APT Loss Functions #58
Conversation
Update point source test to have a consistent ab_zeropoint, ensure bright ps to offset how bright the source at low redshift is.
- Consistent source/ps center_x,center_y - Up the magnitude of the PS
| """Helper function to convert mu, prec_matrix to normalized parameter | ||
| space | ||
|
|
||
| Args: |
There was a problem hiding this comment.
I'm not familiar with this string formatting. Does that shape information translate to the read the docs?
Remove the indentation from Returns (should have same indentation as Args)
| @@ -120,6 +120,37 @@ def unnormalize_outputs(input_norm_path,learning_params,mean,standard_dev=None, | |||
| if cov_mat is not None: | |||
| cov_mat[:,lpi,:] *= param_std | |||
| cov_mat[:,:,lpi] *= param_std | |||
|
|
|||
| # TODO: write test after moving (make sure identity operation w/ unnormalized) | |||
There was a problem hiding this comment.
I think you've added a test for this now so please remove the TODO.
| self.proposal_prec = tf.constant(proposal_prec,dtype=tf.float32) | ||
|
|
||
| @staticmethod | ||
| def log_gauss_full(y_true,y_pred,prec_mat): |
There was a problem hiding this comment.
Is this the same function as the class you're inheriting? If so there's no need to redefine it. Or am I missing an important detail?
| return tf.reduce_min(loss_stack,axis=-1) | ||
|
|
||
|
|
||
| class DiagonalCovarianceAPTLoss(DiagonalCovarianceLoss): |
There was a problem hiding this comment.
I think if you inherit from FullCovarianceAPTLoss you don't have to repeat the init
| norm_dict = Analysis.dataset_generation.normalize_outputs(metadata, | ||
| learning_params,input_norm_path) | ||
|
|
||
| # Actual mean/precision matrix output by a network for these params |
There was a problem hiding this comment.
The formatting of the spacing looks odd here
|
|
||
| # create APT loss object | ||
| input_norm_path = 'test_data/apt_norms.csv' | ||
| #'/Users/smericks/Desktop/StrongLensing/STRIDES14results/sep7_narrow_lognorm/lr_1e-3/norms.csv' |
| mu_prior, prec_prior, mu_prop,prec_prop,input_norm_path=input_norm_path) | ||
| # move to normalized space | ||
| # MU_PRIOR, PREC_PRIOR, MU_PROP, PREC_PROP ALREADY MODIFIED B/C PASS BY REFERENCE | ||
| # (TODO: change how this is handled to avoid further issues) |
| truth1 = tf.constant(truth1,dtype=tf.float32) | ||
| truth1_batched = tf.squeeze(tf.stack([truth1,truth1])) | ||
|
|
||
| # CONFIRMED: the problem is NOT the prefactor |
There was a problem hiding this comment.
Is this comment still relevant?
| # Add point source and validate output | ||
| self.c.sample['point_source_parameters'] = {'x_point_source':0.001, | ||
| 'y_point_source':0.001,'magnitude':22,'output_ab_zeropoint':25.95, | ||
| self.c.sample['point_source_parameters'] = {'x_point_source':0.0, |
| @@ -0,0 +1,9 @@ | |||
| parameter,mean,std | |||
| main_deflector_parameters_theta_E,0.7992458031844867,0.15239538119453927 | |||
There was a problem hiding this comment.
I'm guessing this is pulled from an actual norm file. If you manually change the values to be more exact (0.8, 0.15, etc.) could be easier for testing.
Adding diagonal and full covariance APT loss functions for SNPE training with paltas. Includes test cases.