Fix Hardcoded steps_per_epoch=512 to Dynamic Calculation #391

[ljluestc]

Original Problem:

• Training PSPNet with different dataset sizes (2975 images vs 3 images) always showed "512/512" iterations
• Changing batch_size parameter had no effect on iteration count
• Users reported: "The number of iterations is always 512. Either using a training dataset folder with 2975 images or folder with 3 images."

Root Cause:
The train() function in keras_segmentation/train.py had hardcoded parameters:
def train(..., steps_per_epoch=512, val_steps_per_epoch=512, ...):This meant regardless of dataset size, training would always attempt 512 steps per epoch.

Solution Implementation

Core Fix: Dynamic steps_per_epoch Calculation

Modified keras_segmentation/train.py:

1. Changed function signature to allow None values (dynamic calculation):
   def train(..., steps_per_epoch=None, val_steps_per_epoch=None, ...):2. Added dynamic calculation logic before model.fit():

Calculate steps_per_epoch dynamically if not provided

if steps_per_epoch is None:
from .data_utils.data_loader import get_pairs_from_paths
img_seg_pairs = get_pairs_from_paths(train_images, train_annotations)
total_train_samples = len(img_seg_pairs)
steps_per_epoch = total_train_samples // batch_size
if steps_per_epoch == 0:
steps_per_epoch = 1 # Minimum 1 step per epoch
print(f"Calculated steps_per_epoch: {steps_per_epoch} (from {total_train_samples} samples, batch_size={batch_size})")

Calculate val_steps_per_epoch dynamically if validation enabled

if validate and val_steps_per_epoch is None:
from .data_utils.data_loader import get_pairs_from_paths
val_img_seg_pairs = get_pairs_from_paths(val_images, val_annotations)
total_val_samples = len(val_img_seg_pairs)
val_steps_per_epoch = total_val_samples // val_batch_size
if val_steps_per_epoch == 0:
val_steps_per_epoch = 1 # Minimum 1 step per epoch
print(f"Calculated val_steps_per_epoch: {val_steps_per_epoch} (from {total_val_samples} samples, val_batch_size={val_batch_size})")3. Updated CLI interface (keras_segmentation/cli_interface.py):
parser.add_argument("--steps_per_epoch", type=int, default=None) # Changed from 512### 🧪 Comprehensive Testing Suite

Added test/unit/test_steps_per_epoch.py with 5 comprehensive tests:

1. Small Dataset Test: 10 samples, batch_size=2 → steps_per_epoch=5
2. Large Dataset Test: 100 samples, batch_size=2 → steps_per_epoch=50 (not 512!)
3. Validation Test: Both train and val steps calculated correctly
4. Explicit Override Test: Explicit parameter still works when provided
5. Minimum Steps Test: Ensures minimum 1 step even with tiny datasets