Find Difficulty in Reproduction

[PKUterran]

It seems that your model can only work on netlists with num_cell = 710. I have to make a lot of revisions to fit DeepPlace to other datasets.

[Rogerity]

Thanks for you comments. Could you please give some examples for us to check? We look forward to your commit if it is a common issue.

[PKUTAN]

As for on aspect, the macro netlist for adaptec1,2,4 and other dataset is missing and hard to construct a similar netlist followed by n_egde_710.dat. Once the data is completed , the code in your work is easy to revise

[PKUterran]

https://github.com/PKUterran/DeepPlace/blob/8e339991430989d82012262edba18641109590f7/a2c_ppo_acktr/model.py#L114
For example, if I'm working on a netlist with 1030 macros, should I change n's value to 1029 and train a agent from start, rather than using that one you trained on n_egde_710.dat?

[PKUterran]

So far, I have made these revisions in my fork to make the code flexible to more netlists:

• Some variables in main.py (should be named train.py?) and validation.py can be configurated by argument input.
• In place_env.py, add more input args to class Placement and some functions to fit netlists with num_cell != 710 and grid_size != 32.
• Function build_graph in a2c_ppo_acktr/model.py is not friendly for reproduction. I tried to treat graph as an input argument of the functions in class Policy, but revising Policy.evaluate_actions(...) caused a chain of revisions which is hard to handle.

Some revisions have been update in my fork https://github.com/PKUterran/DeepPlace. I may submit a pull request if it's ok to use.

[PKUterran]

As for on aspect, the macro netlist for adaptec1,2,4 and other dataset is missing and hard to construct a similar netlist followed by n_egde_710.dat. Once the data is completed , the code in your work is easy to revise

I also find this problem. For example, if I mask all the standard cells in netlist ispd2015/mgc_des_perf_1, there is only one macro in most nets. It's too sparse:

{"num_cell": 378, "num_net": 4114, "num_pin": 4147}

[PKUTAN]

https://github.com/PKUterran/DeepPlace/blob/8e339991430989d82012262edba18641109590f7/a2c_ppo_acktr/model.py#L114 For example, if I'm working on a netlist with 1030 macros, should I change n's value to 1029 and train a agent from start, rather than using that one you trained on n_egde_710.dat?

How can you train a chip with 1030 macros using the model trained on n_edge_710? (I think that the parameter number do not match)

[PKUterran]

https://github.com/PKUterran/DeepPlace/blob/8e339991430989d82012262edba18641109590f7/a2c_ppo_acktr/model.py#L114 For example, if I'm working on a netlist with 1030 macros, should I change n's value to 1029 and train a agent from start, rather than using that one you trained on n_egde_710.dat?

How can you train a chip with 1030 macros using the model trained on n_edge_710? (I think that the parameter number do not match)

Indeed