def train(network_specs, training_params, image_path, save_path, ckpt_path):
print('creating datapipe...')
# create images DataPipeline
datapipe = DataPipeline(image_path=image_path,
training_params=training_params)
print('creating network model...')
# create model VAE
model = UNet(network_specs=network_specs,
datapipe=datapipe,
training_params=training_params,
mode='evaluating')
# train the model
# save_config is flexible
print('''
=============
HERE WE GO
=============
''')
images, labels, preds = model.evaluate(ckpt_path=ckpt_path)
for i in range(labels.shape[0]):
# plt.subplots(figsize=[16,12])
for j in range(3):
plt.subplot(2, 6, 2 * j + 1)
plt.imshow(labels[i, :, :, j])
plt.subplot(2, 6, 2 * j + 2)
plt.imshow(preds[i, :, :, j])
plt.subplot(2, 6, 7)
plt.imshow(np.squeeze(images[i]))
plt.show()
_f2s(result[m])
for m in ['mean_iou', 'mean_AP', 'mean_AR', 'f1_score']
]
score_table.append(f"|{'|'.join(average)}|")
report = OUTPUT_FORMAT.format(date=d,
runtime=runtime,
config_table='\n'.join(config_table),
score_table='\n'.join(score_table))
CONFIG.EVAL_DIR.mkdir(parents=True, exist_ok=True)
with open(CONFIG.EVAL_DIR / f'report_{d}.md', 'w') as f:
f.write(report)
def _f2s(v, ndigits=4):
return f'{round(v, ndigits)}'
if __name__ == '__main__':
data_gen_val = DataGenerator('val')
ds_val = data_gen_val.get_one_shot_iterator()
model = UNet(input_shape=(data_gen_val.H, data_gen_val.W, 3),
class_num=CONFIG.CLASS_NUM)
result = model.evaluate(ds=ds_val,
steps=data_gen_val.data_length //
CONFIG.BATCH_SIZE)
report(result=result)
