Comments (13)
yangapku
commented on May 21, 2024
2
@yumulinfeng1 您好，有以下几个要点请您注意下：
- 训练过程中进行的验证，都是在一个batch内部计算Acc，这里计算的结果和最终全局召回的Recall@1/5/10及Mean Recall不是一个指标，仅用于训练过程判断收敛趋势。如果要得到可以和我们汇报的Recall@1/5/10以及Mean Recall对比的指标，请按照我们readme中跨模态检索部分，描述的训练→特征提取→KNN召回→计算Recall这个过程，完整走一遍finetune和测试集全图片池召回的过程。
- finetune的效果与超参数也有关系。能够跑出最优结果的超参数，请参见我们技术报告的附录部分A.3，给出了每个规模、每个数据集的最优超参数，供您参考。
- 请保证预训练ckpt有正常load进来。
还不太清楚您所说的60+具体是哪个指标。建议您可以先尝试对齐zero-shot结果，无须finetune。直接用想要对齐规模的模型预训练ckpt、和我们提供的预处理好的数据集，走完特征提取→KNN召回→计算Recall这个流程，看下Recall指标能否和我们汇报的结果一致，也供您熟悉一下用Chinese-CLIP进行图文检索的一个标准流程。 在此基础上，进一步跑个finetune，观察下效果能否提升。
from chinese-clip.
dengfenglai321
commented on May 21, 2024
建议您可以先尝试对齐zero-shot结果，无须fine
回答得很详细 谢谢
from chinese-clip.
dengfenglai321
commented on May 21, 2024
@yumulinfeng1 您好，有以下几个要点请您注意下：
- 训练过程中进行的验证，都是在一个batch内部计算Acc，这里计算的结果和最终全局召回的Recall@1/5/10及Mean Recall不是一个指标，仅用于训练过程判断收敛趋势。如果要得到可以和我们汇报的Recall@1/5/10以及Mean Recall对比的指标，请按照我们readme中跨模态检索部分，描述的训练→特征提取→KNN召回→计算Recall这个过程，完整走一遍finetune和测试集全图片池召回的过程。
- finetune的效果与超参数也有关系。能够跑出最优结果的超参数，请参见我们技术报告的附录部分A.3，给出了每个规模、每个数据集的最优超参数，供您参考。
- 请保证预训练ckpt有正常load进来。
还不太清楚您所说的60+具体是哪个指标。建议您可以先尝试对齐zero-shot结果，无须finetune。直接用想要对齐规模的模型预训练ckpt、和我们提供的预处理好的数据集，走完特征提取→KNN召回→计算Recall这个流程，看下Recall指标能否和我们汇报的结果一致，也供您熟悉一下用Chinese-CLIP进行图文检索的一个标准流程。 在此基础上，进一步跑个finetune，观察下效果能否提升。
你好 请问一下你仓库里给的验证集的log示例：
2022-06-16,11:06:00 | INFO | Rank 0 | Validation Result (epoch 1 @ 150 steps) | Valid Loss: 0.503617 | Image2Text Acc: 84.76 | Text2Image Acc: 84.37 | logit_scale: 4.605 | Valid Batch Size: 128
是哪个数据集的验证集？
如果是MUGE，那验证集的差距太大了。
我这边跑MUGE验证集的准确率就没变过，一直在32左右，是否有问题
from chinese-clip.
yangapku
commented on May 21, 2024
您好，这个应该是来自MUGE的log，您可以贴一下您实际的运行脚本和log，便于我们判断
from chinese-clip.
dengfenglai321
commented on May 21, 2024
脚本如下：
# Number of GPUs per GPU worker
GPUS_PER_NODE=2
# Number of GPU workers, for single-worker training, please set to 1
WORKER_CNT=1
# The ip address of the rank-0 worker, for single-worker training, please set to localhost
export MASTER_ADDR=10.5.55.18
# The port for communication
export MASTER_PORT=8514
# The rank of this worker, should be in {0, ..., WORKER_CNT-1}, for single-worker training, please set to 0
export RANK=0
export DATAPATH=/xxx/Text_Based_Image_Retrieval/data
export OUTLOGPATH=/xxx/Text_Based_Image_Retrieval/Chinese-CLIP-master/experiments
# data options
train_data=${DATAPATH}/datasets/MUGE/lmdb/train
val_data=${DATAPATH}/datasets/MUGE/lmdb/valid # if val_data is not specified, the validation will be automatically disabled
# restore options
resume=${OUTLOGPATH}/pretrained_weights/clip_cn_vit-b-16.pt # or specify your customed ckpt path to resume
reset_data_offset="--reset-data-offset"
reset_optimizer="--reset-optimizer"
# reset_optimizer=""
# output options
output_base_dir=${OUTLOGPATH}/
name=muge_finetune_vit-b-16_roberta-base_bs32
save_step_frequency=999999 # disable it
save_epoch_frequency=1
log_interval=1
report_training_batch_acc="--report-training-batch-acc"
# report_training_batch_acc=""
# training hyper-params
context_length=52
warmup=100
batch_size=48
valid_batch_size=48
lr=2e-5
wd=0.001
max_epochs=16
valid_step_interval=150
valid_epoch_interval=1
vision_model=ViT-B-16
text_model=RoBERTa-wwm-ext-base-chinese
use_augment="--use-augment"
# use_augment=""
python3 -m torch.distributed.launch --nproc_per_node=${GPUS_PER_NODE} --nnodes=${WORKER_CNT} --node_rank=${RANK} \
--master_addr=${MASTER_ADDR} --master_port=${MASTER_PORT} cn_clip/training/main.py \
--train-data=${train_data} \
--val-data=${val_data} \
--resume=${resume} \
${reset_data_offset} \
${reset_optimizer} \
--logs=${output_base_dir} \
--name=${name} \
--save-step-frequency=${save_step_frequency} \
--save-epoch-frequency=${save_epoch_frequency} \
--log-interval=${log_interval} \
${report_training_batch_acc} \
--context-length=${context_length} \
--warmup=${warmup} \
--batch-size=${batch_size} \
--valid-batch-size=${valid_batch_size} \
--valid-step-interval=${valid_step_interval} \
--valid-epoch-interval=${valid_epoch_interval} \
--lr=${lr} \
--wd=${wd} \
--max-epochs=${max_epochs} \
--vision-model=${vision_model} \
${use_augment} \
--text-model=${text_model}
log如下：
2022-12-09,11:31:47 | INFO | Rank 0 | Global Steps: 25500/41728 | Train Epoch: 10 [194688/250368 (78%)] | Loss: 0.191398 | Image2Text Acc: 92.71 | Text2Image Acc: 94.79 | Data Time: 0.060s | Batch Time: 0.835s | LR: 0.000007 | logit_scale: 4.584 | Global Batch Size: 96
2022-12-09,11:31:47 | INFO | Rank 0 | Begin to eval on validation set (epoch 10 @ 25500 steps)...
2022-12-09,11:32:56 | INFO | Rank 1 | Evaluated 100/319 batches...
2022-12-09,11:32:58 | INFO | Rank 0 | Evaluated 100/319 batches...
2022-12-09,11:34:04 | INFO | Rank 1 | Evaluated 200/319 batches...
2022-12-09,11:34:10 | INFO | Rank 0 | Evaluated 200/319 batches...2022-12-09,11:35:13 | INFO | Rank 1 | Evaluated 300/319 batches... 0
2022-12-09,11:35:21 | INFO | Rank 0 | Evaluated 300/319 batches...
2022-12-09,11:35:35 | INFO | Rank 1 | Validation Result (epoch 10 @ 25500 steps) | Valid Loss: 2.025491 | Image2Text Acc: 32.39 | Text2Image Acc: 33.03 | logit_scale: 4.584 | Valid Batch Size: 48
2022-12-09,11:35:35 | INFO | Rank 0 | Validation Result (epoch 10 @ 25500 steps) | Valid Loss: 2.025491 | Image2Text Acc: 32.39 | Text2Image Acc: 33.03 | logit_scale: 4.584 | Valid Batch Size: 48
from chinese-clip.
yangapku
commented on May 21, 2024
请问方便提供一个完整的训练log文件链接吗？我们初步判断，是没有正确load进来预训练的ckpt参数，建议您也检查下ckpt是否放在--resume指定的位置。
from chinese-clip.
yangapku
commented on May 21, 2024
您也可以看下log中是否有"=> no checkpoint found at"这样的日志打印出来，确认这一点。
from chinese-clip.
dengfenglai321
commented on May 21, 2024
您也可以看下log中是否有"=> no checkpoint found at"这样的日志打印出来，确认这一点。
应该是加载成功权重了
2022-12-08,18:15:40 | INFO | Rank 0 | train LMDB file contains 129380 images and 250314 pairs.
2022-12-08,18:15:40 | INFO | Rank 0 | val LMDB file contains 29806 images and 30588 pairs.
2022-12-08,18:15:40 | INFO | Rank 0 | Params:
2022-12-08,18:15:40 | INFO | Rank 0 | aggregate: True
2022-12-08,18:15:40 | INFO | Rank 0 | batch_size: 48
2022-12-08,18:15:40 | INFO | Rank 0 | bert_weight_path: None
2022-12-08,18:15:40 | INFO | Rank 0 | beta1: 0.9
2022-12-08,18:15:40 | INFO | Rank 0 | beta2: 0.98
2022-12-08,18:15:40 | INFO | Rank 0 | checkpoint_path: /storage1/xxx/Text_Based_Image_Retrieval/Chinese-CLIP-master/experiments/muge_finetune_vit-b-16_roberta-base_bs32/checkpoints
2022-12-08,18:15:40 | INFO | Rank 0 | clip_weight_path: None
2022-12-08,18:15:40 | INFO | Rank 0 | context_length: 52
2022-12-08,18:15:40 | INFO | Rank 0 | debug: False
2022-12-08,18:15:40 | INFO | Rank 0 | device: cuda:0
2022-12-08,18:15:40 | INFO | Rank 0 | eps: 1e-06
2022-12-08,18:15:40 | INFO | Rank 0 | freeze_vision: False
2022-12-08,18:15:40 | INFO | Rank 0 | grad_checkpointing: False
2022-12-08,18:15:40 | INFO | Rank 0 | local_device_rank: 0
2022-12-08,18:15:40 | INFO | Rank 0 | local_rank: 0
2022-12-08,18:15:40 | INFO | Rank 0 | log_interval: 1
2022-12-08,18:15:40 | INFO | Rank 0 | log_level: 20
2022-12-08,18:15:40 | INFO | Rank 0 | log_path: /storage1/xxx/Text_Based_Image_Retrieval/Chinese-CLIP-master/experiments/muge_finetune_vit-b-16_roberta-base_bs32/out_2022-12-08-10-15-34.log
2022-12-08,18:15:40 | INFO | Rank 0 | logs: /storage1/xxx/Text_Based_Image_Retrieval/Chinese-CLIP-master/experiments/
2022-12-08,18:15:40 | INFO | Rank 0 | lr: 2e-05
2022-12-08,18:15:40 | INFO | Rank 0 | max_epochs: 16
2022-12-08,18:15:40 | INFO | Rank 0 | max_steps: 41728
2022-12-08,18:15:40 | INFO | Rank 0 | name: muge_finetune_vit-b-16_roberta-base_bs32
2022-12-08,18:15:40 | INFO | Rank 0 | num_workers: 4
2022-12-08,18:15:40 | INFO | Rank 0 | precision: amp
2022-12-08,18:15:40 | INFO | Rank 0 | rank: 0
2022-12-08,18:15:40 | INFO | Rank 0 | report_training_batch_acc: True
2022-12-08,18:15:40 | INFO | Rank 0 | reset_data_offset: True
2022-12-08,18:15:40 | INFO | Rank 0 | reset_optimizer: True
2022-12-08,18:15:40 | INFO | Rank 0 | resume: /storage1/xxx/Text_Based_Image_Retrieval/Chinese-CLIP-master/experiments/pretrained_weights/clip_cn_vit-b-16.pt
2022-12-08,18:15:40 | INFO | Rank 0 | save_epoch_frequency: 1
2022-12-08,18:15:40 | INFO | Rank 0 | save_step_frequency: 999999
2022-12-08,18:15:40 | INFO | Rank 0 | seed: 123
2022-12-08,18:15:40 | INFO | Rank 0 | skip_aggregate: False
2022-12-08,18:15:40 | INFO | Rank 0 | skip_scheduler: False
2022-12-08,18:15:40 | INFO | Rank 0 | text_model: RoBERTa-wwm-ext-base-chinese
2022-12-08,18:15:40 | INFO | Rank 0 | train_data: /storage1/xxx/Text_Based_Image_Retrieval/data/datasets/MUGE/lmdb/train
2022-12-08,18:15:40 | INFO | Rank 0 | use_augment: True
2022-12-08,18:15:40 | INFO | Rank 0 | use_bn_sync: False
2022-12-08,18:15:40 | INFO | Rank 0 | val_data: /storage1/xxx/Text_Based_Image_Retrieval/data/datasets/MUGE/lmdb/valid
2022-12-08,18:15:40 | INFO | Rank 0 | valid_batch_size: 48
2022-12-08,18:15:40 | INFO | Rank 0 | valid_epoch_interval: 1
2022-12-08,18:15:40 | INFO | Rank 0 | valid_step_interval: 150
2022-12-08,18:15:40 | INFO | Rank 0 | vision_model: ViT-B-16
2022-12-08,18:15:40 | INFO | Rank 0 | warmup: 100
2022-12-08,18:15:40 | INFO | Rank 0 | wd: 0.001
2022-12-08,18:15:40 | INFO | Rank 0 | world_size: 2
2022-12-08,18:15:40 | INFO | Rank 0 | Use GPU: 0 for training
2022-12-08,18:15:40 | INFO | Rank 0 | => begin to load checkpoint '/storage1/xxx/Text_Based_Image_Retrieval/Chinese-CLIP-master/experiments/pretrained_weights/clip_cn_vit-b-16.pt'
2022-12-08,18:15:41 | INFO | Rank 0 | => loaded checkpoint '/storage1/xxx/Text_Based_Image_Retrieval/Chinese-CLIP-master/experiments/pretrained_weights/clip_cn_vit-b-16.pt' (epoch 15 @ 0 steps)
2022-12-08,18:15:43 | INFO | Rank 0 | Global Steps: 1/41728 | Train Epoch: 1 [96/250368 (0%)] | Loss: 0.689587 | Image2Text Acc: 82.29 | Text2Image Acc: 81.25 | Data Time: 0.308s | Batch Time: 1.291s | LR: 0.000000 | logit_scale: 4.605 | Global Batch Size: 96
2022-12-08,18:15:43 | INFO | Rank 0 | Reducer buckets have been rebuilt in this iteration.
2022-12-08,18:15:43 | INFO | Rank 0 | Global Steps: 2/41728 | Train Epoch: 1 [192/250368 (0%)] | Loss: 1.083234 | Image2Text Acc: 71.88 | Text2Image Acc: 76.04 | Data Time: 0.014s | Batch Time: 0.739s | LR: 0.000000 | logit_scale: 4.605 | Global Batch Size: 96
2022-12-08,18:15:44 | INFO | Rank 0 | Global Steps: 3/41728 | Train Epoch: 1 [288/250368 (0%)] | Loss: 0.726163 | Image2Text Acc: 78.12 | Text2Image Acc: 78.12 | Data Time: 0.014s | Batch Time: 0.811s | LR: 0.000001 | logit_scale: 4.605 | Global Batch Size: 96
2022-12-08,18:15:45 | INFO | Rank 0 | Global Steps: 4/41728 | Train Epoch: 1 [384/250368 (0%)] | Loss: 0.866006 | Image2Text Acc: 73.96 | Text2Image Acc: 77.08 | Data Time: 0.057s | Batch Time: 0.824s | LR: 0.000001 | logit_scale: 4.605 | Global Batch Size: 96
2022-12-08,18:15:46 | INFO | Rank 0 | Global Steps: 5/41728 | Train Epoch: 1 [480/250368 (0%)] | Loss: 0.849096 | Image2Text Acc: 79.17 | Text2Image Acc: 82.29 | Data Time: 0.062s | Batch Time: 0.836s | LR: 0.000001 | logit_scale: 4.605 | Global Batch Size: 96
2022-12-08,18:15:47 | INFO | Rank 0 | Global Steps: 6/41728 | Train Epoch: 1 [576/250368 (0%)] | Loss: 1.290957 | Image2Text Acc: 62.50 | Text2Image Acc: 66.67 | Data Time: 0.057s | Batch Time: 0.789s | LR: 0.000001 | logit_scale: 4.605 | Global Batch Size: 96
2022-12-08,18:15:47 | INFO | Rank 0 | Global Steps: 7/41728 | Train Epoch: 1 [672/250368 (0%)] | Loss: 0.592875 | Image2Text Acc: 82.29 | Text2Image Acc: 83.33 | Data Time: 0.014s | Batch Time: 0.805s | LR: 0.000001 | logit_scale: 4.605 | Global Batch Size: 96
2022-12-08,18:15:48 | INFO | Rank 0 | Global Steps: 8/41728 | Train Epoch: 1 [768/250368 (0%)] | Loss: 1.060924 | Image2Text Acc: 75.00 | Text2Image Acc: 76.04 | Data Time: 0.057s | Batch Time: 0.830s | LR: 0.000002 | logit_scale: 4.605 | Global Batch Size: 96
2022-12-08,18:15:49 | INFO | Rank 0 | Global Steps: 9/41728 | Train Epoch: 1 [864/250368 (0%)] | Loss: 0.883194 | Image2Text Acc: 78.12 | Text2Image Acc: 78.12 | Data Time: 0.059s | Batch Time: 0.837s | LR: 0.000002 | logit_scale: 4.605 | Global Batch Size: 96
2022-12-08,18:15:50 | INFO | Rank 0 | Global Steps: 10/41728 | Train Epoch: 1 [960/250368 (0%)] | Loss: 0.976302 | Image2Text Acc: 75.00 | Text2Image Acc: 76.04 | Data Time: 0.058s | Batch Time: 0.831s | LR: 0.000002 | logit_scale: 4.605 | Global Batch Size: 96
2022-12-08,18:15:51 | INFO | Rank 0 | Global Steps: 11/41728 | Train Epoch: 1 [1056/250368 (0%)] | Loss: 1.009400 | Image2Text Acc: 71.88 | Text2Image Acc: 75.00 | Data Time: 0.059s | Batch Time: 0.835s | LR: 0.000002 | logit_scale: 4.605 | Global Batch Size: 96
2022-12-08,18:15:52 | INFO | Rank 0 | Global Steps: 12/41728 | Train Epoch: 1 [1152/250368 (0%)] | Loss: 0.968916 | Image2Text Acc: 78.12 | Text2Image Acc: 77.08 | Data Time: 0.059s | Batch Time: 0.830s | LR: 0.000002 | logit_scale: 4.605 | Global Batch Size: 96
2022-12-08,18:15:52 | INFO | Rank 0 | Global Steps: 13/41728 | Train Epoch: 1 [1248/250368 (0%)] | Loss: 0.952544 | Image2Text Acc: 77.08 | Text2Image Acc: 78.12 | Data Time: 0.058s | Batch Time: 0.836s | LR: 0.000003 | logit_scale: 4.605 | Global Batch Size: 96
2022-12-08,18:15:53 | INFO | Rank 0 | Global Steps: 14/41728 | Train Epoch: 1 [1344/250368 (1%)] | Loss: 1.175281 | Image2Text Acc: 75.00 | Text2Image Acc: 75.00 | Data Time: 0.059s | Batch Time: 0.829s | LR: 0.000003 | logit_scale: 4.605 | Global Batch Size: 96
2022-12-08,18:15:54 | INFO | Rank 0 | Global Steps: 15/41728 | Train Epoch: 1 [1440/250368 (1%)] | Loss: 0.807069 | Image2Text Acc: 75.00 | Text2Image Acc: 77.08 | Data Time: 0.059s | Batch Time: 0.832s | LR: 0.000003 | logit_scale: 4.605 | Global Batch Size: 96
2022-12-08,18:15:55 | INFO | Rank 0 | Global Steps: 16/41728 | Train Epoch: 1 [1536/250368 (1%)] | Loss: 1.177227 | Image2Text Acc: 69.79 | Text2Image Acc: 73.96 | Data Time: 0.058s | Batch Time: 0.833s | LR: 0.000003 | logit_scale: 4.605 | Global Batch Size: 96
2022-12-08,18:15:56 | INFO | Rank 0 | Global Steps: 17/41728 | Train Epoch: 1 [1632/250368 (1%)] | Loss: 0.982769 | Image2Text Acc: 68.75 | Text2Image Acc: 76.04 | Data Time: 0.059s | Batch Time: 0.831s | LR: 0.000003 | logit_scale: 4.605 | Global Batch Size: 96
2022-12-08,18:15:57 | INFO | Rank 0 | Global Steps: 18/41728 | Train Epoch: 1 [1728/250368 (1%)] | Loss: 1.051534 | Image2Text Acc: 68.75 | Text2Image Acc: 70.83 | Data Time: 0.059s | Batch Time: 0.832s | LR: 0.000004 | logit_scale: 4.605 | Global Batch Size: 96
2022-12-08,18:15:57 | INFO | Rank 0 | Global Steps: 19/41728 | Train Epoch: 1 [1824/250368 (1%)] | Loss: 0.981971 | Image2Text Acc: 77.08 | Text2Image Acc: 71.88 | Data Time: 0.058s | Batch Time: 0.833s | LR: 0.000004 | logit_scale: 4.605 | Global Batch Size: 96
2022-12-08,18:15:58 | INFO | Rank 0 | Global Steps: 20/41728 | Train Epoch: 1 [1920/250368 (1%)] | Loss: 0.906785 | Image2Text Acc: 72.92 | Text2Image Acc: 78.12 | Data Time: 0.059s | Batch Time: 0.830s | LR: 0.000004 | logit_scale: 4.605 | Global Batch Size: 96
2022-12-08,18:15:59 | INFO | Rank 0 | Global Steps: 21/41728 | Train Epoch: 1 [2016/250368 (1%)] | Loss: 0.401411 | Image2Text Acc: 92.71 | Text2Image Acc: 85.42 | Data Time: 0.058s | Batch Time: 0.833s | LR: 0.000004 | logit_scale: 4.605 | Global Batch Size: 96
2022-12-08,18:16:00 | INFO | Rank 0 | Global Steps: 22/41728 | Train Epoch: 1 [2112/250368 (1%)] | Loss: 0.765108 | Image2Text Acc: 77.08 | Text2Image Acc: 78.12 | Data Time: 0.058s | Batch Time: 0.839s | LR: 0.000004 | logit_scale: 4.605 | Global Batch Size: 96
2022-12-08,18:16:01 | INFO | Rank 0 | Global Steps: 23/41728 | Train Epoch: 1 [2208/250368 (1%)] | Loss: 0.725298 | Image2Text Acc: 81.25 | Text2Image Acc: 79.17 | Data Time: 0.059s | Batch Time: 0.839s | LR: 0.000005 | logit_scale: 4.605 | Global Batch Size: 96
2022-12-08,18:16:02 | INFO | Rank 0 | Global Steps: 24/41728 | Train Epoch: 1 [2304/250368 (1%)] | Loss: 1.053250 | Image2Text Acc: 71.88 | Text2Image Acc: 71.88 | Data Time: 0.059s | Batch Time: 0.833s | LR: 0.000005 | logit_scale: 4.605 | Global Batch Size: 96
2022-12-08,18:16:02 | INFO | Rank 0 | Global Steps: 25/41728 | Train Epoch: 1 [2400/250368 (1%)] | Loss: 0.892181 | Image2Text Acc: 71.88 | Text2Image Acc: 80.21 | Data Time: 0.059s | Batch Time: 0.833s | LR: 0.000005 | logit_scale: 4.605 | Global Batch Size: 96
2022-12-08,18:16:03 | INFO | Rank 0 | Global Steps: 26/41728 | Train Epoch: 1 [2496/250368 (1%)] | Loss: 0.863705 | Image2Text Acc: 72.92 | Text2Image Acc: 82.29 | Data Time: 0.058s | Batch Time: 0.834s | LR: 0.000005 | logit_scale: 4.605 | Global Batch Size: 96
2022-12-08,18:16:04 | INFO | Rank 0 | Global Steps: 27/41728 | Train Epoch: 1 [2592/250368 (1%)] | Loss: 1.011039 | Image2Text Acc: 75.00 | Text2Image Acc: 73.96 | Data Time: 0.059s | Batch Time: 0.833s | LR: 0.000005 | logit_scale: 4.605 | Global Batch Size: 96
2022-12-08,18:16:05 | INFO | Rank 0 | Global Steps: 28/41728 | Train Epoch: 1 [2688/250368 (1%)] | Loss: 0.954087 | Image2Text Acc: 73.96 | Text2Image Acc: 78.12 | Data Time: 0.058s | Batch Time: 0.834s | LR: 0.000006 | logit_scale: 4.605 | Global Batch Size: 96
2022-12-08,18:16:06 | INFO | Rank 0 | Global Steps: 29/41728 | Train Epoch: 1 [2784/250368 (1%)] | Loss: 0.826310 | Image2Text Acc: 75.00 | Text2Image Acc: 71.88 | Data Time: 0.058s | Batch Time: 0.833s | LR: 0.000006 | logit_scale: 4.605 | Global Batch Size: 96
2022-12-08,18:16:07 | INFO | Rank 0 | Global Steps: 30/41728 | Train Epoch: 1 [2880/250368 (1%)] | Loss: 0.941817 | Image2Text Acc: 77.08 | Text2Image Acc: 73.96 | Data Time: 0.058s | Batch Time: 0.837s | LR: 0.000006 | logit_scale: 4.605 | Global Batch Size: 96
2022-12-08,18:16:07 | INFO | Rank 0 | Global Steps: 31/41728 | Train Epoch: 1 [2976/250368 (1%)] | Loss: 0.758239 | Image2Text Acc: 76.04 | Text2Image Acc: 75.00 | Data Time: 0.058s | Batch Time: 0.841s | LR: 0.000006 | logit_scale: 4.605 | Global Batch Size: 96
2022-12-08,18:16:08 | INFO | Rank 0 | Global Steps: 32/41728 | Train Epoch: 1 [3072/250368 (1%)] | Loss: 0.736244 | Image2Text Acc: 76.04 | Text2Image Acc: 71.88 | Data Time: 0.058s | Batch Time: 0.833s | LR: 0.000006 | logit_scale: 4.605 | Global Batch Size: 96
2022-12-08,18:16:09 | INFO | Rank 0 | Global Steps: 33/41728 | Train Epoch: 1 [3168/250368 (1%)] | Loss: 0.652642 | Image2Text Acc: 78.12 | Text2Image Acc: 81.25 | Data Time: 0.058s | Batch Time: 0.839s | LR: 0.000007 | logit_scale: 4.605 | Global Batch Size: 96
2022-12-08,18:16:10 | INFO | Rank 0 | Global Steps: 34/41728 | Train Epoch: 1 [3264/250368 (1%)] | Loss: 0.621304 | Image2Text Acc: 82.29 | Text2Image Acc: 83.33 | Data Time: 0.059s | Batch Time: 0.840s | LR: 0.000007 | logit_scale: 4.605 | Global Batch Size: 96
2022-12-08,18:16:11 | INFO | Rank 0 | Global Steps: 35/41728 | Train Epoch: 1 [3360/250368 (1%)] | Loss: 0.860351 | Image2Text Acc: 78.12 | Text2Image Acc: 79.17 | Data Time: 0.059s | Batch Time: 0.841s | LR: 0.000007 | logit_scale: 4.605 | Global Batch Size: 96
2022-12-08,18:16:12 | INFO | Rank 0 | Global Steps: 36/41728 | Train Epoch: 1 [3456/250368 (1%)] | Loss: 0.715295 | Image2Text Acc: 81.25 | Text2Image Acc: 78.12 | Data Time: 0.058s | Batch Time: 0.836s | LR: 0.000007 | logit_scale: 4.605 | Global Batch Size: 96
2022-12-08,18:16:12 | INFO | Rank 0 | Global Steps: 37/41728 | Train Epoch: 1 [3552/250368 (1%)] | Loss: 0.674997 | Image2Text Acc: 80.21 | Text2Image Acc: 84.38 | Data Time: 0.058s | Batch Time: 0.836s | LR: 0.000007 | logit_scale: 4.605 | Global Batch Size: 96
2022-12-08,18:16:13 | INFO | Rank 0 | Global Steps: 38/41728 | Train Epoch: 1 [3648/250368 (1%)] | Loss: 0.958120 | Image2Text Acc: 72.92 | Text2Image Acc: 73.96 | Data Time: 0.058s | Batch Time: 0.836s | LR: 0.000008 | logit_scale: 4.605 | Global Batch Size: 96
2022-12-08,18:16:14 | INFO | Rank 0 | Global Steps: 39/41728 | Train Epoch: 1 [3744/250368 (1%)] | Loss: 1.020398 | Image2Text Acc: 71.88 | Text2Image Acc: 75.00 | Data Time: 0.058s | Batch Time: 0.841s | LR: 0.000008 | logit_scale: 4.605 | Global Batch Size: 96
2022-12-08,18:16:15 | INFO | Rank 0 | Global Steps: 40/41728 | Train Epoch: 1 [3840/250368 (2%)] | Loss: 0.860142 | Image2Text Acc: 71.88 | Text2Image Acc: 69.79 | Data Time: 0.058s | Batch Time: 0.832s | LR: 0.000008 | logit_scale: 4.605 | Global Batch Size: 96
2022-12-08,18:16:16 | INFO | Rank 0 | Global Steps: 41/41728 | Train Epoch: 1 [3936/250368 (2%)] | Loss: 0.591362 | Image2Text Acc: 82.29 | Text2Image Acc: 84.38 | Data Time: 0.058s | Batch Time: 0.836s | LR: 0.000008 | logit_scale: 4.605 | Global Batch Size: 96
2022-12-08,18:16:17 | INFO | Rank 0 | Global Steps: 42/41728 | Train Epoch: 1 [4032/250368 (2%)] | Loss: 0.677259 | Image2Text Acc: 76.04 | Text2Image Acc: 80.21 | Data Time: 0.058s | Batch Time: 0.836s | LR: 0.000008 | logit_scale: 4.605 | Global Batch Size: 96
2022-12-08,18:16:17 | INFO | Rank 0 | Global Steps: 43/41728 | Train Epoch: 1 [4128/250368 (2%)] | Loss: 0.857192 | Image2Text Acc: 76.04 | Text2Image Acc: 75.00 | Data Time: 0.058s | Batch Time: 0.837s | LR: 0.000009 | logit_scale: 4.605 | Global Batch Size: 96
2022-12-08,18:16:18 | INFO | Rank 0 | Global Steps: 44/41728 | Train Epoch: 1 [4224/250368 (2%)] | Loss: 0.734577 | Image2Text Acc: 80.21 | Text2Image Acc: 84.38 | Data Time: 0.058s | Batch Time: 0.842s | LR: 0.000009 | logit_scale: 4.605 | Global Batch Size: 96
2022-12-08,18:16:19 | INFO | Rank 0 | Global Steps: 45/41728 | Train Epoch: 1 [4320/250368 (2%)] | Loss: 0.731063 | Image2Text Acc: 80.21 | Text2Image Acc: 78.12 | Data Time: 0.058s | Batch Time: 0.835s | LR: 0.000009 | logit_scale: 4.605 | Global Batch Size: 96
2022-12-08,18:16:20 | INFO | Rank 0 | Global Steps: 46/41728 | Train Epoch: 1 [4416/250368 (2%)] | Loss: 0.641702 | Image2Text Acc: 81.25 | Text2Image Acc: 83.33 | Data Time: 0.058s | Batch Time: 0.839s | LR: 0.000009 | logit_scale: 4.605 | Global Batch Size: 96
2022-12-08,18:16:21 | INFO | Rank 0 | Global Steps: 47/41728 | Train Epoch: 1 [4512/250368 (2%)] | Loss: 0.874175 | Image2Text Acc: 73.96 | Text2Image Acc: 76.04 | Data Time: 0.058s | Batch Time: 0.840s | LR: 0.000009 | logit_scale: 4.605 | Global Batch Size: 96
2022-12-08,18:16:22 | INFO | Rank 0 | Global Steps: 48/41728 | Train Epoch: 1 [4608/250368 (2%)] | Loss: 0.808500 | Image2Text Acc: 79.17 | Text2Image Acc: 81.25 | Data Time: 0.058s | Batch Time: 0.838s | LR: 0.000010 | logit_scale: 4.605 | Global Batch Size: 96
2022-12-08,18:16:22 | INFO | Rank 0 | Global Steps: 49/41728 | Train Epoch: 1 [4704/250368 (2%)] | Loss: 0.682884 | Image2Text Acc: 78.12 | Text2Image Acc: 80.21 | Data Time: 0.059s | Batch Time: 0.841s | LR: 0.000010 | logit_scale: 4.605 | Global Batch Size: 96
2022-12-08,18:16:23 | INFO | Rank 0 | Global Steps: 50/41728 | Train Epoch: 1 [4800/250368 (2%)] | Loss: 0.991166 | Image2Text Acc: 65.62 | Text2Image Acc: 68.75 | Data Time: 0.058s | Batch Time: 0.834s | LR: 0.000010 | logit_scale: 4.605 | Global Batch Size: 96
2022-12-08,18:16:24 | INFO | Rank 0 | Global Steps: 51/41728 | Train Epoch: 1 [4896/250368 (2%)] | Loss: 0.666178 | Image2Text Acc: 80.21 | Text2Image Acc: 81.25 | Data Time: 0.058s | Batch Time: 0.841s | LR: 0.000010 | logit_scale: 4.605 | Global Batch Size: 96
2022-12-08,18:16:25 | INFO | Rank 0 | Global Steps: 52/41728 | Train Epoch: 1 [4992/250368 (2%)] | Loss: 0.741099 | Image2Text Acc: 80.21 | Text2Image Acc: 81.25 | Data Time: 0.059s | Batch Time: 0.838s | LR: 0.000010 | logit_scale: 4.605 | Global Batch Size: 96
2022-12-08,18:16:26 | INFO | Rank 0 | Global Steps: 53/41728 | Train Epoch: 1 [5088/250368 (2%)] | Loss: 0.617694 | Image2Text Acc: 79.17 | Text2Image Acc: 83.33 | Data Time: 0.059s | Batch Time: 0.837s | LR: 0.000011 | logit_scale: 4.605 | Global Batch Size: 96
2022-12-08,18:16:27 | INFO | Rank 0 | Global Steps: 54/41728 | Train Epoch: 1 [5184/250368 (2%)] | Loss: 0.492508 | Image2Text Acc: 85.42 | Text2Image Acc: 86.46 | Data Time: 0.058s | Batch Time: 0.841s | LR: 0.000011 | logit_scale: 4.605 | Global Batch Size: 96
2022-12-08,18:16:28 | INFO | Rank 0 | Global Steps: 55/41728 | Train Epoch: 1 [5280/250368 (2%)] | Loss: 0.768075 | Image2Text Acc: 82.29 | Text2Image Acc: 79.17 | Data Time: 0.058s | Batch Time: 0.841s | LR: 0.000011 | logit_scale: 4.605 | Global Batch Size: 96
2022-12-08,18:16:28 | INFO | Rank 0 | Global Steps: 56/41728 | Train Epoch: 1 [5376/250368 (2%)] | Loss: 0.469336 | Image2Text Acc: 86.46 | Text2Image Acc: 86.46 | Data Time: 0.058s | Batch Time: 0.834s | LR: 0.000011 | logit_scale: 4.605 | Global Batch Size: 96
2022-12-08,18:16:29 | INFO | Rank 0 | Global Steps: 57/41728 | Train Epoch: 1 [5472/250368 (2%)] | Loss: 0.766508 | Image2Text Acc: 79.17 | Text2Image Acc: 76.04 | Data Time: 0.059s | Batch Time: 0.835s | LR: 0.000011 | logit_scale: 4.605 | Global Batch Size: 96
2022-12-08,18:16:30 | INFO | Rank 0 | Global Steps: 58/41728 | Train Epoch: 1 [5568/250368 (2%)] | Loss: 0.836963 | Image2Text Acc: 70.83 | Text2Image Acc: 82.29 | Data Time: 0.058s | Batch Time: 0.844s | LR: 0.000012 | logit_scale: 4.605 | Global Batch Size: 96
2022-12-08,18:16:31 | INFO | Rank 0 | Global Steps: 59/41728 | Train Epoch: 1 [5664/250368 (2%)] | Loss: 0.662819 | Image2Text Acc: 82.29 | Text2Image Acc: 76.04 | Data Time: 0.058s | Batch Time: 0.832s | LR: 0.000012 | logit_scale: 4.605 | Global Batch Size: 96
2022-12-08,18:16:32 | INFO | Rank 0 | Global Steps: 60/41728 | Train Epoch: 1 [5760/250368 (2%)] | Loss: 0.831568 | Image2Text Acc: 77.08 | Text2Image Acc: 77.08 | Data Time: 0.058s | Batch Time: 0.842s | LR: 0.000012 | logit_scale: 4.605 | Global Batch Size: 96
2022-12-08,18:16:33 | INFO | Rank 0 | Global Steps: 61/41728 | Train Epoch: 1 [5856/250368 (2%)] | Loss: 0.987221 | Image2Text Acc: 76.04 | Text2Image Acc: 77.08 | Data Time: 0.059s | Batch Time: 0.838s | LR: 0.000012 | logit_scale: 4.605 | Global Batch Size: 96
2022-12-08,18:16:33 | INFO | Rank 0 | Global Steps: 62/41728 | Train Epoch: 1 [5952/250368 (2%)] | Loss: 0.849950 | Image2Text Acc: 75.00 | Text2Image Acc: 76.04 | Data Time: 0.059s | Batch Time: 0.840s | LR: 0.000012 | logit_scale: 4.605 | Global Batch Size: 96
2022-12-08,18:16:34 | INFO | Rank 0 | Global Steps: 63/41728 | Train Epoch: 1 [6048/250368 (2%)] | Loss: 0.703894 | Image2Text Acc: 79.17 | Text2Image Acc: 82.29 | Data Time: 0.058s | Batch Time: 0.835s | LR: 0.000013 | logit_scale: 4.605 | Global Batch Size: 96
2022-12-08,18:16:35 | INFO | Rank 0 | Global Steps: 64/41728 | Train Epoch: 1 [6144/250368 (2%)] | Loss: 0.706505 | Image2Text Acc: 81.25 | Text2Image Acc: 80.21 | Data Time: 0.058s | Batch Time: 0.840s | LR: 0.000013 | logit_scale: 4.605 | Global Batch Size: 96
2022-12-08,18:16:36 | INFO | Rank 0 | Global Steps: 65/41728 | Train Epoch: 1 [6240/250368 (2%)] | Loss: 0.719888 | Image2Text Acc: 82.29 | Text2Image Acc: 81.25 | Data Time: 0.059s | Batch Time: 0.837s | LR: 0.000013 | logit_scale: 4.605 | Global Batch Size: 96
2022-12-08,18:16:37 | INFO | Rank 0 | Global Steps: 66/41728 | Train Epoch: 1 [6336/250368 (3%)] | Loss: 0.777751 | Image2Text Acc: 75.00 | Text2Image Acc: 76.04 | Data Time: 0.059s | Batch Time: 0.836s | LR: 0.000013 | logit_scale: 4.605 | Global Batch Size: 96
2022-12-08,18:16:38 | INFO | Rank 0 | Global Steps: 67/41728 | Train Epoch: 1 [6432/250368 (3%)] | Loss: 0.822757 | Image2Text Acc: 75.00 | Text2Image Acc: 75.00 | Data Time: 0.058s | Batch Time: 0.839s | LR: 0.000013 | logit_scale: 4.605 | Global Batch Size: 96
2022-12-08,18:16:38 | INFO | Rank 0 | Global Steps: 68/41728 | Train Epoch: 1 [6528/250368 (3%)] | Loss: 0.637543 | Image2Text Acc: 79.17 | Text2Image Acc: 81.25 | Data Time: 0.059s | Batch Time: 0.842s | LR: 0.000014 | logit_scale: 4.605 | Global Batch Size: 96
2022-12-08,18:16:39 | INFO | Rank 0 | Global Steps: 69/41728 | Train Epoch: 1 [6624/250368 (3%)] | Loss: 0.503515 | Image2Text Acc: 83.33 | Text2Image Acc: 83.33 | Data Time: 0.059s | Batch Time: 0.837s | LR: 0.000014 | logit_scale: 4.605 | Global Batch Size: 96
2022-12-08,18:16:40 | INFO | Rank 0 | Global Steps: 70/41728 | Train Epoch: 1 [6720/250368 (3%)] | Loss: 0.861151 | Image2Text Acc: 76.04 | Text2Image Acc: 70.83 | Data Time: 0.059s | Batch Time: 0.831s | LR: 0.000014 | logit_scale: 4.605 | Global Batch Size: 96
2022-12-08,18:16:41 | INFO | Rank 0 | Global Steps: 71/41728 | Train Epoch: 1 [6816/250368 (3%)] | Loss: 0.842519 | Image2Text Acc: 75.00 | Text2Image Acc: 75.00 | Data Time: 0.058s | Batch Time: 0.838s | LR: 0.000014 | logit_scale: 4.605 | Global Batch Size: 96
2022-12-08,18:16:42 | INFO | Rank 0 | Global Steps: 72/41728 | Train Epoch: 1 [6912/250368 (3%)] | Loss: 0.647028 | Image2Text Acc: 80.21 | Text2Image Acc: 80.21 | Data Time: 0.059s | Batch Time: 0.840s | LR: 0.000014 | logit_scale: 4.605 | Global Batch Size: 96
2022-12-08,18:16:43 | INFO | Rank 0 | Global Steps: 73/41728 | Train Epoch: 1 [7008/250368 (3%)] | Loss: 0.533600 | Image2Text Acc: 79.17 | Text2Image Acc: 82.29 | Data Time: 0.059s | Batch Time: 0.835s | LR: 0.000015 | logit_scale: 4.605 | Global Batch Size: 96
2022-12-08,18:16:43 | INFO | Rank 0 | Global Steps: 74/41728 | Train Epoch: 1 [7104/250368 (3%)] | Loss: 0.602674 | Image2Text Acc: 83.33 | Text2Image Acc: 81.25 | Data Time: 0.062s | Batch Time: 0.841s | LR: 0.000015 | logit_scale: 4.605 | Global Batch Size: 96
2022-12-08,18:16:44 | INFO | Rank 0 | Global Steps: 75/41728 | Train Epoch: 1 [7200/250368 (3%)] | Loss: 0.748078 | Image2Text Acc: 77.08 | Text2Image Acc: 83.33 | Data Time: 0.058s | Batch Time: 0.841s | LR: 0.000015 | logit_scale: 4.605 | Global Batch Size: 96
2022-12-08,18:16:45 | INFO | Rank 0 | Global Steps: 76/41728 | Train Epoch: 1 [7296/250368 (3%)] | Loss: 1.266804 | Image2Text Acc: 75.00 | Text2Image Acc: 73.96 | Data Time: 0.059s | Batch Time: 0.838s | LR: 0.000015 | logit_scale: 4.605 | Global Batch Size: 96
2022-12-08,18:16:46 | INFO | Rank 0 | Global Steps: 77/41728 | Train Epoch: 1 [7392/250368 (3%)] | Loss: 0.701536 | Image2Text Acc: 79.17 | Text2Image Acc: 80.21 | Data Time: 0.059s | Batch Time: 0.838s | LR: 0.000015 | logit_scale: 4.605 | Global Batch Size: 96
2022-12-08,18:16:47 | INFO | Rank 0 | Global Steps: 78/41728 | Train Epoch: 1 [7488/250368 (3%)] | Loss: 0.640760 | Image2Text Acc: 80.21 | Text2Image Acc: 81.25 | Data Time: 0.059s | Batch Time: 0.841s | LR: 0.000016 | logit_scale: 4.605 | Global Batch Size: 96
2022-12-08,18:16:48 | INFO | Rank 0 | Global Steps: 79/41728 | Train Epoch: 1 [7584/250368 (3%)] | Loss: 0.555545 | Image2Text Acc: 82.29 | Text2Image Acc: 89.58 | Data Time: 0.059s | Batch Time: 0.836s | LR: 0.000016 | logit_scale: 4.605 | Global Batch Size: 96
2022-12-08,18:16:48 | INFO | Rank 0 | Global Steps: 80/41728 | Train Epoch: 1 [7680/250368 (3%)] | Loss: 0.600125 | Image2Text Acc: 84.38 | Text2Image Acc: 83.33 | Data Time: 0.059s | Batch Time: 0.836s | LR: 0.000016 | logit_scale: 4.605 | Global Batch Size: 96
2022-12-08,18:16:49 | INFO | Rank 0 | Global Steps: 81/41728 | Train Epoch: 1 [7776/250368 (3%)] | Loss: 0.811505 | Image2Text Acc: 76.04 | Text2Image Acc: 78.12 | Data Time: 0.059s | Batch Time: 0.844s | LR: 0.000016 | logit_scale: 4.605 | Global Batch Size: 96
2022-12-08,18:16:50 | INFO | Rank 0 | Global Steps: 82/41728 | Train Epoch: 1 [7872/250368 (3%)] | Loss: 0.555656 | Image2Text Acc: 79.17 | Text2Image Acc: 86.46 | Data Time: 0.059s | Batch Time: 0.833s | LR: 0.000016 | logit_scale: 4.605 | Global Batch Size: 96
2022-12-08,18:16:51 | INFO | Rank 0 | Global Steps: 83/41728 | Train Epoch: 1 [7968/250368 (3%)] | Loss: 0.585370 | Image2Text Acc: 84.38 | Text2Image Acc: 81.25 | Data Time: 0.058s | Batch Time: 0.838s | LR: 0.000017 | logit_scale: 4.605 | Global Batch Size: 96
2022-12-08,18:16:52 | INFO | Rank 0 | Global Steps: 84/41728 | Train Epoch: 1 [8064/250368 (3%)] | Loss: 0.584991 | Image2Text Acc: 86.46 | Text2Image Acc: 82.29 | Data Time: 0.059s | Batch Time: 0.844s | LR: 0.000017 | logit_scale: 4.605 | Global Batch Size: 96
2022-12-08,18:16:53 | INFO | Rank 0 | Global Steps: 85/41728 | Train Epoch: 1 [8160/250368 (3%)] | Loss: 0.515017 | Image2Text Acc: 86.46 | Text2Image Acc: 86.46 | Data Time: 0.059s | Batch Time: 0.840s | LR: 0.000017 | logit_scale: 4.605 | Global Batch Size: 96
2022-12-08,18:16:53 | INFO | Rank 0 | Global Steps: 86/41728 | Train Epoch: 1 [8256/250368 (3%)] | Loss: 0.865349 | Image2Text Acc: 71.88 | Text2Image Acc: 78.12 | Data Time: 0.059s | Batch Time: 0.839s | LR: 0.000017 | logit_scale: 4.605 | Global Batch Size: 96
2022-12-08,18:16:54 | INFO | Rank 0 | Global Steps: 87/41728 | Train Epoch: 1 [8352/250368 (3%)] | Loss: 0.640673 | Image2Text Acc: 79.17 | Text2Image Acc: 79.17 | Data Time: 0.059s | Batch Time: 0.847s | LR: 0.000017 | logit_scale: 4.605 | Global Batch Size: 96
2022-12-08,18:16:55 | INFO | Rank 0 | Global Steps: 88/41728 | Train Epoch: 1 [8448/250368 (3%)] | Loss: 0.806827 | Image2Text Acc: 71.88 | Text2Image Acc: 80.21 | Data Time: 0.059s | Batch Time: 0.838s | LR: 0.000018 | logit_scale: 4.605 | Global Batch Size: 96
2022-12-08,18:16:56 | INFO | Rank 0 | Global Steps: 89/41728 | Train Epoch: 1 [8544/250368 (3%)] | Loss: 0.757677 | Image2Text Acc: 75.00 | Text2Image Acc: 76.04 | Data Time: 0.059s | Batch Time: 0.840s | LR: 0.000018 | logit_scale: 4.605 | Global Batch Size: 96
2022-12-08,18:16:57 | INFO | Rank 0 | Global Steps: 90/41728 | Train Epoch: 1 [8640/250368 (3%)] | Loss: 0.518550 | Image2Text Acc: 84.38 | Text2Image Acc: 86.46 | Data Time: 0.059s | Batch Time: 0.833s | LR: 0.000018 | logit_scale: 4.605 | Global Batch Size: 96
2022-12-08,18:16:58 | INFO | Rank 0 | Global Steps: 91/41728 | Train Epoch: 1 [8736/250368 (3%)] | Loss: 0.763695 | Image2Text Acc: 77.08 | Text2Image Acc: 79.17 | Data Time: 0.058s | Batch Time: 0.838s | LR: 0.000018 | logit_scale: 4.605 | Global Batch Size: 96
2022-12-08,18:16:59 | INFO | Rank 0 | Global Steps: 92/41728 | Train Epoch: 1 [8832/250368 (4%)] | Loss: 0.794845 | Image2Text Acc: 73.96 | Text2Image Acc: 76.04 | Data Time: 0.058s | Batch Time: 0.838s | LR: 0.000018 | logit_scale: 4.605 | Global Batch Size: 96
2022-12-08,18:16:59 | INFO | Rank 0 | Global Steps: 93/41728 | Train Epoch: 1 [8928/250368 (4%)] | Loss: 0.611818 | Image2Text Acc: 76.04 | Text2Image Acc: 79.17 | Data Time: 0.058s | Batch Time: 0.836s | LR: 0.000019 | logit_scale: 4.605 | Global Batch Size: 96
2022-12-08,18:17:00 | INFO | Rank 0 | Global Steps: 94/41728 | Train Epoch: 1 [9024/250368 (4%)] | Loss: 0.778732 | Image2Text Acc: 79.17 | Text2Image Acc: 78.12 | Data Time: 0.059s | Batch Time: 0.839s | LR: 0.000019 | logit_scale: 4.605 | Global Batch Size: 96
2022-12-08,18:17:01 | INFO | Rank 0 | Global Steps: 95/41728 | Train Epoch: 1 [9120/250368 (4%)] | Loss: 1.208666 | Image2Text Acc: 69.79 | Text2Image Acc: 68.75 | Data Time: 0.059s | Batch Time: 0.840s | LR: 0.000019 | logit_scale: 4.605 | Global Batch Size: 96
2022-12-08,18:17:02 | INFO | Rank 0 | Global Steps: 96/41728 | Train Epoch: 1 [9216/250368 (4%)] | Loss: 0.822677 | Image2Text Acc: 75.00 | Text2Image Acc: 80.21 | Data Time: 0.058s | Batch Time: 0.834s | LR: 0.000019 | logit_scale: 4.605 | Global Batch Size: 96
2022-12-08,18:17:03 | INFO | Rank 0 | Global Steps: 97/41728 | Train Epoch: 1 [9312/250368 (4%)] | Loss: 1.062310 | Image2Text Acc: 72.92 | Text2Image Acc: 68.75 | Data Time: 0.059s | Batch Time: 0.837s | LR: 0.000019 | logit_scale: 4.605 | Global Batch Size: 96
2022-12-08,18:17:04 | INFO | Rank 0 | Global Steps: 98/41728 | Train Epoch: 1 [9408/250368 (4%)] | Loss: 0.967333 | Image2Text Acc: 76.04 | Text2Image Acc: 78.12 | Data Time: 0.059s | Batch Time: 0.843s | LR: 0.000020 | logit_scale: 4.605 | Global Batch Size: 96
2022-12-08,18:17:04 | INFO | Rank 0 | Global Steps: 99/41728 | Train Epoch: 1 [9504/250368 (4%)] | Loss: 0.582876 | Image2Text Acc: 82.29 | Text2Image Acc: 81.25 | Data Time: 0.058s | Batch Time: 0.840s | LR: 0.000020 | logit_scale: 4.605 | Global Batch Size: 96
2022-12-08,18:17:05 | INFO | Rank 0 | Global Steps: 100/41728 | Train Epoch: 1 [9600/250368 (4%)] | Loss: 0.706923 | Image2Text Acc: 77.08 | Text2Image Acc: 81.25 | Data Time: 0.059s | Batch Time: 0.831s | LR: 0.000020 | logit_scale: 4.605 | Global Batch Size: 96
2022-12-08,18:17:06 | INFO | Rank 0 | Global Steps: 101/41728 | Train Epoch: 1 [9696/250368 (4%)] | Loss: 0.872719 | Image2Text Acc: 76.04 | Text2Image Acc: 78.12 | Data Time: 0.059s | Batch Time: 0.839s | LR: 0.000020 | logit_scale: 4.605 | Global Batch Size: 96
2022-12-08,18:17:07 | INFO | Rank 0 | Global Steps: 102/41728 | Train Epoch: 1 [9792/250368 (4%)] | Loss: 0.717685 | Image2Text Acc: 73.96 | Text2Image Acc: 76.04 | Data Time: 0.059s | Batch Time: 0.839s | LR: 0.000020 | logit_scale: 4.605 | Global Batch Size: 96
2022-12-08,18:17:08 | INFO | Rank 0 | Global Steps: 103/41728 | Train Epoch: 1 [9888/250368 (4%)] | Loss: 0.957555 | Image2Text Acc: 73.96 | Text2Image Acc: 71.88 | Data Time: 0.059s | Batch Time: 0.842s | LR: 0.000020 | logit_scale: 4.605 | Global Batch Size: 96
2022-12-08,18:17:09 | INFO | Rank 0 | Global Steps: 104/41728 | Train Epoch: 1 [9984/250368 (4%)] | Loss: 0.609131 | Image2Text Acc: 80.21 | Text2Image Acc: 80.21 | Data Time: 0.059s | Batch Time: 0.836s | LR: 0.000020 | logit_scale: 4.605 | Global Batch Size: 96
2022-12-08,18:17:09 | INFO | Rank 0 | Global Steps: 105/41728 | Train Epoch: 1 [10080/250368 (4%)] | Loss: 0.697119 | Image2Text Acc: 75.00 | Text2Image Acc: 82.29 | Data Time: 0.059s | Batch Time: 0.841s | LR: 0.000020 | logit_scale: 4.605 | Global Batch Size: 96
2022-12-08,18:17:10 | INFO | Rank 0 | Global Steps: 106/41728 | Train Epoch: 1 [10176/250368 (4%)] | Loss: 1.170625 | Image2Text Acc: 70.83 | Text2Image Acc: 66.67 | Data Time: 0.059s | Batch Time: 0.840s | LR: 0.000020 | logit_scale: 4.605 | Global Batch Size: 96
2022-12-08,18:17:11 | INFO | Rank 0 | Global Steps: 107/41728 | Train Epoch: 1 [10272/250368 (4%)] | Loss: 0.866036 | Image2Text Acc: 75.00 | Text2Image Acc: 80.21 | Data Time: 0.059s | Batch Time: 0.838s | LR: 0.000020 | logit_scale: 4.605 | Global Batch Size: 96
2022-12-08,18:17:12 | INFO | Rank 0 | Global Steps: 108/41728 | Train Epoch: 1 [10368/250368 (4%)] | Loss: 0.657084 | Image2Text Acc: 81.25 | Text2Image Acc: 81.25 | Data Time: 0.059s | Batch Time: 0.840s | LR: 0.000020 | logit_scale: 4.605 | Global Batch Size: 96
2022-12-08,18:17:13 | INFO | Rank 0 | Global Steps: 109/41728 | Train Epoch: 1 [10464/250368 (4%)] | Loss: 0.803462 | Image2Text Acc: 77.08 | Text2Image Acc: 77.08 | Data Time: 0.058s | Batch Time: 0.842s | LR: 0.000020 | logit_scale: 4.605 | Global Batch Size: 96
2022-12-08,18:17:14 | INFO | Rank 0 | Global Steps: 110/41728 | Train Epoch: 1 [10560/250368 (4%)] | Loss: 0.991874 | Image2Text Acc: 77.08 | Text2Image Acc: 72.92 | Data Time: 0.059s | Batch Time: 0.837s | LR: 0.000020 | logit_scale: 4.605 | Global Batch Size: 96
2022-12-08,18:17:14 | INFO | Rank 0 | Global Steps: 111/41728 | Train Epoch: 1 [10656/250368 (4%)] | Loss: 0.752720 | Image2Text Acc: 78.12 | Text2Image Acc: 79.17 | Data Time: 0.059s | Batch Time: 0.844s | LR: 0.000020 | logit_scale: 4.605 | Global Batch Size: 96
2022-12-08,18:17:15 | INFO | Rank 0 | Global Steps: 112/41728 | Train Epoch: 1 [10752/250368 (4%)] | Loss: 0.643739 | Image2Text Acc: 85.42 | Text2Image Acc: 83.33 | Data Time: 0.059s | Batch Time: 0.832s | LR: 0.000020 | logit_scale: 4.605 | Global Batch Size: 96
2022-12-08,18:17:16 | INFO | Rank 0 | Global Steps: 113/41728 | Train Epoch: 1 [10848/250368 (4%)] | Loss: 0.738462 | Image2Text Acc: 82.29 | Text2Image Acc: 76.04 | Data Time: 0.059s | Batch Time: 0.839s | LR: 0.000020 | logit_scale: 4.605 | Global Batch Size: 96
2022-12-08,18:17:17 | INFO | Rank 0 | Global Steps: 114/41728 | Train Epoch: 1 [10944/250368 (4%)] | Loss: 0.594122 | Image2Text Acc: 80.21 | Text2Image Acc: 84.38 | Data Time: 0.059s | Batch Time: 0.840s | LR: 0.000020 | logit_scale: 4.605 | Global Batch Size: 96
2022-12-08,18:17:18 | INFO | Rank 0 | Global Steps: 115/41728 | Train Epoch: 1 [11040/250368 (4%)] | Loss: 1.017626 | Image2Text Acc: 73.96 | Text2Image Acc: 76.04 | Data Time: 0.059s | Batch Time: 0.841s | LR: 0.000020 | logit_scale: 4.605 | Global Batch Size: 96
2022-12-08,18:17:19 | INFO | Rank 0 | Global Steps: 116/41728 | Train Epoch: 1 [11136/250368 (4%)] | Loss: 0.676917 | Image2Text Acc: 80.21 | Text2Image Acc: 81.25 | Data Time: 0.059s | Batch Time: 0.833s | LR: 0.000020 | logit_scale: 4.605 | Global Batch Size: 96
2022-12-08,18:17:19 | INFO | Rank 0 | Global Steps: 117/41728 | Train Epoch: 1 [11232/250368 (4%)] | Loss: 0.659712 | Image2Text Acc: 82.29 | Text2Image Acc: 80.21 | Data Time: 0.059s | Batch Time: 0.836s | LR: 0.000020 | logit_scale: 4.605 | Global Batch Size: 96
2022-12-08,18:17:20 | INFO | Rank 0 | Global Steps: 118/41728 | Train Epoch: 1 [11328/250368 (5%)] | Loss: 0.583284 | Image2Text Acc: 81.25 | Text2Image Acc: 82.29 | Data Time: 0.059s | Batch Time: 0.839s | LR: 0.000020 | logit_scale: 4.605 | Global Batch Size: 96
2022-12-08,18:17:21 | INFO | Rank 0 | Global Steps: 119/41728 | Train Epoch: 1 [11424/250368 (5%)] | Loss: 0.683597 | Image2Text Acc: 77.08 | Text2Image Acc: 81.25 | Data Time: 0.059s | Batch Time: 0.840s | LR: 0.000020 | logit_scale: 4.605 | Global Batch Size: 96
2022-12-08,18:17:22 | INFO | Rank 0 | Global Steps: 120/41728 | Train Epoch: 1 [11520/250368 (5%)] | Loss: 0.905988 | Image2Text Acc: 72.92 | Text2Image Acc: 73.96 | Data Time: 0.059s | Batch Time: 0.845s | LR: 0.000020 | logit_scale: 4.605 | Global Batch Size: 96
2022-12-08,18:17:23 | INFO | Rank 0 | Global Steps: 121/41728 | Train Epoch: 1 [11616/250368 (5%)] | Loss: 0.906386 | Image2Text Acc: 72.92 | Text2Image Acc: 73.96 | Data Time: 0.059s | Batch Time: 0.837s | LR: 0.000020 | logit_scale: 4.605 | Global Batch Size: 96
2022-12-08,18:17:24 | INFO | Rank 0 | Global Steps: 122/41728 | Train Epoch: 1 [11712/250368 (5%)] | Loss: 0.835916 | Image2Text Acc: 75.00 | Text2Image Acc: 71.88 | Data Time: 0.059s | Batch Time: 0.840s | LR: 0.000020 | logit_scale: 4.605 | Global Batch Size: 96
2022-12-08,18:17:25 | INFO | Rank 0 | Global Steps: 123/41728 | Train Epoch: 1 [11808/250368 (5%)] | Loss: 0.764448 | Image2Text Acc: 77.08 | Text2Image Acc: 79.17 | Data Time: 0.059s | Batch Time: 0.841s | LR: 0.000020 | logit_scale: 4.605 | Global Batch Size: 96
2022-12-08,18:17:25 | INFO | Rank 0 | Global Steps: 124/41728 | Train Epoch: 1 [11904/250368 (5%)] | Loss: 0.849898 | Image2Text Acc: 77.08 | Text2Image Acc: 73.96 | Data Time: 0.059s | Batch Time: 0.837s | LR: 0.000020 | logit_scale: 4.605 | Global Batch Size: 96
2022-12-08,18:17:26 | INFO | Rank 0 | Global Steps: 125/41728 | Train Epoch: 1 [12000/250368 (5%)] | Loss: 0.727899 | Image2Text Acc: 79.17 | Text2Image Acc: 77.08 | Data Time: 0.059s | Batch Time: 0.840s | LR: 0.000020 | logit_scale: 4.605 | Global Batch Size: 96
2022-12-08,18:17:27 | INFO | Rank 0 | Global Steps: 126/41728 | Train Epoch: 1 [12096/250368 (5%)] | Loss: 0.800566 | Image2Text Acc: 81.25 | Text2Image Acc: 80.21 | Data Time: 0.060s | Batch Time: 0.843s | LR: 0.000020 | logit_scale: 4.605 | Global Batch Size: 96
2022-12-08,18:17:28 | INFO | Rank 0 | Global Steps: 127/41728 | Train Epoch: 1 [12192/250368 (5%)] | Loss: 0.960883 | Image2Text Acc: 72.92 | Text2Image Acc: 73.96 | Data Time: 0.059s | Batch Time: 0.840s | LR: 0.000020 | logit_scale: 4.605 | Global Batch Size: 96
2022-12-08,18:17:29 | INFO | Rank 0 | Global Steps: 128/41728 | Train Epoch: 1 [12288/250368 (5%)] | Loss: 0.790139 | Image2Text Acc: 72.92 | Text2Image Acc: 81.25 | Data Time: 0.061s | Batch Time: 0.843s | LR: 0.000020 | logit_scale: 4.605 | Global Batch Size: 96
2022-12-08,18:17:30 | INFO | Rank 0 | Global Steps: 129/41728 | Train Epoch: 1 [12384/250368 (5%)] | Loss: 0.780300 | Image2Text Acc: 82.29 | Text2Image Acc: 79.17 | Data Time: 0.059s | Batch Time: 0.839s | LR: 0.000020 | logit_scale: 4.605 | Global Batch Size: 96
2022-12-08,18:17:30 | INFO | Rank 0 | Global Steps: 130/41728 | Train Epoch: 1 [12480/250368 (5%)] | Loss: 0.850396 | Image2Text Acc: 76.04 | Text2Image Acc: 79.17 | Data Time: 0.060s | Batch Time: 0.842s | LR: 0.000020 | logit_scale: 4.605 | Global Batch Size: 96
2022-12-08,18:17:31 | INFO | Rank 0 | Global Steps: 131/41728 | Train Epoch: 1 [12576/250368 (5%)] | Loss: 0.649198 | Image2Text Acc: 81.25 | Text2Image Acc: 80.21 | Data Time: 0.059s | Batch Time: 0.835s | LR: 0.000020 | logit_scale: 4.605 | Global Batch Size: 96
2022-12-08,18:17:32 | INFO | Rank 0 | Global Steps: 132/41728 | Train Epoch: 1 [12672/250368 (5%)] | Loss: 0.680545 | Image2Text Acc: 77.08 | Text2Image Acc: 83.33 | Data Time: 0.061s | Batch Time: 0.839s | LR: 0.000020 | logit_scale: 4.605 | Global Batch Size: 96
2022-12-08,18:17:33 | INFO | Rank 0 | Global Steps: 133/41728 | Train Epoch: 1 [12768/250368 (5%)] | Loss: 0.597127 | Image2Text Acc: 85.42 | Text2Image Acc: 83.33 | Data Time: 0.059s | Batch Time: 0.847s | LR: 0.000020 | logit_scale: 4.605 | Global Batch Size: 96
2022-12-08,18:17:34 | INFO | Rank 0 | Global Steps: 134/41728 | Train Epoch: 1 [12864/250368 (5%)] | Loss: 0.929113 | Image2Text Acc: 75.00 | Text2Image Acc: 70.83 | Data Time: 0.060s | Batch Time: 0.836s | LR: 0.000020 | logit_scale: 4.605 | Global Batch Size: 96
2022-12-08,18:17:35 | INFO | Rank 0 | Global Steps: 135/41728 | Train Epoch: 1 [12960/250368 (5%)] | Loss: 1.120120 | Image2Text Acc: 68.75 | Text2Image Acc: 70.83 | Data Time: 0.059s | Batch Time: 0.847s | LR: 0.000020 | logit_scale: 4.605 | Global Batch Size: 96
2022-12-08,18:17:35 | INFO | Rank 0 | Global Steps: 136/41728 | Train Epoch: 1 [13056/250368 (5%)] | Loss: 1.008211 | Image2Text Acc: 71.88 | Text2Image Acc: 72.92 | Data Time: 0.061s | Batch Time: 0.835s | LR: 0.000020 | logit_scale: 4.605 | Global Batch Size: 96
2022-12-08,18:17:36 | INFO | Rank 0 | Global Steps: 137/41728 | Train Epoch: 1 [13152/250368 (5%)] | Loss: 0.680133 | Image2Text Acc: 79.17 | Text2Image Acc: 71.88 | Data Time: 0.059s | Batch Time: 0.840s | LR: 0.000020 | logit_scale: 4.605 | Global Batch Size: 96
2022-12-08,18:17:37 | INFO | Rank 0 | Global Steps: 138/41728 | Train Epoch: 1 [13248/250368 (5%)] | Loss: 0.839577 | Image2Text Acc: 78.12 | Text2Image Acc: 76.04 | Data Time: 0.060s | Batch Time: 0.833s | LR: 0.000020 | logit_scale: 4.605 | Global Batch Size: 96
2022-12-08,18:17:38 | INFO | Rank 0 | Global Steps: 139/41728 | Train Epoch: 1 [13344/250368 (5%)] | Loss: 0.828378 | Image2Text Acc: 72.92 | Text2Image Acc: 77.08 | Data Time: 0.059s | Batch Time: 0.841s | LR: 0.000020 | logit_scale: 4.605 | Global Batch Size: 96
2022-12-08,18:17:39 | INFO | Rank 0 | Global Steps: 140/41728 | Train Epoch: 1 [13440/250368 (5%)] | Loss: 0.553494 | Image2Text Acc: 85.42 | Text2Image Acc: 83.33 | Data Time: 0.059s | Batch Time: 0.842s | LR: 0.000020 | logit_scale: 4.605 | Global Batch Size: 96
2022-12-08,18:17:40 | INFO | Rank 0 | Global Steps: 141/41728 | Train Epoch: 1 [13536/250368 (5%)] | Loss: 0.626136 | Image2Text Acc: 79.17 | Text2Image Acc: 81.25 | Data Time: 0.059s | Batch Time: 0.834s | LR: 0.000020 | logit_scale: 4.605 | Global Batch Size: 96
2022-12-08,18:17:40 | INFO | Rank 0 | Global Steps: 142/41728 | Train Epoch: 1 [13632/250368 (5%)] | Loss: 0.748430 | Image2Text Acc: 80.21 | Text2Image Acc: 79.17 | Data Time: 0.060s | Batch Time: 0.837s | LR: 0.000020 | logit_scale: 4.605 | Global Batch Size: 96
2022-12-08,18:17:41 | INFO | Rank 0 | Global Steps: 143/41728 | Train Epoch: 1 [13728/250368 (5%)] | Loss: 0.809352 | Image2Text Acc: 73.96 | Text2Image Acc: 80.21 | Data Time: 0.059s | Batch Time: 0.839s | LR: 0.000020 | logit_scale: 4.605 | Global Batch Size: 96
2022-12-08,18:17:42 | INFO | Rank 0 | Global Steps: 144/41728 | Train Epoch: 1 [13824/250368 (6%)] | Loss: 0.945053 | Image2Text Acc: 71.88 | Text2Image Acc: 73.96 | Data Time: 0.059s | Batch Time: 0.838s | LR: 0.000020 | logit_scale: 4.605 | Global Batch Size: 96
2022-12-08,18:17:43 | INFO | Rank 0 | Global Steps: 145/41728 | Train Epoch: 1 [13920/250368 (6%)] | Loss: 0.878521 | Image2Text Acc: 73.96 | Text2Image Acc: 77.08 | Data Time: 0.059s | Batch Time: 0.833s | LR: 0.000020 | logit_scale: 4.605 | Global Batch Size: 96
2022-12-08,18:17:44 | INFO | Rank 0 | Global Steps: 146/41728 | Train Epoch: 1 [14016/250368 (6%)] | Loss: 0.946872 | Image2Text Acc: 69.79 | Text2Image Acc: 70.83 | Data Time: 0.060s | Batch Time: 0.840s | LR: 0.000020 | logit_scale: 4.605 | Global Batch Size: 96
2022-12-08,18:17:45 | INFO | Rank 0 | Global Steps: 147/41728 | Train Epoch: 1 [14112/250368 (6%)] | Loss: 0.575010 | Image2Text Acc: 86.46 | Text2Image Acc: 82.29 | Data Time: 0.059s | Batch Time: 0.840s | LR: 0.000020 | logit_scale: 4.605 | Global Batch Size: 96
2022-12-08,18:17:46 | INFO | Rank 0 | Global Steps: 148/41728 | Train Epoch: 1 [14208/250368 (6%)] | Loss: 0.860036 | Image2Text Acc: 72.92 | Text2Image Acc: 72.92 | Data Time: 0.059s | Batch Time: 0.838s | LR: 0.000020 | logit_scale: 4.605 | Global Batch Size: 96
2022-12-08,18:17:46 | INFO | Rank 0 | Global Steps: 149/41728 | Train Epoch: 1 [14304/250368 (6%)] | Loss: 0.896334 | Image2Text Acc: 78.12 | Text2Image Acc: 78.12 | Data Time: 0.059s | Batch Time: 0.834s | LR: 0.000020 | logit_scale: 4.605 | Global Batch Size: 96
2022-12-08,18:17:47 | INFO | Rank 0 | Global Steps: 150/41728 | Train Epoch: 1 [14400/250368 (6%)] | Loss: 0.772457 | Image2Text Acc: 78.12 | Text2Image Acc: 75.00 | Data Time: 0.060s | Batch Time: 0.838s | LR: 0.000020 | logit_scale: 4.605 | Global Batch Size: 96
2022-12-08,18:17:47 | INFO | Rank 0 | Begin to eval on validation set (epoch 1 @ 150 steps)...
2022-12-08,18:18:58 | INFO | Rank 0 | Evaluated 100/319 batches...
2022-12-08,18:20:09 | INFO | Rank 0 | Evaluated 200/319 batches...
2022-12-08,18:21:21 | INFO | Rank 0 | Evaluated 300/319 batches...
2022-12-08,18:21:35 | INFO | Rank 0 | Validation Result (epoch 1 @ 150 steps) | Valid Loss: 1.608249 | Image2Text Acc: 32.55 | Text2Image Acc: 33.00 | logit_scale: 4.605 | Valid Batch Size: 48
from chinese-clip.
yangapku
commented on May 21, 2024
收到，我们也去排查下
from chinese-clip.
dengfenglai321
commented on May 21, 2024
2022-12-08,18:15:41 | INFO | Rank 0 | => loaded checkpoint '/storage1/xxx/Text_Based_Image_Retrieval/Chinese-CLIP-master/experiments/pretrained_weights/clip_cn_vit-b-16.pt' (epoch 15 @ 0
好的谢谢。麻烦排查下。从训练日志看，训练loss一直在下降。训练集acc一直在提升。 但是验证集的的loss和acc从一开始就没怎么变过，acc维持在32左右浮动
from chinese-clip.
yangapku
commented on May 21, 2024
@yumulinfeng1 您好，关于您在issue提出的问题，大概有以下几点结论和建议：
- 关于您所提到的finetune时，验证集inbatch acc一直较低的问题，主要是因为MUGE数据集验证集存在文对图一对多。在默认验证集不shuffle、且GPU分布式卡数较少的情况下，一个验证batch里面会有多个图文样本的文本相同的情况。由于我们的val inbatch acc是最简单的实现版本，只把样本自己作为ground truth，所以这种情况下会出现inbatch acc计算结果偏低的现象。实际上模型训练和收敛完全没有影响，最终走完整个评测流程，得到的Recall指标如果你有尝试跑过，应该也没问题。 更具体的描述参见PR #29 ，我们也做了个小改动，把finetune时验证集的sampler也改成shuffle的情况了，规避掉这种特殊情况。现在如果拉取最新代码再跑，不会出现验证集inbatch acc一直较低这个情况了。
- 关于模型finetune的效果对比，建议您和我们汇报结果所对比的效果指标，应该是跑完图文检索的评测流程最终得到的Recall（或者同一个零样本图片分类数据集上的准确率），而不是对比验证集inbatch acc。这是因为验证集inbatch acc显然和
valid_batch_size有关，您所跑的valid_batch_size=48和我们脚本默认的valid_batch_size=128也不相同，所以和我们那个log样例写的inbatch acc对比意义不大，或者说您最新代码跑出来一定会比我们的inbatch acc要高，因为batch小，负例更少。验证集inbatch acc只用于您自己几组实验评估收敛趋势。 - 您finetune使用的训练batch size很小，2卡总计96。根据我们Readme中的描述，对比学习的训练收敛和稳定性和总batch size相关。如您使用更小的batch size（相比脚本默认配置128 per-GPU * 8 GPU），建议使用更小的学习率。您的学习率为2e-5，我们实测这组超参下，finetune最终的Recall甚至比不finetune的zeroshot结果还会要低（MUGE Mean Recall (MR) 68左右，对比我们base规模预训练模型zeroshot的MR 71.1）。建议您试试更小的学习率或者更大的batch（可以考虑用Readme中的重计算策略）。我们有尝试如果您提供这组超参，学习率降到2e-6，最终的MR将达到75.5以上，相比于zeroshot的MR 71.1会有提升。供您参考。
如果有更多问题，欢迎继续留言。如果觉得Chinese-CLIP代码库对您有帮助，请您为我们点点star⭐️并推荐给身边的朋友们！
from chinese-clip.
dengfenglai321
commented on May 21, 2024
好的 十分感谢！！我去试试
from chinese-clip.
6Roy
commented on May 21, 2024
请问您使用的torch版本是多少？我在复现的时候一直出现问题
from chinese-clip.
Related Issues (20)
- 这个支持词性标注吗 HOT 1
- ImportError: cannot import name 'convert_state_dict' from 'cn_clip.clip.model' HOT 2
- AttributeError: 'Namespace' object has no attribute 'use_flash_attention' HOT 2
- load_from_name function logic problem
- torch onnx tensorrt 推理速度和deployment.md评估效果差异较大 HOT 1
- 关于跨模态检索的召回率疑问，请大佬解答一番 HOT 1
- 论文初始化image encoder的模型参数
- 关于训练过程一个batch内，text文本标签存在较多重复，是否会导致训练不收敛 HOT 3
- 转换torchscript报错
- 尝试使用CLIP模型来进行文搜文遇到的问题
- 请问将cn_clip转为onnx格式后，有没有不使用cn_clip库进行预处理的思路，我们在尝试用C++部署
- 使用脚本怎么跑不了了，一直卡着什么输出也没有
- finetune时如何运行sh文件
- 使用flash-Attention训练，loss不下降，acc也不增加
- 请问在chinese-clip获取到图文特征json数据后，如何在clip-retrieval使用呢，想搭建自己的demo HOT 2
- finetune时的问题
- 在Huggingface上使用你们的模型的时候出现无法加载tokenizer的错误
- 请问，imagenet上的数据集精度是如何得到的？
- 几何图形特征
- 关于执行sh脚本文件的报错 RuntimeError: NCCL error in: ../torch/lib/c10d/ProcessGroupNCCL.cpp:911, unhandled system error, NCCL version 2.7.8
Recommend Projects
-
React
A declarative, efficient, and flexible JavaScript library for building user interfaces.
-
Vue.js
🖖 Vue.js is a progressive, incrementally-adoptable JavaScript framework for building UI on the web.
-
Typescript
TypeScript is a superset of JavaScript that compiles to clean JavaScript output.
-
TensorFlow
An Open Source Machine Learning Framework for Everyone
-
Django
The Web framework for perfectionists with deadlines.
-
Laravel
A PHP framework for web artisans
-
D3
Bring data to life with SVG, Canvas and HTML. 📊📈🎉
-
Recommend Topics
-
javascript
JavaScript (JS) is a lightweight interpreted programming language with first-class functions.
-
web
Some thing interesting about web. New door for the world.
-
server
A server is a program made to process requests and deliver data to clients.
-
Machine learning
Machine learning is a way of modeling and interpreting data that allows a piece of software to respond intelligently.
-
Visualization
Some thing interesting about visualization, use data art
-
Game
Some thing interesting about game, make everyone happy.
Recommend Org
-
Facebook
We are working to build community through open source technology. NB: members must have two-factor auth.
-
Microsoft
Open source projects and samples from Microsoft.
-
Google
Google ❤️ Open Source for everyone.
-
Alibaba
Alibaba Open Source for everyone
-
D3
Data-Driven Documents codes.
-
Tencent
China tencent open source team.
from chinese-clip.