Gist for evaluation of xgboost speeds for different hp

hp-speed-xgoobst.py

# Defining the grid of parameters 
n_estimators = [30, 100, 200, 300, 600, 900]
max_depth = [4, 6, 8, 12, 14, 16]

# Number of rows to test on 
nrows = 30000

# Creating a dictionary
gpu_dict = {
    'objective': ['reg:squarederror'], 
    'n_estimators': n_estimators,
    'max_depth': max_depth,
    'tree_method': ['gpu_hist']
}

cpu_dict = {
    'objective': ['reg:squarederror'], 
    'n_estimators': n_estimators,
    'max_depth': max_depth
}

# Creating the grid of hyper parameters 
gpu_hp = ParameterGrid(gpu_dict)
cpu_hp = ParameterGrid(cpu_dict)

# Getting the random subset of data 
subset = d.sample(nrows)

# Iterating through all hte hyper parameters: 
gpu_speeds = []
for params in gpu_hp:
    # Creating X and Y 
    X = pd.get_dummies(subset[features])
    Y = subset['Sales']
    
    # Initiating the model objects
    gpu = xgb.XGBRegressor(**params)
    
    # Training on cpu 
    start = time.time()
    gpu.fit(X, Y)
    gpu_speed = time.time() - start
    
    gpu_speeds.append(gpu_speed)
    
# Iterating through all hte hyper parameters: 
cpu_speeds = []
for params in cpu_hp:
    # Creating X and Y 
    X = pd.get_dummies(subset[features])
    Y = subset['Sales']
    
    # Initiating the model objects
    cpu = xgb.XGBRegressor(**params)
    
    # Training on cpu 
    start = time.time()
    cpu.fit(X, Y)
    cpu_speed = time.time() - start
    
    cpu_speeds.append(cpu_speed)
    
# Creating the speed frame 
speeds = pd.DataFrame({
    'n_estimators': [x.get('n_estimators') for x in list(cpu_hp)],
    'max_depth': [x.get('max_depth') for x in list(cpu_hp)],
    'cpu_speed': cpu_speeds,
    'gpu_speed': gpu_speeds
})

# Diff of speed between GPU and CPU
speeds['diff'] = speeds['cpu_speed'] - speeds['gpu_speed']

# Creating the coordinates for surface plot
X, Y = np.meshgrid(n_estimators, max_depth)
Z = speeds['diff'].values
Z = np.array(Z).reshape((len(n_estimators), len(max_depth)))

# PLoting
fig = plt.figure(figsize=(12, 12))
ax = plt.axes(projection='3d')
ax.plot_surface(X, Y, Z, rstride=1, cstride=1, edgecolor='none', alpha=0.7)
ax.set_xlabel('n_estimators')
ax.set_ylabel('max_depth')
ax.set_zlabel('GPU speed advantage (s)')
plt.show()