the best reconstruct task result has some different with the paper.

[robotzheng]

1. epoch 34;
2. replace 'test' by 'val';
3. self.permutation not used in demo.py.

maybe some different options? please give me some sign!

[robotzheng]

now, the best result is:
Val BLEU (2,3,4): 0.95 0.935 0.921
Iteration 126600: loss 0.092868

[robotzheng]

it has more iterations than paper's, why?

[robotzheng]

same options as the paper:
Use model cnn_deconv
Use 3 conv/deconv layers
{'restore': True, 'layer': 3, 'fix_emb': False, 'log_path': './log', 'sent_len2': 129, 'substitution': 's', 'sent_len4': 30, 'filter_size': 300, 'max_epochs': 100, 'maxlen': 253, 'discrimination': False, 'embed_size': 300, 'filter_shape': 5, 'sent_len3': 63, 'lr': 1e-05, 'save_path': './save/hotel_900_dim_cnn_deconv_s0', 'is_train': True, 'valid_freq': 100, 'dropout': True, 'reuse_w': True, 'L': 100, 'batch_size': 32, 'n_gan': 900, 'n_words': 25002, 'H_dis': 300, 'dropout_ratio': 1.0, 'reuse_cnn': False, 'W_emb': None, 'cnn_W': None, 'tanh': True, 'reuse_discrimination': True, 'print_freq': 100, 'cnn_b': None, 'sent_len': 261, 'stride': [2, 2, 2], 'cnn_layer_dropout': False, 'permutation': 0, 'model': 'cnn_deconv', 'batch_norm': False}
Total words: 25002
No embedding file found.
Encoder len 261 Decoder len 261 Output len 261

[robotzheng]

now, the best result is:
Val BLEU (2,3,4): 0.985 0.98 0.975
Iteration 396000: loss 0.022921

[wwmmqq]

An onlooker. 🙃

[albertom-pirovano]

@robotzheng how did you improve the performance from Val BLEU (2,3,4): 0.669 0.579 0.505 to Val BLEU (2,3,4): 0.985 0.98 0.975 ?
Thank you