# Copyright 2025 - Oumi
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
"""The CNNClassifier model provides a basic example how to use ConvNets in Oumi."""
from typing import Callable, Optional
import torch
import torch.nn as nn
from torch.nn import functional as F
from oumi.core import registry
from oumi.core.models.base_model import BaseModel
[docs]
@registry.register("CnnClassifier", registry.RegistryType.MODEL)
class CNNClassifier(BaseModel):
"""A simple ConvNet for classification of small fixed-size images."""
def __init__(
self,
image_width: int,
image_height: int,
*,
in_channels: int = 3,
output_dim: int = 10,
kernel_size: int = 5,
):
"""Initialize the ConvNet for image classification.
Args:
image_width: Width of input images in pixels.
image_height: Height of input images in pixels.
in_channels: The number of input channels e.g., 3 for RGB, 1 for greyscale.
output_dim: The output dimension i.e., the number of classes.
kernel_size: Convolutional kernel size.
"""
super().__init__()
self.conv1 = nn.Conv2d(in_channels, 32, kernel_size=kernel_size)
w, h = self._compute_next_level_image_size(
image_width, image_height, kernel_size=kernel_size, halve=False
)
self.conv2 = nn.Conv2d(32, 32, kernel_size=kernel_size)
w, h = self._compute_next_level_image_size(
w, h, kernel_size=kernel_size, halve=True
)
self.conv3 = nn.Conv2d(32, 64, kernel_size=kernel_size)
w, h = self._compute_next_level_image_size(
w, h, kernel_size=kernel_size, halve=True
)
self._final_image_width = w
self._final_image_height = h
self.fc1 = nn.Linear(
self._final_image_width * self._final_image_height * 64, 256
)
self.fc2 = nn.Linear(256, output_dim)
@staticmethod
def _compute_next_level_image_size(
w: int, h: int, kernel_size: int, halve: bool
) -> tuple[int, int]:
w, h = (w - (kernel_size - 1)), (h - (kernel_size - 1))
if halve:
w, h = (w // 2), (h // 2)
if w <= 0 or h <= 0:
raise ValueError(f"Image is too small for kernel_size={kernel_size}")
return (w, h)
[docs]
def forward(
self,
images: torch.Tensor,
labels: Optional[torch.LongTensor] = None,
**kwargs,
) -> dict[str, torch.Tensor]:
"""Forward pass of the model."""
# Whether to apply dropout. `False` corresponds to inference mode.
training_mode = labels is not None
x = F.relu(self.conv1(images))
x = F.relu(F.max_pool2d(self.conv2(x), 2))
x = F.dropout(x, p=0.5, training=training_mode)
x = F.relu(F.max_pool2d(self.conv3(x), 2))
x = F.dropout(x, p=0.5, training=training_mode)
x = x.view(-1, self._final_image_width * self._final_image_height * 64)
x = F.relu(self.fc1(x))
x = F.dropout(x, training=training_mode)
logits = self.fc2(x)
outputs = {"logits": logits}
if training_mode:
targets = F.log_softmax(logits, dim=1)
loss = self.criterion(targets, labels)
outputs["loss"] = loss
return outputs
@property
def criterion(self) -> Callable:
"""Returns the criterion function to compute loss."""
return F.nll_loss
