diff options
| -rw-r--r-- | include/ct1975.c (renamed from include/ct1973.c) | 22 | ||||
| -rw-r--r-- | include/ct1975.h (renamed from include/ct1973.h) | 1 | ||||
| -rw-r--r-- | include/weights.h | 7 | ||||
| -rw-r--r-- | src/ctaklm.c | 2 | ||||
| -rw-r--r-- | src/train5.py | 91 |
5 files changed, 71 insertions, 52 deletions
diff --git a/include/ct1973.c b/include/ct1975.c index 592b0b1..760fcd1 100644 --- a/include/ct1973.c +++ b/include/ct1975.c @@ -1,9 +1,8 @@ -#include "ct1973.h" +#include "ct1975.h" float flattened[CONV_NUM+2]; float dense1[DENSE1_NUM]; float dense2[DENSE2_NUM]; -float output[OUTPUT_NUM]; #define RELU(x) ((x) = ((x)<0)?0:(x)) @@ -76,14 +75,15 @@ evaluate_black_win(void) { /* ------------- * * Output layer * * ------------- */ - float norm = 0; - for (uint8_t out = 0; out < OUTPUT_NUM; out++) { - output[out] = output_biases[out]; - for (uint8_t d2 = 0; d2 < DENSE2_NUM; d2++) { - output[out] += dense2[d2]*output_weights[out][d2]; - } - output[out] = exp(output[out]); - norm += output[out]; + float output = output_bias; + for (uint8_t d2 = 0; d2 < DENSE2_NUM; d2++) { + output += dense2[d2]*output_weights[d2]; } - return output[1]/norm; + + // Truncated Pade approximant of logistic function + output = (12.0+output+50.0*output/(output*output+10.0))/24.0; + if (output > 1.0) return 1.0; + else if (output < 0.0) return 0.0; + + return output; } diff --git a/include/ct1973.h b/include/ct1975.h index 6350ff3..79c20d2 100644 --- a/include/ct1973.h +++ b/include/ct1975.h @@ -1,5 +1,4 @@ #include <stdint.h> -#include <math.h> #include "tak.h" #include "weights.h" diff --git a/include/weights.h b/include/weights.h index 3c83010..18079ba 100644 --- a/include/weights.h +++ b/include/weights.h @@ -5,8 +5,7 @@ #define CONV_NUM (KERN_NUM * KERN_OSIZE * KERN_OSIZE) // 108 #define DENSE1_NUM 9 -#define DENSE2_NUM 8 -#define OUTPUT_NUM 2 +#define DENSE2_NUM 9 extern const float conv2d_weights[KERN_NUM][KERN_SIZE][KERN_SIZE][KERN_CHAN]; extern const float conv2d_biases[KERN_NUM]; @@ -14,5 +13,5 @@ extern const float dense1_weights[DENSE1_NUM][CONV_NUM+2]; extern const float dense1_biases[DENSE1_NUM]; extern const float dense2_weights[DENSE2_NUM][DENSE1_NUM]; extern const float dense2_biases[DENSE2_NUM]; -extern const float output_weights[OUTPUT_NUM][DENSE2_NUM]; -extern const float output_biases[2]; +extern const float output_weights[DENSE2_NUM]; +extern const float output_bias; diff --git a/src/ctaklm.c b/src/ctaklm.c index a5bfe05..5df23c9 100644 --- a/src/ctaklm.c +++ b/src/ctaklm.c @@ -4,7 +4,7 @@ #include <linenoise.h> #include <tak.h> -#include <ct1973.h> +#include <ct1975.h> const char *blk = "\033[41m", *wht = "\033[44m"; const char *rev = "\033[7m", *und = "\033[4m", *rst = "\033[0m"; diff --git a/src/train5.py b/src/train5.py index 03e9669..214ad88 100644 --- a/src/train5.py +++ b/src/train5.py @@ -1,10 +1,19 @@ -# 30 10(11) 5 # 27 17 4 +# Loading data import pandas as pd import numpy as np +# Output of weights +import re from tensorflow import transpose +# Custom activation function +from tensorflow import constant as k +from tensorflow.math import add, divide, maximum, minimum, multiply, square + +# Computing size from tensorflow.keras.backend import get_value + +# Building models from tensorflow.keras.models import Model from tensorflow.keras.layers import Concatenate from tensorflow.keras.layers import Input @@ -13,28 +22,16 @@ from tensorflow.keras.layers import Convolution2D from tensorflow.keras.layers import Flatten -def load_data(size): - shape = (-1, size, size, 8) - - tr_fn = "training-"+str(size)+".csv" - training_csv = pd.read_csv(tr_fn) # .head(80000) - training_data = training_csv - training_stack_input = np.array(training_data.iloc[:, 2:-2]).reshape(shape, order='F') - training_flats_input = np.array(training_data.iloc[:, 0:2]) - training_input = [training_stack_input, training_flats_input] - training_outcome = training_data.iloc[:, -2:] - - val_fn = "validation-"+str(size)+".csv" - val_data = pd.read_csv(val_fn).tail(20000) - val_stack_input = np.array(val_data.iloc[:, 2:-2]).reshape(shape, order='F') - val_flats_input = np.array(val_data.iloc[:, 0:2]) - val_input = [val_stack_input, val_flats_input] - val_outcome = val_data.iloc[:, -2:] - - return [(training_input, training_outcome), (val_input, val_outcome)] - +# We need something that's close to logistic, but cheaper to compute. +# (12.0+x+50.0*x/(x*x+10.0))/24.0, clamped between 0 and 1 +# as it would otherwise exceed this range at +- 4.6 or so +def truncated_pade_logistic(x): + val = add(k(12.0), + add(x, multiply(k(50.0), + divide(x, add(square(x), k(10.0)))))) + return minimum(k(1.0), maximum(k(0.0), divide(val, k(24.0)))) -def make_model(size, magic=[9, 18, 9]): +def make_model(size, magic=[12, 9, 9]): # Our model for the stacks, a small CNN stack_shape = (size, size, 8) stack_input = Input(shape=stack_shape) @@ -43,25 +40,43 @@ def make_model(size, magic=[9, 18, 9]): use_bias=True)(stack_input) stack_model = Flatten()(stack_model) stack_model = Model(inputs=stack_input, outputs=stack_model) - # The overall model flats_input = Input(shape=(2,)) combn_input = Concatenate()([stack_model.output, flats_input]) model = Dense(magic[1], activation="relu", use_bias=True)(combn_input) model = Dense(magic[2], activation="relu", use_bias=True)(model) - model = Dense(2, activation="softmax", use_bias=True)(model) + model = Dense(1, activation=truncated_pade_logistic, use_bias=True)(model) model = Model(inputs=[stack_model.input, flats_input], outputs=model) - model.compile(optimizer='adam', - loss='binary_crossentropy', + # loss='binary_crossentropy', # loss='categorical_crossentropy', - # loss='mean_squared_error', + loss='mean_squared_error', + # loss='mean_absolute_error', metrics=['accuracy']) - model.summary() return model +def load_data(size): + shape = (-1, size, size, 8) + + tr_fn = "training-"+str(size)+".csv" + training_csv = pd.read_csv(tr_fn) #.head(100000) + training_data = training_csv + training_stack_input = np.array(training_data.iloc[:, 2:-1]).reshape(shape, order='F') + training_flats_input = np.array(training_data.iloc[:, 0:2]) + training_input = [training_stack_input, training_flats_input] + training_outcome = training_data.iloc[:, -1:] + + val_fn = "validation-"+str(size)+".csv" + val_data = pd.read_csv(val_fn).tail(20000) + val_stack_input = np.array(val_data.iloc[:, 2:-1]).reshape(shape, order='F') + val_flats_input = np.array(val_data.iloc[:, 0:2]) + val_input = [val_stack_input, val_flats_input] + val_outcome = val_data.iloc[:, -1:] + + return [(training_input, training_outcome), (val_input, val_outcome)] + def train(size, model, data, iterations=1, epochs=10): (training_input, training_outcome), (val_input, val_outcome) = data results = [] @@ -84,21 +99,26 @@ def model_size(model): def magic_search(data): results = [] - for width in range(10, 16): + for width in range(12, 16): for dense1 in range(8, 24): for dense2 in range(6, dense1+1): m = make_model(5, [width, dense1, dense2]) if model_size(m) <= 2000: results += [([width, dense1, dense2], - train(5, m, data, iterations=1, epochs=10))] + train(5, m, data, iterations=1, epochs=20))] print("\nSummary") for r in results: - print("Parameters {0}: {1}".format(r[0], r[1][-1:][0])) + print("Parameters {0}: {1}".format(r[0], r[1][0][-1:][0])) + return results def write_weights(model): def fix(string): - return string.replace("[", "{").replace("]", "}") + string = string.replace("[", "{").replace("]", "}") + string = re.sub(r'}\n', '},\n', string) + string = re.sub(r'([0-9]+)\n', r'\1,\n', string) + string = re.sub(r'([0-9]+) ', r'\1, ', string) + return string f = open("weights.txt", "w") conv2d_weights = np.array(transpose(model.trainable_variables[0], perm=[3, 1, 0, 2])) conv2d_biases = np.array(model.trainable_variables[1]) @@ -112,16 +132,17 @@ def write_weights(model): dense1_weights, dense1_biases, dense2_weights, dense2_biases, output_weights, output_biases]: - f.write(fix(str(v))+"\n") + f.write(fix(str(v))+"\n\n") f.close() data = load_data(5) -model = make_model(5, [12, 9, 8]) +model = make_model(5, [12, 9, 9]) results, model = train(5, model, data, iterations=10, epochs=20) print("\nScores") for i in range(len(results)): print("Iteration {0}: {1}".format(i+1, results[i])) + write_weights(model) -# magic_search(data) +#results = magic_search(data) |