minor changes to improve visualization

kirayuta · May 8, 2017 · 4360b34 · 4360b34
1 parent 3878ba2
commit 4360b34
Show file tree

Hide file tree

Showing 3 changed files with 82 additions and 17 deletions.
diff --git a/.gitignore b/.gitignore
@@ -3,7 +3,7 @@ data/
 logs/
 checkpoints/
 out/
-out_boundary/
+out_boundary_*/
 results/
 
 experiments/BBBC022_hand/GT_labels

diff --git a/code/helper/visualize.py b/code/helper/visualize.py
@@ -90,14 +90,19 @@ def visualize_boundary_hard(pred_y, true_x, true_y, out_dir='./', label=''):
 
     for sampleIndex in range(nSamples):
 
-        nCols = 4
-        figure, axes = plt.subplots(ncols=nCols, nrows=1, figsize=(nCols*5, 5))
-        figure.tight_layout(pad = 1)
+        nCols = 3
+        nRows = 2
+        figure, axes = plt.subplots(ncols=nCols, nrows=2, figsize=(nCols*5+2, nRows*5+2))
+        # figure.tight_layout(pad = 1)
 
-        predFig = axes[0]
-        trueFig = axes[1]
-        compFig = axes[2]
-        cmatFig = axes[3]
+        origFig = axes[0,0]
+        trueFig = axes[0,1]
+
+        predProbMapFig = axes[1,0]
+        predFig = axes[1,1]
+
+        compFig = axes[0,2]
+        cmatFig = axes[1,2]
 
         pred_prob_map = pred_y[sampleIndex,:,:,0]
         pred = pred_prob_map >= 0.5
@@ -113,14 +118,25 @@ def visualize_boundary_hard(pred_y, true_x, true_y, out_dir='./', label=''):
 
         cmat = sklearn.metrics.confusion_matrix(y_true = true.flatten(), y_pred = pred.flatten()) #, labels=[0,1])
 
-        predFig.imshow(skimage.color.label2rgb(pred, image=true_x[sampleIndex,:,:,0]))
+        mappable = origFig.imshow(true_x[sampleIndex,:,:,0])
+        figure.colorbar(mappable, ax=origFig)
+
         trueFig.imshow(skimage.color.label2rgb(true, image=true_x[sampleIndex,:,:,0]))
+
+        mappable = predProbMapFig.imshow(pred_prob_map)
+        figure.colorbar(mappable, ax=predProbMapFig)
+
+        predFig.imshow(skimage.color.label2rgb(pred, image=true_x[sampleIndex,:,:,0]))
+
         compFig.imshow(skimage.color.label2rgb(comp, image=true_x[sampleIndex,:,:,0]))
         cmatFig.matshow(cmat, cmap = "cool")
-
-        predFig.set_title('Prediction')
+
+        predProbMapFig.set_title('Prediction (not thresholded)')
+        origFig.set_title('Image')
+        predFig.set_title('Prediction (thresholded)')
         trueFig.set_title('Truth')
         compFig.set_title('Errors')
+        cmatFig.set_title('Confusion Matrix')
 
         predFig.axis('off')
         trueFig.axis('off')
@@ -145,6 +161,55 @@ def visualize_boundary_hard(pred_y, true_x, true_y, out_dir='./', label=''):
         f = open(out_dir + '/' + label + '_' + str(sampleIndex) + '.txt', 'w')
         f.write('Cross Entropy: ' + str(ce) + '\n')
         f.close()
+
+def visualize_boundary_soft(pred_y, true_x, true_y, out_dir='./', label=''):
+
+    plt.figure()
+    plt.hist(pred_y.flatten())
+    plt.savefig(out_dir + label + '_' + 'hist_probmap_boundary')
+
+    # print all samples for visual inspection
+    nSamples = pred_y.shape[0]
+
+    for sampleIndex in range(nSamples):
+
+        nCols = 4
+        figure, axes = plt.subplots(ncols=nCols, nrows=1, figsize=(nCols*5+2, 5+2))
+
+        origFig = axes[0]
+        predFig = axes[1]
+        trueFig = axes[2]
+        compFig = axes[3]
+
+        pred_prob_map = pred_y[sampleIndex,:,:,0]
+
+        true_prob_map = true_y[sampleIndex,:,:,0]
+
+        comp = pred_prob_map - true_prob_map
+
+        origFig.imshow(true_x[sampleIndex,:,:,0])
+        predFig.imshow(pred_prob_map) 
+        trueFig.imshow(true_prob_map)
+        compFig.imshow(comp)
+
+        origFig.set_title('Image')
+        predFig.set_title('Prediction')
+        trueFig.set_title('Truth')
+        compFig.set_title('Errors (MSE)')
+
+        predFig.axis('off')
+        trueFig.axis('off')
+        compFig.axis('off')
+
+        plt.savefig(out_dir + label + '_' + str(sampleIndex) + '_vis')
+        classNames = ['background', 'boundary']
+
+        # write mean squared error
+        ce = sklearn.metrics.mean_squared_error(y_pred = pred_prob_map.flatten(), y_true = true_prob_map.flatten())
+
+        f = open(out_dir + '/' + label + '_' + str(sampleIndex) + '.txt', 'w')
+        f.write('Mean Squared Error: ' + str(ce) + '\n')
+        f.close()
 
 def visualize_learning_stats(statistics, out_dir, metrics):
     plt.figure()

diff --git a/code/training-hand-200-boundary.py b/code/training-hand-200-boundary.py
@@ -24,22 +24,22 @@
 import numpy as np
 
 # constants
-const_lr = 1e-3
+const_lr = 1e-4
 
-chkpt_file = "../checkpoints/checkpoint_boundary_2.hdf5"
+chkpt_file = "../checkpoints/checkpoint_boundary_6.hdf5"
 
-out_dir = "../out_boundary_2/"
+out_dir = "../out_boundary_6/"
 tb_log_dir = "../logs/logs_tensorboard_boundary/"
 
 train_dir_x = '/home/jr0th/github/segmentation/data/BBBC022_hand_200/training/x'
-train_dir_y = '/home/jr0th/github/segmentation/data/BBBC022_hand_200/training/y_boundary_2'
+train_dir_y = '/home/jr0th/github/segmentation/data/BBBC022_hand_200/training/y_boundary_6'
 
 val_dir_x = '/home/jr0th/github/segmentation/data/BBBC022_hand_200/validation/x'
-val_dir_y = '/home/jr0th/github/segmentation/data/BBBC022_hand_200/validation/y_boundary_2'
+val_dir_y = '/home/jr0th/github/segmentation/data/BBBC022_hand_200/validation/y_boundary_6'
 
 data_type = "images" # "images" or "array"
 
-nb_epoch =  10 # 500
+nb_epoch =  30 # 500
 batch_size = 10
 nb_batches = int(400 / batch_size) # 100 images, 400 patches