Test #8
Test #8
Conversation
…txt, change generate vector formation
kochkov92
left a comment
kochkov92
left a comment
There was a problem hiding this comment.
Thanks for fixing some typos. I think having bonds_file is a good idea.
-) I suggest to stay compatible with the ED code on the wave function format at least for now.
-) I think rescaling at psi_target is more intuitive for someone who is reading the code.
cgs_vmc/evaluation.py
Outdated
| config_ind = 0 | ||
| for i in range(0, batch_size): | ||
| out_file.write('({},{})\n'.format(psi[i].real, psi[i].imag)) | ||
| out_file.write('{} {}\n'.format(psi[i].real, psi[i].imag)) |
There was a problem hiding this comment.
I think for the purpose of testing vectors with python we can just use pandas to read-in files. In this case we maintain uniformity with ED code, which is a nice tool to test and compare against. Here's a possible way to read the current format:
def get_wf(file_name):
df = pd.read_csv(file_name)
df = df.apply(lambda col: col.apply(lambda val: complex(val.strip('()'))))
return df.values[:, 0]
cgs_vmc/evaluation.py
Outdated
| psi = session.run(wavefunction_value, feed_dict=feed_dict) | ||
| for i in range(0, config_ind): | ||
| out_file.write('({},{})\n'.format(psi[i].real, psi[i].imag)) | ||
| out_file.write('{} {}\n'.format(psi[i].real, psi[i].imag)) |
| for _ in range(normalization_iterations): | ||
| session.run(update_config) | ||
| session.run(update_norm_on_batch) | ||
|
|
There was a problem hiding this comment.
Since we are not deploying this part yet, I would move it to prototypes; then we can recover this part of the code when needed.
cgs_vmc/normalizer.py
Outdated
|
|
||
|
|
||
| def run_normalization_ops( | ||
| max_value: Any, |
There was a problem hiding this comment.
I think max_value is probably a np.float or something.
cgs_vmc/run_energy_evaluation.py
Outdated
| 'Full path to the checkpoint directory.') | ||
|
|
||
| flags.DEFINE_string( | ||
| 'bond_file', '', |
There was a problem hiding this comment.
nit: bonds_file
Name of the file that contains bonds of the Hamiltonian.
cgs_vmc/run_training.py
Outdated
| 'checkpoint_dir', '', | ||
| 'Full path to the checkpoint directory.') | ||
| flags.DEFINE_string( | ||
| 'bond_file', '', |
There was a problem hiding this comment.
Same as for run_supervised_training.py
cgs_vmc/training.py
Outdated
|
|
||
| psi = wavefunction(configs) | ||
| psi_target = target_wavefunction(configs) | ||
| psi_target = target_wavefunction(configs) * hparams.scale |
There was a problem hiding this comment.
Since LogOverlapSWO doesn't depend on the norm mismatch there is no reason to scale psi_target here.
cgs_vmc/training.py
Outdated
|
|
||
| loss = tf.reduce_mean( | ||
| tf.squared_difference(psi, psi_target) | ||
| tf.squared_difference(psi, psi_target * hparams.scale) |
There was a problem hiding this comment.
I would move scaling up in the code to improve readability. One naturally would expect target to be target, not a target before scaling.
cgs_vmc/training.py
Outdated
|
|
||
| loss = tf.reduce_mean( | ||
| tf.squared_difference(psi, psi_target * np.sqrt(2 ** n_sites))) | ||
| tf.squared_difference(psi, psi_target * hparams.scale)) |
There was a problem hiding this comment.
I think here you also want to move the the scaling up, so that psi_target has the scale you want.
cgs_vmc/training.py
Outdated
|
|
||
| loss = tf.reduce_mean( | ||
| tf.squared_difference(psi, psi_target * np.sqrt(2**n_sites)) / | ||
| tf.squared_difference(psi, psi_target * hparams.scale) / |
There was a problem hiding this comment.
I haven't thought about it before, but it makes more sense to put scale up to the psi_target, simply improves readability. (otherwise it's confusing that we are matching scaled target)
|
For normalizer.py, do you suggest to put the whole file into prototypes, or just keep the first function? |
No description provided.