Logging and checkpointing are important comonents of model training workflow. It allows users to keep a record of the model hyper-parameters and its performance on training.
In this tutorial we will look at some of the utilities from Modulus to simplify this important aspect of model training.
Modulus provides utilities to standardize the logs of different training runs. Using the logging utilites from Modulus, you would have the flexibility of choosing between the good-old console logging to more advanced ML experiments trackers like MLFlow and Weights & Biases. You can always implement these loggers yourself, but in this example, we will use the utilites from Modulus that will not only simplify this process but also provide a standardized output format. Let’s get started.
Console logging
The below example shows a simple setup using the console logging.
import torch
import modulus
from modulus.datapipes.benchmarks.darcy import Darcy2D
from modulus.launch.logging import LaunchLogger, PythonLogger
from modulus.metrics.general.mse import mse
from modulus.models.fno.fno import FNO
normaliser = {
"permeability": (1.25, 0.75),
"darcy": (4.52e-2, 2.79e-2),
}
dataloader = Darcy2D(
resolution=256, batch_size=64, nr_permeability_freq=5, normaliser=normaliser
)
model = FNO(
in_channels=1,
out_channels=1,
decoder_layers=1,
decoder_layer_size=32,
dimension=2,
latent_channels=32,
num_fno_layers=4,
num_fno_modes=12,
padding=5,
).to("cuda")
optimizer = torch.optim.Adam(model.parameters(), lr=0.01)
scheduler = torch.optim.lr_scheduler.LambdaLR(
optimizer, lr_lambda=lambda step: 0.85**step
)
# Initialize the logger
logger = PythonLogger("main") # General python logger
LaunchLogger.initialize()
# Use logger methods to track various information during training
logger.info("Starting Training!")
# we will setup the training to run for 20 epochs each epoch running for 5 iterations
for i in range(20):
# wrap the epoch in launch logger to control frequency of output for console logs
with LaunchLogger("train", epoch=i) as launchlog:
# this would be iterations through different batches
for _ in range(5):
batch = next(iter(dataloader))
true = batch["darcy"]
pred = model(batch["permeability"])
loss = mse(pred, true)
loss.backward()
optimizer.step()
scheduler.step()
launchlog.log_minibatch({"Loss": loss.detach().cpu().numpy()})
launchlog.log_epoch({"Learning Rate": optimizer.param_groups[0]["lr"]})
logger.info("Finished Training!")
The logger output can be seen below.
Warp 0.10.1 initialized:
CUDA Toolkit: 11.5, Driver: 12.2
Devices:
"cpu" | x86_64
"cuda:0" | Tesla V100-SXM2-16GB-N (sm_70)
"cuda:1" | Tesla V100-SXM2-16GB-N (sm_70)
"cuda:2" | Tesla V100-SXM2-16GB-N (sm_70)
"cuda:3" | Tesla V100-SXM2-16GB-N (sm_70)
"cuda:4" | Tesla V100-SXM2-16GB-N (sm_70)
"cuda:5" | Tesla V100-SXM2-16GB-N (sm_70)
"cuda:6" | Tesla V100-SXM2-16GB-N (sm_70)
"cuda:7" | Tesla V100-SXM2-16GB-N (sm_70)
Kernel cache: /root/.cache/warp/0.10.1
/usr/local/lib/python3.10/dist-packages/pydantic/_internal/_fields.py:128: UserWarning: Field "model_server_url" has conflict with protected namespace "model_".
You may be able to resolve this warning by setting `model_config['protected_namespaces'] = ()`.
warnings.warn(
/usr/local/lib/python3.10/dist-packages/pydantic/_internal/_config.py:317: UserWarning: Valid config keys have changed in V2:
* 'schema_extra' has been renamed to 'json_schema_extra'
warnings.warn(message, UserWarning)
[21:23:57 - main - INFO] Starting Training!
Module modulus.datapipes.benchmarks.kernels.initialization load on device 'cuda:0' took 73.06 ms
Module modulus.datapipes.benchmarks.kernels.utils load on device 'cuda:0' took 314.91 ms
Module modulus.datapipes.benchmarks.kernels.finite_difference load on device 'cuda:0' took 149.86 ms
[21:24:02 - train - INFO] Epoch 0 Metrics: Learning Rate = 4.437e-03, Loss = 1.009e+00
[21:24:02 - train - INFO] Epoch Execution Time: 5.664e+00s, Time/Iter: 1.133e+03ms
[21:24:06 - train - INFO] Epoch 1 Metrics: Learning Rate = 1.969e-03, Loss = 6.040e-01
[21:24:06 - train - INFO] Epoch Execution Time: 4.013e+00s, Time/Iter: 8.025e+02ms
...
[21:25:32 - train - INFO] Epoch 19 Metrics: Learning Rate = 8.748e-10, Loss = 1.384e-01
[21:25:32 - train - INFO] Epoch Execution Time: 4.010e+00s, Time/Iter: 8.020e+02ms
[21:25:32 - main - INFO] Finished Training!
MLFlow logging
The below example shows a simple setup using the MLFlow logging. The only difference from
the previous example is that, we will use initialize_mlflow function to initialize
the MLFlow client and also set use_mlflow=True when initializing the LaunchLogger.
import torch
import modulus
from modulus.datapipes.benchmarks.darcy import Darcy2D
from modulus.launch.logging import LaunchLogger, PythonLogger, initialize_mlflow
from modulus.metrics.general.mse import mse
from modulus.models.fno.fno import FNO
normaliser = {
"permeability": (1.25, 0.75),
"darcy": (4.52e-2, 2.79e-2),
}
dataloader = Darcy2D(
resolution=256, batch_size=64, nr_permeability_freq=5, normaliser=normaliser
)
model = FNO(
in_channels=1,
out_channels=1,
decoder_layers=1,
decoder_layer_size=32,
dimension=2,
latent_channels=32,
num_fno_layers=4,
num_fno_modes=12,
padding=5,
).to("cuda")
optimizer = torch.optim.Adam(model.parameters(), lr=0.01)
scheduler = torch.optim.lr_scheduler.LambdaLR(
optimizer, lr_lambda=lambda step: 0.85**step
)
# Initialize the console logger
logger = PythonLogger("main") # General python logger
# Initialize the MLFlow logger
initialize_mlflow(
experiment_name="Modulus Tutorials",
experiment_desc="Simple Modulus Tutorials",
run_name="Modulus MLFLow Tutorial",
run_desc="Modulus Tutorial Training",
user_name="Modulus User",
mode="offline",
)
LaunchLogger.initialize(use_mlflow=True)
# Use logger methods to track various information during training
logger.info("Starting Training!")
# we will setup the training to run for 20 epochs each epoch running for 5 iterations
for i in range(20):
# wrap the epoch in launch logger to control frequency of output for console logs
with LaunchLogger("train", epoch=i) as launchlog:
for _ in range(5):
batch = next(iter(dataloader))
true = batch["darcy"]
pred = model(batch["permeability"])
loss = mse(pred, true)
loss.backward()
optimizer.step()
scheduler.step()
launchlog.log_minibatch({"Loss": loss.detach().cpu().numpy()})
launchlog.log_epoch({"Learning Rate": optimizer.param_groups[0]["lr"]})
logger.info("Finished Training!")
During the run, you will notice a directory named as mlruns_0 created which stores
the mlflow logs. To visulaize the logs interactively, you can run the following:
mlflow ui --backend-store-uri mlruns_0/
And then navigate to localhost:5000 in your favorite browser.
Currently the MLFlow logger will log the output of each processor separately. So in multi-processor runs, you will see multiple directories being created. This is a known issue and will be fixed in the future releases.
Weight and Biases logging
The below example shows a simple setup using the Weights and Biases logging. The only
difference from the previous example is that, we will use initialize_wandb function
to initialize the Weights and Biases logger and also set use_wandb=True when
initializing the LaunchLogger.
import torch
import modulus
from modulus.datapipes.benchmarks.darcy import Darcy2D
from modulus.launch.logging import LaunchLogger, PythonLogger, initialize_wandb
from modulus.metrics.general.mse import mse
from modulus.models.fno.fno import FNO
normaliser = {
"permeability": (1.25, 0.75),
"darcy": (4.52e-2, 2.79e-2),
}
dataloader = Darcy2D(
resolution=256, batch_size=64, nr_permeability_freq=5, normaliser=normaliser
)
model = FNO(
in_channels=1,
out_channels=1,
decoder_layers=1,
decoder_layer_size=32,
dimension=2,
latent_channels=32,
num_fno_layers=4,
num_fno_modes=12,
padding=5,
).to("cuda")
optimizer = torch.optim.Adam(model.parameters(), lr=0.01)
scheduler = torch.optim.lr_scheduler.LambdaLR(
optimizer, lr_lambda=lambda step: 0.85**step
)
# Initialize the console logger
logger = PythonLogger("main") # General python logger
# Initialize the MLFlow logger
initialize_wandb(
project="Modulus Tutorials",
name="Simple Modulus Tutorials",
entity="Modulus MLFLow Tutorial",
mode="offline",
)
LaunchLogger.initialize(use_wandb=True)
# Use logger methods to track various information during training
logger.info("Starting Training!")
# we will setup the training to run for 20 epochs each epoch running for 10 iterations
for i in range(20):
# wrap the epoch in launch logger to control frequency of output for console logs
with LaunchLogger("train", epoch=i) as launchlog:
# this would be iterations through different batches
for _ in range(10):
batch = next(iter(dataloader))
true = batch["darcy"]
pred = model(batch["permeability"])
loss = mse(pred, true)
loss.backward()
optimizer.step()
scheduler.step()
launchlog.log_minibatch({"Loss": loss.detach().cpu().numpy()})
launchlog.log_epoch({"Learning Rate": optimizer.param_groups[0]["lr"]})
logger.info("Finished Training!")
During the run, you will notice a directory named as wandb created which stores
the wandb logs.
The logger output can also be seen below.
Warp 0.10.1 initialized:
CUDA Toolkit: 11.5, Driver: 12.2
Devices:
"cpu" | x86_64
"cuda:0" | Tesla V100-SXM2-16GB-N (sm_70)
"cuda:1" | Tesla V100-SXM2-16GB-N (sm_70)
"cuda:2" | Tesla V100-SXM2-16GB-N (sm_70)
"cuda:3" | Tesla V100-SXM2-16GB-N (sm_70)
"cuda:4" | Tesla V100-SXM2-16GB-N (sm_70)
"cuda:5" | Tesla V100-SXM2-16GB-N (sm_70)
"cuda:6" | Tesla V100-SXM2-16GB-N (sm_70)
"cuda:7" | Tesla V100-SXM2-16GB-N (sm_70)
Kernel cache: /root/.cache/warp/0.10.1
/usr/local/lib/python3.10/dist-packages/pydantic/_internal/_fields.py:128: UserWarning: Field "model_server_url" has conflict with protected namespace "model_".
You may be able to resolve this warning by setting `model_config['protected_namespaces'] = ()`.
warnings.warn(
/usr/local/lib/python3.10/dist-packages/pydantic/_internal/_config.py:317: UserWarning: Valid config keys have changed in V2:
* 'schema_extra' has been renamed to 'json_schema_extra'
warnings.warn(message, UserWarning)
wandb: Tracking run with wandb version 0.15.12
wandb: W&B syncing is set to `offline` in this directory.
wandb: Run `wandb online` or set WANDB_MODE=online to enable cloud syncing.
[21:26:38 - main - INFO] Starting Training!
Module modulus.datapipes.benchmarks.kernels.initialization load on device 'cuda:0' took 74.11 ms
Module modulus.datapipes.benchmarks.kernels.utils load on device 'cuda:0' took 310.06 ms
Module modulus.datapipes.benchmarks.kernels.finite_difference load on device 'cuda:0' took 151.24 ms
[21:26:48 - train - INFO] Epoch 0 Metrics: Learning Rate = 1.969e-03, Loss = 7.164e-01
[21:26:48 - train - INFO] Epoch Execution Time: 9.703e+00s, Time/Iter: 9.703e+02ms
...
[21:29:47 - train - INFO] Epoch 19 Metrics: Learning Rate = 7.652e-17, Loss = 3.519e-01
[21:29:47 - train - INFO] Epoch Execution Time: 1.125e+01s, Time/Iter: 1.125e+03ms
[21:29:47 - main - INFO] Finished Training!
wandb: Waiting for W&B process to finish... (success).
wandb:
wandb: Run history:
wandb: epoch ▁▁▂▂▂▃▃▄▄▄▅▅▅▆▆▇▇▇██
wandb: train/Epoch Time (s) ▃▁▃▃▃▃▁█▁▁▁▃▃▃▃▆▁▃▃▆
wandb: train/Learning Rate █▂▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁
wandb: train/Loss █▁▂▃▃▃▃▃▃▃▃▃▃▃▃▃▃▃▃▃
wandb: train/Time per iter (ms) ▃▁▃▃▃▃▁█▁▁▁▃▃▃▃▆▁▃▃▆
wandb:
wandb: Run summary:
wandb: epoch 19
wandb: train/Epoch Time (s) 11.24806
wandb: train/Learning Rate 0.0
wandb: train/Loss 0.35193
wandb: train/Time per iter (ms) 1124.80645
wandb:
wandb: You can sync this run to the cloud by running:
wandb: wandb sync /workspace/modulus/docs/test_scripts/wandb/wandb/offline-run-20231115_212638-ib4ylq4e
wandb: Find logs at: ./wandb/wandb/offline-run-20231115_212638-ib4ylq4e/logs
To visulaize the logs interactively, simply follow the instructions printed in the outputs.
Modulus provides easy utilities to save and load the checkpoints of the model, optimizer, scheduler, and scaler during training and inference. Similar to logging, custom implementation can be used, but in this example we will see the utilites from Modulus and some of its benefits.
Loading and saving checkpoints during training
The example below shows how you can save and load a checkpoint during training. The implementation
allows the model training to be resumed from the last saved checkpoint. Here, we will
demonstrate the use of load_checkpoint and the save_checkpoint functions.
import torch
import modulus
from modulus.datapipes.benchmarks.darcy import Darcy2D
from modulus.launch.utils import load_checkpoint, save_checkpoint
from modulus.metrics.general.mse import mse
from modulus.models.fno.fno import FNO
normaliser = {
"permeability": (1.25, 0.75),
"darcy": (4.52e-2, 2.79e-2),
}
dataloader = Darcy2D(
resolution=256, batch_size=64, nr_permeability_freq=5, normaliser=normaliser
)
model = FNO(
in_channels=1,
out_channels=1,
decoder_layers=1,
decoder_layer_size=32,
dimension=2,
latent_channels=32,
num_fno_layers=4,
num_fno_modes=12,
padding=5,
).to("cuda")
optimizer = torch.optim.Adam(model.parameters(), lr=0.01)
scheduler = torch.optim.lr_scheduler.LambdaLR(
optimizer, lr_lambda=lambda step: 0.85**step
)
# load the epoch and optimizer, model ans scheduler parameters from the checkpoint if
# it exists
loaded_epoch = load_checkpoint(
"./checkpoints",
models=model,
optimizer=optimizer,
scheduler=scheduler,
device="cuda",
)
# we will setup the training to run for 20 epochs each epoch running for 5 iterations
# starting with the loaded epoch
for i in range(max(1, loaded_epoch), 20):
# this would be iterations through different batches
for _ in range(5):
batch = next(iter(dataloader))
true = batch["darcy"]
pred = model(batch["permeability"])
loss = mse(pred, true)
loss.backward()
optimizer.step()
scheduler.step()
# save checkpoint every 5th epoch
if i % 5 == 0:
save_checkpoint(
"./checkpoints",
models=model,
optimizer=optimizer,
scheduler=scheduler,
epoch=i,
)
The output of the above script when loaded from a partially trained model will be something like below.
>>> python test_scripts/test_basic_checkpointing.py
...
[23:11:09 - checkpoint - INFO] Loaded model state dictionary /workspace/release_23.11/docs_upgrade/modulus/docs/checkpoints/FourierNeuralOperator.0.10.mdlus to device cuda
[23:11:09 - checkpoint - INFO] Loaded checkpoint file /workspace/release_23.11/docs_upgrade/modulus/docs/checkpoints/checkpoint.0.10.pt to device cuda
[23:11:09 - checkpoint - INFO] Loaded optimizer state dictionary
[23:11:09 - checkpoint - INFO] Loaded scheduler state dictionary
...
[23:11:11 - checkpoint - INFO] Saved model state dictionary: /workspace/release_23.11/docs_upgrade/modulus/docs/checkpoints/FourierNeuralOperator.0.10.mdlus
[23:11:12 - checkpoint - INFO] Saved training checkpoint: /workspace/release_23.11/docs_upgrade/modulus/docs/checkpoints/checkpoint.0.10.pt
[23:11:16 - checkpoint - INFO] Saved model state dictionary: /workspace/release_23.11/docs_upgrade/modulus/docs/checkpoints/FourierNeuralOperator.0.15.mdlus
[23:11:16 - checkpoint - INFO] Saved training checkpoint: /workspace/release_23.11/docs_upgrade/modulus/docs/checkpoints/checkpoint.0.15.pt
[23:11:21 - checkpoint - INFO] Saved model state dictionary: /workspace/release_23.11/docs_upgrade/modulus/docs/checkpoints/FourierNeuralOperator.0.20.mdlus
[23:11:21 - checkpoint - INFO] Saved training checkpoint: /workspace/release_23.11/docs_upgrade/modulus/docs/checkpoints/checkpoint.0.20.pt
Loading checkpoints during inference
For loading the checkpoint in inference, the process is simple and you can refer the samples provided in Running infernece on trained models and Saving and Loading Modulus Models .