""" To visualize the differences between rfmp4a (white and black) and rfmp4c7o (color bars). Notes on abbrevations in this script: a- : Achromatic c- : Chromatic/Color SP : Stimulus Parameters CR : Center Responses Tony Fu, Aug 11, 2022 """ import os import sys import math import concurrent.futures import numpy as np import pandas as pd from scipy.stats import pearsonr import torchvision.models as models import matplotlib.pyplot as plt from matplotlib.backends.backend_pdf import PdfPages from tqdm import tqdm sys.path.append('../../..') import src.rf_mapping.constants as c from src.rf_mapping.hook import ConvUnitCounter from src.rf_mapping.spatial import get_rf_sizes from src.rf_mapping.result_txt_format import (GtGaussian as GT, Rfmp4aTB1 as TB1, Rfmp4aTBn as TBn, Rfmp4aNonOverlap as NO, Rfmp4aWeighted as W, Rfmp4aSplist as aSP, Rfmp4c7oSplist as cSP, CenterReponses as CR,) # Please specify the model model = models.alexnet() model_name = 'alexnet' # model = models.vgg16() # model_name = 'vgg16' # model = models.resnet18() # model_name = 'resnet18' top_n = 1000 this_is_a_test_run = False # Source directories rfmp4a_mapping_dir = os.path.join(c.REPO_DIR, 'results', 'rfmp4a', 'mapping') rfmp4a_fit_dir = os.path.join(c.REPO_DIR, 'results', 'rfmp4a', 'gaussian_fit') rfmp4c7o_mapping_dir = os.path.join(c.REPO_DIR, 'results', 'rfmp4c7o', 'mapping') rfmp4c7o_fit_dir = os.path.join(c.REPO_DIR, 'results', 'rfmp4c7o', 'gaussian_fit') # Result directories if this_is_a_test_run: result_dir = os.path.join(c.REPO_DIR, 'results', 'compare', 'rfmp4a_vs_rfmp4c7o', 'test') else: result_dir = os.path.join(c.REPO_DIR, 'results', 'compare', 'rfmp4a_vs_rfmp4c7o', model_name) ############################################################################### # Load the data of RFMP4a (achromatic bars) into pandas DFs. a_tb1_path = os.path.join(rfmp4a_mapping_dir, model_name, f"{model_name}_rfmp4a_tb1.txt") a_tb20_path = os.path.join(rfmp4a_mapping_dir, model_name, f"{model_name}_rfmp4a_tb20.txt") a_tb100_path = os.path.join(rfmp4a_mapping_dir, model_name, f"{model_name}_rfmp4a_tb100.txt") a_no_path = os.path.join(rfmp4a_fit_dir, model_name, f"non_overlap.txt") a_w_t_path = os.path.join(rfmp4a_fit_dir, model_name, f"weighted_top.txt") a_w_b_path = os.path.join(rfmp4a_fit_dir, model_name, f"weighted_bot.txt") # Data source Abbreviation(s) # ----------------------- --------------- a_tb1_df = pd.read_csv(a_tb1_path, sep=" ", header=None) # Top bar tb1 a_tb20_df = pd.read_csv(a_tb20_path, sep=" ", header=None) # Avg of top 20 bars tb20 a_tb100_df = pd.read_csv(a_tb100_path, sep=" ", header=None) # Avg of top 100 bars tb100 a_no_df = pd.read_csv(a_no_path, sep=" ", header=None) # Non-overlap bar maps no a_w_t_df = pd.read_csv(a_w_t_path, sep=" ", header=None) # Weighted top bar maps w_t a_w_b_df = pd.read_csv(a_w_b_path, sep=" ", header=None) # Weighted bottom bar maps w_b # Load the data of RFMP4c7o (Chromatic bars) into pandas DFs. c_tb1_path = os.path.join(rfmp4c7o_mapping_dir, model_name, f"{model_name}_rfmp4c7o_tb1.txt") c_tb20_path = os.path.join(rfmp4c7o_mapping_dir, model_name, f"{model_name}_rfmp4c7o_tb20.txt") c_tb100_path = os.path.join(rfmp4c7o_mapping_dir, model_name, f"{model_name}_rfmp4c7o_tb100.txt") c_no_path = os.path.join(rfmp4c7o_fit_dir, model_name, f"non_overlap.txt") c_w_t_path = os.path.join(rfmp4c7o_fit_dir, model_name, f"weighted_top.txt") c_w_b_path = os.path.join(rfmp4c7o_fit_dir, model_name, f"weighted_bot.txt") # Data source Abbreviation(s) # ----------------------- --------------- c_tb1_df = pd.read_csv(c_tb1_path, sep=" ", header=None) # Top bar tb1 c_tb20_df = pd.read_csv(c_tb20_path, sep=" ", header=None) # Avg of top 20 bars tb20 c_tb100_df = pd.read_csv(c_tb100_path, sep=" ", header=None) # Avg of top 100 bars tb100 c_no_df = pd.read_csv(c_no_path, sep=" ", header=None) # Non-overlap bar maps no c_w_t_df = pd.read_csv(c_w_t_path, sep=" ", header=None) # Weighted top bar maps w_t c_w_b_df = pd.read_csv(c_w_b_path, sep=" ", header=None) # Weighted bottom bar maps w_b # Name the columns. def set_column_names(df, Format): """Name the columns of the pandas DF according to Format.""" df.columns = [e.name for e in Format] set_column_names(a_tb1_df, TB1) set_column_names(a_tb20_df, TBn) set_column_names(a_tb100_df, TBn) set_column_names(a_no_df, NO) set_column_names(a_w_t_df, W) set_column_names(a_w_b_df, W) set_column_names(c_tb1_df, TB1) set_column_names(c_tb20_df, TBn) set_column_names(c_tb100_df, TBn) set_column_names(c_no_df, NO) set_column_names(c_w_t_df, W) set_column_names(c_w_b_df, W) # Pad the missing layers with NAN because not all layers are mapped. gt_dir = os.path.join(c.REPO_DIR, 'results', 'ground_truth', 'gaussian_fit') gt_top_path = os.path.join(gt_dir, model_name, 'abs', f"{model_name}_gt_gaussian_top.txt") gt_t_df = pd.read_csv(gt_top_path, sep=" ", header=None) set_column_names(gt_t_df, GT) gt_no_data = gt_t_df[['LAYER', 'UNIT']].copy() # template df used for padding def pad_missing_layers(df): return pd.merge(gt_no_data, df, how='left') a_tb1_df = pad_missing_layers(a_tb1_df) a_tb20_df = pad_missing_layers(a_tb20_df) a_tb100_df = pad_missing_layers(a_tb100_df) a_no_df = pad_missing_layers(a_no_df) a_w_t_df = pad_missing_layers(a_w_t_df) a_w_b_df = pad_missing_layers(a_w_b_df) c_tb1_df = pad_missing_layers(c_tb1_df) c_tb20_df = pad_missing_layers(c_tb20_df) c_tb100_df = pad_missing_layers(c_tb100_df) c_no_df = pad_missing_layers(c_no_df) c_w_t_df = pad_missing_layers(c_w_t_df) c_w_b_df = pad_missing_layers(c_w_b_df) # Get/set some model-specific information. layer_indices, rf_sizes = get_rf_sizes(model, (227, 227)) unit_counter = ConvUnitCounter(model) _, nums_units = unit_counter.count() num_layers = len(rf_sizes) fxvar_thres = 0.8 #######################################.####################################### # # # PDF NO.1 DISTRIBUTION OF BAR LENGTH IN TOP/BOTTOM MAPS # # # # To see if it is possible to reduce the number of bars by eliminating the # # the small bars. # # # ############################################################################### def make_blen_color_pdf(conv_i): layer_name = f"conv{conv_i+1}" num_units = nums_units[conv_i] # Load layer-specific RFMP4a (Achromatic) data that couldn't be loaded earlier. a_splist_path = os.path.join(rfmp4a_mapping_dir, model_name, f"{layer_name}_splist.txt") a_t5000_path = os.path.join(rfmp4a_mapping_dir, model_name, f"{layer_name}_top5000_responses.txt") a_b5000_path = os.path.join(rfmp4a_mapping_dir, model_name, f"{layer_name}_bot5000_responses.txt") if os.path.exists(a_splist_path): a_splist_df = pd.read_csv(a_splist_path, sep=" ", header=None) else: return # In case this layer was not mapped. if os.path.exists(a_t5000_path): a_t5000_df = pd.read_csv(a_t5000_path, sep=" ", header=None) if os.path.exists(a_b5000_path): a_b5000_df = pd.read_csv(a_b5000_path, sep=" ", header=None) # Give the dataframes meaningful headers. set_column_names(a_splist_df, aSP) set_column_names(a_t5000_df, CR) set_column_names(a_b5000_df, CR) # Load layer-specific RFMP4c7o (Chromatic) data that couldn't be loaded earlier. c_splist_path = os.path.join(rfmp4c7o_mapping_dir, model_name, f"{layer_name}_splist.txt") c_t5000_path = os.path.join(rfmp4c7o_mapping_dir, model_name, f"{layer_name}_top5000_responses.txt") c_b5000_path = os.path.join(rfmp4c7o_mapping_dir, model_name, f"{layer_name}_bot5000_responses.txt") if os.path.exists(c_splist_path): c_splist_df = pd.read_csv(c_splist_path, sep=" ", header=None) else: return # In case this layer was not mapped. if os.path.exists(c_t5000_path): c_t5000_df = pd.read_csv(c_t5000_path, sep=" ", header=None) if os.path.exists(c_b5000_path): c_b5000_df = pd.read_csv(c_b5000_path, sep=" ", header=None) # Give the dataframes meaningful headers. set_column_names(c_splist_df, cSP) set_column_names(c_t5000_df, CR) set_column_names(c_b5000_df, CR) pdf_path = os.path.join(result_dir, f"{model_name}_{layer_name}_rfmp4a_rfmp4c7o_blen_color.pdf") with PdfPages(pdf_path) as pdf: for unit_i in range(num_units): plt.figure(figsize=(24,12)) plt.suptitle(f"Distribution of Bar Lengths of Top and Bottom {top_n} Bars ({model_name} {layer_name} no.{unit_i})", fontsize=18) plt.subplot(2,4,1) plt.grid() blens = sorted(a_splist_df.LEN.unique()) xlim = [min(blens)-2, max(blens)+2] response_avg = [] yerror = [] for blen in blens: stim_i = a_splist_df.loc[a_splist_df.LEN == blen, 'STIM_I'] responses_of_this_blen = a_t5000_df.loc[(a_t5000_df.UNIT == unit_i) & (a_t5000_df.STIM_I.isin(stim_i)) & (a_t5000_df.RANK < top_n), 'R'] response_avg.append(np.mean(responses_of_this_blen)) if len(responses_of_this_blen) == 0: yerror.append(0) else: yerror.append(np.std(responses_of_this_blen)/math.sqrt(len(responses_of_this_blen))) plt.errorbar(blens, response_avg, yerr=yerror, ecolor='black') plt.xlabel('bar length (pix)', fontsize=12) plt.ylabel('avg response', fontsize=12) plt.xlim(xlim) plt.title(f"RFMP4a top", fontsize=18) plt.subplot(2,4,2) plt.grid() stim_i = a_t5000_df.loc[(a_t5000_df.UNIT == unit_i) & (a_t5000_df.RANK < top_n), 'STIM_I'] all_blens = a_splist_df.loc[a_splist_df.STIM_I.isin(stim_i), 'LEN'] plt.hist(all_blens) plt.xlabel('bar length (pix)', fontsize=12) plt.ylabel('counts', fontsize=12) plt.xlim(xlim) plt.title(f"RFMP4a top", fontsize=18) plt.subplot(2,4,3) plt.grid() blens = sorted(c_splist_df.LEN.unique()) response_avg = [] yerror = [] for blen in blens: stim_i = c_splist_df.loc[c_splist_df.LEN == blen, 'STIM_I'] responses_of_this_blen = c_t5000_df.loc[(c_t5000_df.UNIT == unit_i) & (c_t5000_df.STIM_I.isin(stim_i)) & (c_t5000_df.RANK < top_n), 'R'] response_avg.append(np.mean(responses_of_this_blen)) if len(responses_of_this_blen) == 0: yerror.append(0) else: yerror.append(np.std(responses_of_this_blen)/math.sqrt(len(responses_of_this_blen))) plt.errorbar(blens, response_avg, yerr=yerror, ecolor='black') plt.xlabel('bar length (pix)', fontsize=12) plt.ylabel('avg response', fontsize=12) plt.xlim(xlim) plt.title(f"RFMP4c7o top", fontsize=18) plt.subplot(2,4,4) plt.grid() stim_i = c_t5000_df.loc[(c_t5000_df.UNIT == unit_i) & (c_t5000_df.RANK < top_n), 'STIM_I'] all_blens = c_splist_df.loc[c_splist_df.STIM_I.isin(stim_i), 'LEN'] plt.hist(all_blens) plt.xlabel('bar length (pix)', fontsize=12) plt.ylabel('counts', fontsize=12) plt.xlim(xlim) plt.title(f"RFMP4c7o top", fontsize=18) plt.subplot(2,4,5) plt.grid() blens = sorted(a_splist_df.LEN.unique()) response_avg = [] yerror = [] for blen in blens: stim_i = a_splist_df.loc[a_splist_df.LEN == blen, 'STIM_I'] responses_of_this_blen = a_b5000_df.loc[(a_b5000_df.UNIT == unit_i) & (a_b5000_df.STIM_I.isin(stim_i)) & (a_b5000_df.RANK < top_n), 'R'] response_avg.append(np.mean(responses_of_this_blen)) if len(responses_of_this_blen) == 0: yerror.append(0) else: yerror.append(np.std(responses_of_this_blen)/math.sqrt(len(responses_of_this_blen))) plt.errorbar(blens, response_avg, yerr=yerror, ecolor='black') plt.xlabel('bar length (pix)', fontsize=12) plt.ylabel('avg response', fontsize=12) plt.xlim(xlim) plt.title(f"RFMP4a bottom", fontsize=18) plt.subplot(2,4,6) plt.grid() stim_i = a_b5000_df.loc[(a_b5000_df.UNIT == unit_i) & (a_b5000_df.RANK < top_n), 'STIM_I'] all_blens = a_splist_df.loc[a_splist_df.STIM_I.isin(stim_i), 'LEN'] plt.hist(all_blens) plt.xlabel('bar length (pix)', fontsize=12) plt.ylabel('counts', fontsize=12) plt.xlim(xlim) plt.title(f"RFMP4a bottom", fontsize=18) plt.subplot(2,4,7) plt.grid() blens = sorted(c_splist_df.LEN.unique()) response_avg = [] yerror = [] for blen in blens: stim_i = c_splist_df.loc[c_splist_df.LEN == blen, 'STIM_I'] responses_of_this_blen = c_b5000_df.loc[(c_b5000_df.UNIT == unit_i) & (c_b5000_df.STIM_I.isin(stim_i)) & (c_b5000_df.RANK < top_n), 'R'] response_avg.append(np.mean(responses_of_this_blen)) if len(responses_of_this_blen) == 0: yerror.append(0) else: yerror.append(np.std(responses_of_this_blen)/math.sqrt(len(responses_of_this_blen))) plt.errorbar(blens, response_avg, yerr=yerror, ecolor='black') plt.xlabel('bar length (pix)', fontsize=12) plt.ylabel('avg response', fontsize=12) plt.xlim(xlim) plt.title(f"RFMP4c7o bottom", fontsize=18) plt.subplot(2,4,8) plt.grid() stim_i = c_b5000_df.loc[(c_b5000_df.UNIT == unit_i) & (c_b5000_df.RANK < top_n), 'STIM_I'] all_blens = c_splist_df.loc[c_splist_df.STIM_I.isin(stim_i), 'LEN'] plt.hist(all_blens) plt.xlabel('bar length (pix)', fontsize=12) plt.ylabel('counts', fontsize=12) plt.xlim(xlim) plt.title(f"RFMP4c7o bottom", fontsize=18) pdf.savefig() plt.close() print(f"{pdf_path} done.") if __name__ == "__main__": # batch_size = 5 # conv_i = 0 # while (conv_i < num_layers): # real_batch_size = min(batch_size, num_layers - conv_i) # with concurrent.futures.ProcessPoolExecutor() as executor: # executor.map(make_blen_color_pdf, [i for i in range(conv_i, conv_i + real_batch_size)]) # conv_i += real_batch_size for conv_i in range(num_layers): make_blen_color_pdf(conv_i) pass #######################################.####################################### # # # PDF NO.2 DISTRIBUTION OF BAR WIDTH IN TOP/BOTTOM MAPS # # # # To see if it is possible to reduce the number of bars by eliminating the # # the some aspect ratios. # # # ############################################################################### def make_bwid_color_pdf(conv_i): layer_name = f"conv{conv_i+1}" num_units = nums_units[conv_i] # Load layer-specific RFMP4a (Achromatic) data that couldn't be loaded earlier. a_splist_path = os.path.join(rfmp4a_mapping_dir, model_name, f"{layer_name}_splist.txt") a_t5000_path = os.path.join(rfmp4a_mapping_dir, model_name, f"{layer_name}_top5000_responses.txt") a_b5000_path = os.path.join(rfmp4a_mapping_dir, model_name, f"{layer_name}_bot5000_responses.txt") if os.path.exists(a_splist_path): a_splist_df = pd.read_csv(a_splist_path, sep=" ", header=None) else: return # In case this layer was not mapped. if os.path.exists(a_t5000_path): a_t5000_df = pd.read_csv(a_t5000_path, sep=" ", header=None) if os.path.exists(a_b5000_path): a_b5000_df = pd.read_csv(a_b5000_path, sep=" ", header=None) # Give the dataframes meaningful headers. set_column_names(a_splist_df, aSP) set_column_names(a_t5000_df, CR) set_column_names(a_b5000_df, CR) # Load layer-specific RFMP4c7o (Chromatic) data that couldn't be loaded earlier. c_splist_path = os.path.join(rfmp4c7o_mapping_dir, model_name, f"{layer_name}_splist.txt") c_t5000_path = os.path.join(rfmp4c7o_mapping_dir, model_name, f"{layer_name}_top5000_responses.txt") c_b5000_path = os.path.join(rfmp4c7o_mapping_dir, model_name, f"{layer_name}_bot5000_responses.txt") if os.path.exists(c_splist_path): c_splist_df = pd.read_csv(c_splist_path, sep=" ", header=None) else: return # In case this layer was not mapped. if os.path.exists(c_t5000_path): c_t5000_df = pd.read_csv(c_t5000_path, sep=" ", header=None) if os.path.exists(c_b5000_path): c_b5000_df = pd.read_csv(c_b5000_path, sep=" ", header=None) # Give the dataframes meaningful headers. set_column_names(c_splist_df, cSP) set_column_names(c_t5000_df, CR) set_column_names(c_b5000_df, CR) pdf_path = os.path.join(result_dir, f"{model_name}_{layer_name}_rfmp4a_rfmp4c7o_bwid_color.pdf") with PdfPages(pdf_path) as pdf: for unit_i in range(num_units): plt.figure(figsize=(24,12)) plt.suptitle(f"Distribution of Bar Widths of Top and Bottom {top_n} Bars ({model_name} {layer_name} no.{unit_i})", fontsize=18) plt.subplot(2,4,1) plt.grid() bwids = sorted(a_splist_df.WID.unique()) xlim = [min(bwids)-2, max(bwids)+2] response_avg = [] yerror = [] for bwid in bwids: stim_i = a_splist_df.loc[a_splist_df.WID == bwid, 'STIM_I'] responses_of_this_bwid = a_t5000_df.loc[(a_t5000_df.UNIT == unit_i) & (a_t5000_df.STIM_I.isin(stim_i)) & (a_t5000_df.RANK < top_n), 'R'] response_avg.append(np.mean(responses_of_this_bwid)) if len(responses_of_this_bwid) == 0: yerror.append(0) else: yerror.append(np.std(responses_of_this_bwid)/math.sqrt(len(responses_of_this_bwid))) plt.errorbar(bwids, response_avg, yerr=yerror, ecolor='black') plt.xlabel('bar width (pix)', fontsize=12) plt.ylabel('avg response', fontsize=12) plt.xlim(xlim) plt.title(f"RFMP4a top", fontsize=18) plt.subplot(2,4,2) plt.grid() stim_i = a_t5000_df.loc[(a_t5000_df.UNIT == unit_i) & (a_t5000_df.RANK < top_n), 'STIM_I'] all_bwids = a_splist_df.loc[a_splist_df.STIM_I.isin(stim_i), 'WID'] plt.hist(all_bwids) plt.xlabel('bar width (pix)', fontsize=12) plt.ylabel('counts', fontsize=12) plt.xlim(xlim) plt.title(f"RFMP4a top", fontsize=18) plt.subplot(2,4,3) plt.grid() bwids = sorted(c_splist_df.WID.unique()) response_avg = [] yerror = [] for bwid in bwids: stim_i = c_splist_df.loc[c_splist_df.WID == bwid, 'STIM_I'] responses_of_this_bwid = c_t5000_df.loc[(c_t5000_df.UNIT == unit_i) & (c_t5000_df.STIM_I.isin(stim_i)) & (c_t5000_df.RANK < top_n), 'R'] response_avg.append(np.mean(responses_of_this_bwid)) if len(responses_of_this_bwid) == 0: yerror.append(0) else: yerror.append(np.std(responses_of_this_bwid)/math.sqrt(len(responses_of_this_bwid))) plt.errorbar(bwids, response_avg, yerr=yerror, ecolor='black') plt.xlabel('bar width (pix)', fontsize=12) plt.ylabel('avg response', fontsize=12) plt.xlim(xlim) plt.title(f"RFMP4c7o top", fontsize=18) plt.subplot(2,4,4) plt.grid() stim_i = c_t5000_df.loc[(c_t5000_df.UNIT == unit_i) & (c_t5000_df.RANK < top_n), 'STIM_I'] all_bwids = c_splist_df.loc[c_splist_df.STIM_I.isin(stim_i), 'WID'] plt.hist(all_bwids) plt.xlabel('bar width (pix)', fontsize=12) plt.ylabel('counts', fontsize=12) plt.xlim(xlim) plt.title(f"RFMP4c7o top", fontsize=18) plt.subplot(2,4,5) plt.grid() bwids = sorted(a_splist_df.WID.unique()) response_avg = [] yerror = [] for bwid in bwids: stim_i = a_splist_df.loc[a_splist_df.WID == bwid, 'STIM_I'] responses_of_this_bwid = a_b5000_df.loc[(a_b5000_df.UNIT == unit_i) & (a_b5000_df.STIM_I.isin(stim_i)) & (a_b5000_df.RANK < top_n), 'R'] response_avg.append(np.mean(responses_of_this_bwid)) if len(responses_of_this_bwid) == 0: yerror.append(0) else: yerror.append(np.std(responses_of_this_bwid)/math.sqrt(len(responses_of_this_bwid))) plt.errorbar(bwids, response_avg, yerr=yerror, ecolor='black') plt.xlabel('bar width (pix)', fontsize=12) plt.ylabel('avg response', fontsize=12) plt.xlim(xlim) plt.title(f"RFMP4a bottom", fontsize=18) plt.subplot(2,4,6) plt.grid() stim_i = a_b5000_df.loc[(a_b5000_df.UNIT == unit_i) & (a_b5000_df.RANK < top_n), 'STIM_I'] all_bwids = a_splist_df.loc[a_splist_df.STIM_I.isin(stim_i), 'WID'] plt.hist(all_bwids) plt.xlabel('bar width (pix)', fontsize=12) plt.ylabel('counts', fontsize=12) plt.xlim(xlim) plt.title(f"RFMP4a bottom", fontsize=18) plt.subplot(2,4,7) plt.grid() bwids = sorted(c_splist_df.WID.unique()) response_avg = [] yerror = [] for bwid in bwids: stim_i = c_splist_df.loc[c_splist_df.WID == bwid, 'STIM_I'] responses_of_this_bwid = c_b5000_df.loc[(c_b5000_df.UNIT == unit_i) & (c_b5000_df.STIM_I.isin(stim_i)) & (c_b5000_df.RANK < top_n), 'R'] response_avg.append(np.mean(responses_of_this_bwid)) if len(responses_of_this_bwid) == 0: yerror.append(0) else: yerror.append(np.std(responses_of_this_bwid)/math.sqrt(len(responses_of_this_bwid))) plt.errorbar(bwids, response_avg, yerr=yerror, ecolor='black') plt.xlabel('bar width (pix)', fontsize=12) plt.ylabel('avg response', fontsize=12) plt.xlim(xlim) plt.title(f"RFMP4c7o bottom", fontsize=18) plt.subplot(2,4,8) plt.grid() stim_i = c_b5000_df.loc[(c_b5000_df.UNIT == unit_i) & (c_b5000_df.RANK < top_n), 'STIM_I'] all_bwids = c_splist_df.loc[c_splist_df.STIM_I.isin(stim_i), 'WID'] plt.hist(all_bwids) plt.xlabel('bar width (pix)', fontsize=12) plt.ylabel('counts', fontsize=12) plt.xlim(xlim) plt.title(f"RFMP4c7o bottom", fontsize=18) pdf.savefig() plt.close() print(f"{pdf_path} done.") if __name__ == "__main__": # batch_size = 5 # conv_i = 0 # while (conv_i < num_layers): # real_batch_size = min(batch_size, num_layers - conv_i) # with concurrent.futures.ProcessPoolExecutor() as executor: # executor.map(make_bwid_color_pdf, [i for i in range(conv_i, conv_i + real_batch_size)]) # conv_i += real_batch_size for conv_i in range(num_layers): make_bwid_color_pdf(conv_i) pass #######################################.####################################### # # # PDF NO.3 COMPARING MAX and MIN ACTIVATION OF RFMP4a and RFMP4c7o # # # # To see how important colors are. # # # ############################################################################### annotate_threshold = 20 # Display unit number if (Chromatic_r / Achromatic_r) # is greater than this threshold. def annotate_cr_ar_ratio(a_r, c_r): # Display unit idx if c_top_r/a_top_r > annotate_threshold. ax = plt.gca() for unit_i, (ar, cr) in enumerate(zip(a_r, c_r)): if cr/ar > annotate_threshold: ax.annotate(unit_i, (ar, cr), fontsize=5) def config_plot(limits): plt.axhline(0, color=(0, 0, 0, 0.5)) plt.axvline(0, color=(0, 0, 0, 0.5)) plt.xlabel('RFMP4a') plt.ylabel('RFMP4c7o') plt.xlim(limits) plt.ylim(limits) plt.plot(limits, limits, '-', color=(0, 0, 0, 0.5)) ax = plt.gca() ax.set_aspect('equal') def make_tb1_r_color_pdf(): pdf_path = os.path.join(result_dir, f"{model_name}_rfmp4a_vs_rfmp4c7o_tb1_r.pdf") with PdfPages(pdf_path) as pdf: plt.figure(figsize=(5*num_layers,10)) plt.suptitle(f"Top and bottom 5000 Responses ({model_name})", fontsize=18) for conv_i in range(num_layers): layer_name = f"conv{conv_i+1}" a_top_r = a_tb1_df.loc[a_tb1_df.LAYER == layer_name, 'TOP_R'] a_bot_r = a_tb1_df.loc[a_tb1_df.LAYER == layer_name, 'BOT_R'] c_top_r = c_tb1_df.loc[c_tb1_df.LAYER == layer_name, 'TOP_R'] c_bot_r = c_tb1_df.loc[c_tb1_df.LAYER == layer_name, 'BOT_R'] plt.subplot(2,num_layers,conv_i+1) plt.scatter(a_top_r, c_top_r, alpha=0.4) limits = (min(a_top_r.min(), c_top_r.min())-2, max(a_top_r.max(), c_top_r.max())+2) config_plot(limits) try: r_val, p_val = pearsonr(a_top_r, c_top_r) except: r_val = np.NaN plt.title(f"{layer_name} (top, r = {r_val:.2f})", fontsize=16) annotate_cr_ar_ratio(a_top_r, c_top_r) plt.subplot(2,num_layers,conv_i+num_layers+1) plt.scatter(a_bot_r, c_bot_r, alpha=0.4) limits = (min(a_bot_r.min(), c_bot_r.min())-2, max(a_bot_r.max(), c_bot_r.max())+2) config_plot(limits) try: r_val, p_val = pearsonr(a_bot_r, c_bot_r) except: r_val = np.NaN plt.title(f"{layer_name} (bottom, r = {r_val:.2f})", fontsize=16) annotate_cr_ar_ratio(a_bot_r, c_bot_r) pdf.savefig() plt.close() if __name__ == "__main__": make_tb1_r_color_pdf() pass #######################################.####################################### # # # PDF NO.4 COMPARING COORINATES OF RFMP4a and RFMP4c7o # # # ############################################################################### def config_plot(limits): plt.plot(limits, limits, '-', color=(0, 0, 0, 0.5)) plt.axhline(0, color=(0, 0, 0, 0.5)) plt.axvline(0, color=(0, 0, 0, 0.5)) plt.xlim(limits) plt.ylim(limits) ax = plt.gca() ax.set_aspect('equal') def make_error_coords_pdf(): pdf_path = os.path.join(result_dir, f"{model_name}_rfmp4a_vs_rfmp4c7o_coords.pdf") with PdfPages(pdf_path) as pdf: for conv_i, rf_size in enumerate(rf_sizes): # Get some layer-specific information. layer_name = f'conv{conv_i+1}' num_units_total = len(a_tb100_df.loc[(a_tb100_df.LAYER == layer_name)]) limits = (-100, 100) plt.figure(figsize=(25,20)) plt.suptitle(f"RF center coordinates ({model_name} {layer_name}, n = {num_units_total})", fontsize=18) # if sum(np.isfinite(xddata)) == 0: # continue # Skip this layer if no data plt.subplot(4,5,1) config_plot(limits) a_data = a_tb1_df.loc[(a_tb1_df.LAYER == layer_name), 'TOP_X'] c_data = c_tb1_df.loc[(c_tb1_df.LAYER == layer_name), 'TOP_X'] num_units_included = len(a_data) plt.scatter(a_data, c_data, alpha=0.4) plt.xlabel('RFMP4a top x') plt.ylabel('RFMP4c7o top x') r_val, p_val = pearsonr(a_data, c_data) plt.title(f'top 1 x-coord (n={num_units_included}, r={r_val:.2f})') a_data = a_tb1_df.loc[(a_tb1_df.LAYER == layer_name), 'TOP_Y'] c_data = c_tb1_df.loc[(c_tb1_df.LAYER == layer_name), 'TOP_Y'] num_units_included = len(a_data) plt.subplot(4,5,6) config_plot(limits) plt.scatter(a_data, c_data, alpha=0.4) plt.xlabel('RFMP4a top y') plt.ylabel('RFMP4c7o top y') r_val, p_val = pearsonr(a_data, c_data) plt.title(f'top 1 y-coord (n={num_units_included}, r={r_val:.2f})') a_data = a_tb20_df.loc[(a_tb20_df.LAYER == layer_name), 'TOP_MUX'] c_data = c_tb20_df.loc[(c_tb20_df.LAYER == layer_name), 'TOP_MUX'] num_units_included = len(a_data) plt.subplot(4,5,2) config_plot(limits) plt.scatter(a_data, c_data, alpha=0.4) plt.xlabel('RFMP4a top mu_x') plt.ylabel('RFMP4c7o top mu_x') r_val, p_val = pearsonr(a_data, c_data) plt.title(f'top 20 x-coord (n={num_units_included}, r={r_val:.2f})') a_data = a_tb20_df.loc[(a_tb20_df.LAYER == layer_name), 'TOP_MUY'] c_data = c_tb20_df.loc[(c_tb20_df.LAYER == layer_name), 'TOP_MUY'] num_units_included = len(a_data) plt.subplot(4,5,7) config_plot(limits) plt.scatter(a_data, c_data, alpha=0.4) plt.xlabel('RFMP4a top mu_y') plt.ylabel('RFMP4c7o top mu_y') r_val, p_val = pearsonr(a_data, c_data) plt.title(f'top 20 y-coord (n={num_units_included}, r={r_val:.2f})') a_data = a_tb100_df.loc[(a_tb100_df.LAYER == layer_name), 'TOP_MUX'] c_data = c_tb100_df.loc[(c_tb100_df.LAYER == layer_name), 'TOP_MUX'] num_units_included = len(a_data) plt.subplot(4,5,3) config_plot(limits) plt.scatter(a_data, c_data, alpha=0.4) plt.xlabel('RFMP4a top mu_x') plt.ylabel('RFMP4c7o top mu_x') r_val, p_val = pearsonr(a_data, c_data) plt.title(f'top 100 x-coord (n={num_units_included}, r={r_val:.2f})') a_data = a_tb100_df.loc[(a_tb100_df.LAYER == layer_name), 'TOP_MUY'] c_data = c_tb100_df.loc[(c_tb100_df.LAYER == layer_name), 'TOP_MUY'] num_units_included = len(a_data) plt.subplot(4,5,8) config_plot(limits) plt.scatter(a_data, c_data, alpha=0.4) plt.xlabel('RFMP4a top mu_y') plt.ylabel('RFMP4c7o top mu_y') r_val, p_val = pearsonr(a_data, c_data) plt.title(f'top 100 y-coord (n={num_units_included}, r={r_val:.2f})') a_data = a_no_df.loc[(a_no_df.LAYER == layer_name) & (a_no_df.TOP_RAD_10 != -1) & (c_no_df.TOP_RAD_10 != -1), 'TOP_X'] c_data = c_no_df.loc[(c_no_df.LAYER == layer_name) & (a_no_df.TOP_RAD_10 != -1) & (c_no_df.TOP_RAD_10 != -1), 'TOP_X'] num_units_included = len(a_data) plt.subplot(4,5,4) config_plot(limits) plt.scatter(a_data, c_data, alpha=0.4) plt.xlabel('RFMP4a top x') plt.ylabel('RFMP4c7o top x') try: r_val, p_val = pearsonr(a_data, c_data) except: r_val = 0 plt.title(f'top non-overlap x-coord (n={num_units_included}, r={r_val:.2f})') a_data = a_no_df.loc[(a_no_df.LAYER == layer_name) & (a_no_df.TOP_RAD_10 != -1) & (c_no_df.TOP_RAD_10 != -1), 'TOP_Y'] c_data = c_no_df.loc[(c_no_df.LAYER == layer_name) & (a_no_df.TOP_RAD_10 != -1) & (c_no_df.TOP_RAD_10 != -1), 'TOP_Y'] num_units_included = len(a_data) plt.subplot(4,5,9) config_plot(limits) plt.scatter(a_data, c_data, alpha=0.4) plt.xlabel('RFMP4a top y') plt.ylabel('RFMP4c7o top y') try: r_val, p_val = pearsonr(a_data, c_data) except: r_val = 0 plt.title(f'top non-overlap y-coord (n={num_units_included}, r={r_val:.2f})') a_data = a_w_t_df.loc[(a_w_t_df.LAYER == layer_name) & (a_w_t_df.FXVAR > fxvar_thres) & (c_w_t_df.FXVAR > fxvar_thres), 'MUX'] c_data = c_w_t_df.loc[(c_w_t_df.LAYER == layer_name) & (a_w_t_df.FXVAR > fxvar_thres) & (c_w_t_df.FXVAR > fxvar_thres), 'MUX'] num_units_included = len(a_data) plt.subplot(4,5,5) config_plot(limits) plt.scatter(a_data, c_data, alpha=0.4) plt.xlabel('RFMP4a top x') plt.ylabel('RFMP4c7o top x') try: r_val, p_val = pearsonr(a_data, c_data) except: r_val = 0 plt.title(f'top weighted x-coord (n={num_units_included}, r={r_val:.2f})') a_data = a_w_t_df.loc[(a_w_t_df.LAYER == layer_name) & (a_w_t_df.FXVAR > fxvar_thres) & (c_w_t_df.FXVAR > fxvar_thres), 'MUY'] c_data = c_w_t_df.loc[(c_w_t_df.LAYER == layer_name) & (a_w_t_df.FXVAR > fxvar_thres) & (c_w_t_df.FXVAR > fxvar_thres), 'MUY'] num_units_included = len(a_data) plt.subplot(4,5,10) config_plot(limits) plt.scatter(a_data, c_data, alpha=0.4) plt.xlabel('RFMP4a top y') plt.ylabel('RFMP4c7o top y') try: r_val, p_val = pearsonr(a_data, c_data) except: r_val = 0 plt.title(f'top weighted y-coord (n={num_units_included}, r={r_val:.2f})') plt.subplot(4,5,11) config_plot(limits) a_data = a_tb1_df.loc[(a_tb1_df.LAYER == layer_name), 'TOP_Y'] c_data = c_tb1_df.loc[(c_tb1_df.LAYER == layer_name), 'TOP_Y'] num_units_included = len(a_data) plt.scatter(a_data, c_data, alpha=0.4) plt.xlabel('RFMP4a bottom x') plt.ylabel('RFMP4c7o bottom x') r_val, p_val = pearsonr(a_data, c_data) plt.title(f'bottom 1 x-coord (n={num_units_included}, r={r_val:.2f})') a_data = a_tb1_df.loc[(a_tb1_df.LAYER == layer_name), 'BOT_Y'] c_data = c_tb1_df.loc[(c_tb1_df.LAYER == layer_name), 'BOT_Y'] num_units_included = len(a_data) plt.subplot(4,5,16) config_plot(limits) plt.scatter(a_data, c_data, alpha=0.4) plt.xlabel('RFMP4a bottom y') plt.ylabel('RFMP4c7o bottom y') r_val, p_val = pearsonr(a_data, c_data) plt.title(f'bottom 1 y-coord (n={num_units_included}, r={r_val:.2f})') a_data = a_tb20_df.loc[(a_tb20_df.LAYER == layer_name), 'BOT_MUX'] c_data = c_tb20_df.loc[(c_tb20_df.LAYER == layer_name), 'BOT_MUX'] num_units_included = len(a_data) plt.subplot(4,5,12) config_plot(limits) plt.scatter(a_data, c_data, alpha=0.4) plt.xlabel('RFMP4a bot mu_x') plt.ylabel('RFMP4c7o bot mu_x') r_val, p_val = pearsonr(a_data, c_data) plt.title(f'bottom 20 x-coord (n={num_units_included}, r={r_val:.2f})') a_data = a_tb20_df.loc[(a_tb20_df.LAYER == layer_name), 'BOT_MUY'] c_data = c_tb20_df.loc[(c_tb20_df.LAYER == layer_name), 'BOT_MUY'] num_units_included = len(a_data) plt.subplot(4,5,17) config_plot(limits) plt.scatter(a_data, c_data, alpha=0.4) plt.xlabel('RFMP4a bot mu_y') plt.ylabel('RFMP4c7o bot mu_y') r_val, p_val = pearsonr(a_data, c_data) plt.title(f'bottom 20 y-coord (n={num_units_included}, r={r_val:.2f})') a_data = a_tb100_df.loc[(a_tb100_df.LAYER == layer_name), 'BOT_MUX'] c_data = c_tb100_df.loc[(c_tb100_df.LAYER == layer_name), 'BOT_MUX'] num_units_included = len(a_data) plt.subplot(4,5,13) config_plot(limits) plt.scatter(a_data, c_data, alpha=0.4) plt.xlabel('RFMP4a bot mu_x') plt.ylabel('RFMP4c7o bot mu_x') r_val, p_val = pearsonr(a_data, c_data) plt.title(f'bot 100 x-coord (n={num_units_included}, r={r_val:.2f})') a_data = a_tb100_df.loc[(a_tb100_df.LAYER == layer_name), 'BOT_MUY'] c_data = c_tb100_df.loc[(c_tb100_df.LAYER == layer_name), 'BOT_MUY'] num_units_included = len(a_data) plt.subplot(4,5,18) config_plot(limits) plt.scatter(a_data, c_data, alpha=0.4) plt.xlabel('RFMP4a bot mu_y') plt.ylabel('RFMP4c7o bot mu_y') r_val, p_val = pearsonr(a_data, c_data) plt.title(f'bottom 100 y-coord (n={num_units_included}, r={r_val:.2f})') a_data = a_no_df.loc[(a_no_df.LAYER == layer_name) & (a_no_df.TOP_RAD_10 != -1) & (c_no_df.TOP_RAD_10 != -1), 'BOT_X'] c_data = c_no_df.loc[(c_no_df.LAYER == layer_name) & (a_no_df.TOP_RAD_10 != -1) & (c_no_df.TOP_RAD_10 != -1), 'BOT_X'] num_units_included = len(a_data) plt.subplot(4,5,14) config_plot(limits) plt.scatter(a_data, c_data, alpha=0.4) plt.xlabel('RFMP4a bot x') plt.ylabel('RFMP4c7o bot x') try: r_val, p_val = pearsonr(a_data, c_data) except: r_val = 0 plt.title(f'bottom non-overlap x-coord (n={num_units_included}, r={r_val:.2f})') a_data = a_no_df.loc[(a_no_df.LAYER == layer_name) & (a_no_df.TOP_RAD_10 != -1) & (c_no_df.TOP_RAD_10 != -1), 'BOT_Y'] c_data = c_no_df.loc[(c_no_df.LAYER == layer_name) & (a_no_df.TOP_RAD_10 != -1) & (c_no_df.TOP_RAD_10 != -1), 'BOT_Y'] num_units_included = len(a_data) plt.subplot(4,5,19) config_plot(limits) plt.scatter(a_data, c_data, alpha=0.4) plt.xlabel('RFMP4a bot y') plt.ylabel('RFMP4c7o bot y') try: r_val, p_val = pearsonr(a_data, c_data) except: r_val = 0 plt.title(f'bottom non-overlap y-coord (n={num_units_included}, r={r_val:.2f})') a_data = a_w_b_df.loc[(a_w_b_df.LAYER == layer_name) & (a_w_b_df.FXVAR > fxvar_thres) & (c_w_b_df.FXVAR > fxvar_thres), 'MUX'] c_data = c_w_b_df.loc[(c_w_b_df.LAYER == layer_name) & (a_w_b_df.FXVAR > fxvar_thres) & (c_w_b_df.FXVAR > fxvar_thres), 'MUX'] num_units_included = len(a_data) plt.subplot(4,5,15) config_plot(limits) plt.scatter(a_data, c_data, alpha=0.4) plt.xlabel('RFMP4a bot x') plt.ylabel('RFMP4c7o bot x') try: r_val, p_val = pearsonr(a_data, c_data) except: r_val = 0 plt.title(f'bottom weighted x-coord (n={num_units_included}, r={r_val:.2f})') a_data = a_w_b_df.loc[(a_w_b_df.LAYER == layer_name) & (a_w_b_df.FXVAR > fxvar_thres) & (c_w_b_df.FXVAR > fxvar_thres), 'MUY'] c_data = c_w_b_df.loc[(c_w_b_df.LAYER == layer_name) & (a_w_b_df.FXVAR > fxvar_thres) & (c_w_b_df.FXVAR > fxvar_thres), 'MUY'] num_units_included = len(a_data) plt.subplot(4,5,20) config_plot(limits) plt.scatter(a_data, c_data, alpha=0.4) plt.xlabel('RFMP4a bot y') plt.ylabel('RFMP4c7o bot y') try: r_val, p_val = pearsonr(a_data, c_data) except: r_val = 0 plt.title(f'bottom weighted y-coord (n={num_units_included}, r={r_val:.2f})') pdf.savefig() plt.close() if __name__ == '__main__': make_error_coords_pdf() pass #######################################.####################################### # # # PDF NO.5 COMPARING RADII OF RFMP4a and RFMP4c7o # # # ############################################################################### def config_plot(limits): line = np.linspace(min(limits), max(limits), 100) plt.plot(line, line, 'k', alpha=0.4) plt.xlim(limits) plt.ylim(limits) ax = plt.gca() ax.set_aspect('equal') def geo_mean(sd1, sd2): return np.sqrt(np.power(sd1, 2) + np.power(sd2, 2)) def del_outliers(radius_1, radius_2, rf_size): new_radius_1 = [] new_radius_2 = [] for i in range(len(radius_1)): if radius_1.iloc[i] < rf_size and radius_2.iloc[i] < rf_size: new_radius_1.append(radius_1.iloc[i]) new_radius_2.append(radius_2.iloc[i]) return np.array(new_radius_1), np.array(new_radius_2) def make_radius_color_pdf(): pdf_path = os.path.join(result_dir, f"{model_name}_rfmp4a_vs_rfmp4c7o_radius.pdf") with PdfPages(pdf_path) as pdf: for conv_i, rf_size in enumerate(rf_sizes): # Get some layer-specific information. layer_name = f'conv{conv_i+1}' num_units_total = len(a_tb100_df.loc[(a_tb100_df.LAYER == layer_name)]) limits = (0, 150) rf_size = rf_size[0] plt.figure(figsize=(20,10)) plt.suptitle(f"RF Radii ({model_name} {layer_name}, n = {num_units_total}, ERF = {rf_size})", fontsize=24) a_radius = a_no_df.loc[(a_no_df.LAYER == layer_name) & (a_no_df.TOP_RAD_10 != -1) & (c_no_df.TOP_RAD_10 != -1), 'TOP_RAD_10'] c_radius = c_no_df.loc[(c_no_df.LAYER == layer_name) & (a_no_df.TOP_RAD_10 != -1) & (c_no_df.TOP_RAD_10 != -1), 'TOP_RAD_10'] if sum(np.isfinite(a_radius)) == 0: continue # Skip this layer if no data plt.subplot(2,4,1) config_plot(limits) plt.scatter(a_radius, c_radius, alpha=0.4) plt.xlabel('RFMP4a radius') plt.ylabel('RFMP4c7o radius') try: r_val, p_val = pearsonr(a_radius, c_radius) except: r_val = np.NaN plt.title(f'Non-overlap (top 10% mass, n={len(a_radius)}, r={r_val:.2f})') a_radius = a_no_df.loc[(a_no_df.LAYER == layer_name) & (a_no_df.TOP_RAD_50 != -1) & (c_no_df.TOP_RAD_50 != -1), 'TOP_RAD_50'] c_radius = c_no_df.loc[(c_no_df.LAYER == layer_name) & (a_no_df.TOP_RAD_50 != -1) & (c_no_df.TOP_RAD_50 != -1), 'TOP_RAD_50'] if sum(np.isfinite(a_radius)) == 0: continue # Skip this layer if no data plt.subplot(2,4,2) config_plot(limits) plt.scatter(a_radius, c_radius, alpha=0.4) plt.xlabel('RFMP4a radius') plt.ylabel('RFMP4c7o radius') try: r_val, p_val = pearsonr(a_radius, c_radius) except: r_val = np.NaN plt.title(f'Non-overlap (top 50% mass, n={len(a_radius)}, r={r_val:.2f})') a_radius = a_no_df.loc[(a_no_df.LAYER == layer_name) & (a_no_df.TOP_RAD_90 != -1) & (c_no_df.TOP_RAD_90 != -1), 'TOP_RAD_90'] c_radius = c_no_df.loc[(c_no_df.LAYER == layer_name) & (a_no_df.TOP_RAD_90 != -1) & (c_no_df.TOP_RAD_90 != -1), 'TOP_RAD_90'] if sum(np.isfinite(a_radius)) == 0: continue # Skip this layer if no data plt.subplot(2,4,3) config_plot(limits) plt.scatter(a_radius, c_radius, alpha=0.4) plt.xlabel('RFMP4a radius') plt.ylabel('RFMP4c7o radius') try: r_val, p_val = pearsonr(a_radius, c_radius) except: r_val = np.NaN plt.title(f'Non-overlap (top 90% mass, n={len(a_radius)}, r={r_val:.2f})') a_sd1 = a_w_t_df.loc[(a_w_t_df.LAYER == layer_name) & (a_w_t_df.FXVAR > fxvar_thres) & (c_w_t_df.FXVAR > fxvar_thres), 'SD1'] a_sd2 = a_w_t_df.loc[(a_w_t_df.LAYER == layer_name) & (a_w_t_df.FXVAR > fxvar_thres) & (c_w_t_df.FXVAR > fxvar_thres), 'SD2'] a_radius = geo_mean(a_sd1, a_sd2) c_sd1 = c_w_t_df.loc[(c_w_t_df.LAYER == layer_name) & (a_w_t_df.FXVAR > fxvar_thres) & (c_w_t_df.FXVAR > fxvar_thres), 'SD1'] c_sd2 = c_w_t_df.loc[(c_w_t_df.LAYER == layer_name) & (a_w_t_df.FXVAR > fxvar_thres) & (c_w_t_df.FXVAR > fxvar_thres), 'SD2'] c_radius = geo_mean(c_sd1, c_sd2) a_radius, c_radius = del_outliers(a_radius, c_radius, rf_size) plt.subplot(2,4,4) config_plot(limits) plt.scatter(a_radius, c_radius, alpha=0.4) plt.xlabel('RFMP4a $\sqrt{sd_1^2+sd_2^2}$') plt.ylabel('RFMP4c7o $\sqrt{sd_1^2+sd_2^2}$') try: r_val, p_val = pearsonr(a_radius, c_radius) except: r_val = np.NaN plt.title(f'Weighted (top, n={len(a_radius)}, r={r_val:.2f})') a_radius = a_no_df.loc[(a_no_df.LAYER == layer_name) & (a_no_df.BOT_RAD_10 != -1) & (c_no_df.BOT_RAD_10 != -1), 'BOT_RAD_10'] c_radius = c_no_df.loc[(c_no_df.LAYER == layer_name) & (a_no_df.BOT_RAD_10 != -1) & (c_no_df.BOT_RAD_10 != -1), 'BOT_RAD_10'] if sum(np.isfinite(a_radius)) == 0: continue # Skip this layer if no data plt.subplot(2,4,5) config_plot(limits) plt.scatter(a_radius, c_radius, alpha=0.4) plt.xlabel('RFMP4a radius') plt.ylabel('RFMP4c7o radius') try: r_val, p_val = pearsonr(a_radius, c_radius) except: r_val = np.NaN plt.title(f'Non-overlap (bottom 10% mass, n={len(a_radius)}, r={r_val:.2f})') a_radius = a_no_df.loc[(a_no_df.LAYER == layer_name) & (a_no_df.BOT_RAD_50 != -1) & (c_no_df.BOT_RAD_50 != -1), 'BOT_RAD_50'] c_radius = c_no_df.loc[(c_no_df.LAYER == layer_name) & (a_no_df.BOT_RAD_50 != -1) & (c_no_df.BOT_RAD_50 != -1), 'BOT_RAD_50'] if sum(np.isfinite(a_radius)) == 0: continue # Skip this layer if no data plt.subplot(2,4,6) config_plot(limits) plt.scatter(a_radius, c_radius, alpha=0.4) plt.xlabel('RFMP4a radius') plt.ylabel('RFMP4c7o radius') try: r_val, p_val = pearsonr(a_radius, c_radius) except: r_val = np.NaN plt.title(f'Non-overlap (bottom 50% mass, n={len(a_radius)}, r={r_val:.2f})') a_radius = a_no_df.loc[(a_no_df.LAYER == layer_name) & (a_no_df.BOT_RAD_90 != -1) & (c_no_df.BOT_RAD_90 != -1), 'BOT_RAD_90'] c_radius = c_no_df.loc[(c_no_df.LAYER == layer_name) & (a_no_df.BOT_RAD_90 != -1) & (c_no_df.BOT_RAD_90 != -1), 'BOT_RAD_90'] if sum(np.isfinite(a_radius)) == 0: continue # Skip this layer if no data plt.subplot(2,4,7) config_plot(limits) plt.scatter(a_radius, c_radius, alpha=0.4) plt.xlabel('RFMP4a radius') plt.ylabel('RFMP4c7o radius') try: r_val, p_val = pearsonr(a_radius, c_radius) except: r_val = np.NaN plt.title(f'Non-overlap (bottom 90% mass, n={len(a_radius)}, r={r_val:.2f})') a_sd1 = a_w_b_df.loc[(a_w_b_df.LAYER == layer_name) & (a_w_b_df.FXVAR > fxvar_thres) & (c_w_b_df.FXVAR > fxvar_thres), 'SD1'] a_sd2 = a_w_b_df.loc[(a_w_b_df.LAYER == layer_name) & (a_w_b_df.FXVAR > fxvar_thres) & (c_w_b_df.FXVAR > fxvar_thres), 'SD2'] a_radius = geo_mean(a_sd1, a_sd2) c_sd1 = c_w_b_df.loc[(c_w_b_df.LAYER == layer_name) & (a_w_b_df.FXVAR > fxvar_thres) & (c_w_b_df.FXVAR > fxvar_thres), 'SD1'] c_sd2 = c_w_b_df.loc[(c_w_b_df.LAYER == layer_name) & (a_w_b_df.FXVAR > fxvar_thres) & (c_w_b_df.FXVAR > fxvar_thres), 'SD2'] c_radius = geo_mean(c_sd1, c_sd2) # Eliminate outliers (too big) a_radius, c_radius = del_outliers(a_radius, c_radius, rf_size) plt.subplot(2,4,8) config_plot(limits) plt.scatter(a_radius, c_radius, alpha=0.4) plt.xlabel('RFMP4a $\sqrt{sd_1^2+sd_2^2}$') plt.ylabel('RFMP4c7o $\sqrt{sd_1^2+sd_2^2}$') try: r_val, p_val = pearsonr(a_radius, c_radius) except: r_val = np.NaN plt.title(f'Weighted (bottom, n={len(a_radius)}, r={r_val:.2f})') pdf.savefig() plt.close() if __name__ == '__main__': make_radius_color_pdf() pass ######################################.####################################### # # # PDF NO.6 ERROR ORIENTATION # # # ############################################################################## def eccentricity(sd1, sd2): short = np.minimum(sd1, sd2) long = np.maximum(sd1, sd2) # ecc = np.sqrt(1 - np.power(short, 2)/np.power(long, 2)) ecc = long/short return ecc def config_plot(): plt.xlim([-5, 95]) # plt.ylim([0, 7]) plt.xlabel('$\Delta \Theta $ (°)') def delta_ori(ori_1, ori_2): # Note: this function assumes 0 <= ori < 180. theta_small = np.minimum(ori_1, ori_2) theta_large = np.maximum(ori_1, ori_2) # Because angles wraps around 0 and 180 deg, we need to consider two cases: delta_theta_a = theta_large - theta_small delta_theta_b = (theta_small + 180) - theta_large return np.minimum(delta_theta_a, delta_theta_b) annotate_eccentricity_threshold = 3 def annotate_eccentricity(units, angle_diff, eccentricities): # Display unid indices for those that have large eccentricity values. ax = plt.gca() for unit_i, angle, ecc in zip(units, angle_diff, eccentricities): if ecc > annotate_eccentricity_threshold: ax.annotate(unit_i, (angle, ecc), fontsize=5) def make_error_ori_pdf(): pdf_path = os.path.join(result_dir, f"{model_name}_rfmp4a_vs_rfmp4c7o_ori.pdf") with PdfPages(pdf_path) as pdf: for conv_i, rf_size in enumerate(rf_sizes): # Get some layer-specific information. layer_name = f'conv{conv_i+1}' num_units_total = len(gt_t_df.loc[(gt_t_df.LAYER == layer_name)]) # Get RFMP4a (Achromatic) data (top and bottom) a_w_t_data = a_w_t_df.loc[(a_w_t_df.LAYER == layer_name) & (a_w_t_df.FXVAR > fxvar_thres) & (c_w_t_df.FXVAR > fxvar_thres)] a_w_t_ecc = eccentricity(a_w_t_data['SD1'], a_w_t_data['SD2']) a_w_t_ori = a_w_t_data['ORI'] a_w_b_data = a_w_b_df.loc[(a_w_b_df.LAYER == layer_name) & (a_w_b_df.FXVAR > fxvar_thres) & (c_w_b_df.FXVAR > fxvar_thres)] a_w_b_ecc = eccentricity(a_w_b_data['SD1'], a_w_b_data['SD2']) a_w_b_ori = a_w_b_data['ORI'] # Get RFMP4c7o (Chromatic) data (top and bottom) c_w_t_data = c_w_t_df.loc[(c_w_t_df.LAYER == layer_name) & (a_w_t_df.FXVAR > fxvar_thres) & (c_w_t_df.FXVAR > fxvar_thres)] c_w_t_ecc = eccentricity(c_w_t_data['SD1'], c_w_t_data['SD2']) c_w_t_ori = c_w_t_data['ORI'] c_w_b_data = c_w_b_df.loc[(c_w_b_df.LAYER == layer_name) & (a_w_b_df.FXVAR > fxvar_thres) & (c_w_b_df.FXVAR > fxvar_thres)] c_w_b_ecc = eccentricity(c_w_b_data['SD1'], c_w_b_data['SD2']) c_w_b_ori = c_w_b_data['ORI'] plt.figure(figsize=(10,11)) plt.suptitle(f"{model_name} {layer_name} RF orientations: RFMP4a vs. RFMP4c7o\n(n = {num_units_total})", fontsize=16) plt.subplot(2,2,1) angle_diff = delta_ori(a_w_t_ori, c_w_t_ori) plt.scatter(angle_diff , a_w_t_ecc, alpha=0.4) config_plot() annotate_eccentricity(a_w_t_data['UNIT'], angle_diff, a_w_t_ecc) plt.ylabel('RFMP4a eccentricity') plt.title(f'top, n = {len(a_w_t_ori)}') plt.subplot(2,2,2) angle_diff = delta_ori(a_w_t_ori, c_w_t_ori) plt.scatter(angle_diff, c_w_t_ecc, alpha=0.4) config_plot() annotate_eccentricity(a_w_t_data['UNIT'], angle_diff, c_w_t_ecc) plt.ylabel('RFMP4c7o eccentricity') plt.title(f'top, n = {len(a_w_t_ori)}') plt.subplot(2,2,3) angle_diff = delta_ori(a_w_b_ori, c_w_b_ori) plt.scatter(angle_diff, a_w_b_ecc, alpha=0.4) config_plot() annotate_eccentricity(a_w_b_data['UNIT'], angle_diff, a_w_b_ecc) plt.ylabel('RFMP4a eccentricity') plt.title(f'bottom, n = {len(a_w_b_ori)}') plt.subplot(2,2,4) angle_diff = delta_ori(a_w_b_ori, c_w_b_ori) plt.scatter(angle_diff, c_w_b_ecc, alpha=0.4) config_plot() annotate_eccentricity(a_w_b_data['UNIT'], angle_diff, c_w_b_ecc) plt.ylabel('RFMPc7o eccentricity') plt.title(f'bottom, n = {len(a_w_b_ori)}') pdf.savefig() plt.close() if __name__ == '__main__': make_error_ori_pdf() pass