Some changes to scripts

Author: neubig

chainer/bilstm-tagger-withchar.py

@@ -161,7 +161,7 @@ def tag_sent(self, words):
 print("startup time: %r" % (time.time() - start))
 start = time.time()
 i = all_time = dev_time = all_tagged = this_tagged = this_loss = 0
-for ITER in xrange(10):
+for ITER in xrange(100):
   random.shuffle(train)
   for s in train:
     i += 1

chainer/bilstm-tagger.py

@@ -140,7 +140,7 @@ def tag_sent(self, words):
 print ("startup time: %r" % (time.time() - start))
 start = time.time()
 i = all_time = dev_time = all_tagged = this_tagged = this_loss = 0
-for ITER in xrange(50):
+for ITER in xrange(100):
   random.shuffle(train)
   for s in train:
     i += 1

chainer/rnnlm-batch.py

@@ -127,7 +127,7 @@ def calc_lm_loss(sents):
 print ("startup time: %r" % (time.time() - start))
 start = time.time()
 i = all_time = dev_time = all_tagged = this_words = this_loss = 0
-for ITER in xrange(10):
+for ITER in xrange(100):
   random.shuffle(train_order)
   for sid in train_order:
     i += 1

chainer/treenn.py

@@ -179,7 +179,7 @@ def classify(self, e):
 print("startup time: %r" % (time.time() - start))
 sents = 0
 all_time = 0
-for ITER in range(50):
+for ITER in range(100):
   random.shuffle(train)
   closs = 0.0
   cwords = 0

dynet-cpp/bilstm-tagger-withchar.cc

@@ -189,7 +189,7 @@ int main(int argc, char**argv) {
   start = system_clock::now();
   int i = 0, all_tagged = 0, this_words = 0;
   float this_loss = 0.f, all_time = 0.f;
-  for(int iter = 0; iter < 10; iter++) {
+  for(int iter = 0; iter < 100; iter++) {
     shuffle(train.begin(), train.end(), *dynet::rndeng);
     for(auto & s : train) {
       i++;

dynet-cpp/bilstm-tagger.cc

@@ -167,7 +167,7 @@ int main(int argc, char**argv) {
   start = system_clock::now();
   int i = 0, all_tagged = 0, this_words = 0;
   float this_loss = 0.f, all_time = 0.f;
-  for(int iter = 0; iter < 10; iter++) {
+  for(int iter = 0; iter < 100; iter++) {
     shuffle(train.begin(), train.end(), *dynet::rndeng);
     for(auto & s : train) {
       i++;

dynet-cpp/rnnlm-batch.cc

@@ -154,7 +154,7 @@ int main(int argc, char** argv) {
   start = system_clock::now();
   int i = 0, all_words = 0, this_words = 0;
   float this_loss = 0.f, all_time = 0.f;
-  for(int iter = 0; iter < 10; iter++) {
+  for(int iter = 0; iter < 100; iter++) {
     shuffle(train_ids.begin(), train_ids.end(), *dynet::rndeng);
     for(auto sid : train_ids) {
       i++;

dynet-cpp/rnnlm-seq.cc

@@ -139,7 +139,7 @@ int main(int argc, char** argv) {
   start = system_clock::now();
   int i = 0, all_words = 0, this_words = 0;
   float this_loss = 0.f, all_time = 0.f;
-  for(int iter = 0; iter < 10; iter++) {
+  for(int iter = 0; iter < 100; iter++) {
     shuffle(train_ids.begin(), train_ids.end(), *dynet::rndeng);
     for(auto sid : train_ids) {
       i++;

dynet-cpp/treenn.cc

@@ -207,7 +207,7 @@ int main(int argc, char**argv) {
 
   int i = 0, all_tagged = 0, this_nodes = 0;
   float this_loss = 0.f, all_time = 0.f;
-  for(int iter = 0; iter < 50; iter++) {
+  for(int iter = 0; iter < 100; iter++) {
     shuffle(train.begin(), train.end(), *dynet::rndeng);
     start = system_clock::now();
     for(auto tree : train) {

dynet-py/bilstm-tagger-withchar.py

@@ -168,7 +168,7 @@ def tag_sent(words):
 print ("startup time: %r" % (time.time() - start))
 start = time.time()
 i = all_time = dev_time = all_tagged = this_tagged = this_loss = 0
-for ITER in range(10):
+for ITER in range(100):
   random.shuffle(train)
   for s in train:
     i += 1

dynet-py/bilstm-tagger.py

@@ -146,7 +146,7 @@ def tag_sent(words):
 print ("startup time: %r" % (time.time() - start))
 start = time.time()
 i = all_time = dev_time = all_tagged = this_tagged = this_loss = 0
-for ITER in range(50):
+for ITER in range(100):
   random.shuffle(train)
   for s in train:
     i += 1

dynet-py/rnnlm-batch-batch.py

@@ -131,7 +131,7 @@ def calc_lm_loss(sents):
 # Perform training
 start = time.time()
 i = all_time = dev_time = all_tagged = this_words = this_loss = 0
-for ITER in range(10):
+for ITER in range(100):
   random.shuffle(train_order)
   for sid in train_order: 
     i += 1

dynet-py/rnnlm-batch.py

@@ -116,7 +116,7 @@ def calc_lm_loss(sents):
 # Perform training
 start = time.time()
 i = all_time = dev_time = all_tagged = this_words = this_loss = 0
-for ITER in range(10):
+for ITER in range(100):
   random.shuffle(train_order)
   for sid in train_order: 
     i += 1

dynet-py/treenn.py

@@ -142,7 +142,7 @@ def expr_for_tree(self, tree, decorate=False):
 print ("startup time: %r" % (time.time() - start))
 sents = 0
 all_time = 0
-for ITER in range(50):
+for ITER in range(100):
     random.shuffle(train)
     closs = 0.0
     cwords = 0

run-tests.sh

@@ -22,7 +22,7 @@ else
 fi
 
 TIMEOUT=600
-LONGTIMEOUT=600
+LONGTIMEOUT=3600
 
 runcmd() {
   LFILE=log/$2$GPUSUF/$4.log
@@ -32,6 +32,8 @@ runcmd() {
       mycmd="$1/$2$GPUSUF $DYFLAGS"
       if [[ $4 =~ dynet-cpp-bs01-ws128-hs256-.* ]] || [[ $4 =~ dynet-cpp-bs16-ws128-hs256-.* ]] || [[ $2 =~ bilstm.* ]] || [[ $2 =~ treenn ]]; then 
         MYTIMEOUT=$LONGTIMEOUT
+      else
+        return
       fi
     elif [[ $1 == "dynet-py" ]]; then
       mycmd="$PYTHON -u $1/$2.py $DYFLAGS"
@@ -48,11 +50,38 @@ runcmd() {
 
 for trial in 1 2 3; do
 
+  # Run rnnlm-batch
+  for embsize in 128; do
+    hidsize=$(($embsize*2))
+    # for mbsize in 64 32 16 08 04 02 01; do
+    for mbsize in 64 16 04 01; do
+      # for f in dynet-cpp dynet-py chainer theano tensorflow; do
+      for f in dynet-cpp dynet-py; do
+        runcmd $f rnnlm-batch "$mbsize $embsize $hidsize 0" $f-ms$mbsize-es$embsize-hs$hidsize-sp0-t$trial
+        if [[ $f == "dynet-cpp" ]]; then
+          runcmd $f rnnlm-seq "$mbsize $embsize $hidsize 0" $f-ms$mbsize-es$embsize-hs$hidsize-sp0-t$trial
+        fi
+      done
+    done
+  done
+
+  # run sparse rnnlm-seq on a subset
+  for embsize in 128; do
+    hidsize=$(($embsize*2))
+    for mbsize in 16 01; do
+      # for f in dynet-cpp dynet-py; do
+      for f in dynet-cpp; do
+        runcmd $f rnnlm-seq "$mbsize $embsize $hidsize 1" $f-ms$mbsize-es$embsize-hs$hidsize-sp0-t$trial
+      done
+    done
+  done
+
   # Run bilstm-tagger
   wembsize=128
   hidsize=50
   mlpsize=32
-  for f in dynet-cpp dynet-py theano chainer tensorflow; do
+  # for f in dynet-cpp dynet-py chainer theano tensorflow; do
+  for f in dynet-cpp dynet-py; do
     runcmd $f bilstm-tagger "$wembsize $hidsize $mlpsize 0" $f-ws$wembsize-hs$hidsize-mlps$mlpsize-su0-t$trial
     if [[ $f == dynet* ]]; then
       runcmd $f bilstm-tagger "$wembsize $hidsize $mlpsize 1" $f-ws$wembsize-hs$hidsize-mlps$mlpsize-su1-t$trial
@@ -64,7 +93,8 @@ for trial in 1 2 3; do
   wembsize=128
   hidsize=50
   mlpsize=32
-  for f in dynet-cpp dynet-py theano chainer; do
+  # for f in dynet-cpp dynet-py theano chainer; do
+  for f in dynet-cpp dynet-py; do
     runcmd $f bilstm-tagger-withchar "$cembsize $wembsize $hidsize $mlpsize 0" $f-cs$cembsize-ws$wembsize-hs$hidsize-mlps$mlpsize-su0-t$trial
     if [[ $f == dynet* ]]; then
       runcmd $f bilstm-tagger-withchar "$cembsize $wembsize $hidsize $mlpsize 1" $f-cs$cembsize-ws$wembsize-hs$hidsize-mlps$mlpsize-su1-t$trial
@@ -74,32 +104,12 @@ for trial in 1 2 3; do
   # Run treenn
   wembsize=128
   hidsize=128
-  for f in dynet-cpp dynet-py chainer; do
+  # for f in dynet-cpp dynet-py chainer; do
+  for f in dynet-cpp dynet-py; do
     runcmd $f treenn "$wembsize $hidsize 0" $f-ws$wembsize-hs$hidsize-su0-t$trial
     if [[ $f == dynet* ]]; then
       runcmd $f treenn "$wembsize $hidsize 1" $f-ws$wembsize-hs$hidsize-su1-t$trial
     fi
   done
 
-  # Run rnnlm-batch
-  for embsize in 64 128 256; do
-    hidsize=$(($embsize*2))
-    # for mbsize in 64 32 16 08 04 02 01; do
-    for mbsize in 64 16 04 01; do
-      for f in dynet-cpp dynet-py theano chainer tensorflow; do
-        runcmd $f rnnlm-batch "$mbsize $embsize $hidsize 0" $f-ms$mbsize-es$embsize-hs$hidsize-sp0-t$trial
-      done
-    done
-  done
-
-  # run sparse rnnlm-batch on a subset
-  for embsize in 128; do
-    hidsize=$(($embsize*2))
-    for mbsize in 16 01; do
-      for f in dynet-cpp dynet-py; do
-        runcmd $f rnnlm-batch "$mbsize $embsize $hidsize 1" $f-ms$mbsize-es$embsize-hs$hidsize-sp0-t$trial
-      done
-    done
-  done
-
 done

tensorflow/bilstm-tagger.py

@@ -137,7 +137,7 @@ def get_tags(log_probs):
 start = time.time()
 i = all_time = dev_time = all_tagged = this_tagged = this_loss = 0
 
-for ITER in range(50):
+for ITER in range(100):
   random.shuffle(train)
   for s in train:
     i += 1

theano/bilstm-tagger-withchar.py

@@ -218,7 +218,7 @@ def tag_sent(sent, decode_func):
 start = time.time()
 i = all_time = dev_time = all_tagged = this_tagged = this_loss = 0
 
-for ITER in range(50):
+for ITER in range(100):
   random.shuffle(train)
   for s in train:
     i += 1

theano/bilstm-tagger.py

@@ -195,7 +195,7 @@ def tag_sent(batch_sents, decode_func):
 start = time.time()
 i = all_time = dev_time = all_tagged = this_tagged = this_loss = 0
 
-for ITER in range(50):
+for ITER in range(100):
   random.shuffle(train)
   for batch_id, (batch_start, batch_end) in enumerate(batches):
     i += MB_SIZE

theano/rnnlm-batch.py

@@ -124,7 +124,7 @@ def build_graph():
 print("startup time: %r" % (time.time() - start))
 start = time.time()
 i = all_time = dev_time = all_tagged = this_words = this_loss = 0
-for ITER in range(10):
+for ITER in range(100):
   random.shuffle(train_order)
   for sid in train_order:
     i += 1