Module 5 Test: Deep Learning
Ye tumhara progress tracker hai. 20 sawal: 5 Easy, 10 Moderate, 5 Hard . Pehle sab sawal khud hal karo (code ko Colab me chalane se pehle apne dimagh se output sochna), phir Answer Key se check karo.
| Score | Matlab |
|---|---|
| 18 se 20 | Zabardast! Module 5 par tumhari grip mazboot hai. |
| 12 se 17 | Acha hai. Jin sawalon me ghalti hui, un chapters ka Key Takeaways dobara parho. |
| 12 se kam | Koi baat nahi. Chapters dobara parho, code khud chalao, aur test dobara do. |
EASY
Q1. Ek neuron apna output kaise banata hai?
-
A) Sirf inputs ko jama karta hai
-
B) Weighted sum + bias nikal kar activation function se guzarta hai
-
C) Sirf activation function lagata hai
-
D) Loss ko ghata kar
Q2. Hidden layers me aam taur par pehli pasand kaun sa activation function hota hai?
- A) Softmax
B) Sigmoid C) ReLU
D) Koi nahi
Q3. 10 classes wale multiclass classification (integer labels) ki aakhri layer kaun si honi chahiye?
-
A) Dense(1, activation="sigmoid")
-
B) Dense(10, activation="softmax")
-
C) Dense(10) (koi activation nahi)
D) Dense(1, activation="relu")
Q4. CNN me Max Pooling ka kaam kya hai?
A) Weights ko zero karna
B) Labels ko encode karna
- C) Overfitting ko barhana
D) Feature map ka size chhota karna aur har hisse ka sab se bara number rakhna
Q5. LSTM me kaun sa gate tay karta hai ke purani yaaddasht (cell state) me se kitna bhool jana hai?
A) Forget gate
B) Input gate
- C) Output gate
D) Reset gate
MODERATE
Q6. Ek Dense(16) layer ko 20 features wala input milta hai. Is layer ke kitne parameters hain?
A) 320 B) 336 C) 16 D) 356
Q7. Conv2D(32, (3, 3)) layer (padding="valid") par input shape (28, 28, 1) hai. Is layer ke parameters kitne hain?
A) 288 B) 296 C) 320
D) 896
Q8. Conv layer ke output size ka formula lagao: input 28x28, filter 5x5, padding 0, stride 1. Output size kya hoga?
A) 24x24 B) 23x23 C) 28x28
D) 25x25
Q9. Training loss girta ja raha hai magar validation loss ek point ke baad uper jane laga. Ye kis cheez ki nishani hai?
A) Underfitting B) Overfitting
C) Learning rate ka bilkul zero hona
D) Data ka scale hona
Q10. Dropout(0.3) lagane ka sahi matlab kya hai?
A) 70% weights hamesha ke liye hata diye jate hain
B) Testing me 30% neurons band hote hain
C) Learning rate 0.3 ho jata hai
D) Training me har step par taqreeban 30% neurons random band hote hain
Q11. Time series forecasting me train/test split ka sahi tareeqa kaun sa hai?
A) Random shuffle kar ke 80/20
B) Purana data train aur uske baad ka naya data test
C) Test pehle aur train baad me
D) K-Fold shuffle ke saath
Q12. Is code ka output kya hoga?
scores = np.array([2.0, 1.0, 0.1])
e = np.exp(scores)
p = e / e.sum()
print(p.argmax(), round(p.sum(), 1))
A) 2 1.0 B) 0 3.1 C) 0 1.0 D) 1 1.0
**Q13. Transfer Learning me
base.trainable = False
likhne se kya hota hai?**
A) Pre-trained base ke weights freeze ho jate hain, sirf naya head train hota hai
B) Pura model delete ho jata hai
C) Base ke weights random ho jate hain
D) Data augmentation shuru ho jati hai
Q14. Autoencoder ko sirf normal data par train kiya gaya. Test me anomaly (anokhi cheez) ka reconstruction error aam taur par kaisa hota hai?
A) Normal data se kam
B) Bilkul zero
C) Hamesha barabar
D) Normal data se zyada
Q15. Adam optimizer ke baare me kaun sa bayan sahi hai?
A) Ye sirf seedha SGD hai
B) Momentum aur RMSprop ke idea ko mila kar banta hai
C) Isme learning rate hota hi nahi
D) Sirf CNN ke liye hai
HARD
Q16. Conv2D(64, (3, 3)) layer ka input 32 channels ka hai. Is layer ke parameters kitne hain?
A) 18,432 B) 2,080 C) 18,496 D) 576
Q17. LSTM(32) layer ko har time step par sirf 1 feature milta hai. Is layer ke parameters kitne hain? A) 4,352 B) 3,360 C) 1,088 D) 4,224
Q18. Flatten layer ka output size kitna hoga?
model = keras.Sequential([
keras.Input(shape=(28, 28, 1)),
layers.Conv2D(8, 3, padding="same", activation="relu"),
layers.MaxPooling2D(2),
layers.Conv2D(16, 3, activation="relu"),
layers.MaxPooling2D(2),
layers.Flatten(),
])
A) 784 B) 1024 C) 400 D) 576
Q19. Ek GAN ki training me Generator ka loss kam hota ja raha hai. Is se kya natija nikala ja sakta hai?
A) Generated samples yaqeenan behtareen hain
B) Discriminator kharab ho gaya hai
C) Kuch pakka nahi keh sakte: dono loss ek doosre se jude hain, samples khud dekhne chahiyen D) Training rokne ka waqt aa gaya
Q20. Is code se X aur y ki shapes kya hongi?
def make_windows(values, window):
X, y = [], []
for i in range(window, len(values)):
X.append(values[i - window:i])
y.append(values[i])
return np.array(X), np.array(y)
X, y = make_windows(np.arange(10), 4)
print(X.shape, y.shape)
A) (10, 4) aur (10,) B) (6, 4) aur (6,) C) (4, 6) aur (6,) D) (6, 4) aur (6, 4)
Module 5 Answer Key (Jawabat aur wazahat)
| # | Level | Jawab | Wazahat |
|---|---|---|---|
| 1 | Easy | B) Weighted sum + bias nikal kar activation function se guzarta hai | Neuron pehle weights se guna kar jama karta hai, bias jodta hai, phir activation (jaise ReLU/sigmoid) lagata hai. |