Scheduler losses now are plotted

egs/iemocap/local/scoring.py

@@ -8,6 +8,7 @@ import numpy as np
 import torchaudio
 import sklearn.metrics as metrics
 import matplotlib.pyplot as plt
+from matplotlib.ticker import FormatStrFormatter
 from sidekit.nnet.xvector import Xtractor
 import argparse
 
@@ -158,7 +159,7 @@ fil.close()
 
 # Search for "Loss:", "Validation Loss", "Epoch" and "reducing" in all lines
 valid_loss = [line for line in file if "Validation Loss" in line]
-
+lr_scheduler = [float(line.rsplit(":",1)[1].replace("\n","")) for line in file if "Scheduer" in line]
 # "Loss:" and "Epoch" in same line
 if model_type == "custom":
     loss_epoch = [line for line in file if "Epoch" in line]
@@ -183,6 +184,7 @@ for linev, linel in zip(valid_loss, loss_epoch):
     vacc.append(round(float(linev[1].split("%")[0].replace(" ","")),2))
     aepoch.append(linel[3].split(" ")[1])
     tloss.append(round(float(linel[4].split("\t")[0].replace(" ", "")), 2))
+
 assert len(aepoch) == len(tloss) == len(vloss) == len(vacc)
 
 print("--------------\n")
@@ -203,7 +205,7 @@ if not os.path.isdir(path.rsplit("/", 1)[0]):
 if not os.path.isdir(path):
     os.mkdir(path)
 
-# Plot confusion matrice
+## Plot confusion matrice ##
 sns.heatmap(confMatrix, annot=annot, fmt="10", cmap="Blues", vmin=0, vmax=350, xticklabels=labels, yticklabels=labels)
 plt.title("Model: {}{}{}_".format(model_type, model_2nd, freeze) + str(config["speaker_number"]) + "emo_{}batch\nepoch: {}  lr: {}  Data: Test-IEMOCAP {}".format(nb_batch, aepoch[-1], lr, ses_nb) + "  UAR = " + str(UARPercent) + "%")
 plt.xlabel("Prediction")
@@ -211,26 +213,21 @@ plt.ylabel("Ground truth")
 plt.savefig(os.path.join(path, "confusion_matrix_{}{}{}_".format(model_type, model_2nd, freeze) + str(config["speaker_number"]) + "emo_{}batch_epoch-{}_lr-{}_Test-IEMOCAP{}.png".format(nb_batch, aepoch[-1], lr, ses_nb)))
 plt.show()
 plt.clf()
-print("\nConfusion matrix done!")
+print("\nConfusion matrix plotted!")
 
-# Plot losses
-fig, ax = plt.subplots()
-twin = ax.twinx() # Add a 3rd axis on the right
+## Plot losses ##
+fig, axs = plt.subplots(2, 1, constrained_layout=True)
+ax = axs[0]
+twin = ax.twinx()
+ax2 = axs[1]
 
+# 1st subplot #
 plot_tl, = ax.plot(tloss, label="Training loss")
 plot_vl, = ax.plot(vloss, label="Validation loss")
 plot_vac, = twin.plot(vacc, "g-.",  label="Validation accuracy")
+handles = [plot_tl, plot_vl, plot_vac] # For the legend
 
-handles = [plot_tl, plot_vl, plot_vac]
-
-ax.set_xlabel("Epochs")
-ax.set_ylabel("CrossEntropyLoss()")
-twin.set_ylabel("Validation accuracy (%)")
-
-twin.set_ylim(0, 100)
-ax.set_ylim(0,9)
-ax.set_xlim(0, len(aepoch))
-
+# Change of learning rate during the training (can be disused)
 if len(reduce_lr_list) != 0:
     reduce_lr_list = [re.sub(r"\s+", " ", elmt) for elmt in reduce_lr_list]
     reduce_lr_list = [elmt.split(" ") for elmt in reduce_lr_list]
@@ -243,8 +240,32 @@ if len(reduce_lr_list) != 0:
         handles.append(plt.axvline(x=key, color=colors[i], linestyle='--', label=label))
         i += 1
 
+# 2nd subplot #
+plot_lr = ax2.plot(lr_scheduler, "m")
+maxlr, minlr = max(lr_scheduler), min(lr_scheduler) # Used for legend
+maxlr_index, minlr_index = lr_scheduler.index(maxlr), lr_scheduler.index(minlr) # Used for axvline
+lr_max = ax2.axvline(x=maxlr_index, color="r", linestyle='--', label = "Max: {}".format(format((maxlr),".1e")))
+lr_min = ax2.axvline(x=minlr_index, color="b", linestyle='--', label = "Min: {}".format(format((minlr),".1e")))
+handles_lr = [lr_max, lr_min] # For the legend
+ax2.legend(handles=handles_lr) # Show the legend
+
+# General settings #
+
+# 1st subplot #
+ax.set_xlabel("Epochs")
+ax.set_ylabel("CrossEntropyLoss()")
+twin.set_ylabel("Validation accuracy (%)")
+twin.set_ylim(0, 100)
+ax.set_ylim(0,9)
+ax.set_xlim(0, len(aepoch))
 ax.legend(handles=handles)
-#plt.gca().xaxis.set_major_locator(matplotlib.ticker.MaxNLocator(20))
+
+# 2nd subplot #
+ax2.set_title("Scheduler learning rate") 
+ax2.get_xaxis().set_visible(False) # Do not show the x axis
+ax2.yaxis.set_major_formatter(FormatStrFormatter('%.e')) # Format of the y ticks
+
+# Figure #
 plt.title("Model: {}{}{}_{}emo_{}batch\nEpoch: {}  lr: {}  Data: Test-IEMOCAP {}".format(model_type, model_2nd, freeze, cates, nb_batch, aepoch[-1], lr, ses_nb))
 plt.savefig(os.path.join(path, "losses_{}{}{}_{}emo_{}batch_epoch-{}_lr-{}_Test-IEMOCAP{}.png".format(model_type, model_2nd, freeze, cates, nb_batch, aepoch[-1], lr, ses_nb)))
 plt.show()