训练问题

[weiweili123]

您好！我用resnet18提取了数据集的特征，但是不知道怎样训练，好像训练代码是针对于bottom-up-attention的。同时也不是很清楚如何用bottom-up-attention提出AI Challenger训练集的特征

[ruotianluo]

应该是兼容的，可以直接训练

[weiweili123]

您好！我把run_train.sh修改成如下： #! /bin/sh

#larger batch

id="dense_box_bn"$1 ckpt_path="log_"$id if [ ! -d $ckpt_path ]; then mkdir $ckpt_path fi if [ ! -f $ckpt_path"/infos_"$id".pkl" ]; then start_from="" else start_from="--start_from "$ckpt_path fi

python train.py --caption_model denseatt --input_json data/chinese_talk.json --input_label_h5 data/chinese_talk_label.h5 --input_fc_dir data/chinese_talk_fc --input_att_dir data/chinese_talk_att --seq_per_img 512 --batch_size 50 --beam_size 1 --learning_rate 5e-4 --learning_rate_decay_start 0 --scheduled_sampling_start 0 --checkpoint_path $ckpt_path $start_from --save_checkpoint_every 3000 --language_eval 1 --val_images_use 10000 --max_epoch 37 --rnn_size 1300 --use_bn 1

if [ ! -d xe/$ckpt_path ]; then cp -r $ckpt_path xe/ fi

python train.py --caption_model denseatt --input_json data/chinese_talk.json --input_label_h5 data/chinese_talk_label.h5 --input_fc_dir data/chinese_talk_fc --input_att_dir data/chinese_talk_att --seq_per_img 5 --batch_size 50 --beam_size 1 --learning_rate 5e-5 --learning_rate_decay_start 0 --learning_rate_decay_every 55 --learning_rate_decay_rate 0.1 --scheduled_sampling_start 0 --checkpoint_path $ckpt_path --start_from $ckpt_path --save_checkpoint_every 3000 --language_eval 1 --val_images_use 10000 --self_critical_after 37 --rnn_size 1300 --use_bn 1

运行结果： zou@zou:~/Image_Captioning_chinese$ sh run_train.sh /usr/local/lib/python2.7/dist-packages/h5py/init.py:36: FutureWarning: Conversion of the second argument of issubdtype from float to np.floating is deprecated. In future, it will be treated as np.float64 == np.dtype(float).type. from ._conv import register_converters as _register_converters

DataLoader loading json file: data/chinese_talk.json vocab size is 4461 DataLoader loading h5 file: data/chinese_talk_fc data/chinese_talk_att data/cocotalk_box data/chinese_talk_label.h5 max sequence length in data is 20 read 240000 image features assigned 220000 images to split train assigned 10000 images to split val assigned 10000 images to split test Traceback (most recent call last): File "train.py", line 228, in train(opt) File "train.py", line 114, in train data = loader.get_batch('train') File "/home/zou/Image_Captioning_chinese/dataloader.py", line 164, in get_batch data['att_feats'][i*seq_per_img:(i+1)*seq_per_img, :att_batch[i].shape[0]] = att_batch[i] ValueError: could not broadcast input array from shape (14,14,512) into shape (512,14,14) Terminating BlobFetcher /usr/local/lib/python2.7/dist-packages/h5py/init.py:36: FutureWarning: Conversion of the second argument of issubdtype from float to np.floating is deprecated. In future, it will be treated as np.float64 == np.dtype(float).type. from .conv import register_converters as register_converters DataLoader loading json file: data/chinese_talk.json vocab size is 4461 DataLoader loading h5 file: data/chinese_talk_fc data/chinese_talk_att data/cocotalk_box data/chinese_talk_label.h5 max sequence length in data is 20 read 240000 image features assigned 220000 images to split train assigned 10000 images to split val assigned 10000 images to split test Traceback (most recent call last): File "train.py", line 228, in train(opt) File "train.py", line 48, in train with open(os.path.join(opt.start_from, 'infos'+opt.id+'.pkl')) as f: IOError: [Errno 2] No such file or directory: 'log_dense_box_bn/infos.pkl'

不知道是训练参数设置问题，还是使用resnet18提取特征导致输出的参数跟后面不能匹配？

[ruotianluo]

哦，好像不完全兼容。。seq per img不用改。你可以在dataloader的load not时，把14x14x512 reshape成196x512

[weiweili123]

好的，还想再问一个问题：用resnet101提取特征的时候，我的1080显卡（8.1G)）每次处理到3000张图片的时候就爆掉了，您当时提取特征的显存多少？以及代码是否设置了多显卡选项？

[ruotianluo]

1080ti。没有多卡

[weiweili123]

好的，谢谢！

[zhufeijuanjuan]

您好！我把run_train.sh修改成如下： #! /bin/sh

#larger batch

id="dense_box_bn"$1 ckpt_path="log_"$id if [ ! -d $ckpt_path ]; then mkdir $ckpt_path fi if [ ! -f $ckpt_path"/infos_"$id".pkl" ]; then start_from="" else start_from="--start_from "$ckpt_path fi

python train.py --caption_model denseatt --input_json data/chinese_talk.json --input_label_h5 data/chinese_talk_label.h5 --input_fc_dir data/chinese_talk_fc --input_att_dir data/chinese_talk_att --seq_per_img 512 --batch_size 50 --beam_size 1 --learning_rate 5e-4 --learning_rate_decay_start 0 --scheduled_sampling_start 0 --checkpoint_path $ckpt_path $start_from --save_checkpoint_every 3000 --language_eval 1 --val_images_use 10000 --max_epoch 37 --rnn_size 1300 --use_bn 1

if [ ! -d xe/$ckpt_path ]; then cp -r $ckpt_path xe/ fi

python train.py --caption_model denseatt --input_json data/chinese_talk.json --input_label_h5 data/chinese_talk_label.h5 --input_fc_dir data/chinese_talk_fc --input_att_dir data/chinese_talk_att --seq_per_img 5 --batch_size 50 --beam_size 1 --learning_rate 5e-5 --learning_rate_decay_start 0 --learning_rate_decay_every 55 --learning_rate_decay_rate 0.1 --scheduled_sampling_start 0 --checkpoint_path $ckpt_path --start_from $ckpt_path --save_checkpoint_every 3000 --language_eval 1 --val_images_use 10000 --self_critical_after 37 --rnn_size 1300 --use_bn 1

运行结果： zou@zou:~/Image_Captioning_chinese$ sh run_train.sh /usr/local/lib/python2.7/dist-packages/h5py/init.py:36: FutureWarning: Conversion of the second argument of issubdtype from float to np.floating is deprecated. In future, it will be treated as np.float64 == np.dtype(float).type. from ._conv import register_converters as _register_converters

DataLoader loading json file: data/chinese_talk.json vocab size is 4461 DataLoader loading h5 file: data/chinese_talk_fc data/chinese_talk_att data/cocotalk_box data/chinese_talk_label.h5 max sequence length in data is 20 read 240000 image features assigned 220000 images to split train assigned 10000 images to split val assigned 10000 images to split test Traceback (most recent call last): File "train.py", line 228, in train(opt) File "train.py", line 114, in train data = loader.get_batch('train') File "/home/zou/Image_Captioning_chinese/dataloader.py", line 164, in get_batch data['att_feats'][i*seq_per_img:(i+1)*seq_per_img, :att_batch[i].shape[0]] = att_batch[i] ValueError: could not broadcast input array from shape (14,14,512) into shape (512,14,14) Terminating BlobFetcher /usr/local/lib/python2.7/dist-packages/h5py/init.py:36: FutureWarning: Conversion of the second argument of issubdtype from float to np.floating is deprecated. In future, it will be treated as np.float64 == np.dtype(float).type. from .conv import register_converters as register_converters DataLoader loading json file: data/chinese_talk.json vocab size is 4461 DataLoader loading h5 file: data/chinese_talk_fc data/chinese_talk_att data/cocotalk_box data/chinese_talk_label.h5 max sequence length in data is 20 read 240000 image features assigned 220000 images to split train assigned 10000 images to split val assigned 10000 images to split test Traceback (most recent call last): File "train.py", line 228, in train(opt) File "train.py", line 48, in train with open(os.path.join(opt.start_from, 'infos'+opt.id+'.pkl')) as f: IOError: [Errno 2] No such file or directory: 'log_dense_box_bn/infos.pkl'

不知道是训练参数设置问题，还是使用resnet18提取特征导致输出的参数跟后面不能匹配？

你好，我也遇到了这个问题，请问是如何解决的，谢谢。