Merge pull request #281 from PanduhBeer/master

Add ThumbyAI Neural Network Demo

Author: TinyCircuits

ThumbyAI/ThumbyAI.py

@@ -0,0 +1,365 @@
+import thumby
+import time
+import math
+import random
+import gc
+
+#
+# --------------------------------------------------------------------
+# FONT & SCREEN SETUP
+# --------------------------------------------------------------------
+# Thumby includes a 72x40 pixel display. The 3x5 font in "/lib/font3x5.bin"
+# saves space when drawing text on screen.
+thumby.display.setFont("/lib/font3x5.bin", 3, 5, 1)
+SCREEN_W = 72
+SCREEN_H = 40
+
+# Each 3x5 character occupies about 4 pixels in width (3 for the glyph + 1 spacing),
+# so a single line can hold ~18 characters in 72 pixels of width.
+MAX_CHARS_PER_LINE = 18
+
+#
+# --------------------------------------------------------------------
+# DATASET: CELSIUS -> FAHRENHEIT (21 SAMPLES)
+# --------------------------------------------------------------------
+# Celsius values range from 0 to 100 in steps of 5, producing a list of 21 total 
+# samples as inputs for training.
+#
+# Fahrenheit is computed by F = (9/5)*C + 32, producing a corresponding list of 21
+# samples as targets for training.
+#  
+# The network sees only the numeric training pairs.
+c_values = list(range(0, 101, 5))
+f_values = [(9/5)*c + 32 for c in c_values]
+
+# Normalization is required because the hidden layer uses sigmoid,
+# which expects inputs and outputs in or near [0..1].
+def normalize_c(c):
+    return c / 100.0
+
+def normalize_f(f):
+    return (f - 32.0) / 180.0
+
+def denormalize_f(fn):
+    return fn * 180.0 + 32.0
+
+# Build the training set by normalizing c_values and f_values.
+training_inputs  = [[normalize_c(c)] for c in c_values]
+training_targets = [[normalize_f(f)] for f in f_values]
+
+#
+# --------------------------------------------------------------------
+# NEURAL NETWORK: 1 -> 5 -> 1 (LINEAR OUTPUT)
+# --------------------------------------------------------------------
+# This network has:
+#   - 1 input neuron (normalized C)
+#   - 5 hidden neurons (using sigmoid)
+#   - 1 output neuron (using a linear activation)
+# The final output is unconstrained, which helps a linear mapping like C->F.
+INPUT_SIZE  = 1
+HIDDEN_SIZE = 5
+OUTPUT_SIZE = 1
+
+W1 = []
+B1 = []
+W2 = []
+B2 = []
+
+# Training parameters
+LEARNING_RATE = 0.2   # Lower rate for stability
+EPOCHS = 500          # Increase to improve convergence
+
+def init_network():
+    # Initializes all weights/biases randomly in [-1, 1].
+    # Runs garbage collection to free memory on the Thumby.
+    global W1, B1, W2, B2
+    gc.collect()
+
+    def rand_weight():
+        return random.uniform(-1, 1)
+
+    # W1 is (HIDDEN_SIZE x INPUT_SIZE), B1 is (HIDDEN_SIZE,)
+    W1 = [[rand_weight() for _ in range(INPUT_SIZE)] for _ in range(HIDDEN_SIZE)]
+    B1 = [rand_weight() for _ in range(HIDDEN_SIZE)]
+
+    # W2 is (OUTPUT_SIZE x HIDDEN_SIZE), B2 is (OUTPUT_SIZE,)
+    W2 = [[rand_weight() for _ in range(HIDDEN_SIZE)] for _ in range(OUTPUT_SIZE)]
+    B2 = [rand_weight() for _ in range(OUTPUT_SIZE)]
+
+    gc.collect()
+
+def sigmoid(x):
+    # Applies a sigmoid clamp to avoid overflow on exponent.
+    if x > 10:
+        x = 10
+    if x < -10:
+        x = -10
+    return 1.0 / (1.0 + math.e**(-x))
+
+def dsigmoid(y):
+    # Computes derivative of sigmoid if y=sigmoid(x).
+    return y * (1.0 - y)
+
+#
+# --------------------------------------------------------------------
+# FORWARD PASS (HIDDEN: SIGMOID, OUTPUT: LINEAR)
+# --------------------------------------------------------------------
+def forward_pass(x):
+    # Passes input x through the hidden layer using sigmoid,
+    # then computes a linear output with no final sigmoid.
+    global W1, B1, W2, B2
+
+    # Hidden layer
+    h = []
+    for i in range(HIDDEN_SIZE):
+        sum_h = B1[i]
+        for j in range(INPUT_SIZE):
+            sum_h += W1[i][j] * x[j]
+        h.append(sigmoid(sum_h))
+
+    # Output layer (linear)
+    o = []
+    for i in range(OUTPUT_SIZE):
+        sum_o = B2[i]
+        for j in range(HIDDEN_SIZE):
+            sum_o += W2[i][j] * h[j]
+        o.append(sum_o)  # raw linear output
+    return h, o
+
+#
+# --------------------------------------------------------------------
+# BACKPROP (FINAL LAYER IS LINEAR)
+# --------------------------------------------------------------------
+def backprop(x, h, o, t):
+    # Adjusts weights/biases in two stages:
+    #  1) Linear output error
+    #  2) Sigmoid hidden layer error
+    global W1, B1, W2, B2
+
+    # Output layer error: (t - o) for a linear final neuron
+    output_errors = []
+    for i in range(OUTPUT_SIZE):
+        error = t[i] - o[i]
+        output_errors.append(error)
+
+    # Hidden layer errors (uses dsigmoid on the hidden output)
+    hidden_errors = []
+    for i in range(HIDDEN_SIZE):
+        err_sum = 0.0
+        for j in range(OUTPUT_SIZE):
+            err_sum += output_errors[j] * W2[j][i]
+        hidden_errors.append(err_sum * dsigmoid(h[i]))
+
+    # Update W2, B2 (linear output layer)
+    for i in range(OUTPUT_SIZE):
+        for j in range(HIDDEN_SIZE):
+            W2[i][j] += LEARNING_RATE * output_errors[i] * h[j]
+        B2[i] += LEARNING_RATE * output_errors[i]
+
+    # Update W1, B1 (sigmoid hidden layer)
+    for i in range(HIDDEN_SIZE):
+        for j in range(INPUT_SIZE):
+            W1[i][j] += LEARNING_RATE * hidden_errors[i] * x[j]
+        B1[i] += LEARNING_RATE * hidden_errors[i]
+
+#
+# --------------------------------------------------------------------
+# TRAIN NETWORK
+# --------------------------------------------------------------------
+def train_network():
+    # Runs EPOCHS times through the entire dataset.
+    for epoch in range(EPOCHS):
+        for i, inp in enumerate(training_inputs):
+            t = training_targets[i]
+            h, o = forward_pass(inp)
+            backprop(inp, h, o, t)
+
+#
+# --------------------------------------------------------------------
+# TRAINING SCREEN
+# --------------------------------------------------------------------
+def display_training_screen():
+    # Clears the screen, draws "TRAINING..." centered, then updates.
+    thumby.display.fill(0)
+    text_str = "TRAINING..."
+    text_str_2 = "Please Wait"
+    text_width = len(text_str) * 4
+    x = (SCREEN_W - text_width) // 2
+    y = (SCREEN_H - 7) // 2
+    thumby.display.drawText(text_str, x, y - 2, 1)
+    thumby.display.drawText(text_str_2, x, y + 6, 1)
+    thumby.display.update()
+
+#
+# --------------------------------------------------------------------
+# WRAPPING FUNCTIONS FOR TUTORIAL
+# --------------------------------------------------------------------
+def wrap_text(line, max_chars=18):
+    wrapped = []
+    start = 0
+    while start < len(line):
+        wrapped.append(line[start:start+max_chars])
+        start += max_chars
+    return wrapped
+
+def wrap_tutorial_lines(lines, max_chars=18):
+    final = []
+    for l in lines:
+        if len(l) > max_chars:
+            sub_lines = wrap_text(l, max_chars)
+            final.extend(sub_lines)
+        else:
+            final.append(l)
+    return final
+
+#
+# --------------------------------------------------------------------
+# SCROLLABLE TUTORIAL SCREEN
+# --------------------------------------------------------------------
+def display_tutorial():
+    raw_tutorial_lines = [
+        "C->F",
+        "",
+        "1->5->1 network,",
+        "sigmoid hidden,",
+        "linear output.",
+        "Steps of 5C from",
+        "0 to 100.",
+        "500 epochs,",
+        "LR=0.2.",
+        "Should predict",
+        "100C near 212F.",
+        "",
+        "Press Up/Down to",
+        "scroll. Press B to exit."
+    ]
+
+    tutorial_lines = wrap_tutorial_lines(raw_tutorial_lines, MAX_CHARS_PER_LINE)
+    scroll_offset = 0
+
+    while True:
+        thumby.display.fill(0)
+        for i in range(5):
+            line_index = scroll_offset + i
+            if line_index < len(tutorial_lines):
+                thumby.display.drawText(tutorial_lines[line_index], 0, i*7, 1)
+        thumby.display.update()
+
+        if thumby.buttonD.justPressed():
+            if scroll_offset < (len(tutorial_lines) - 5):
+                scroll_offset += 1
+            time.sleep(0.1)
+        if thumby.buttonU.justPressed():
+            if scroll_offset > 0:
+                scroll_offset -= 1
+            time.sleep(0.1)
+        if thumby.buttonB.justPressed():
+            break
+        time.sleep(0.05)
+
+#
+# --------------------------------------------------------------------
+# SCROLLABLE RESULTS SCREEN WITH WEIGHTS
+# --------------------------------------------------------------------
+def display_results():
+    # Displays final predictions and then shows final weights/biases.
+    global W1, B1, W2, B2
+
+    result_lines = []
+    result_lines.append("TRAINED NN RESULTS")
+
+    # Predict F for each Celsius in c_values
+    for c_val in c_values:
+        c_norm = [normalize_c(c_val)]
+        _, out = forward_pass(c_norm)
+        f_pred = denormalize_f(out[0])
+        line = f"{c_val}C => {f_pred:.1f}F"
+        result_lines.append(line)
+
+    # Show final weights and biases. Round to 2 decimals to fit display space.
+    # W1: shape (5,1) since there is 1 input and 5 hidden
+    # B1: shape (5,)
+    # W2: shape (1,5) since there is 1 output and 5 hidden
+    # B2: shape (1,)
+
+    # For W1, gather each row's single value (since input=1)
+    w1_values = [round(W1[i][0], 2) for i in range(HIDDEN_SIZE)]
+    b1_values = [round(B1[i], 2) for i in range(HIDDEN_SIZE)]
+    w2_values = [round(W2[0][i], 2) for i in range(HIDDEN_SIZE)]  # row 0, col i
+    b2_value  = round(B2[0], 2)
+
+    result_lines.append("----------")
+    result_lines.append("W1:"+str(w1_values))
+    result_lines.append("B1:"+str(b1_values))
+    result_lines.append("W2:"+str(w2_values))
+    result_lines.append("B2:"+str([b2_value]))
+    result_lines.append("PRESS B TO EXIT")
+    
+    scroll_offset = 0
+    while True:
+        thumby.display.fill(0)
+        for i in range(5):
+            line_index = scroll_offset + i
+            if line_index < len(result_lines):
+                thumby.display.drawText(result_lines[line_index], 0, i*7, 1)
+        thumby.display.update()
+
+        if thumby.buttonD.justPressed():
+            if scroll_offset < (len(result_lines) - 5):
+                scroll_offset += 1
+            time.sleep(0.1)
+        if thumby.buttonU.justPressed():
+            if scroll_offset > 0:
+                scroll_offset -= 1
+            time.sleep(0.1)
+        if thumby.buttonB.justPressed():
+            break
+        time.sleep(0.05)
+
+#
+# --------------------------------------------------------------------
+# MAIN MENU
+# --------------------------------------------------------------------
+menu_items = ["Train NN", "Tutorial", "Exit"]
+selected_index = 0
+
+def draw_menu():
+    thumby.display.fill(0)
+    thumby.display.drawText("C->F AI DEMO", 0, 0, 1)
+    y = 10
+    for i, item in enumerate(menu_items):
+        marker = ">" if i == selected_index else " "
+        line = marker + " " + item
+        thumby.display.drawText(line, 0, y, 1)
+        y += 8
+    thumby.display.update()
+
+def main_menu():
+    global selected_index
+    while True:
+        draw_menu()
+        if thumby.buttonU.justPressed():
+            selected_index = max(0, selected_index - 1)
+            time.sleep(0.1)
+        if thumby.buttonD.justPressed():
+            selected_index = min(len(menu_items)-1, selected_index + 1)
+            time.sleep(0.1)
+
+        if thumby.buttonA.justPressed():
+            choice = menu_items[selected_index]
+            if choice == "Train NN":
+                init_network()
+                display_training_screen()  # Shows TRAINING... in the center
+                train_network()
+                display_results()
+            elif choice == "Tutorial":
+                display_tutorial()
+            elif choice == "Exit":
+                break
+        time.sleep(0.1)
+
+#
+# --------------------------------------------------------------------
+# START PROGRAM
+# --------------------------------------------------------------------
+main_menu()