Open gc by default

[sneaxiy]

Enable GC when PADDLE_ON_INFERENCE is off.

This PR also fixes bug of GC when using conditional_block op.

See CE Debug here.

There are 3 models in CE Debug failed.

[22:30:44][Step 3/3] task: model_cycle_gan
[22:30:44][Step 3/3] passed:  False
[22:30:44][Step 3/3] infos [d_train_cost] pass
[22:30:44][Step 3/3] [g_train_cost] pass
[22:30:44][Step 3/3] [duration] failed, diff ratio: 0.06092244084538961 larger than 0.05.
[22:30:44][Step 3/3] kpis keys ['d_train_cost', 'g_train_cost', 'duration']
[22:30:44][Step 3/3] kpis values [[[2.12036180496]], [[12.1869783401]], [[1.05452570915]]]
[22:30:44][Step 3/3] task: model_icnet
[22:30:44][Step 3/3] passed:  False
[22:30:44][Step 3/3] infos [train_cost] pass
[22:30:44][Step 3/3] [train_duration] failed, diff ratio: 0.04116670226811877 larger than 0.02.
[22:30:44][Step 3/3] kpis keys ['train_cost', 'train_duration']
[22:30:44][Step 3/3] kpis values [[[2.1491]], [[41.469499]]]
[22:30:44][Step 3/3] task: model_ocr_recognition
[22:30:44][Step 3/3] passed:  False
[22:30:44][Step 3/3] infos [train_cost] pass
[22:30:44][Step 3/3] [train_duration] failed, diff ratio: 0.06440406463978306 larger than 0.06.
[22:30:44][Step 3/3] kpis keys ['train_cost', 'train_duration']
[22:30:44][Step 3/3] kpis values [[[58.981466]], [[46.207872]]]
[22:31:02][Step 3/3] model_cycle_gan model_icnet model_ocr_recognition 

• model_cycle_gan: CE Debug uses legacy CycleGAN code in release/1.1 branch, which uses Executor and there is no cache. In develop branch of models, GC is enabled in CycleGAN model.
• model_icnet: I find that this model fails randomly because I have tested a PR which does nothing to do with memory optimization. It also fails with 4.48% differences in train duration. See here.
• model_ocr_recognition: do not know why. The results I tested on my machine is 39.386209s(develop) vs 39.1990187s(This PR). The difference is very small.

I have tried to run CE Debug for the second time. All tasks have passed. Please see here.

[chengduoZH]

Do you mean that if the max_memory_size_ is little 1, there doesn't need a lock?

[sneaxiy]

Exactly.

[chengduoZH]

LGTM

[chengduoZH]

In fact, this checking is unnecessary.