How to use with CPU

[haowenli]

I am using the transformer model which is an example in README. I haven't NVIDIA card in my computer. How can I use HParams to train transformer model in README? HParams (by model):

• attention: ['attention_lm_base', 'attention_lm_moe_base', 'attention_lm_moe_large', 'attention_lm_moe_small', 'attention_lm_moe_tiny', 'attention_lm_no_moe_small', 'attention_lm_small', 'attention_lm_translation']
• basic: ['basic_1']
• bluenet: ['bluenet_base', 'bluenet_tiny']
• bytenet: ['bytenet_base']
• cycle: ['cycle_gan_small']
• gene: ['gene_expression_conv_base']
• lstm: ['lstm_attention']
• multimodel: ['multimodel_base', 'multimodel_tiny']
• neuralgpu: ['neuralgpu_1']
• shakeshake: ['shakeshake_cifar10']
• slicenet: ['slicenet_1', 'slicenet_1noam', 'slicenet_1tiny']
• transformer: ['transformer_alt', 'transformer_base', 'transformer_base_single_gpu', 'transformer_big', 'transformer_big_dr1', 'transformer_big_dr2', 'transformer_big_enfr', 'transformer_big_single_gpu', 'transformer_dr0', 'transformer_dr2', 'transformer_ff1024', 'transformer_ff4096', 'transformer_h1', 'transformer_h16', 'transformer_h32', 'transformer_h4', 'transformer_hs1024', 'transformer_hs256', 'transformer_k128', 'transformer_k256', 'transformer_l10', 'transformer_l2', 'transformer_l4', 'transformer_l8', 'transformer_ls0', 'transformer_ls2', 'transformer_moe_1b', 'transformer_moe_base', 'transformer_n_da', 'transformer_n_da_l10', 'transformer_no_moe', 'transformer_parameter_attention_a', 'transformer_parameter_attention_b', 'transformer_parsing_base', 'transformer_parsing_big', 'transformer_parsing_ice', 'transformer_small', 'transformer_tiny', 'transformer_vae_base', 'transformer_vae_small']
• xception: ['xception_base', 'xception_tiny']

[martinpopel]

Without GPU, the training (and decoding) will be very slow, even when using a fast CPU with e.g. 32 cores. I would suggest to use some of the toy problems (algorithmic). If you really want to try machine translation, use transformer_tiny and the TranslateEndeWmt8k problem (which used to be called wmt_ende_tokens_8k in older t2t versions), which is likely to train fast (but with lower final BLEU, of course). If TensorFlow does not find a GPU, it automatically switches to CPU.