Random Forest always predicts same class - machine-learning

Question

I have been trying to make a Random Forest model for Sign Language classification. The Data Set is balanced and the accuracy is 98%, but it always predicts the same class.
This is how I get and process the data:
train_dir = "../input/asl-alphabet/asl_alphabet_train/asl_alphabet_train/"
test_dir = "../input/asl-alphabet/asl_alphabet_test/asl_alphabet_test/"
train_len = 87000
image_size=50
def get_data(folder):
X = np.empty((train_len, image_size, image_size, 3), dtype = np.float32)
y = np.empty((train_len), dtype = int)
cnt = 0
for folderName in os.listdir(folder):
if not folderName.startswith('.'):
if folderName in ['A']:
label = 0
elif folderName in ['B']:
label = 1
elif folderName in ['C']:
label = 2
elif folderName in ['D']:
label = 3
elif folderName in ['E']:
label = 4
elif folderName in ['F']:
label = 5
elif folderName in ['G']:
label = 6
elif folderName in ['H']:
label = 7
elif folderName in ['I']:
label = 8
elif folderName in ['J']:
label = 9
elif folderName in ['K']:
label = 10
elif folderName in ['L']:
label = 11
elif folderName in ['M']:
label = 12
elif folderName in ['N']:
label = 13
elif folderName in ['O']:
label = 14
elif folderName in ['P']:
label = 15
elif folderName in ['Q']:
label = 16
elif folderName in ['R']:
label = 17
elif folderName in ['S']:
label = 18
elif folderName in ['T']:
label = 19
elif folderName in ['U']:
label = 20
elif folderName in ['V']:
label = 21
elif folderName in ['W']:
label = 22
elif folderName in ['X']:
label = 23
elif folderName in ['Y']:
label = 24
elif folderName in ['Z']:
label = 25
elif folderName in ['del']:
label = 26
elif folderName in ['nothing']:
label = 27
elif folderName in ['space']:
label = 28
else:
label = 29
for image_filename in tqdm(os.listdir(folder + folderName)):
img_file = cv2.imread(folder + folderName + '/' + image_filename)
if img_file is not None:
img_file = skimage.transform.resize(img_file, (image_size, image_size, 3))
img_arr = np.asarray(img_file).reshape((-1, image_size, image_size, 3))
X[cnt] = img_arr
y[cnt] = label
cnt += 1
return X,y
letters, labels= get_data(train_dir)
print("The shape of letters is : ", letters.shape)
#preprocessing
import tensorflow as tf
X_train, X_test, y_train, y_test = train_test_split(letters, labels, test_size=0.3, random_state=42, stratify=labels)
X_valid, X_train = X_train[:20000] / 255., X_train[20000:] / 255.
y_valid, y_train = y_train[:20000], y_train[20000:]
X_test = X_test / 255.
#reshape to 2d array
nsamples, nx, ny, nrgb = X_train.shape
x_train2 = X_train.reshape((nsamples,nx*ny*nrgb))
#so,eventually,model.predict() should also be a 2d input
nsamples, nx, ny, nrgb = X_test.shape
x_test2 = X_test.reshape((nsamples,nx*ny*nrgb))
This is how I made the model:
from sklearn.ensemble import RandomForestClassifier
model=RandomForestClassifier(random_state=42, n_jobs=-1, max_depth=20,
n_estimators=100, oob_score=True)
model.fit(x_train2,y_train)
y_pred=model.predict(x_test2)
y_pred
#y_pred = clf.predict(X_test)
accuracy_score(y_test, y_pred)
accuracy_score(y_pred,y_test)
print(classification_report(y_pred,y_test))
To test the predictions I use:
label_names = ['A','B','C','D','E','F','G','H','I','J','K','L','M','N',
'O','P','Q','R','S','T','U','V','W','X','Y','Z','del','nothing','space']
img_path='../input/asl-alphabet/asl_alphabet_test/asl_alphabet_test/L_test.jpg'
img_arr=cv2.imread(img_path)
img_arr=cv2.resize(img_arr,(64,64))
#so,eventually,model.predict() should also be a 2d input
nx, ny, nrgb = img_arr.shape
img_arr2 = img_arr.reshape(1,(nx*ny*nrgb))
classes = ["A","B","C","D","E","F","G","H","I","J", "K", "L", "M" "N", "O", "P", "Q", "R", "S", "T", "U", "V", "W", "X", "Y", "Z", "nothing", "space"]
ans=model.predict(img_arr2)
print(label_names[ans[0]])
I have tried to implement Grid Search but I wasn't able to make it work. (I am writing it on Kaggle and it tells "Your notebook tried to allocate more memory than is available")
I am quite new to this, so I am not completely certain about anything in my code.

Answer 1

You are missing the preprocessing of the images between training and testing. You need to be applying the same steps on the test set so your model is getting same kind of inputs. Try the following:
label_names = ['A','B','C','D','E','F','G','H','I','J','K','L','M','N',
'O','P','Q','R','S','T','U','V','W','X','Y','Z','del','nothing','space']
img_path='../input/asl-alphabet/asl_alphabet_test/asl_alphabet_test/L_test.jpg'
img_arr=cv2.imread(img_path)
# Preprocess test image
img_arr = skimage.transform.resize(img_arr, (50, 50, 3))
img_arr = img_arr/255.
#so,eventually,model.predict() should also be a 2d input
nx, ny, nrgb = img_arr.shape
img_arr2 = img_arr.reshape(1,(nx*ny*nrgb))
classes = ["A","B","C","D","E","F","G","H","I","J", "K", "L", "M" "N", "O", "P", "Q", "R", "S", "T", "U", "V", "W", "X", "Y", "Z", "nothing", "space"]
ans=model.predict(img_arr2)
print(label_names[ans[0]])