# Training ## Training on a Single GPU You can use `tools/train.py` to train a model on a single machine with a CPU and optionally a GPU. Here is the full usage of the script: ```shell python tools/train.py ${CONFIG_FILE} [ARGS] ``` :::{note} By default, MMOCR prefers GPU to CPU. If you want to train a model on CPU, please empty `CUDA_VISIBLE_DEVICES` or set it to -1 to make GPU invisible to the program. Note that CPU training requires **MMCV >= 1.4.4**. ```bash CUDA_VISIBLE_DEVICES= python tools/train.py ${CONFIG_FILE} [ARGS] ``` ::: | ARGS | Type | Description | | ----------------- | --------------------------------- | -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | | `--work-dir` | str | The target folder to save logs and checkpoints. Defaults to `./work_dirs`. | | `--load-from` | str | Path to the pre-trained model, which will be used to initialize the network parameters. | | `--resume-from` | str | Resume training from a previously saved checkpoint, which will inherit the training epoch and optimizer parameters. | | `--no-validate` | bool | Disable checkpoint evaluation during training. Defaults to `False`. | | `--gpus` | int | **Deprecated, please use --gpu-id.** Numbers of gpus to use. Only applicable to non-distributed training. | | `--gpu-ids` | int*N | **Deprecated, please use --gpu-id.** A list of GPU ids to use. Only applicable to non-distributed training. | | `--gpu-id` | int | The GPU id to use. Only applicable to non-distributed training. | | `--seed` | int | Random seed. | | `--diff_seed` | bool | Whether or not set different seeds for different ranks. | | `--deterministic` | bool | Whether to set deterministic options for CUDNN backend. | | `--cfg-options` | str | Override some settings in the used config, the key-value pair in xxx=yyy format will be merged into the config file. If the value to be overwritten is a list, it should be of the form of either key="[a,b]" or key=a,b. The argument also allows nested list/tuple values, e.g. key="[(a,b),(c,d)]". Note that the quotation marks are necessary and that no white space is allowed. | | `--launcher` | 'none', 'pytorch', 'slurm', 'mpi' | Options for job launcher. | | `--local_rank` | int | Used for distributed training. | | `--mc-config` | str | Memory cache config for image loading speed-up during training. | ## Training on Multiple GPUs MMOCR implements **distributed** training with `MMDistributedDataParallel`. (Please refer to [datasets.md](datasets.md) to prepare your datasets) ```shell [PORT={PORT}] ./tools/dist_train.sh ${CONFIG_FILE} ${WORK_DIR} ${GPU_NUM} [PY_ARGS] ``` | Arguments | Type | Description | | --------- | ---- | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ | | `PORT` | int | The master port that will be used by the machine with rank 0. Defaults to 29500. **Note:** If you are launching multiple distrbuted training jobs on a single machine, you need to specify different ports for each job to avoid port conflicts. | | `PY_ARGS` | str | Arguments to be parsed by `tools/train.py`. | ## Training on Multiple Machines MMOCR relies on torch.distributed package for distributed training. Thus, as a basic usage, one can launch distributed training via PyTorch’s [launch utility](https://pytorch.org/docs/stable/distributed.html#launch-utility). ## Training with Slurm If you run MMOCR on a cluster managed with [Slurm](https://slurm.schedmd.com/), you can use the script `slurm_train.sh`. ```shell [GPUS=${GPUS}] [GPUS_PER_NODE=${GPUS_PER_NODE}] [CPUS_PER_TASK=${CPUS_PER_TASK}] [SRUN_ARGS=${SRUN_ARGS}] ./tools/slurm_train.sh ${PARTITION} ${JOB_NAME} ${CONFIG_FILE} ${WORK_DIR} [PY_ARGS] ``` | Arguments | Type | Description | | --------------- | ---- | ----------------------------------------------------------------------------------------------------------- | | `GPUS` | int | The number of GPUs to be used by this task. Defaults to 8. | | `GPUS_PER_NODE` | int | The number of GPUs to be allocated per node. Defaults to 8. | | `CPUS_PER_TASK` | int | The number of CPUs to be allocated per task. Defaults to 5. | | `SRUN_ARGS` | str | Arguments to be parsed by srun. Available options can be found [here](https://slurm.schedmd.com/srun.html). | | `PY_ARGS` | str | Arguments to be parsed by `tools/train.py`. | Here is an example of using 8 GPUs to train a text detection model on the dev partition. ```shell ./tools/slurm_train.sh dev psenet-ic15 configs/textdet/psenet/psenet_r50_fpnf_sbn_1x_icdar2015.py /nfs/xxxx/psenet-ic15 ``` ### Running Multiple Training Jobs on a Single Machine If you are launching multiple training jobs on a single machine with Slurm, you may need to modify the port in configs to avoid communication conflicts. For example, in `config1.py`, ```python dist_params = dict(backend='nccl', port=29500) ``` In `config2.py`, ```python dist_params = dict(backend='nccl', port=29501) ``` Then you can launch two jobs with `config1.py` ang `config2.py`. ```shell CUDA_VISIBLE_DEVICES=0,1,2,3 GPUS=4 ./tools/slurm_train.sh ${PARTITION} ${JOB_NAME} config1.py ${WORK_DIR} CUDA_VISIBLE_DEVICES=4,5,6,7 GPUS=4 ./tools/slurm_train.sh ${PARTITION} ${JOB_NAME} config2.py ${WORK_DIR} ``` ## Commonly Used Training Configs Here we list some configs that are frequently used during training for quick reference. ```python total_epochs = 1200 data = dict( # Note: User can configure general settings of train, val and test dataloader by specifying them here. However, their values can be overridden in dataloader's config. samples_per_gpu=8, # Batch size per GPU workers_per_gpu=4, # Number of workers to process data for each GPU train_dataloader=dict(samples_per_gpu=10, drop_last=True), # Batch size = 10, workers_per_gpu = 4 val_dataloader=dict(samples_per_gpu=6, workers_per_gpu=1), # Batch size = 6, workers_per_gpu = 1 test_dataloader=dict(workers_per_gpu=16), # Batch size = 8, workers_per_gpu = 16 ... ) # Evaluation evaluation = dict(interval=1, by_epoch=True) # Evaluate the model every epoch # Saving and Logging checkpoint_config = dict(interval=1) # Save a checkpoint every epoch log_config = dict( interval=5, # Print out the model's performance every 5 iterations hooks=[ dict(type='TextLoggerHook') ]) # Optimizer optimizer = dict(type='SGD', lr=0.02, momentum=0.9, weight_decay=0.0001) # Supports all optimizers in PyTorch and shares the same parameters optimizer_config = dict(grad_clip=None) # Parameters for the optimizer hook. See https://github.com/open-mmlab/mmcv/blob/master/mmcv/runner/hooks/optimizer.py for implementation details # Learning policy lr_config = dict(policy='poly', power=0.9, min_lr=1e-7, by_epoch=True) ```