Code Tweaks for GZ COSMOS - May 2025 (#206)

* update caesar config script

* training - only train on first frame in multi-frame subject

* prediction - only predict based on first frame for multi-frame subject

* add n_blocks support to metadata and bajor calls

---------

Co-authored-by: tooyosi <oluwatoyosi.oyegoke@gmail.com>

Author: lcjohnso

app/services/batch/prediction/export_manifest.rb

@@ -94,23 +94,22 @@ def create_manifest_data(panoptes_client, slices = MANIFEST_SUBJECT_SET_BATCH_SI
               end
             end.map(&:wait)
             subject_responses.each do |subject|
-              # Create a data row for each image URL in the Subject
-              # this will duplicate the subject information for each image URL
-              subject['locations'].each_with_index do |location, frame_id|
-                manifest_data << [
-                  location.values[0], # image_url
-                  # The subject's JSON information is stored as a string,
-                  # Yes, really - this is the format that hamlet sets up.
-                  JSON.dump(
-                    {
-                      project_id: project_id,
-                      subject_set_id: subject_set_id.to_s,
-                      subject_id: subject['id'],
-                      frame_id: frame_id.to_s
-                    }
-                  )
-                ]
-              end
+              # Create a data row for the first frame in the Subject
+              location = subject['locations'].first
+              frame_id = 0
+              manifest_data << [
+                location.values[0], # image_url
+                # The subject's JSON information is stored as a string,
+                # Yes, really - this is the format that hamlet sets up.
+                JSON.dump(
+                  {
+                    project_id: project_id,
+                    subject_set_id: subject_set_id.to_s,
+                    subject_id: subject['id'],
+                    frame_id: frame_id.to_s
+                  }
+                )
+              ]
             end
           ensure
             Faraday.default_connection.close

app/services/batch/training/create_job.rb

@@ -33,9 +33,14 @@ def training_options
           context.metadata['fixed_crop']
         end
 
+        n_blocks = if context.metadata.is_a?(Hash) && context.metadata['n_blocks']
+          context.metadata['n_blocks']
+        end
+
         {
           workflow_name: context.extractor_name,
           fixed_crop: fixed_crop,
+          n_blocks: n_blocks
         }.compact
       end
     end

app/services/format/training_data_csv.rb

@@ -31,19 +31,18 @@ def run
         raise MissingLocationData, "For subject id: #{reduced_subject.id}" if reduced_subject.locations.blank?
 
         # Ensure we handle multi image subjects here
-        # include 1 line per image for use in training catalogues
-        reduced_subject.locations.each do |location|
-          # each location is an object containing only 1 mimetype key and an image URL
-          image_url = location.values.first
-          csv << [
-            grouped_reduction.unique_id,
-            Zoobot::Storage.container_image_path(image_url),
-            # fetch all the reduction's saved question:answer values
-            # ensure we add 0's to the missing column headers - Zoobot demands this!
-            # https://zoobot.readthedocs.io/guides/training_from_scratch.html#creating-a-catalog
-            *grouped_reduction.labels.fetch_values(*label_column_headers) { |_key| 0 }
-          ]
-        end
+        # assume we will only train on first image
+        location = reduced_subject.locations.first
+        # each location is an object containing only 1 mimetype key and an image URL
+        image_url = location.values.first
+        csv << [
+          grouped_reduction.unique_id,
+          Zoobot::Storage.container_image_path(image_url),
+          # fetch all the reduction's saved question:answer values
+          # ensure we add 0's to the missing column headers - Zoobot demands this!
+          # https://zoobot.readthedocs.io/guides/training_from_scratch.html#creating-a-catalog
+          *grouped_reduction.labels.fetch_values(*label_column_headers) { |_key| 0 }
+        ]
       end
       temp_file.rewind
       temp_file

caesar/setup_ceaser_workflow.py caesar/setup_caesar_workflow.pycaesar/setup_ceaser_workflow.py renamed to caesar/setup_caesar_workflow.py

@@ -94,7 +94,7 @@ def create_workflow_subject_rules_and_effects():
 
         # create the subject rule effect (send to KaDE) config
         effect_config = {
-            'url': f'{KADE_ENDPOINT}/reductions/galaxy_zoo_cosmic_dawn_{subject_rule_task_key.lower()}',
+            'url': f'{KADE_ENDPOINT}/reductions/galaxy_zoo_jwst_cosmos_{subject_rule_task_key.lower()}',
             'reducer_key': f'{subject_rule_task_key}_sum',
             'password': KADE_API_BASIC_AUTH_PASSWORD,
             'username': KADE_API_BASIC_AUTH_USERNAME
@@ -114,7 +114,7 @@ def create_workflow_subject_rules_and_effects():
 if __name__ == '__main__':
     """
     Setup a an Active Learning Loop workflow for Caesar
-    For the Cosmic Dawn Survey dataset
+    For the JWST COSMOS Survey dataset
     """
     FORMAT = '%(asctime)s - %(levelname)s - %(message)s'
     # set level to DEBUG to get the panoptes network traffic
@@ -161,7 +161,8 @@ def create_workflow_subject_rules_and_effects():
     caesar = Caesar(endpoint=caesar_endpoint)
 
     # currently all extractors are 'question' type, if this changes we can add them here for task key lookup tuple
-    GZ_DECISION_TREE_TASK_KEYS = {'T0':True,'T1':True,'T2':True,'T3':True,'T4':True,'T5':True,'T6':True,'T7':True,'T8':True,'T11':True,'T12':True,'T13':True,'T14':True,'T15':True}
+    #GZ_DECISION_TREE_TASK_KEYS = {'T0':True,'T1':True,'T2':True,'T3':True,'T4':True,'T5':True,'T6':True,'T7':True,'T8':True,'T11':True,'T12':True,'T13':True,'T14':True,'T15':True}
+    GZ_DECISION_TREE_TASK_KEYS = {'T0':True,'T2':True,'T3':True,'T4':True,'T5':True,'T6':True,'T7':True,'T8':True,'T11':True,'T12':True,'T19':True}
     # setup known count reducer keys
     COUNT_REDUCER_KEYS = {f'{task_key}_count': task_key for task_key in GZ_DECISION_TREE_TASK_KEYS}
     # setup known sum reducer keys
@@ -172,7 +173,7 @@ def create_workflow_subject_rules_and_effects():
     # enough information to augment the trained ML system
     # each new classification will send data to kade till retirement
     # finally - these numbers can be easily adjusted post setup in the caesar UI system, https://caesar.zooniverse.org/workflows
-    NUM_CLASSIFICATIONS_BEFORE_SEND_TO_KADE = 3 if args.caesar_env == 'production' else 1
+    NUM_CLASSIFICATIONS_BEFORE_SEND_TO_KADE = 5 if args.caesar_env == 'production' else 1 # was: 3 else 1
 
     # lookup the worklfow
     zoo_api_workflow = Workflow.find(args.workflow_id)

lib/bajor/client.rb

@@ -138,7 +138,7 @@ def bajor_service_host
 
     def build_opts(options, include_schema=true)
       raw = options.with_indifferent_access
-      overrides = raw.symbolize_keys.slice(:workflow_name, :fixed_crop)
+      overrides = raw.symbolize_keys.slice(:workflow_name, :fixed_crop, :n_blocks)
 
       DEFAULT_OPTIONS
       .merge(overrides)
@@ -149,6 +149,10 @@ def build_opts(options, include_schema=true)
           run_opts << "--fixed-crop '#{o[:fixed_crop].to_json}'"
           o.delete(:fixed_crop)
         end
+        if o[:n_blocks].present?
+          run_opts << "--n-blocks #{o[:n_blocks]}"
+          o.delete(:n_blocks)
+        end
         o[:run_opts] = run_opts.join(' ') if run_opts.any?
       end
     end

spec/lib/bajor/client_spec.rb

@@ -3,14 +3,15 @@
 require 'bajor/client'
 require 'rails_helper'
 
-def build_expected_body(manifest_url: nil, manifest_path: nil, workflow_name:, fixed_crop: nil)
+def build_expected_body(manifest_url: nil, manifest_path: nil, workflow_name:, fixed_crop: nil, n_blocks: nil)
   opts = {
     workflow_name: workflow_name
   }
 
   run_opts = []
   run_opts << "--schema #{workflow_name}" if manifest_path
   run_opts << "--fixed-crop '#{fixed_crop.to_json}'" if fixed_crop
+  run_opts << "--n-blocks #{n_blocks}" if n_blocks
   opts[:run_opts] = run_opts.join(' ') unless run_opts.empty?
 
   if manifest_url
@@ -154,6 +155,35 @@ def build_expected_body(manifest_url: nil, manifest_path: nil, workflow_name:, f
       end
     end
 
+    context 'with jswt_cosmos workflow and n_blocks' do
+      let(:workflow_name) { 'jswt_cosmos' }
+      let(:n_blocks) { 2 }
+
+      let(:expected_body) { build_expected_body(manifest_path: catalogue_manifest_path, workflow_name: workflow_name, n_blocks: n_blocks) }
+      let(:request) do
+        stub_request(:post, request_url)
+          .with(
+            body: expected_body.to_json,
+            headers: request_headers
+          )
+      end
+
+      before do
+        request.to_return(status: 201, body: expected_body.to_json, headers: { content_type: 'application/json' })
+      end
+
+      it 'sends jswt_cosmos workflow and n_blocks settings' do
+        bajor_client.create_training_job(
+          catalogue_manifest_path,
+          { workflow_name: workflow_name, n_blocks: n_blocks }
+        )
+        expect(
+          a_request(:post, request_url)
+            .with(body: expected_body, headers: request_headers)
+        ).to have_been_made.once
+      end
+    end
+
     context 'with jswt_cosmos workflow and no fixed crop' do
       let(:workflow_name) { 'jswt_cosmos' }
       let(:expected_body) { build_expected_body(manifest_path: catalogue_manifest_path, workflow_name: workflow_name) }
@@ -315,6 +345,36 @@ def build_expected_body(manifest_url: nil, manifest_path: nil, workflow_name:, f
       end
     end
 
+    context 'with jswt_cosmos workflow and n_blocks' do
+      let(:workflow_name) { 'jswt_cosmos' }
+      let(:n_blocks) { 2 }
+
+      let(:expected_body) { build_expected_body(manifest_url: manifest_url, workflow_name: workflow_name, n_blocks: n_blocks) }
+
+      let(:request) do
+        stub_request(:post, request_url)
+          .with(
+            body: expected_body.to_json,
+            headers: request_headers
+          )
+      end
+
+      before do
+        request.to_return(status: 201, body: bajor_response_body.to_json, headers: { content_type: 'application/json' })
+      end
+
+      it 'sends jswt_cosmos workflow and n_blocks settings' do
+        bajor_client.create_prediction_job(
+          manifest_url,
+          { workflow_name: workflow_name, n_blocks: n_blocks }
+        )
+        expect(
+          a_request(:post, request_url)
+            .with(body: expected_body.to_json, headers: request_headers)
+        ).to have_been_made.once
+      end
+    end
+
     context 'with jswt_cosmos workflow and no fixed crop' do
       let(:workflow_name) { 'jswt_cosmos' }
       let(:expected_body) { build_expected_body(manifest_url: manifest_url, workflow_name: workflow_name) }

spec/services/format/training_data_csv_spec.rb

@@ -86,8 +86,7 @@
       end
       let(:expected_lines) do
         [
-          '8000_231121_468,/test/2f2490b4-65c1-4dca-ba25-c44128aa7a39.jpeg,3,9,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0',
-          '8000_231121_468,/test/fdccb1cf-0fc9-49b5-b054-62c83bccb9cd.jpeg,3,9,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0'
+          '8000_231121_468,/test/2f2490b4-65c1-4dca-ba25-c44128aa7a39.jpeg,3,9,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0'
         ].join("\n")
       end