recent GPU (starting from NVIDIA Volta series), you should try **16-bit fine-tuning** (FP16).
Here is an example of hyper-parameters for a FP16 run we tried:
```bash
python ./run_squad.py \
--vocab_file $BERT_LARGE_DIR/vocab.txt \
--bert_config_file $BERT_LARGE_DIR/bert_config.json \
--init_checkpoint $BERT_LARGE_DIR/pytorch_model.bin \
--do_lower_case \
--do_train \
--do_predict \
--train_file $SQUAD_TRAIN \
--predict_file $SQUAD_EVAL \
--learning_rate 3e-5 \
--num_train_epochs 2 \
--max_seq_length 384 \
--doc_stride 128 \
--output_dir $OUTPUT_DIR \
--train_batch_size 24 \
--fp16 \
--loss_scale 128
```
The results were similar to the above FP32 results (actually slightly higher):
```bash
{"exact_match": 84.65468306527909, "f1": 91.238669287002}
```
convert_tf_checkpoint_to_pytorch.py 100644 → 100755
View file @ f4d79f44
...@@ -47,11 +47,17 @@ def convert(config_path, ckpt_path, out_path=None): ...@@ -47,11 +47,17 @@ def convert(config_path, ckpt_path, out_path=None):
arrays.append(array) arrays.append(array)
for name, array in zip(names, arrays): for name, array in zip(names, arrays):
name = name[5:] # skip "bert/" if not name.startswith("bert"):
print("Skipping {}".format(name))
continue
else:
name = name.replace("bert/", "") # skip "bert/"
print("Loading {}".format(name)) print("Loading {}".format(name))
name = name.split('/') name = name.split('/')
if name[0] in ['redictions', 'eq_relationship']: # adam_v and adam_m are variables used in AdamWeightDecayOptimizer to calculated m and v
print("Skipping") # which are not required for using pretrained model
if name[0] in ['redictions', 'eq_relationship'] or name[-1] == "adam_v" or name[-1] == "adam_m":
print("Skipping {}".format("/".join(name)))
continue continue
pointer = model pointer = model
for m_name in name: for m_name in name:
......
modeling.py
View file @ f4d79f44
...@@ -342,8 +342,8 @@ class BertModel(nn.Module): ...@@ -342,8 +342,8 @@ class BertModel(nn.Module):
token_type_ids = torch.zeros_like(input_ids) token_type_ids = torch.zeros_like(input_ids)
# We create a 3D attention mask from a 2D tensor mask. # We create a 3D attention mask from a 2D tensor mask.
# Sizes are [batch_size, 1, 1, from_seq_length] # Sizes are [batch_size, 1, 1, to_seq_length]
# So we can broadcast to [batch_size, num_heads, to_seq_length, from_seq_length] # So we can broadcast to [batch_size, num_heads, from_seq_length, to_seq_length]
# this attention mask is more simple than the triangular masking of causal attention # this attention mask is more simple than the triangular masking of causal attention
# used in OpenAI GPT, we just need to prepare the broadcast dimension here. # used in OpenAI GPT, we just need to prepare the broadcast dimension here.
extended_attention_mask = attention_mask.unsqueeze(1).unsqueeze(2) extended_attention_mask = attention_mask.unsqueeze(1).unsqueeze(2)
...@@ -353,7 +353,7 @@ class BertModel(nn.Module): ...@@ -353,7 +353,7 @@ class BertModel(nn.Module):
# positions we want to attend and -10000.0 for masked positions. # positions we want to attend and -10000.0 for masked positions.
# Since we are adding it to the raw scores before the softmax, this is # Since we are adding it to the raw scores before the softmax, this is
# effectively the same as removing these entirely. # effectively the same as removing these entirely.
extended_attention_mask = extended_attention_mask.float() extended_attention_mask = extended_attention_mask.to(dtype=next(self.parameters()).dtype) # fp16 compatibility
extended_attention_mask = (1.0 - extended_attention_mask) * -10000.0 extended_attention_mask = (1.0 - extended_attention_mask) * -10000.0
embedding_output = self.embeddings(input_ids, token_type_ids) embedding_output = self.embeddings(input_ids, token_type_ids)
...@@ -393,10 +393,10 @@ class BertForSequenceClassification(nn.Module): ...@@ -393,10 +393,10 @@ class BertForSequenceClassification(nn.Module):
if isinstance(module, (nn.Linear, nn.Embedding)): if isinstance(module, (nn.Linear, nn.Embedding)):
# Slightly different from the TF version which uses truncated_normal for initialization # Slightly different from the TF version which uses truncated_normal for initialization
# cf https://github.com/pytorch/pytorch/pull/5617 # cf https://github.com/pytorch/pytorch/pull/5617
module.weight.data.normal_(config.initializer_range) module.weight.data.normal_(mean=0.0, std=config.initializer_range)
elif isinstance(module, BERTLayerNorm): elif isinstance(module, BERTLayerNorm):
module.beta.data.normal_(config.initializer_range) module.beta.data.normal_(mean=0.0, std=config.initializer_range)
module.gamma.data.normal_(config.initializer_range) module.gamma.data.normal_(mean=0.0, std=config.initializer_range)
if isinstance(module, nn.Linear): if isinstance(module, nn.Linear):
module.bias.data.zero_() module.bias.data.zero_()
self.apply(init_weights) self.apply(init_weights)
...@@ -443,10 +443,10 @@ class BertForQuestionAnswering(nn.Module): ...@@ -443,10 +443,10 @@ class BertForQuestionAnswering(nn.Module):
if isinstance(module, (nn.Linear, nn.Embedding)): if isinstance(module, (nn.Linear, nn.Embedding)):
# Slightly different from the TF version which uses truncated_normal for initialization # Slightly different from the TF version which uses truncated_normal for initialization
# cf https://github.com/pytorch/pytorch/pull/5617 # cf https://github.com/pytorch/pytorch/pull/5617
module.weight.data.normal_(config.initializer_range) module.weight.data.normal_(mean=0.0, std=config.initializer_range)
elif isinstance(module, BERTLayerNorm): elif isinstance(module, BERTLayerNorm):
module.beta.data.normal_(config.initializer_range) module.beta.data.normal_(mean=0.0, std=config.initializer_range)
module.gamma.data.normal_(config.initializer_range) module.gamma.data.normal_(mean=0.0, std=config.initializer_range)
if isinstance(module, nn.Linear): if isinstance(module, nn.Linear):
module.bias.data.zero_() module.bias.data.zero_()
self.apply(init_weights) self.apply(init_weights)
...@@ -460,11 +460,13 @@ class BertForQuestionAnswering(nn.Module): ...@@ -460,11 +460,13 @@ class BertForQuestionAnswering(nn.Module):
end_logits = end_logits.squeeze(-1) end_logits = end_logits.squeeze(-1)
if start_positions is not None and end_positions is not None: if start_positions is not None and end_positions is not None:
# If we are on multi-GPU, split add a dimension - if not this is a no-op # If we are on multi-GPU, split add a dimension
if len(start_positions.size()) > 1:
start_positions = start_positions.squeeze(-1) start_positions = start_positions.squeeze(-1)
if len(end_positions.size()) > 1:
end_positions = end_positions.squeeze(-1) end_positions = end_positions.squeeze(-1)
# sometimes the start/end positions are outside our model inputs, we ignore these terms # sometimes the start/end positions are outside our model inputs, we ignore these terms
ignored_index = start_logits.size(1) + 1 ignored_index = start_logits.size(1)
start_positions.clamp_(0, ignored_index) start_positions.clamp_(0, ignored_index)
end_positions.clamp_(0, ignored_index) end_positions.clamp_(0, ignored_index)
...@@ -472,6 +474,6 @@ class BertForQuestionAnswering(nn.Module): ...@@ -472,6 +474,6 @@ class BertForQuestionAnswering(nn.Module):
start_loss = loss_fct(start_logits, start_positions) start_loss = loss_fct(start_logits, start_positions)
end_loss = loss_fct(end_logits, end_positions) end_loss = loss_fct(end_logits, end_positions)
total_loss = (start_loss + end_loss) / 2 total_loss = (start_loss + end_loss) / 2
return total_loss, (start_logits, end_logits) return total_loss
else: else:
return start_logits, end_logits return start_logits, end_logits
optimization.py
View file @ f4d79f44
...@@ -90,27 +90,6 @@ class BERTAdam(Optimizer): ...@@ -90,27 +90,6 @@ class BERTAdam(Optimizer):
lr.append(lr_scheduled) lr.append(lr_scheduled)
return lr return lr
def to(self, device):
""" Move the optimizer state to a specified device"""
for state in self.state.values():
state['exp_avg'].to(device)
state['exp_avg_sq'].to(device)
def initialize_step(self, initial_step):
"""Initialize state with a defined step (but we don't have stored averaged).
Arguments:
initial_step (int): Initial step number.
"""
for group in self.param_groups:
for p in group['params']:
state = self.state[p]
# State initialization
state['step'] = initial_step
# Exponential moving average of gradient values
state['exp_avg'] = torch.zeros_like(p.data)
# Exponential moving average of squared gradient values
state['exp_avg_sq'] = torch.zeros_like(p.data)
def step(self, closure=None): def step(self, closure=None):
"""Performs a single optimization step. """Performs a single optimization step.
......
run_classifier.py
View file @ f4d79f44
...@@ -309,6 +309,32 @@ def accuracy(out, labels): ...@@ -309,6 +309,32 @@ def accuracy(out, labels):
outputs = np.argmax(out, axis=1) outputs = np.argmax(out, axis=1)
return np.sum(outputs==labels) return np.sum(outputs==labels)
def copy_optimizer_params_to_model(named_params_model, named_params_optimizer):
""" Utility function for optimize_on_cpu and 16-bits training.
Copy the parameters optimized on CPU/RAM back to the model on GPU
"""
for (name_opti, param_opti), (name_model, param_model) in zip(named_params_optimizer, named_params_model):
if name_opti != name_model:
logger.error("name_opti != name_model: {} {}".format(name_opti, name_model))
raise ValueError
param_model.data.copy_(param_opti.data)
def set_optimizer_params_grad(named_params_optimizer, named_params_model, test_nan=False):
""" Utility function for optimize_on_cpu and 16-bits training.
Copy the gradient of the GPU parameters to the CPU/RAMM copy of the model
"""
is_nan = False
for (name_opti, param_opti), (name_model, param_model) in zip(named_params_optimizer, named_params_model):
if name_opti != name_model:
logger.error("name_opti != name_model: {} {}".format(name_opti, name_model))
raise ValueError
if test_nan and torch.isnan(param_model.grad).sum() > 0:
is_nan = True
if param_opti.grad is None:
param_opti.grad = torch.nn.Parameter(param_opti.data.new().resize_(*param_opti.data.size()))
param_opti.grad.data.copy_(param_model.grad.data)
return is_nan
def main(): def main():
parser = argparse.ArgumentParser() parser = argparse.ArgumentParser()
...@@ -392,10 +418,6 @@ def main(): ...@@ -392,10 +418,6 @@ def main():
default=False, default=False,
action='store_true', action='store_true',
help="Whether not to use CUDA when available") help="Whether not to use CUDA when available")
parser.add_argument("--accumulate_gradients",
type=int,
default=1,
help="Number of steps to accumulate gradient on (divide the batch_size and accumulate)")
parser.add_argument("--local_rank", parser.add_argument("--local_rank",
type=int, type=int,
default=-1, default=-1,
...@@ -408,6 +430,18 @@ def main(): ...@@ -408,6 +430,18 @@ def main():
type=int, type=int,
default=1, default=1,
help="Number of updates steps to accumualte before performing a backward/update pass.") help="Number of updates steps to accumualte before performing a backward/update pass.")
parser.add_argument('--optimize_on_cpu',
default=False,
action='store_true',
help="Whether to perform optimization and keep the optimizer averages on CPU")
parser.add_argument('--fp16',
default=False,
action='store_true',
help="Whether to use 16-bit float precision instead of 32-bit")
parser.add_argument('--loss_scale',
type=float, default=128,
help='Loss scaling, positive power of 2 values can improve fp16 convergence.')
args = parser.parse_args() args = parser.parse_args()
processors = { processors = {
...@@ -424,13 +458,16 @@ def main(): ...@@ -424,13 +458,16 @@ def main():
n_gpu = 1 n_gpu = 1
# Initializes the distributed backend which will take care of sychronizing nodes/GPUs # Initializes the distributed backend which will take care of sychronizing nodes/GPUs
torch.distributed.init_process_group(backend='nccl') torch.distributed.init_process_group(backend='nccl')
if args.fp16:
logger.info("16-bits training currently not supported in distributed training")
args.fp16 = False # (see https://github.com/pytorch/pytorch/pull/13496)
logger.info("device %s n_gpu %d distributed training %r", device, n_gpu, bool(args.local_rank != -1)) logger.info("device %s n_gpu %d distributed training %r", device, n_gpu, bool(args.local_rank != -1))
if args.accumulate_gradients < 1: if args.gradient_accumulation_steps < 1:
raise ValueError("Invalid accumulate_gradients parameter: {}, should be >= 1".format( raise ValueError("Invalid gradient_accumulation_steps parameter: {}, should be >= 1".format(
args.accumulate_gradients)) args.gradient_accumulation_steps))
args.train_batch_size = int(args.train_batch_size / args.accumulate_gradients) args.train_batch_size = int(args.train_batch_size / args.gradient_accumulation_steps)
random.seed(args.seed) random.seed(args.seed)
np.random.seed(args.seed) np.random.seed(args.seed)
...@@ -458,7 +495,6 @@ def main(): ...@@ -458,7 +495,6 @@ def main():
raise ValueError("Task not found: %s" % (task_name)) raise ValueError("Task not found: %s" % (task_name))
processor = processors[task_name]() processor = processors[task_name]()
label_list = processor.get_labels() label_list = processor.get_labels()
tokenizer = tokenization.FullTokenizer( tokenizer = tokenization.FullTokenizer(
...@@ -469,26 +505,36 @@ def main(): ...@@ -469,26 +505,36 @@ def main():
if args.do_train: if args.do_train:
train_examples = processor.get_train_examples(args.data_dir) train_examples = processor.get_train_examples(args.data_dir)
num_train_steps = int( num_train_steps = int(
len(train_examples) / args.train_batch_size * args.num_train_epochs) len(train_examples) / args.train_batch_size / args.gradient_accumulation_steps * args.num_train_epochs)
# Prepare model
model = BertForSequenceClassification(bert_config, len(label_list)) model = BertForSequenceClassification(bert_config, len(label_list))
if args.init_checkpoint is not None: if args.init_checkpoint is not None:
model.bert.load_state_dict(torch.load(args.init_checkpoint, map_location='cpu')) model.bert.load_state_dict(torch.load(args.init_checkpoint, map_location='cpu'))
if args.fp16:
model.half()
model.to(device) model.to(device)
if args.local_rank != -1: if args.local_rank != -1:
model = torch.nn.parallel.DistributedDataParallel(model, device_ids=[args.local_rank], model = torch.nn.parallel.DistributedDataParallel(model, device_ids=[args.local_rank],
output_device=args.local_rank) output_device=args.local_rank)
elif n_gpu > 1: elif n_gpu > 1:
model = torch.nn.DataParallel(model) model = torch.nn.DataParallel(model)
# Prepare optimizer
if args.fp16:
param_optimizer = [(n, param.clone().detach().to('cpu').float().requires_grad_()) \
for n, param in model.named_parameters()]
elif args.optimize_on_cpu:
param_optimizer = [(n, param.clone().detach().to('cpu').requires_grad_()) \
for n, param in model.named_parameters()]
else:
param_optimizer = list(model.named_parameters())
no_decay = ['bias', 'gamma', 'beta'] no_decay = ['bias', 'gamma', 'beta']
optimizer_parameters = [ optimizer_grouped_parameters = [
{'params': [p for n, p in model.named_parameters() if n not in no_decay], 'weight_decay_rate': 0.01}, {'params': [p for n, p in param_optimizer if n not in no_decay], 'weight_decay_rate': 0.01},
{'params': [p for n, p in model.named_parameters() if n in no_decay], 'weight_decay_rate': 0.0} {'params': [p for n, p in param_optimizer if n in no_decay], 'weight_decay_rate': 0.0}
] ]
optimizer = BERTAdam(optimizer_grouped_parameters,
optimizer = BERTAdam(optimizer_parameters,
lr=args.learning_rate, lr=args.learning_rate,
warmup=args.warmup_proportion, warmup=args.warmup_proportion,
t_total=num_train_steps) t_total=num_train_steps)
...@@ -501,12 +547,10 @@ def main(): ...@@ -501,12 +547,10 @@ def main():
logger.info(" Num examples = %d", len(train_examples)) logger.info(" Num examples = %d", len(train_examples))
logger.info(" Batch size = %d", args.train_batch_size) logger.info(" Batch size = %d", args.train_batch_size)
logger.info(" Num steps = %d", num_train_steps) logger.info(" Num steps = %d", num_train_steps)
all_input_ids = torch.tensor([f.input_ids for f in train_features], dtype=torch.long) all_input_ids = torch.tensor([f.input_ids for f in train_features], dtype=torch.long)
all_input_mask = torch.tensor([f.input_mask for f in train_features], dtype=torch.long) all_input_mask = torch.tensor([f.input_mask for f in train_features], dtype=torch.long)
all_segment_ids = torch.tensor([f.segment_ids for f in train_features], dtype=torch.long) all_segment_ids = torch.tensor([f.segment_ids for f in train_features], dtype=torch.long)
all_label_ids = torch.tensor([f.label_id for f in train_features], dtype=torch.long) all_label_ids = torch.tensor([f.label_id for f in train_features], dtype=torch.long)
train_data = TensorDataset(all_input_ids, all_input_mask, all_segment_ids, all_label_ids) train_data = TensorDataset(all_input_ids, all_input_mask, all_segment_ids, all_label_ids)
if args.local_rank == -1: if args.local_rank == -1:
train_sampler = RandomSampler(train_data) train_sampler = RandomSampler(train_data)
...@@ -515,25 +559,41 @@ def main(): ...@@ -515,25 +559,41 @@ def main():
train_dataloader = DataLoader(train_data, sampler=train_sampler, batch_size=args.train_batch_size) train_dataloader = DataLoader(train_data, sampler=train_sampler, batch_size=args.train_batch_size)
model.train() model.train()
for epoch in trange(int(args.num_train_epochs), desc="Epoch"): for _ in trange(int(args.num_train_epochs), desc="Epoch"):
tr_loss = 0 tr_loss = 0
nb_tr_examples, nb_tr_steps = 0, 0 nb_tr_examples, nb_tr_steps = 0, 0
for step, (input_ids, input_mask, segment_ids, label_ids) in enumerate(tqdm(train_dataloader, desc="Iteration")): for step, batch in enumerate(tqdm(train_dataloader, desc="Iteration")):
input_ids = input_ids.to(device) batch = tuple(t.to(device) for t in batch)
input_mask = input_mask.to(device) input_ids, input_mask, segment_ids, label_ids = batch
segment_ids = segment_ids.to(device)
label_ids = label_ids.to(device)
loss, _ = model(input_ids, segment_ids, input_mask, label_ids) loss, _ = model(input_ids, segment_ids, input_mask, label_ids)
if n_gpu > 1: if n_gpu > 1:
loss = loss.mean() # mean() to average on multi-gpu. loss = loss.mean() # mean() to average on multi-gpu.
if args.fp16 and args.loss_scale != 1.0:
# rescale loss for fp16 training
# see https://docs.nvidia.com/deeplearning/sdk/mixed-precision-training/index.html
loss = loss * args.loss_scale
if args.gradient_accumulation_steps > 1:
loss = loss / args.gradient_accumulation_steps
loss.backward()
tr_loss += loss.item() tr_loss += loss.item()
nb_tr_examples += input_ids.size(0) nb_tr_examples += input_ids.size(0)
nb_tr_steps += 1 nb_tr_steps += 1
loss.backward()
if (step + 1) % args.gradient_accumulation_steps == 0: if (step + 1) % args.gradient_accumulation_steps == 0:
optimizer.step() # We have accumulated enought gradients if args.fp16 or args.optimize_on_cpu:
if args.fp16 and args.loss_scale != 1.0:
# scale down gradients for fp16 training
for param in model.parameters():
param.grad.data = param.grad.data / args.loss_scale
is_nan = set_optimizer_params_grad(param_optimizer, model.named_parameters(), test_nan=True)
if is_nan:
logger.info("FP16 TRAINING: Nan in gradients, reducing loss scaling")
args.loss_scale = args.loss_scale / 2
model.zero_grad()
continue
optimizer.step()
copy_optimizer_params_to_model(model.named_parameters(), param_optimizer)
else:
optimizer.step()
model.zero_grad() model.zero_grad()
global_step += 1 global_step += 1
...@@ -541,16 +601,13 @@ def main(): ...@@ -541,16 +601,13 @@ def main():
eval_examples = processor.get_dev_examples(args.data_dir) eval_examples = processor.get_dev_examples(args.data_dir)
eval_features = convert_examples_to_features( eval_features = convert_examples_to_features(
eval_examples, label_list, args.max_seq_length, tokenizer) eval_examples, label_list, args.max_seq_length, tokenizer)
logger.info("***** Running evaluation *****") logger.info("***** Running evaluation *****")
logger.info(" Num examples = %d", len(eval_examples)) logger.info(" Num examples = %d", len(eval_examples))
logger.info(" Batch size = %d", args.eval_batch_size) logger.info(" Batch size = %d", args.eval_batch_size)
all_input_ids = torch.tensor([f.input_ids for f in eval_features], dtype=torch.long) all_input_ids = torch.tensor([f.input_ids for f in eval_features], dtype=torch.long)
all_input_mask = torch.tensor([f.input_mask for f in eval_features], dtype=torch.long) all_input_mask = torch.tensor([f.input_mask for f in eval_features], dtype=torch.long)
all_segment_ids = torch.tensor([f.segment_ids for f in eval_features], dtype=torch.long) all_segment_ids = torch.tensor([f.segment_ids for f in eval_features], dtype=torch.long)
all_label_ids = torch.tensor([f.label_id for f in eval_features], dtype=torch.long) all_label_ids = torch.tensor([f.label_id for f in eval_features], dtype=torch.long)
eval_data = TensorDataset(all_input_ids, all_input_mask, all_segment_ids, all_label_ids) eval_data = TensorDataset(all_input_ids, all_input_mask, all_segment_ids, all_label_ids)
if args.local_rank == -1: if args.local_rank == -1:
eval_sampler = SequentialSampler(eval_data) eval_sampler = SequentialSampler(eval_data)
...@@ -567,6 +624,7 @@ def main(): ...@@ -567,6 +624,7 @@ def main():
segment_ids = segment_ids.to(device) segment_ids = segment_ids.to(device)
label_ids = label_ids.to(device) label_ids = label_ids.to(device)
with torch.no_grad():
tmp_eval_loss, logits = model(input_ids, segment_ids, input_mask, label_ids) tmp_eval_loss, logits = model(input_ids, segment_ids, input_mask, label_ids)
logits = logits.detach().cpu().numpy() logits = logits.detach().cpu().numpy()
...@@ -579,13 +637,13 @@ def main(): ...@@ -579,13 +637,13 @@ def main():
nb_eval_examples += input_ids.size(0) nb_eval_examples += input_ids.size(0)
nb_eval_steps += 1 nb_eval_steps += 1
eval_loss = eval_loss / nb_eval_steps #len(eval_dataloader) eval_loss = eval_loss / nb_eval_steps
eval_accuracy = eval_accuracy / nb_eval_examples #len(eval_dataloader) eval_accuracy = eval_accuracy / nb_eval_examples
result = {'eval_loss': eval_loss, result = {'eval_loss': eval_loss,
'eval_accuracy': eval_accuracy, 'eval_accuracy': eval_accuracy,
'global_step': global_step, 'global_step': global_step,
'loss': tr_loss/nb_tr_steps}#'loss': loss.item()} 'loss': tr_loss/nb_tr_steps}
output_eval_file = os.path.join(args.output_dir, "eval_results.txt") output_eval_file = os.path.join(args.output_dir, "eval_results.txt")
with open(output_eval_file, "w") as writer: with open(output_eval_file, "w") as writer:
......
run_squad.py
View file @ f4d79f44
...@@ -669,6 +669,31 @@ def _compute_softmax(scores): ...@@ -669,6 +669,31 @@ def _compute_softmax(scores):
probs.append(score / total_sum) probs.append(score / total_sum)
return probs return probs
def copy_optimizer_params_to_model(named_params_model, named_params_optimizer):
""" Utility function for optimize_on_cpu and 16-bits training.
Copy the parameters optimized on CPU/RAM back to the model on GPU
"""
for (name_opti, param_opti), (name_model, param_model) in zip(named_params_optimizer, named_params_model):
if name_opti != name_model:
logger.error("name_opti != name_model: {} {}".format(name_opti, name_model))
raise ValueError
param_model.data.copy_(param_opti.data)
def set_optimizer_params_grad(named_params_optimizer, named_params_model, test_nan=False):
""" Utility function for optimize_on_cpu and 16-bits training.
Copy the gradient of the GPU parameters to the CPU/RAMM copy of the model
"""
is_nan = False
for (name_opti, param_opti), (name_model, param_model) in zip(named_params_optimizer, named_params_model):
if name_opti != name_model:
logger.error("name_opti != name_model: {} {}".format(name_opti, name_model))
raise ValueError
if test_nan and torch.isnan(param_model.grad).sum() > 0:
is_nan = True
if param_opti.grad is None:
param_opti.grad = torch.nn.Parameter(param_opti.data.new().resize_(*param_opti.data.size()))
param_opti.grad.data.copy_(param_model.grad.data)
return is_nan
def main(): def main():
parser = argparse.ArgumentParser() parser = argparse.ArgumentParser()
...@@ -719,7 +744,6 @@ def main(): ...@@ -719,7 +744,6 @@ def main():
parser.add_argument("--max_answer_length", default=30, type=int, parser.add_argument("--max_answer_length", default=30, type=int,
help="The maximum length of an answer that can be generated. This is needed because the start " help="The maximum length of an answer that can be generated. This is needed because the start "
"and end predictions are not conditioned on one another.") "and end predictions are not conditioned on one another.")
parser.add_argument("--verbose_logging", default=False, action='store_true', parser.add_argument("--verbose_logging", default=False, action='store_true',
help="If true, all of the warnings related to data processing will be printed. " help="If true, all of the warnings related to data processing will be printed. "
"A number of warnings are expected for a normal SQuAD evaluation.") "A number of warnings are expected for a normal SQuAD evaluation.")
...@@ -727,14 +751,6 @@ def main(): ...@@ -727,14 +751,6 @@ def main():
default=False, default=False,
action='store_true', action='store_true',
help="Whether not to use CUDA when available") help="Whether not to use CUDA when available")
parser.add_argument("--local_rank",
type=int,
default=-1,
help="local_rank for distributed training on gpus")
parser.add_argument("--accumulate_gradients",
type=int,
default=1,
help="Number of steps to accumulate gradient on (divide the batch_size and accumulate)")
parser.add_argument('--seed', parser.add_argument('--seed',
type=int, type=int,
default=42, default=42,
...@@ -742,7 +758,22 @@ def main(): ...@@ -742,7 +758,22 @@ def main():
parser.add_argument('--gradient_accumulation_steps', parser.add_argument('--gradient_accumulation_steps',
type=int, type=int,
default=1, default=1,
help="Number of updates steps to accumualte before performing a backward/update pass.") help="Number of updates steps to accumulate before performing a backward/update pass.")
parser.add_argument("--local_rank",
type=int,
default=-1,
help="local_rank for distributed training on gpus")
parser.add_argument('--optimize_on_cpu',
default=False,
action='store_true',
help="Whether to perform optimization and keep the optimizer averages on CPU")
parser.add_argument('--fp16',
default=False,
action='store_true',
help="Whether to use 16-bit float precision instead of 32-bit")
parser.add_argument('--loss_scale',
type=float, default=128,
help='Loss scaling, positive power of 2 values can improve fp16 convergence.')
args = parser.parse_args() args = parser.parse_args()
...@@ -754,13 +785,17 @@ def main(): ...@@ -754,13 +785,17 @@ def main():
n_gpu = 1 n_gpu = 1
# Initializes the distributed backend which will take care of sychronizing nodes/GPUs # Initializes the distributed backend which will take care of sychronizing nodes/GPUs
torch.distributed.init_process_group(backend='nccl') torch.distributed.init_process_group(backend='nccl')
logger.info("device %s n_gpu %d distributed training %r", device, n_gpu, bool(args.local_rank != -1)) if args.fp16:
logger.info("16-bits training currently not supported in distributed training")
args.fp16 = False # (see https://github.com/pytorch/pytorch/pull/13496)
logger.info("device: {} n_gpu: {}, distributed training: {}, 16-bits trainiing: {}".format(
device, n_gpu, bool(args.local_rank != -1), args.fp16))
if args.accumulate_gradients < 1: if args.gradient_accumulation_steps < 1:
raise ValueError("Invalid accumulate_gradients parameter: {}, should be >= 1".format( raise ValueError("Invalid gradient_accumulation_steps parameter: {}, should be >= 1".format(
args.accumulate_gradients)) args.gradient_accumulation_steps))
args.train_batch_size = int(args.train_batch_size / args.accumulate_gradients) args.train_batch_size = int(args.train_batch_size / args.gradient_accumulation_steps)
random.seed(args.seed) random.seed(args.seed)
np.random.seed(args.seed) np.random.seed(args.seed)
...@@ -801,26 +836,36 @@ def main(): ...@@ -801,26 +836,36 @@ def main():
train_examples = read_squad_examples( train_examples = read_squad_examples(
input_file=args.train_file, is_training=True) input_file=args.train_file, is_training=True)
num_train_steps = int( num_train_steps = int(
len(train_examples) / args.train_batch_size * args.num_train_epochs) len(train_examples) / args.train_batch_size / args.gradient_accumulation_steps * args.num_train_epochs)
# Prepare model
model = BertForQuestionAnswering(bert_config) model = BertForQuestionAnswering(bert_config)
if args.init_checkpoint is not None: if args.init_checkpoint is not None:
model.bert.load_state_dict(torch.load(args.init_checkpoint, map_location='cpu')) model.bert.load_state_dict(torch.load(args.init_checkpoint, map_location='cpu'))
if args.fp16:
model.half()
model.to(device) model.to(device)
if args.local_rank != -1: if args.local_rank != -1:
model = torch.nn.parallel.DistributedDataParallel(model, device_ids=[args.local_rank], model = torch.nn.parallel.DistributedDataParallel(model, device_ids=[args.local_rank],
output_device=args.local_rank) output_device=args.local_rank)
elif n_gpu > 1: elif n_gpu > 1:
model = torch.nn.DataParallel(model) model = torch.nn.DataParallel(model)
# Prepare optimizer
if args.fp16:
param_optimizer = [(n, param.clone().detach().to('cpu').float().requires_grad_()) \
for n, param in model.named_parameters()]
elif args.optimize_on_cpu:
param_optimizer = [(n, param.clone().detach().to('cpu').requires_grad_()) \
for n, param in model.named_parameters()]
else:
param_optimizer = list(model.named_parameters())
no_decay = ['bias', 'gamma', 'beta'] no_decay = ['bias', 'gamma', 'beta']
optimizer_parameters = [ optimizer_grouped_parameters = [
{'params': [p for n, p in model.named_parameters() if n not in no_decay], 'weight_decay_rate': 0.01}, {'params': [p for n, p in param_optimizer if n not in no_decay], 'weight_decay_rate': 0.01},
{'params': [p for n, p in model.named_parameters() if n in no_decay], 'weight_decay_rate': 0.0} {'params': [p for n, p in param_optimizer if n in no_decay], 'weight_decay_rate': 0.0}
] ]
optimizer = BERTAdam(optimizer_grouped_parameters,
optimizer = BERTAdam(optimizer_parameters,
lr=args.learning_rate, lr=args.learning_rate,
warmup=args.warmup_proportion, warmup=args.warmup_proportion,
t_total=num_train_steps) t_total=num_train_steps)
...@@ -839,13 +884,11 @@ def main(): ...@@ -839,13 +884,11 @@ def main():
logger.info(" Num split examples = %d", len(train_features)) logger.info(" Num split examples = %d", len(train_features))
logger.info(" Batch size = %d", args.train_batch_size) logger.info(" Batch size = %d", args.train_batch_size)
logger.info(" Num steps = %d", num_train_steps) logger.info(" Num steps = %d", num_train_steps)
all_input_ids = torch.tensor([f.input_ids for f in train_features], dtype=torch.long) all_input_ids = torch.tensor([f.input_ids for f in train_features], dtype=torch.long)
all_input_mask = torch.tensor([f.input_mask for f in train_features], dtype=torch.long) all_input_mask = torch.tensor([f.input_mask for f in train_features], dtype=torch.long)
all_segment_ids = torch.tensor([f.segment_ids for f in train_features], dtype=torch.long) all_segment_ids = torch.tensor([f.segment_ids for f in train_features], dtype=torch.long)
all_start_positions = torch.tensor([f.start_position for f in train_features], dtype=torch.long) all_start_positions = torch.tensor([f.start_position for f in train_features], dtype=torch.long)
all_end_positions = torch.tensor([f.end_position for f in train_features], dtype=torch.long) all_end_positions = torch.tensor([f.end_position for f in train_features], dtype=torch.long)
train_data = TensorDataset(all_input_ids, all_input_mask, all_segment_ids, train_data = TensorDataset(all_input_ids, all_input_mask, all_segment_ids,
all_start_positions, all_end_positions) all_start_positions, all_end_positions)
if args.local_rank == -1: if args.local_rank == -1:
...@@ -855,25 +898,37 @@ def main(): ...@@ -855,25 +898,37 @@ def main():
train_dataloader = DataLoader(train_data, sampler=train_sampler, batch_size=args.train_batch_size) train_dataloader = DataLoader(train_data, sampler=train_sampler, batch_size=args.train_batch_size)
model.train() model.train()
for epoch in trange(int(args.num_train_epochs), desc="Epoch"): for _ in trange(int(args.num_train_epochs), desc="Epoch"):
for step, batch in enumerate(tqdm(train_dataloader, desc="Iteration")): for step, batch in enumerate(tqdm(train_dataloader, desc="Iteration")):
if n_gpu == 1:
batch = tuple(t.to(device) for t in batch) # multi-gpu does scattering it-self
input_ids, input_mask, segment_ids, start_positions, end_positions = batch input_ids, input_mask, segment_ids, start_positions, end_positions = batch
input_ids = input_ids.to(device) loss = model(input_ids, segment_ids, input_mask, start_positions, end_positions)
input_mask = input_mask.to(device)
segment_ids = segment_ids.to(device)
start_positions = start_positions.to(device)
end_positions = start_positions.to(device)
start_positions = start_positions.view(-1, 1)
end_positions = end_positions.view(-1, 1)
loss, _ = model(input_ids, segment_ids, input_mask, start_positions, end_positions)
if n_gpu > 1: if n_gpu > 1:
loss = loss.mean() # mean() to average on multi-gpu. loss = loss.mean() # mean() to average on multi-gpu.
if args.fp16 and args.loss_scale != 1.0:
# rescale loss for fp16 training
# see https://docs.nvidia.com/deeplearning/sdk/mixed-precision-training/index.html
loss = loss * args.loss_scale
if args.gradient_accumulation_steps > 1:
loss = loss / args.gradient_accumulation_steps
loss.backward() loss.backward()
if (step + 1) % args.gradient_accumulation_steps == 0: if (step + 1) % args.gradient_accumulation_steps == 0:
optimizer.step() # We have accumulated enought gradients if args.fp16 or args.optimize_on_cpu:
if args.fp16 and args.loss_scale != 1.0:
# scale down gradients for fp16 training
for param in model.parameters():
param.grad.data = param.grad.data / args.loss_scale
is_nan = set_optimizer_params_grad(param_optimizer, model.named_parameters(), test_nan=True)
if is_nan:
logger.info("FP16 TRAINING: Nan in gradients, reducing loss scaling")
args.loss_scale = args.loss_scale / 2
model.zero_grad()
continue
optimizer.step()
copy_optimizer_params_to_model(model.named_parameters(), param_optimizer)
else:
optimizer.step()
model.zero_grad() model.zero_grad()
global_step += 1 global_step += 1
...@@ -897,7 +952,6 @@ def main(): ...@@ -897,7 +952,6 @@ def main():
all_input_mask = torch.tensor([f.input_mask for f in eval_features], dtype=torch.long) all_input_mask = torch.tensor([f.input_mask for f in eval_features], dtype=torch.long)
all_segment_ids = torch.tensor([f.segment_ids for f in eval_features], dtype=torch.long) all_segment_ids = torch.tensor([f.segment_ids for f in eval_features], dtype=torch.long)
all_example_index = torch.arange(all_input_ids.size(0), dtype=torch.long) all_example_index = torch.arange(all_input_ids.size(0), dtype=torch.long)
eval_data = TensorDataset(all_input_ids, all_input_mask, all_segment_ids, all_example_index) eval_data = TensorDataset(all_input_ids, all_input_mask, all_segment_ids, all_example_index)
if args.local_rank == -1: if args.local_rank == -1:
eval_sampler = SequentialSampler(eval_data) eval_sampler = SequentialSampler(eval_data)
...@@ -908,30 +962,22 @@ def main(): ...@@ -908,30 +962,22 @@ def main():
model.eval() model.eval()
all_results = [] all_results = []
logger.info("Start evaluating") logger.info("Start evaluating")
for input_ids, input_mask, segment_ids, example_index in tqdm(eval_dataloader, desc="Evaluating"): for input_ids, input_mask, segment_ids, example_indices in tqdm(eval_dataloader, desc="Evaluating"):
if len(all_results) % 1000 == 0: if len(all_results) % 1000 == 0:
logger.info("Processing example: %d" % (len(all_results))) logger.info("Processing example: %d" % (len(all_results)))
input_ids = input_ids.to(device) input_ids = input_ids.to(device)
input_mask = input_mask.to(device) input_mask = input_mask.to(device)
segment_ids = segment_ids.to(device) segment_ids = segment_ids.to(device)
with torch.no_grad():
start_logits, end_logits = model(input_ids, segment_ids, input_mask) batch_start_logits, batch_end_logits = model(input_ids, segment_ids, input_mask)
for i, example_index in enumerate(example_indices):
unique_id = [int(eval_features[e.item()].unique_id) for e in example_index] start_logits = batch_start_logits[i].detach().cpu().tolist()
start_logits = [x.view(-1).detach().cpu().numpy() for x in start_logits] end_logits = batch_end_logits[i].detach().cpu().tolist()
end_logits = [x.view(-1).detach().cpu().numpy() for x in end_logits] eval_feature = eval_features[example_index.item()]
for idx, i in enumerate(unique_id): unique_id = int(eval_feature.unique_id)
s = [float(x) for x in start_logits[idx]] all_results.append(RawResult(unique_id=unique_id,
e = [float(x) for x in end_logits[idx]] start_logits=start_logits,
all_results.append( end_logits=end_logits))
RawResult(
unique_id=i,
start_logits=s,
end_logits=e
)
)
output_prediction_file = os.path.join(args.output_dir, "predictions.json") output_prediction_file = os.path.join(args.output_dir, "predictions.json")
output_nbest_file = os.path.join(args.output_dir, "nbest_predictions.json") output_nbest_file = os.path.join(args.output_dir, "nbest_predictions.json")
write_predictions(eval_examples, eval_features, all_results, write_predictions(eval_examples, eval_features, all_results,
......
tests/tokenization_test.py
View file @ f4d79f44
...@@ -43,6 +43,13 @@ class TokenizationTest(unittest.TestCase): ...@@ -43,6 +43,13 @@ class TokenizationTest(unittest.TestCase):
self.assertListEqual( self.assertListEqual(
tokenizer.convert_tokens_to_ids(tokens), [7, 4, 5, 10, 8, 9]) tokenizer.convert_tokens_to_ids(tokens), [7, 4, 5, 10, 8, 9])
def test_chinese(self):
tokenizer = tokenization.BasicTokenizer()
self.assertListEqual(
tokenizer.tokenize(u"ah\u535A\u63A8zz"),
[u"ah", u"\u535A", u"\u63A8", u"zz"])
def test_basic_tokenizer_lower(self): def test_basic_tokenizer_lower(self):
tokenizer = tokenization.BasicTokenizer(do_lower_case=True) tokenizer = tokenization.BasicTokenizer(do_lower_case=True)
......
tokenization.py
View file @ f4d79f44
...@@ -133,6 +133,13 @@ class BasicTokenizer(object): ...@@ -133,6 +133,13 @@ class BasicTokenizer(object):
"""Tokenizes a piece of text.""" """Tokenizes a piece of text."""
text = convert_to_unicode(text) text = convert_to_unicode(text)
text = self._clean_text(text) text = self._clean_text(text)
# This was added on November 1st, 2018 for the multilingual and Chinese
# models. This is also applied to the English models now, but it doesn't
# matter since the English models were not trained on any Chinese data
# and generally don't have any Chinese data in them (there are Chinese
# characters in the vocabulary because Wikipedia does have some Chinese
# words in the English Wikipedia.).
text = self._tokenize_chinese_chars(text)
orig_tokens = whitespace_tokenize(text) orig_tokens = whitespace_tokenize(text)
split_tokens = [] split_tokens = []
for token in orig_tokens: for token in orig_tokens:
...@@ -175,6 +182,41 @@ class BasicTokenizer(object): ...@@ -175,6 +182,41 @@ class BasicTokenizer(object):
return ["".join(x) for x in output] return ["".join(x) for x in output]
def _tokenize_chinese_chars(self, text):
"""Adds whitespace around any CJK character."""
output = []
for char in text:
cp = ord(char)
if self._is_chinese_char(cp):
output.append(" ")
output.append(char)
output.append(" ")
else:
output.append(char)
return "".join(output)
def _is_chinese_char(self, cp):
"""Checks whether CP is the codepoint of a CJK character."""
# This defines a "chinese character" as anything in the CJK Unicode block:
# https://en.wikipedia.org/wiki/CJK_Unified_Ideographs_(Unicode_block)
#
# Note that the CJK Unicode block is NOT all Japanese and Korean characters,
# despite its name. The modern Korean Hangul alphabet is a different block,
# as is Japanese Hiragana and Katakana. Those alphabets are used to write
# space-separated words, so they are not treated specially and handled
# like the all of the other languages.
if ((cp >= 0x4E00 and cp <= 0x9FFF) or #
(cp >= 0x3400 and cp <= 0x4DBF) or #
(cp >= 0x20000 and cp <= 0x2A6DF) or #
(cp >= 0x2A700 and cp <= 0x2B73F) or #
(cp >= 0x2B740 and cp <= 0x2B81F) or #
(cp >= 0x2B820 and cp <= 0x2CEAF) or
(cp >= 0xF900 and cp <= 0xFAFF) or #
(cp >= 0x2F800 and cp <= 0x2FA1F)): #
return True
return False
def _clean_text(self, text): def _clean_text(self, text):
"""Performs invalid character removal and whitespace cleanup on text.""" """Performs invalid character removal and whitespace cleanup on text."""
output = [] output = []
......
Markdown is supported
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment