add yfinance LSTM example

examples/ml/lstm/lstm.kg

@@ -18,7 +18,7 @@ trainx::((#windows)-1)#windows;trainy::((#labels)-1)#labels
 testx::windows@-1;testy::labels@-1
 
 :" print the first window and label "
-.d("series ");.d(1#trainw);.d(" predicts ");.p(1#trainl)
+.d("series ");.d(1#trainx);.d(" predicts ");.p(1#trainy)
 
 :" create the stateful model trainer "
 t::trainer(m)

examples/stocks/yfinance/fetch.kg

@@ -1,6 +1,5 @@
 .py("klongpy.db")
-.pyf("numpy";"std")
-.pyf("numpy";"mean")
+.pyf("numpy";["std" "mean"])
 
 ;" pip3 install yfinance "
 .pyf("yfinance";"download")

examples/stocks/yfinance/lstm.kg

@@ -0,0 +1,51 @@
+.py("klongpy.db")
+.py("lstm.py")
+.py("mkwindows.py")
+
+;" pip3 install yfinance "
+.pyf("yfinance";"download")
+fetch::{[args];args:::{};args,"tickers",,x;args,"start",,y;args,"end",,z;.pyc(download;[];args)}
+
+:" yfinance returns a pandas dataframe so we need to convert it to a KlongPy table "
+data::fetch("MSFT";"2020-01-01"; "2023-01-01"])
+T::.table(data)
+
+:" get close prices "
+close::T?"Close"
+
+k::10
+windows::mkwindows(close;k)
+
+:" create the model wrapper "
+m::model(1;100;1)
+
+:" create windowed data "
+windows::mkwindows(close;k)
+labels::{x@0}'1_windows
+windows::(-1)_windows
+
+.d("series ");.d(1#windows);.d(" predicts ");.p(1#labels)
+
+:" normalize windows and labels "
+wmax::{(x@>x)@0}'windows
+windows::wmax{y%x}'windows
+labels::wmax{y%x}'labels
+
+trainx::(-1)_windows;trainy::(-1)_labels
+testx::windows@-1;testy::labels@-1
+
+:" print the first window and label "
+.d("series ");.d(1#trainx);.d(" predicts ");.p(1#trainy)
+
+:" create the stateful model trainer "
+t::trainer(m)
+
+:" train the model using the Each-2 operator which calls t(x;y) for each window and label "
+loss::trainx t'trainy
+
+:" show that the training loss decreases "
+.d("training loss: ");.d(loss@0);.d(" --> ");.p(loss@-1)
+
+:" predict using the model "
+p::predictor(m)
+.d("prediction: ");.d((wmax@-1)*p(testx));.d(" should be close to ");.p((wmax@-1)*testy)

examples/stocks/yfinance/lstm.py

@@ -0,0 +1,62 @@
+import torch
+import torch.nn as nn
+import numpy as np
+
+class LSTMModel(nn.Module):
+    def __init__(self, input_size, hidden_layer_size, output_size):
+        super(LSTMModel, self).__init__()
+        self.lstm = nn.LSTM(input_size, hidden_layer_size)
+        self.linear = nn.Linear(hidden_layer_size, output_size)
+
+    def forward(self, input_seq):
+        # Reshape input_seq to (seq_len, batch, input_size)
+        input_seq = input_seq.view(len(input_seq), 1, -1)
+        lstm_out, _ = self.lstm(input_seq)
+        predictions = self.linear(lstm_out[-1].view(1, -1))
+        return predictions[:, 0]
+
+
+class LSTMWrapper:
+    """
+    Use a wrapper class to prevent KlongPy from converting the callable model to a Klong function.
+    """
+    def __init__(self, input_size, hidden_layer_size, output_size):
+        self.model = LSTMModel(input_size, hidden_layer_size, output_size)
+
+
+class Trainer:
+
+    def __init__(self, wrapper):
+        self.model = wrapper.model
+        self.loss_function = nn.MSELoss()
+        self.optimizer = torch.optim.Adam(self.model.parameters(), lr=0.001)
+
+    def __call__(self, x, y):
+        x = torch.tensor(x).float()
+        y = torch.tensor([y]).float()
+        self.optimizer.zero_grad()
+        y_pred = self.model(x)
+        single_loss = self.loss_function(y_pred, y)
+        single_loss.backward()
+        self.optimizer.step()
+        return single_loss.item()
+
+class Predictor:
+    def __init__(self, wrapper):
+        self.model = wrapper.model
+
+    def __call__(self, x):
+        x = torch.tensor(x).float()
+        with torch.no_grad():
+            return self.model(x).item()
+
+def load(filename):
+    model =torch.load(filename)
+    model.eval()
+    return model
+
+def save(filename, model):
+    torch.save(model, filename)
+
+
+klongpy_exports = {"load": load, "model": LSTMWrapper, "trainer": Trainer, "predictor": Predictor}