def test_expand_ratio(self):
model = models.create_efficientnet("efficientnet-b0", expand_ratio=2)
for i, block in enumerate(model.blocks):
if i == 0:
continue
for layer in block:
conv = layer.conv_pw
assert conv.out_channels == conv.in_channels * 2
def train():
print("[INFO] Loading and cropping all raw training receipt images...")
images, quality = dataset.load_data(config.TRAIN_IMAGE_FOLDER_PATH, config.TRAIN_CSV_PATH, config.INPUT_SIZE)
split = train_test_split(quality, images, test_size=0.25, random_state=42)
(trainY, validY, trainX, validX) = split
mc = ModelCheckpoint(config.OUTPUT_MODEL_PATH, monitor='val_loss', mode='min', verbose=1, save_best_only=True)
model = models.create_efficientnet(config.INPUT_SIZE, config.INPUT_SIZE, 3, config.MODEL_BASE, config.FIRST_LAYERS_TO_FREEZE)
opt = Adam(lr=config.LEARNING_RATE, decay=config.LEARNING_RATE / config.EPOCHS)
model.compile(loss="mean_squared_error", optimizer=opt)
# Print the training config
print(f"[INFO] Training config: \n\
- MODEL_BASE: {config.MODEL_BASE}\n\
- INPUT_SIZE: {str(config.INPUT_SIZE)}x{str(config.INPUT_SIZE)}\n\
- EPOCHS: {str(config.EPOCHS)}\n\
- BATCH_SIZE: {str(config.BATCH_SIZE)}\n\
- LEARNING_RATE: {str(config.LEARNING_RATE)}\n")
# Train the model
print("[INFO] Training model...")
H = model.fit(x=trainX, y=trainY,
validation_data=(validX, validY),
epochs=config.EPOCHS, batch_size=config.BATCH_SIZE, callbacks=[mc])
# Save Figure
print(f"[INFO] Saving the training firgure...")
save_trainging_plot(H, config.OUTPUT_TRAINING_FIG_PATH)
# Load saved model
model = load_model(config.OUTPUT_MODEL_PATH)
# make predictions on the testing data
print("[INFO] Predicting receipt quality on validation data...")
preds = model.predict(validX)
rmse = math.sqrt(metrics.mean_squared_error(validY, preds))
print(f"[INFO] RMSE on validation data: {rmse}")
def test_group_dim(self):
model = models.create_efficientnet("efficientnet-b0", group_dim=2)
for block in model._blocks:
conv = block._depthwise_conv
assert conv.groups * 2 == conv.in_channels
def test_expand_ratio(self):
model = models.create_efficientnet("efficientnet-b0", expand_ratio=2)
for idx, block in enumerate(model._blocks):
if idx > 0:
conv = block._expand_conv
assert conv.out_channels == conv.in_channels * 2
def test_group_dim(self):
model = models.create_efficientnet("efficientnet-b0", group_dim=2)
for block in model.blocks:
for layer in block:
conv = layer.conv_dw
assert conv.groups * 2 == conv.in_channels