purged the visualizer

Dockerfile

@@ -9,9 +9,9 @@ RUN pip3 install --no-cache \
         meson==0.51.2 \
         numpy==1.16.4 \
         tensorflow==1.14.0 \
-        Flask==1.0.2 \
         cython==0.29.14
+
 RUN mkdir /workspace
-COPY bridge.pyx library.py server.py meson.build main.c /workspace/
+COPY bridge.pyx library.py meson.build main.c /workspace/
 RUN cd /workspace && meson build && cd build && ninja
 WORKDIR /workspace

bridge.pyx

@@ -7,7 +7,6 @@ from libc.stdlib cimport malloc, realloc
 from libc.string cimport memcpy
 
 import library as nn
-import server as srv
 
 
 tokenizers = {}
@@ -43,16 +42,6 @@ cdef public char *greeting():
     return f'The value is {3**3**3}'.encode('utf-8')
 
 
-cdef public void serve():
-    srv.serve()
-
-
-cdef public void server_update(float *emb):
-    embeddings = np.asarray(<float[:getvocsize(),:getemb()]>emb)
-    low_dim = nn.calc_TSNE(embeddings)
-    srv.emb_map = dict(zip(nn.inv_vocab, low_dim))
-
-
 cdef public size_t getwin():
     return nn.WIN
 

docs/report.latex

@@ -244,14 +244,14 @@ for the whole application. Therefore, the formula for the minimum number of
 processes to be requested from \verb|mpiexec| looks like the following:
 
 \begin{lstlisting}
-  NUM_PROC >= (4 * num_text_files) + 2
+  NUM_PROC >= (4 * num_text_files) + 1
 \end{lstlisting}
 
 To figure out how many Learners will be created, the following formula can be
 used:
 
 \begin{lstlisting}
-  num_learners = NUM_PROC - 2 - (3 * num_text_files)
+  num_learners = NUM_PROC - 1 - (3 * num_text_files)
 \end{lstlisting}
 
 During running, the program will create the folder \verb|trained| in the
@@ -367,7 +367,8 @@ statistics and exit.
 \section{Evaluation}
 
 The main focus of evaluation was to determine if executing several neural
-network training nodes in parallel can speed-up the training process. The
+network training nodes in parallel can speed-up the training process.
+The
 employed approach was to define a \textit{target loss} that the network has to
 achieve and then to measure \textit{the number of context windows} that each
 Learner node has to process and, secondarily, the time it takes for the system
@@ -457,6 +458,14 @@ computationally viable not to store the data as one big file but rather have it
 split across multiple nodes, this mode of operation should be investigated
 further and possibly preferred for large-scale training.
 
+As a last note, the learned embeddings themselves were not of high importance
+for the evaluation, since it is known that in order to obtain high quality
+embeddings a much higher amount of data (a dataset of \mbox{$>$ 100B words})
+and computation time is needed than it was feasible to do as a part of the
+project. However, the learning outcomes were empirically evaluated and it was
+found that even with relatively short training runs the networks could capture
+some meaningful relationships between the vocabulary words.
+
 \begin{figure}
   \centering
   \includegraphics[width=\linewidth]{fig/datasets.pdf}

main.c

@@ -37,7 +37,6 @@ typedef enum {
     FILTER,
     BATCHER,
     LEARNER,
-    VISUALIZER,
     DISPATCHER
 } Role;
 
@@ -66,10 +65,7 @@ size_t number_of(Role what) {
                 - number_of(TOKENIZER)
                 - number_of(FILTER)
                 - number_of(BATCHER)
-                - number_of(DISPATCHER)
+                - number_of(DISPATCHER);
-                - number_of(VISUALIZER);
-        case VISUALIZER:
-            return 0;
         case DISPATCHER:
             return 1;
     }
@@ -364,7 +360,6 @@ void learner() {
 void dispatcher() {
     INFO_PRINTF("Starting dispatcher %d\n", getpid());
     int go = 1;
-    // int visualizer = mpi_id_from_role_id(VISUALIZER, 0);
     size_t bs = getbs();
     size_t bpe = getbpe();
     float target = gettarget();
@@ -404,9 +399,6 @@ void dispatcher() {
         crt_loss = eval_net(frank);
         min_loss = crt_loss < min_loss ? crt_loss : min_loss;
         INFO_PRINTF("Round %ld, validation loss %f\n", rounds, crt_loss);
-        // MPI_Send(&go, 1, MPI_INT, visualizer, TAG_INSTR, MPI_COMM_WORLD);
-        // MPI_Send(wl.weights[0].W, emb_mat_size, MPI_FLOAT,
-                // visualizer, TAG_EMBED, MPI_COMM_WORLD);
 
         ckpt_net(frank);
 
@@ -423,9 +415,6 @@ void dispatcher() {
         MPI_Send(&go, 1, MPI_INT, mpi_id_from_role_id(LEARNER, l),
                 TAG_INSTR, MPI_COMM_WORLD);
     }
-    // MPI_Send(&go, 1, MPI_INT, mpi_id_from_role_id(VISUALIZER, 0),
-            // TAG_INSTR, MPI_COMM_WORLD);
-
     save_emb(frank);
 
     float delta_t = finish - start;
@@ -445,32 +434,6 @@ void dispatcher() {
     free(wls);
     free(round);
     INFO_PRINTF("Finishing dispatcher %d\n", getpid());
-    // sleep(4);
-    // INFO_PRINTLN("Visualization server is still running on port 8448\n"
-            // "To terminate, press Ctrl-C");
-}
-
-void visualizer() {
-    INFO_PRINTF("Starting visualizer %d\n", getpid());
-    serve();
-
-    int dispatcher = mpi_id_from_role_id(DISPATCHER, 0);
-    int go_on = 1;
-
-    size_t emb_mat_size = getvocsize() * getemb();
-    float* embeddings = malloc(emb_mat_size * sizeof(float));
-
-    MPI_Recv(&go_on, 1, MPI_INT, dispatcher, TAG_INSTR, MPI_COMM_WORLD,
-            MPI_STATUS_IGNORE);
-
-    while(go_on != -1) {
-        MPI_Recv(embeddings, emb_mat_size, MPI_FLOAT, dispatcher, TAG_EMBED,
-                MPI_COMM_WORLD, MPI_STATUS_IGNORE);
-        server_update(embeddings);
-        MPI_Recv(&go_on, 1, MPI_INT, dispatcher, TAG_INSTR, MPI_COMM_WORLD,
-                MPI_STATUS_IGNORE);
-    }
-    INFO_PRINTF("Exiting visualizer node %d\n", getpid());
 }
 
 int main (int argc, const char **argv) {
@@ -483,7 +446,7 @@ int main (int argc, const char **argv) {
             MPI_Abort(MPI_COMM_WORLD, 1);
         }
         int pipelines = argc - 1;
-        int min_nodes = 4 * pipelines + 2;
+        int min_nodes = 4 * pipelines + 1;
         if (world_size() < min_nodes) {
             INFO_PRINTF("You requested %d pipeline(s) "
                     "but only provided %d procs "
@@ -519,9 +482,6 @@ int main (int argc, const char **argv) {
         case DISPATCHER:
             dispatcher();
             break;
-        case VISUALIZER:
-            visualizer();
-            break;
         default:
             INFO_PRINTLN("DYING HORRIBLY!");
     }

server.py

@@ -1,34 +0,0 @@
-from threading import Thread
-
-import flask
-
-
-t = None
-app = flask.Flask(__name__)
-emb_map = None
-
-
-
-import logging
-log = logging.getLogger('werkzeug')
-log.setLevel(logging.ERROR)
-app.logger.setLevel(logging.ERROR)
-
-
-@app.route('/')
-def main():
-    if emb_map is None:
-        return 'Hello World!'
-    else:
-        return '\n'.join(f'{w}: {vec}' for w, vec in emb_map.items())
-
-
-def serve():
-    global t
-    if t is None:
-        t = Thread(target=app.run, kwargs={'port': 8448})
-    t.start()
-
-
-if __name__ == '__main__':
-    serve()