RuntimeError: Incoming model is an instance of torch.nn.parallel.DataParallel. Parallel wrappers should only be applied to the model(s) AFTER the model(s) have been returned from amp.initialize. about textbrewer

[Youarerare]

我把这行代码删了能运行了，但是根本问题在哪还不知道

TextBrewer/examples/mnli_example/main.distill.py

Line 43 in ba8ad6d

torch.distributed.init_process_group(backend='nccl')

[airaria]

是在同时使用多卡计算和fp16时出现问题吗？使用的是数据并行方式DataParallel还是分布式并行DistributedDataParallel？

[Youarerare]

采用如下配置蒸馏自定义的三层模型同样报上面的错误
device : cuda:5
fp16 : True
fp16_opt_level : O1
data_parallel : False
local_rank : -1

完整错误信息：
2021/03/30 17:49:30 - INFO - Main - output_dir:/data/private/syk/zyb/TextMatch/Bert/mnli_t8_TbaseST3tiny_L3SmmdMSE_lr10e40_bs16
2021/03/30 17:49:30 - INFO - Main - data_dir:/data/private/syk/zyb/TextMatch/data/MNLI
2021/03/30 17:49:30 - INFO - Main - max_seq_length:512
2021/03/30 17:49:30 - INFO - Main - do_train:True
2021/03/30 17:49:30 - INFO - Main - do_predict:True
2021/03/30 17:49:30 - INFO - Main - train_batch_size:16
2021/03/30 17:49:30 - INFO - Main - predict_batch_size:8
2021/03/30 17:49:30 - INFO - Main - learning_rate:0.0001
2021/03/30 17:49:30 - INFO - Main - num_train_epochs:40.0
2021/03/30 17:49:30 - INFO - Main - warmup_proportion:0.1
2021/03/30 17:49:30 - INFO - Main - no_cuda:False
np_resource = np.dtype([("resource", np.ubyte, 1)])
2021/03/30 17:49:30 - INFO - faiss - Loading faiss with AVX2 support.
2021/03/30 17:49:30 - INFO - faiss - Loading faiss.
2021/03/30 17:49:30 - INFO - Main - output_dir:/data/private/syk/zyb/TextMatch/Bert/mnli_t8_TbaseST3tiny_L3SmmdMSE_lr10e40_bs16
2021/03/30 17:49:30 - INFO - Main - data_dir:/data/private/syk/zyb/TextMatch/data/MNLI
2021/03/30 17:49:30 - INFO - Main - max_seq_length:512
2021/03/30 17:49:30 - INFO - Main - do_train:True
2021/03/30 17:49:30 - INFO - Main - do_predict:True
2021/03/30 17:49:30 - INFO - Main - train_batch_size:16
2021/03/30 17:49:30 - INFO - Main - predict_batch_size:8
2021/03/30 17:49:30 - INFO - Main - learning_rate:0.0001
2021/03/30 17:49:30 - INFO - Main - gradient_accumulation_steps:1
2021/03/30 17:49:30 - INFO - Main - local_rank:-1
2021/03/30 17:49:30 - INFO - Main - fp16:True
2021/03/30 17:49:30 - INFO - Main - random_seed:9580
2021/03/30 17:49:30 - INFO - Main - weight_decay_rate:0.01
2021/03/30 17:49:30 - INFO - Main - do_eval:True
2021/03/30 17:49:30 - INFO - Main - data_dir:/data/private/syk/zyb/TextMatch/data/MNLI
2021/03/30 17:49:30 - INFO - Main - max_seq_length:512
2021/03/30 17:49:30 - INFO - Main - do_train:True
2021/03/30 17:49:30 - INFO - Main - do_predict:True
2021/03/30 17:49:30 - INFO - Main - train_batch_size:16
2021/03/30 17:49:30 - INFO - Main - predict_batch_size:8
2021/03/30 17:49:30 - INFO - Main - learning_rate:0.0001
2021/03/30 17:49:30 - INFO - Main - num_train_epochs:40.0
2021/03/30 17:49:30 - INFO - Main - warmup_proportion:0.1
2021/03/30 17:49:30 - INFO - Main - no_cuda:False
2021/03/30 17:49:30 - INFO - Main - gradient_accumulation_steps:1
2021/03/30 17:49:30 - INFO - Main - local_rank:-1
2021/03/30 17:49:30 - INFO - Main - fp16:True
2021/03/30 17:49:30 - INFO - Main - random_seed:9580
2021/03/30 17:49:30 - INFO - Main - weight_decay_rate:0.01
2021/03/30 17:49:30 - INFO - Main - do_eval:True
2021/03/30 17:49:30 - INFO - Main - PRINT_EVERY:200
2021/03/30 17:49:30 - INFO - Main - ckpt_frequency:1
2021/03/30 17:49:30 - INFO - Main - temperature:8.0
2021/03/30 17:49:30 - INFO - Main - teacher_cached:False
2021/03/30 17:49:30 - INFO - Main - task_name:mnli
2021/03/30 17:49:30 - INFO - Main - aux_task_name:None
2021/03/30 17:49:30 - INFO - Main - aux_data_dir:None
2021/03/30 17:49:30 - INFO - Main - matches:['L3_attention_mse', 'L3_hidden_smmd']
2021/03/30 17:49:30 - INFO - Main - model_config_json:/data/private/syk/zyb/TextMatch/Bert/mnli_t8_TbaseST3tiny_L3SmmdMSE_lr10e40_bs16/DIstillBertToT3.json.run
2021/03/30 17:49:30 - INFO - Main - do_test:False
2021/03/30 17:49:30 - WARNING - Main - Output directory () already exists and is not empty.
2021/03/30 17:49:30 - INFO - Main - device cuda:5 n_gpu 8 distributed training False
2021/03/30 17:49:30 - INFO - utils - Loading features from cached file /data/private/syk/zyb/TextMatch/data/MNLI/rbt_3_train_512_mnli
2021/03/30 17:50:07 - INFO - utils - Loading features from cached file /data/private/syk/zyb/TextMatch/data/MNLI/rbt_3_dev_512_mnli
2021/03/30 17:50:09 - INFO - utils - Loading features from cached file /data/private/syk/zyb/TextMatch/data/MNLI/rbt_3_dev_512_mnli-mm
2021/03/30 17:50:12 - INFO - Main - Data loaded
2021/03/30 17:50:14 - INFO - Main - Teacher Model bert loaded
2021/03/30 17:50:20 - INFO - Main - missing keys:['bert.embeddings.position_ids', 'classifier.weight', 'classifier.bias']
2021/03/30 17:50:20 - INFO - Main - unexpected keys:['cls.predictions.bias', 'cls.predictions.transform.dense.weight', 'cls.predictions.transform.dense.bias', 'cls.predictions.transform.LayerNorm.weight', 'cls.predictions.transform.LayerNorm.bias', 'cls.predictions.decoder.weight', 'cls.seq_relationship.weight', 'cls.seq_relationship.bias']
2021/03/30 17:50:20 - INFO - Main - Student Model loaded
2021/03/30 17:50:20 - INFO - Main - Length of all_trainable_params: 2
2021/03/30 17:50:20 - INFO - Main - [{'layer_T': 4, 'layer_S': 1, 'feature': 'attention', 'loss': 'attention_mse', 'weight': 1}, {'layer_T': 8, 'layer_S': 2, 'feature': 'attention', 'loss': 'attention_mse', 'weight': 1}, {'layer_T': 12, 'layer_S': 3, 'feature': 'attention', 'loss': 'attention_mse', 'weight': 1}, {'layer_T': [0, 0], 'layer_S': [0, 0], 'feature': 'hidden', 'loss': 'mmd', 'weight': 1}, {'layer_T': [4, 4], 'layer_S': [1, 1], 'feature': 'hidden', 'loss': 'mmd', 'weight': 1}, {'layer_T': [8, 8], 'layer_S': [2, 2], 'feature': 'hidden', 'loss': 'mmd', 'weight': 1}, {'layer_T': [12, 12], 'layer_S': [3, 3], 'feature': 'hidden', 'loss': 'mmd', 'weight': 1}]
2021/03/30 17:50:20 - INFO - Main - gradient_accumulation_steps : 1
ckpt_frequency : 1
ckpt_epoch_frequency : 1
ckpt_steps : None
log_dir : /data/private/syk/zyb/TextMatch/Bert/mnli_t8_TbaseST3tiny_L3SmmdMSE_lr10e40_bs16
output_dir : /data/private/syk/zyb/TextMatch/Bert/mnli_t8_TbaseST3tiny_L3SmmdMSE_lr10e40_bs16
device : cuda:5
fp16 : True
fp16_opt_level : O1
data_parallel : False
local_rank : -1

2021/03/30 17:50:20 - INFO - Main - temperature : 8.0
temperature_scheduler : None
hard_label_weight : 0
hard_label_weight_scheduler : None
kd_loss_type : ce
kd_loss_weight : 1
kd_loss_weight_scheduler : None
probability_shift : False
intermediate_matches : [
IntermediateMatch: layer_T : 4, layer_S : 1, feature : attention, weight : 1, loss : attention_mse, proj : None,
IntermediateMatch: layer_T : 8, layer_S : 2, feature : attention, weight : 1, loss : attention_mse, proj : None,
IntermediateMatch: layer_T : 12, layer_S : 3, feature : attention, weight : 1, loss : attention_mse, proj : None,
IntermediateMatch: layer_T : [0, 0], layer_S : [0, 0], feature : hidden, weight : 1, loss : mmd, proj : None,
IntermediateMatch: layer_T : [4, 4], layer_S : [1, 1], feature : hidden, weight : 1, loss : mmd, proj : None,
IntermediateMatch: layer_T : [8, 8], layer_S : [2, 2], feature : hidden, weight : 1, loss : mmd, proj : None,
IntermediateMatch: layer_T : [12, 12], layer_S : [3, 3], feature : hidden, weight : 1, loss : mmd, proj : None]
is_caching_logits : False

2021/03/30 17:50:20 - INFO - Main - ***** Running training *****
2021/03/30 17:50:20 - INFO - Main - Num examples = 300000
2021/03/30 17:50:20 - INFO - Main - Forward batch size = 16
2021/03/30 17:50:20 - INFO - Main - Num backward steps = 750000
Selected optimization level O1: Insert automatic casts around Pytorch functions and Tensor methods.

Defaults for this optimization level are:
enabled : True
opt_level : O1
cast_model_type : None
patch_torch_functions : True
keep_batchnorm_fp32 : None
master_weights : None
loss_scale : dynamic
Processing user overrides (additional kwargs that are not None)...
After processing overrides, optimization options are:
enabled : True
opt_level : O1
cast_model_type : None
patch_torch_functions : True
cast_model_type : None
patch_torch_functions : True
keep_batchnorm_fp32 : None
master_weights : None
loss_scale : dynamic
Processing user overrides (additional kwargs that are not None)...
After processing overrides, optimization options are:
enabled : True
opt_level : O1
cast_model_type : None
patch_torch_functions : True
cast_model_type : None
patch_torch_functions : True
keep_batchnorm_fp32 : None
master_weights : None
loss_scale : dynamic
Processing user overrides (additional kwargs that are not None)...
After processing overrides, optimization options are:
enabled : True
opt_level : O1
cast_model_type : None
patch_torch_functions : True
cast_model_type : None
patch_torch_functions : True
keep_batchnorm_fp32 : None
master_weights : None
loss_scale : dynamic
Processing user overrides (additional kwargs that are not None)...
After processing overrides, optimization options are:
enabled : True
opt_level : O1
cast_model_type : None
patch_torch_functions : True
keep_batchnorm_fp32 : None
master_weights : None
loss_scale : dynamic
Traceback (most recent call last):
File "main.distill.py", line 199, in
main()
File "main.distill.py", line 192, in main
num_epochs = args.num_train_epochs, callback=callback_func,max_grad_norm=1)
File "/home/syk/anaconda3/lib/python3.7/site-packages/textbrewer/distiller_basic.py", line 277, in train
optimizer, scheduler, tqdm_disable = self.initialize_training(optimizer, scheduler_class, scheduler_args, scheduler)
File "/home/syk/anaconda3/lib/python3.7/site-packages/textbrewer/distiller_basic.py", line 89, in initialize_training
(self.model_S, self.model_T), optimizer = amp.initialize([self.model_S, self.model_T], optimizer, opt_level=self.t_config.fp16_opt_level)
File "/home/syk/anaconda3/lib/python3.7/site-packages/apex-0.1-py3.7.egg/apex/amp/frontend.py", line 358, in initialize
return _initialize(models, optimizers, _amp_state.opt_properties, num_losses, cast_model_outputs)
File "/home/syk/anaconda3/lib/python3.7/site-packages/apex-0.1-py3.7.egg/apex/amp/_initialize.py", line 168, in _initialize
check_models(models)
File "/home/syk/anaconda3/lib/python3.7/site-packages/apex-0.1-py3.7.egg/apex/amp/_initialize.py", line 75, in check_models
"Parallel wrappers should only be applied to the model(s) AFTER \n"
RuntimeError: Incoming model is an instance of torch.nn.parallel.DataParallel. Parallel wrappers should only be applied to the model(s) AFTER
the model(s) have been returned from amp.initialize.

[Youarerare]

破案了。。。 有多张卡的时候最好加上cuda_visible_device你用的那几张，不然n_gpus 是默认全部卡。。。