rename demonstrative_noun to determiner_noun

phueb · Jun 14, 2021 · b2af8d5 · b2af8d5
1 parent 303dccf
commit b2af8d5
Show file tree

Hide file tree

Showing 45 changed files with 92,015 additions and 11 deletions.
diff --git a/images/exp1a_average.png b/images/exp1a_average.png
diff --git a/images/exp2.png b/images/exp2.png
diff --git a/...orced_choice/8192/probing_agreement_determiner_noun-across_1_adjective_results_260000.txt b/...orced_choice/8192/probing_agreement_determiner_noun-across_1_adjective_results_260000.txt
diff --git a/...forced_choice/8192/probing_agreement_determiner_noun-between_neighbors_results_260000.txt b/...forced_choice/8192/probing_agreement_determiner_noun-between_neighbors_results_260000.txt
diff --git a/...choice/8192/probing_agreement_subject_verb-across_prepositional_phrase_results_260000.txt b/...choice/8192/probing_agreement_subject_verb-across_prepositional_phrase_results_260000.txt
diff --git a/...rced_choice/8192/probing_agreement_subject_verb-across_relative_clause_results_260000.txt b/...rced_choice/8192/probing_agreement_subject_verb-across_relative_clause_results_260000.txt
diff --git a/...forced_choice/8192/probing_agreement_subject_verb-in_question_with_aux_results_260000.txt b/...forced_choice/8192/probing_agreement_subject_verb-in_question_with_aux_results_260000.txt
diff --git a/...s/forced_choice/8192/probing_agreement_subject_verb-in_simple_question_results_260000.txt b/...s/forced_choice/8192/probing_agreement_subject_verb-in_simple_question_results_260000.txt
diff --git a/...5M/2/saves/forced_choice/8192/probing_anaphor_agreement-pronoun_gender_results_260000.txt b/...5M/2/saves/forced_choice/8192/probing_anaphor_agreement-pronoun_gender_results_260000.txt
diff --git a/...2/saves/forced_choice/8192/probing_argument_structure-dropped_argument_results_260000.txt b/...2/saves/forced_choice/8192/probing_argument_structure-dropped_argument_results_260000.txt
diff --git a/.../saves/forced_choice/8192/probing_argument_structure-swapped_arguments_results_260000.txt b/.../saves/forced_choice/8192/probing_argument_structure-swapped_arguments_results_260000.txt
diff --git a/...se_5M/2/saves/forced_choice/8192/probing_argument_structure-transitive_results_260000.txt b/...se_5M/2/saves/forced_choice/8192/probing_argument_structure-transitive_results_260000.txt
diff --git a/...Roberta-base_5M/2/saves/forced_choice/8192/probing_binding-principle_a_results_260000.txt b/...Roberta-base_5M/2/saves/forced_choice/8192/probing_binding-principle_a_results_260000.txt
diff --git a/...rta-base_5M/2/saves/forced_choice/8192/probing_case-subjective_pronoun_results_260000.txt b/...rta-base_5M/2/saves/forced_choice/8192/probing_case-subjective_pronoun_results_260000.txt
diff --git a/...rseq_Roberta-base_5M/2/saves/forced_choice/8192/probing_ellipsis-n_bar_results_260000.txt b/...rseq_Roberta-base_5M/2/saves/forced_choice/8192/probing_ellipsis-n_bar_results_260000.txt
diff --git a/...se_5M/2/saves/forced_choice/8192/probing_filler-gap-wh_question_object_results_260000.txt b/...se_5M/2/saves/forced_choice/8192/probing_filler-gap-wh_question_object_results_260000.txt
diff --git a/...e_5M/2/saves/forced_choice/8192/probing_filler-gap-wh_question_subject_results_260000.txt b/...e_5M/2/saves/forced_choice/8192/probing_filler-gap-wh_question_subject_results_260000.txt
diff --git a/...rseq_Roberta-base_5M/2/saves/forced_choice/8192/probing_irregular-verb_results_260000.txt b/...rseq_Roberta-base_5M/2/saves/forced_choice/8192/probing_irregular-verb_results_260000.txt
diff --git a/...se_5M/2/saves/forced_choice/8192/probing_island-effects-adjunct_island_results_260000.txt b/...se_5M/2/saves/forced_choice/8192/probing_island-effects-adjunct_island_results_260000.txt
diff --git a/...ced_choice/8192/probing_island-effects-coordinate_structure_constraint_results_260000.txt b/...ced_choice/8192/probing_island-effects-coordinate_structure_constraint_results_260000.txt
diff --git a/.../saves/forced_choice/8192/probing_local_attractor-in_question_with_aux_results_260000.txt b/.../saves/forced_choice/8192/probing_local_attractor-in_question_with_aux_results_260000.txt
diff --git a/...se_5M/2/saves/forced_choice/8192/probing_npi_licensing-matrix_question_results_260000.txt b/...se_5M/2/saves/forced_choice/8192/probing_npi_licensing-matrix_question_results_260000.txt
diff --git a/..._5M/2/saves/forced_choice/8192/probing_npi_licensing-only_npi_licensor_results_260000.txt b/..._5M/2/saves/forced_choice/8192/probing_npi_licensing-only_npi_licensor_results_260000.txt
diff --git a/...se_5M/2/saves/forced_choice/8192/probing_quantifiers-existential_there_results_260000.txt b/...se_5M/2/saves/forced_choice/8192/probing_quantifiers-existential_there_results_260000.txt
diff --git a/...rta-base_5M/2/saves/forced_choice/8192/probing_quantifiers-superlative_results_260000.txt b/...rta-base_5M/2/saves/forced_choice/8192/probing_quantifiers-superlative_results_260000.txt
diff --git a/...ect-across_1_adjective_results_260000.txt → ...oun-across_1_adjective_results_260000.txt b/...ect-across_1_adjective_results_260000.txt → ...oun-across_1_adjective_results_260000.txt
diff --git a/...ject-between_neighbors_results_260000.txt → ...noun-between_neighbors_results_260000.txt b/...ject-between_neighbors_results_260000.txt → ...noun-between_neighbors_results_260000.txt
diff --git a/...ect-across_1_adjective_results_260000.txt → ...oun-across_1_adjective_results_260000.txt b/...ect-across_1_adjective_results_260000.txt → ...oun-across_1_adjective_results_260000.txt
diff --git a/...ject-between_neighbors_results_260000.txt → ...noun-between_neighbors_results_260000.txt b/...ject-between_neighbors_results_260000.txt → ...noun-between_neighbors_results_260000.txt
diff --git a/...ect-across_1_adjective_results_260000.txt → ...oun-across_1_adjective_results_260000.txt b/...ect-across_1_adjective_results_260000.txt → ...oun-across_1_adjective_results_260000.txt
diff --git a/...ject-between_neighbors_results_260000.txt → ...noun-between_neighbors_results_260000.txt b/...ject-between_neighbors_results_260000.txt → ...noun-between_neighbors_results_260000.txt
diff --git a/...ect-across_1_adjective_results_260000.txt → ...oun-across_1_adjective_results_260000.txt b/...ect-across_1_adjective_results_260000.txt → ...oun-across_1_adjective_results_260000.txt
diff --git a/...ject-between_neighbors_results_260000.txt → ...noun-between_neighbors_results_260000.txt b/...ject-between_neighbors_results_260000.txt → ...noun-between_neighbors_results_260000.txt
diff --git a/...ect-across_1_adjective_results_500000.txt → ...oun-across_1_adjective_results_500000.txt b/...ect-across_1_adjective_results_500000.txt → ...oun-across_1_adjective_results_500000.txt
diff --git a/...ject-between_neighbors_results_500000.txt → ...noun-between_neighbors_results_500000.txt b/...ject-between_neighbors_results_500000.txt → ...noun-between_neighbors_results_500000.txt
diff --git a/...ect-across_1_adjective_results_500000.txt → ...oun-across_1_adjective_results_500000.txt b/...ect-across_1_adjective_results_500000.txt → ...oun-across_1_adjective_results_500000.txt
diff --git a/...ject-between_neighbors_results_500000.txt → ...noun-between_neighbors_results_500000.txt b/...ject-between_neighbors_results_500000.txt → ...noun-between_neighbors_results_500000.txt
diff --git a/scripts/plot_accuracy_boxplot.py b/scripts/plot_accuracy_boxplot.py
@@ -15,8 +15,14 @@
 STEP = '*'
 IS_LOCAL = True
 REP = 0
-PARAM_NAMES: Optional[List[str]] = None  # [f'param_{i:03}' for i in [1, 2]]
-CONDITIONS = ['leave_unmasked_prob', ]
+PARAM_NAMES: Optional[List[str]] = [
+    'huggingface_Roberta-base_30B',
+    'huggingface_Roberta-base_10M',
+    'fairseq_Roberta-base_5M',
+    'huggingface_BabyBERTa_5M+standard-unmasking',
+    'huggingface_BabyBERTa_5M',
+]
+CONDITIONS = ['corpora', 'leave_unmasked_prob', ]
 
 if IS_LOCAL:
     configs.Eval.local_runs = True

diff --git a/...onstrative_subject-across_1_adjective.txt → ...nt_determiner_noun-across_1_adjective.txt b/...onstrative_subject-across_1_adjective.txt → ...nt_determiner_noun-across_1_adjective.txt
diff --git a/...monstrative_subject-between_neighbors.txt → ...ent_determiner_noun-between_neighbors.txt b/...monstrative_subject-between_neighbors.txt → ...ent_determiner_noun-between_neighbors.txt
diff --git a/...monstrative_subject/across_1_adjective.py → ...ent_determiner_noun/across_1_adjective.py b/...monstrative_subject/across_1_adjective.py → ...ent_determiner_noun/across_1_adjective.py
diff --git a/...onstrative_subject/across_2_adjectives.py → ...nt_determiner_noun/across_2_adjectives.py b/...onstrative_subject/across_2_adjectives.py → ...nt_determiner_noun/across_2_adjectives.py
diff --git a/...emonstrative_subject/between_neighbors.py → ...ment_determiner_noun/between_neighbors.py b/...emonstrative_subject/between_neighbors.py → ...ment_determiner_noun/between_neighbors.py
diff --git a/zorro/utils.py b/zorro/utils.py
@@ -87,7 +87,7 @@ def get_phenomena_and_paradigms(excluded_paradigms: Optional[List[str]] = None,
         # 4
         'agreement_subject_verb',
         # 2
-        'agreement_demonstrative_subject',
+        'agreement_determiner_noun',
         'filler-gap',
         'island-effects',
         'quantifiers',
@@ -205,15 +205,12 @@ def get_legend_label(group_name,
     res = res.replace("corpora=('wikipedia1', 'wikipedia2', 'wikipedia3')", 'Wiki-1 + Wiki-2 + Wiki-3')
     res = res.replace("corpora=('aochildes', 'aonewsela', 'wikipedia3')", 'AO-CHILDES + AO-Newsela + Wiki-3')
 
-    res = res.replace("corpora=('aochildes',)", 'AO-CHILDES')
-    res = res.replace("corpora=('aonewsela',)", 'AO-Newsela')
-    res = res.replace("corpora=('wikipedia1',)", 'Wikipedia-1')
-
     res = res.replace("('aochildes',)", 'AO-CHILDES')
     res = res.replace("('aonewsela',)", 'AO-Newsela')
     res = res.replace("('wikipedia1',)", 'Wikipedia-1')
 
     res = res.replace("data_size=", '')
+    res = res.replace("corpora=", '')
 
     return res
 

diff --git a/zorro/visualizer.py b/zorro/visualizer.py
@@ -27,7 +27,7 @@ def make_ax_title(name: str):
     ax_title = ax_title.replace('structure', 'struct.')
     ax_title = ax_title.replace('prepositional', 'prep.')
     ax_title = ax_title.replace('agreement', 'agreem.')
-    ax_title = ax_title.replace('demonstrative', 'det.')
+    ax_title = ax_title.replace('determiner', 'det.')
     return ax_title
 
 
@@ -505,9 +505,9 @@ def _plot_boxplot_summary_on_axis(self,
 
         # x axis
         x_ticks = [0.0, 0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8, 0.9, 1.0]
-        ax.set_xlabel('Average Accuracy', fontsize=configs.Figs.ax_font_size)
+        ax.set_xlabel('Average Accuracy', fontsize=8)
         ax.set_xticks(x_ticks)
-        ax.set_xticklabels(x_ticks, fontsize=configs.Figs.tick_font_size)
+        ax.set_xticklabels(x_ticks, fontsize=8)
 
         # collect last_accuracy for each replication across all paradigms
         gn2rep2accuracies_by_pd = defaultdict(dict)
@@ -539,7 +539,7 @@ def _plot_boxplot_summary_on_axis(self,
             data_for_one_group = accuracies_by_pd_by_rep.mean(axis=0)
             boxplot_data.append(data_for_one_group)
         # boxplot plots IQR and line at median, not mean
-        positions = [n for n, _ in enumerate(group_names)]
+        positions = [n for n, _ in enumerate(group_names)][::-1]
         box_plot = ax.boxplot(boxplot_data,
                               positions=positions,
                               vert=False,
@@ -548,6 +548,7 @@ def _plot_boxplot_summary_on_axis(self,
                               flierprops={'markersize': 2},
                               boxprops=dict(linewidth=1),
                               zorder=2,
+                              showfliers = False,
                               )
         # mark the mean
         means = [np.mean(x) for x in boxplot_data]