Performance of frrnA on Cityscapes

[jthsieh]

Thank you for providing the code! I ran the frrnA model on Cityscapes and was hoping to replicate the results reported in the paper (FRRN A: 0.630 mean IoU), python train.py --arch frrnA --dataset cityscapes --img_rows 256 --img_cols 512 --n_epoch 35 --batch_size 3 --l_rate 0.001 And then I continue training with learning rate 1e-4 for 10 epochs.

This is roughly the same training setup and hyperparameters as in the paper. However, I only got 0.50 mean IoU. I'm wondering if I'm doing something wrong. Were you able to replicate the results?

[meetps]

Nope, I wasn't able to match the paper results.

After careful perusal of the paper, the only difference seems be that of the optimizer (Adam [We use SGD with momentum]) and that the paper uses nearest neighbour interpolation in the FRRUs, we use bilinear interpolation.

[victorjdcom]

i am using this config file for the cityscapes:

model: arch: fcn8s data: dataset: cityscapes train_split: train val_split: val img_rows: '256' img_cols: '512' path: /home/victor/data/cityscapes training: train_iters: 300000 batch_size: 1 val_interval: 1000 n_workers: 16 print_interval: 50 optimizer: name: 'sgd' lr: 1.0e-10 weight_decay: 0.0005 momentum: 0.99 loss: name: 'cross_entropy' size_average: False lr_schedule: resume: fcn8s_cityscape_best_model.pkl

got error for: "No files for split=[%s] found in %s" % (split, self.images_base)

any idea? thanks