McIntosh Classification Code #148
Conversation
| loss_z = criterion_z(output_z, labels_z) | ||
| loss_p = criterion_p(output_p, labels_p) | ||
| loss_c = criterion_c(output_c, labels_c) | ||
| loss = loss_z + loss_p + loss_c |
There was a problem hiding this comment.
So not all combinations are valid classification how can we heavily penalise invalid combination during training and exclude them at inference? E.g the table about 1/3 way down this page https://www.stce.be/educational/classification
| # Hierarchical classification heads | ||
| # Z -> P -> C | ||
| self.fc_Z = nn.Linear(224, num_classes_Z) | ||
| self.fc_P = nn.Linear(224 + num_classes_Z, num_classes_P) | ||
| self.fc_C = nn.Linear(224 + num_classes_Z + num_classes_P, num_classes_C) |
There was a problem hiding this comment.
Nice is there publication associated with this approach?
| for z, p, c in valid_combined_classes: | ||
| z_idx = encoders["Z_encoder"].transform([z])[0] | ||
| p_idx = encoders["p_encoder"].transform([p])[0] | ||
| c_idx = encoders["c_encoder"].transform([c])[0] | ||
|
|
||
| valid_p_for_z[z_idx].add(p_idx) # Map Z to valid P | ||
| valid_c_for_zp[(z_idx, p_idx)].add(c_idx) # Map (Z, P) to valid C |
There was a problem hiding this comment.
I don't know if the data set will contain all valid classes:
- It could have invalid classes need to check the code we use to filter
- It could be missing a valid class due to split or rarity e.t.c
| "D": "LG", # Merge D, E, F into LG (LargeGroup) | ||
| "E": "LG", | ||
| "F": "LG", | ||
| "H": "H", | ||
| } | ||
|
|
||
| p_component_mapping = { | ||
| "x": "x", | ||
| "r": "r", | ||
| "s": "sym", # Merge s and h into sym | ||
| "h": "sym", | ||
| "a": "asym", # Merge a and k into asym |
There was a problem hiding this comment.
Is there not enough example for these classes?
samaloney
left a comment
samaloney
left a comment
There was a problem hiding this comment.
Looks good apart from setup.cfg conflicts
There was a problem hiding this comment.
Think the conflicts with this file need to be resolved - also the dev section should be for tool required to develop the code e.g. black, flake8 if it required to run should be in model or install sections or similar but we can tackle this in a different PR
|
@edoardolegnaro I'd like to get this merged can you resolve the conflicts? |
2 participants
Hierarchical Multi-Task Learning for McIntosh classification of magnetogram cutouts
(currently prints values, will be updated with logging).