Implement a gradient boosting execution time predictor

Generates a set of random options and tries to predict
the execution time.

Author: Jules Pondard

python/experimental/options_search/predict_time.py

@@ -0,0 +1,56 @@
+import time
+import torch
+import tensor_comprehensions as tc
+#import sklearn
+#from sklearn.linear_model import LinearRegression
+#from sklearn.ensemble import GradientBoostingRegressor
+import numpy as np
+#from sklearn.model_selection import train_test_split
+#from tensor_comprehensions.mapping_options import Options
+from multiprocessing import Pool
+from itertools import repeat
+import utils
+#from tqdm import tqdm
+
+(tc_code, tc_name, inp, init_input_sz) = utils.get_convolution_example(size_type="input", inp_sz_list=[8,2,28,28,8,1,1])
+
+NB_HYPERPARAMS, INIT_INPUT_SZ = utils.NB_HYPERPARAMS, utils.INIT_INPUT_SZ
+
+def createY(x):
+    y = utils.evalTime(x)
+    return y
+
+def getRandom():
+    opt_v = np.zeros(NB_HYPERPARAMS).astype(int)
+    for i in range(opt_v.shape[0]):
+        opt_v[i] = np.random.randint(utils.cat_sz[i])
+    return opt_v
+
+def makeDataset():
+    from tqdm import tqdm
+    sz = 500
+    datasetX, datasetY = [], []
+    for _ in tqdm(range(sz)):
+        opt = getRandom()
+        yi = createY(opt)
+        datasetX.append(opt)
+        datasetY.append(yi)
+    #with Pool(sz) as p:
+    #    datasetY = p.starmap(createY, datasetX)
+    return np.array(datasetX), np.array(datasetY)
+
+def learn():
+    #from sklearn.linear_model import LinearRegression
+    from sklearn.ensemble import GradientBoostingRegressor
+    from sklearn.model_selection import train_test_split
+    datasetX, datasetY = makeDataset()
+    print(min(datasetY))
+    Xtrain, Xtest, Ytrain, Ytest = train_test_split(datasetX, datasetY, test_size=0.2, random_state = 42)
+    model1 = GradientBoostingRegressor(n_estimators=1000)
+    model1.fit(Xtrain, Ytrain)
+    pred0 = model1.predict(Xtrain)
+    pred1 = model1.predict(Xtest)
+    print(np.corrcoef(pred0, Ytrain)[0, 1]**2)
+    print(np.corrcoef(pred1, Ytest)[0,1]**2)
+
+#learn()