Add RemoteMonitor callback

keras/callbacks.py

@@ -4,7 +4,7 @@ import theano
 import theano.tensor as T
 import numpy as np
 import warnings
-import time
+import time, json
 from collections import deque
 
 from .utils.generic_utils import Progbar
@@ -216,7 +216,7 @@ class ModelCheckpoint(Callback):
 
 
 class EarlyStopping(Callback):
-    def __init__(self, patience=1, verbose=0):
+    def __init__(self, patience=0, verbose=0):
         super(Callback, self).__init__()
 
         self.patience = patience
@@ -238,3 +238,29 @@ class EarlyStopping(Callback):
                     print("Epoch %05d: early stopping" % (epoch))
                 self.model.stop_training = True
             self.wait += 1
+
+
+class RemoteMonitor(Callback):
+    def __init__(self, root='http://localhost:9000'):
+        self.root = root
+        self.seen = 0
+        self.tot_loss = 0.
+        self.tot_accuracy = 0.
+
+    def on_epoch_begin(self, epoch, logs={}):
+        self.seen = 0
+        self.tot_loss = 0.
+        self.tot_accuracy = 0.
+
+    def on_batch_end(self, batch, logs={}):
+        batch_size = logs.get('size', 0)
+        self.seen += batch_size
+        self.tot_loss += logs.get('loss', 0.) * batch_size
+        if self.params['show_accuracy']:
+            self.tot_accuracy += logs.get('accuracy', 0.) * batch_size
+
+    def on_epoch_end(self, epoch, logs={}):
+        import requests
+        logs['epoch'] = epoch
+        logs['loss'] = self.tot_loss / self.seen
+        r = requests.post(self.root + '/publish/epoch/end/', {'data':json.dumps(logs)})