Model Core Utils

core classes & helpers

Init

Apply init to layers with relu activations

---

source

weight_init

 weight_init (m:torch.nn.modules.module.Module, leaky:int=0)

	Type	Default	Details
m	Module		the module to initialize
leaky	int	0	if leaky relu used

x = torch.randn(1, 1, 32, 32)
in_channels = x.shape[1]
out_channels = 3
kernel_size = 3
stride = 1
c1 = nn.Conv2d(in_channels, out_channels, kernel_size, stride, kernel_size//2)
x1 = c1(x)
print(f"after conv: {x1.shape}")
print("x flat dim:", x1.flatten(2).shape)
l1 = nn.Linear(32*32, 64)
x2 = l1(x1.flatten(2))
print("after linear:", x2.shape)

leaky = 0.01
nnet = nn.Sequential(c1, nn.LeakyReLU(negative_slope=leaky), nn.Flatten(2), l1)
nnet.eval().cpu()
fig, ax = plt.subplots(1, 3)
ax[0].imshow(x.permute(0,2,3,1).squeeze())
ax[0].set_title("x")
y = nnet(x)
ax[1].imshow(y.detach().squeeze(0).permute(1,0).reshape(8, 8, 3))
ax[1].set_title("no init")
wi = partial(weight_init, leaky=leaky)
nnet.apply(wi)
y = nnet(x)
ax[2].imshow(y.detach().squeeze(0).permute(1,0).reshape(8, 8, 3))
ax[2].set_title("kaiming init")

after conv: torch.Size([1, 3, 32, 32])

x flat dim:
torch.Size([1, 3, 1024])

after linear:
torch.Size([1, 3, 64])

Text(0.5, 1.0, 'kaiming init')

Classifier Abstract Base Class

---

source

Classifier

 Classifier (nnet:torch.nn.modules.module.Module, num_classes:int,
             optimizer:Callable[...,torch.optim.optimizer.Optimizer],
             scheduler:Optional[Callable[...,Any]]=None)

Helper class that provides a standard way to create an ABC using inheritance.

	Type	Default	Details
nnet	Module
num_classes	int
optimizer	Callable		partial of optimizer
scheduler	Optional	None	partial of scheduler

---

source

plot_classifier_metrics_from_csv

 plot_classifier_metrics_from_csv (metrics_csv_path:str|os.PathLike)

Regressor Abstract Class

---

source

Regressor

 Regressor (nnet:lightning.pytorch.core.module.LightningModule,
            optimizer:Callable[...,torch.optim.optimizer.Optimizer],
            scheduler:Optional[Callable[...,Any]]=None)

Helper class that provides a standard way to create an ABC using inheritance.

	Type	Default	Details
nnet	LightningModule
optimizer	Callable		partial of optimizer
scheduler	Optional	None	partial of scheduler

Diffuser Abstract Class

---

source

Diffuser

 Diffuser (nnet:lightning.pytorch.core.module.LightningModule,
           optimizer:Callable[...,torch.optim.optimizer.Optimizer],
           scheduler:Optional[Callable[...,Any]]=None)

Helper class that provides a standard way to create an ABC using inheritance.

	Type	Default	Details
nnet	LightningModule
optimizer	Callable		partial of optimizer
scheduler	Optional	None	partial of scheduler

Sequential Model

---

source

SequentialModelX

 SequentialModelX (modules:List[torch.nn.modules.module.Module], *args,
                   **kwargs)

Helper class that provides a standard way to create an ABC using inheritance.

LR Finder Helper

# use LRFinder pythonm module (other version with lightning)

def find_optimal_lr(model, train_loader, criterion=None, optimizer=None, device='cuda'):
    # If no criterion provided, use default CrossEntropyLoss
    if criterion is None:
        criterion = nn.CrossEntropyLoss()
    
    # If no optimizer provided, use Adam
    if optimizer is None:
        optimizer = torch.optim.Adam(model.parameters(), lr=1e-7, weight_decay=1e-2)
    
    # Initialize LR Finder
    lr_finder = LRFinder(model, optimizer, criterion, device=device)
    
    # Run LR range test
    lr_finder.range_test(
        train_loader, 
        start_lr=1e-7,  # Very small starting learning rate
        end_lr=10,      # Large ending learning rate
        num_iter=100,   # Number of iterations to test
        smooth_f=0.05   # Smoothing factor for the loss
    )
    
    # Plot the learning rate vs loss
    lr_finder.plot(log_lr=True)
    
    # Suggest optimal learning rate
    suggested_lr = lr_finder.reset()
    
    print(f"Suggested Learning Rate: {suggested_lr}")
    
    return suggested_lr

---

source

lr_finder

 lr_finder (model:Callable[...,torch.nn.modules.module.Module],
            datamodule:nimrod.image.datasets.ImageDataModule,
            num_training:int=100, plot:bool=True)

	Type	Default	Details
model	Callable		partial model (missing optim & sched)
datamodule	ImageDataModule		data module
num_training	int	100	number of iterations
plot	bool	True	plot the learning rate vs loss

1-cycle train helper

---

source

train_one_cycle

 train_one_cycle (model:Callable[...,torch.nn.modules.module.Module],
                  datamodule:nimrod.image.datasets.ImageDataModule,
                  max_lr:float=0.1, weight_decay=1e-05, n_epochs:int=5,
                  project_name:str='MNIST-Classifier', tags=['arch',
                  'dev'], test:bool=True, run_name:str=None,
                  model_summary:bool=True, logger_cb:str='wandb',
                  precision='32-true')

train one cycle, adamW optim with wandb logging & learning rate monitor by default

	Type	Default	Details
model	Callable		partial model (missing optim & sched)
datamodule	ImageDataModule
max_lr	float	0.1
weight_decay	float	1e-05
n_epochs	int	5
project_name	str	MNIST-Classifier
tags	list	[‘arch’, ‘dev’]
test	bool	True
run_name	str	None
model_summary	bool	True
logger_cb	str	wandb
precision	str	32-true	16-mixed, 32-true

# data
cfg = OmegaConf.load('../config/data/image/fashion_mnist.yaml')
cfg.data_dir = "../data/image"
cfg.batch_size = 128
cfg.num_workers = 0
dm = instantiate(cfg)
dm.prepare_data()
dm.setup()

[21:54:44] INFO - Init ImageDataModule for fashion_mnist
[21:54:46] INFO - loading dataset fashion_mnist with args () from split train
[21:54:46] INFO - loading dataset fashion_mnist from split train
Overwrite dataset info from restored data version if exists.
[21:54:48] INFO - Overwrite dataset info from restored data version if exists.
Loading Dataset info from ../data/image/fashion_mnist/fashion_mnist/0.0.0/531be5e2ccc9dba0c201ad3ae567a4f3d16ecdd2
[21:54:48] INFO - Loading Dataset info from ../data/image/fashion_mnist/fashion_mnist/0.0.0/531be5e2ccc9dba0c201ad3ae567a4f3d16ecdd2
Found cached dataset fashion_mnist (/user/s/slegroux/Projects/nimrod/nbs/../data/image/fashion_mnist/fashion_mnist/0.0.0/531be5e2ccc9dba0c201ad3ae567a4f3d16ecdd2)
[21:54:48] INFO - Found cached dataset fashion_mnist (/user/s/slegroux/Projects/nimrod/nbs/../data/image/fashion_mnist/fashion_mnist/0.0.0/531be5e2ccc9dba0c201ad3ae567a4f3d16ecdd2)
Loading Dataset info from /user/s/slegroux/Projects/nimrod/nbs/../data/image/fashion_mnist/fashion_mnist/0.0.0/531be5e2ccc9dba0c201ad3ae567a4f3d16ecdd2
[21:54:48] INFO - Loading Dataset info from /user/s/slegroux/Projects/nimrod/nbs/../data/image/fashion_mnist/fashion_mnist/0.0.0/531be5e2ccc9dba0c201ad3ae567a4f3d16ecdd2
[21:54:52] INFO - loading dataset fashion_mnist with args () from split test
[21:54:52] INFO - loading dataset fashion_mnist from split test
Overwrite dataset info from restored data version if exists.
[21:54:53] INFO - Overwrite dataset info from restored data version if exists.
Loading Dataset info from ../data/image/fashion_mnist/fashion_mnist/0.0.0/531be5e2ccc9dba0c201ad3ae567a4f3d16ecdd2
[21:54:53] INFO - Loading Dataset info from ../data/image/fashion_mnist/fashion_mnist/0.0.0/531be5e2ccc9dba0c201ad3ae567a4f3d16ecdd2
Found cached dataset fashion_mnist (/user/s/slegroux/Projects/nimrod/nbs/../data/image/fashion_mnist/fashion_mnist/0.0.0/531be5e2ccc9dba0c201ad3ae567a4f3d16ecdd2)
[21:54:53] INFO - Found cached dataset fashion_mnist (/user/s/slegroux/Projects/nimrod/nbs/../data/image/fashion_mnist/fashion_mnist/0.0.0/531be5e2ccc9dba0c201ad3ae567a4f3d16ecdd2)
Loading Dataset info from /user/s/slegroux/Projects/nimrod/nbs/../data/image/fashion_mnist/fashion_mnist/0.0.0/531be5e2ccc9dba0c201ad3ae567a4f3d16ecdd2
[21:54:53] INFO - Loading Dataset info from /user/s/slegroux/Projects/nimrod/nbs/../data/image/fashion_mnist/fashion_mnist/0.0.0/531be5e2ccc9dba0c201ad3ae567a4f3d16ecdd2
[21:54:53] INFO - split train into train/val [0.8, 0.2]
[21:54:53] INFO - train: 48000 val: 12000, test: 10000

# model
cfg_model = OmegaConf.load('../config/model/image/convnetx.yaml')
feats_dim = [1, 8, 16, 32, 64, 128]
# feats_dim = [1, 4, 8, 16, 8]
# feats_dim = [1, 16, 32, 64, 32]
cfg_model.nnet.n_features = feats_dim
model = instantiate(cfg_model) #partial
do_lr_finder = False

if do_lr_finder:
    suggested_lr = lr_finder(model=model, datamodule=dm, plot=True)
else:
    suggested_lr = 1e-3

# train
N_EPOCHS = 1

project_name = "FASHION-MNIST-Classifier"
run_name = f"{model.func.__name__}-bs:{dm.batch_size}-epochs:{N_EPOCHS}"
tags = [f"feats:{feats_dim}", f"bs:{dm.batch_size}", f"epochs:{N_EPOCHS}"]

trained_model, best_ckpt = train_one_cycle(
    model,
    dm,
    n_epochs=N_EPOCHS,
    max_lr=suggested_lr,
    project_name=project_name,
    tags=tags,
    run_name=run_name
    )

GPU available: True (cuda), used: True
TPU available: False, using: 0 TPU cores
HPU available: False, using: 0 HPUs
[21:54:54] INFO - ConvNetX: init
[21:54:54] INFO - Classifier: init
/user/s/slegroux/miniconda3/envs/nimrod/lib/python3.11/site-packages/lightning/pytorch/utilities/parsing.py:209: Attribute 'nnet' is an instance of `nn.Module` and is already saved during checkpointing. It is recommended to ignore them using `self.save_hyperparameters(ignore=['nnet'])`.

==========================================================================================
Layer (type:depth-idx)                   Output Shape              Param #
==========================================================================================
ConvNet                                  [128, 10]                 --
├─Sequential: 1-1                        [128, 10]                 --
│    └─ConvLayer: 2-1                    [128, 8, 32, 32]          --
│    │    └─Sequential: 3-1              [128, 8, 32, 32]          88
│    └─ConvLayer: 2-2                    [128, 16, 16, 16]         --
│    │    └─Sequential: 3-2              [128, 16, 16, 16]         1,184
│    └─ConvLayer: 2-3                    [128, 32, 8, 8]           --
│    │    └─Sequential: 3-3              [128, 32, 8, 8]           4,672
│    └─ConvLayer: 2-4                    [128, 64, 4, 4]           --
│    │    └─Sequential: 3-4              [128, 64, 4, 4]           18,560
│    └─ConvLayer: 2-5                    [128, 128, 2, 2]          --
│    │    └─Sequential: 3-5              [128, 128, 2, 2]          73,984
│    └─ConvLayer: 2-6                    [128, 10, 1, 1]           --
│    │    └─Sequential: 3-6              [128, 10, 1, 1]           11,540
│    └─Flatten: 2-7                      [128, 10]                 --
==========================================================================================
Total params: 110,028
Trainable params: 110,028
Non-trainable params: 0
Total mult-adds (Units.MEGABYTES): 161.97
==========================================================================================
Input size (MB): 0.52
Forward/backward pass size (MB): 32.53
Params size (MB): 0.44
Estimated Total Size (MB): 33.49
==========================================================================================

Tracking run with wandb version 0.19.1

Run data is saved locally in /tmp/wandb/run-20250206_215454-1fc1d7re

Syncing run ConvNetX-bs:128-epochs:1 to Weights & Biases (docs)

View project at https://wandb.ai/slegroux/FASHION-MNIST-Classifier

View run at https://wandb.ai/slegroux/FASHION-MNIST-Classifier/runs/1fc1d7re

/user/s/slegroux/miniconda3/envs/nimrod/lib/python3.11/site-packages/lightning/pytorch/callbacks/model_checkpoint.py:654: Checkpoint directory /user/s/slegroux/Projects/nimrod/nbs/checkpoints/FASHION-MNIST-Classifier/ConvNetX-bs:128-epochs:1 exists and is not empty.
LOCAL_RANK: 0 - CUDA_VISIBLE_DEVICES: [0]
[21:54:54] INFO - Optimizer: <class 'torch.optim.adamw.AdamW'>
[21:54:54] INFO - Scheduler: <class 'torch.optim.lr_scheduler.OneCycleLR'>

  | Name         | Type               | Params | Mode 
------------------------------------------------------------
0 | nnet         | ConvNet            | 110 K  | train
1 | loss         | CrossEntropyLoss   | 0      | train
2 | train_acc    | MulticlassAccuracy | 0      | train
3 | val_acc      | MulticlassAccuracy | 0      | train
4 | test_acc     | MulticlassAccuracy | 0      | train
5 | train_loss   | MeanMetric         | 0      | train
6 | val_loss     | MeanMetric         | 0      | train
7 | test_loss    | MeanMetric         | 0      | train
8 | val_acc_best | MaxMetric          | 0      | train
------------------------------------------------------------
110 K     Trainable params
0         Non-trainable params
110 K     Total params
0.440     Total estimated model params size (MB)
41        Modules in train mode
0         Modules in eval mode

/user/s/slegroux/miniconda3/envs/nimrod/lib/python3.11/site-packages/lightning/pytorch/trainer/connectors/data_connector.py:425: The 'val_dataloader' does not have many workers which may be a bottleneck. Consider increasing the value of the `num_workers` argument` to `num_workers=23` in the `DataLoader` to improve performance.
/user/s/slegroux/miniconda3/envs/nimrod/lib/python3.11/site-packages/lightning/pytorch/trainer/connectors/data_connector.py:425: The 'train_dataloader' does not have many workers which may be a bottleneck. Consider increasing the value of the `num_workers` argument` to `num_workers=23` in the `DataLoader` to improve performance.

`Trainer.fit` stopped: `max_epochs=1` reached.
LOCAL_RANK: 0 - CUDA_VISIBLE_DEVICES: [0]
/user/s/slegroux/miniconda3/envs/nimrod/lib/python3.11/site-packages/lightning/pytorch/trainer/connectors/data_connector.py:425: The 'test_dataloader' does not have many workers which may be a bottleneck. Consider increasing the value of the `num_workers` argument` to `num_workers=23` in the `DataLoader` to improve performance.

┏━━━━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━━━━━━━━━━━━━┓
┃        Test metric        ┃       DataLoader 0        ┃
┡━━━━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━━━━━━━━━━━━━┩
│         test/acc          │    0.8658000230789185     │
│         test/loss         │    0.5964178442955017     │
└───────────────────────────┴───────────────────────────┘

[21:55:01] INFO - Best ckpt path: /user/s/slegroux/Projects/nimrod/nbs/checkpoints/FASHION-MNIST-Classifier/ConvNetX-bs:128-epochs:1/0-0.57.ckpt

Run history:

epoch	▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁▁
lr-AdamW	▁▁▂▂▃▃▄▅▅▅▆▆▇██████████▇▇▆▅▅▅▅▄▄▄▃▃▂▁▁▁▁
test/acc	▁
test/loss	▁
train/acc_epoch	▁
train/acc_step	▁▂▄▅▅▇▇▇▆▇▇▇▇▇▇▇▇▇█▇▇█▇▇██▇▇███▇▇██▇██▇█
train/loss_epoch	▁
train/loss_step	█▇▆▆▆▄▃▄▃▃▂▂▂▂▂▂▂▂▁▂▁▁▁▁▁▁▁▁▂▁▁▁▁▁▁▁▁▁▁▁
trainer/global_step	▁▁▁▁▂▃▃▃▃▃▄▄▄▄▄▄▄▄▅▅▅▅▆▆▆▆▆▆▇▇▇▇▇███████
val/acc	▁
val/acc_best	▁
val/loss	▁

Run summary:

epoch	1
lr-AdamW	0.0
test/acc	0.8658
test/loss	0.59642
train/acc_epoch	0.79585
train/acc_step	0.89844
train/loss_epoch	0.84797
train/loss_step	0.56174
trainer/global_step	375
val/acc	0.87608
val/acc_best	0.87608
val/loss	0.5739

View run ConvNetX-bs:128-epochs:1 at: https://wandb.ai/slegroux/FASHION-MNIST-Classifier/runs/1fc1d7re
View project at: https://wandb.ai/slegroux/FASHION-MNIST-Classifier
Synced 6 W&B file(s), 0 media file(s), 7 artifact file(s) and 0 other file(s)

Find logs at: /tmp/wandb/run-20250206_215454-1fc1d7re/logs

print(best_ckpt)
x = torch.randn(1, 1, 32, 32)
trained_model.eval()
trained_model(x)

/user/s/slegroux/Projects/nimrod/nbs/checkpoints/FASHION-MNIST-Classifier/ConvNetX-bs:128-epochs:1/0-0.58.ckpt

tensor([[0.0000, 0.0000, 0.5579, 0.1776, 0.0000, 0.3419, 0.0000, 0.0000, 2.2810,
         0.0000]], grad_fn=<ViewBackward0>)