Some conv-lstm, sgdr and other fixes

README.md

@@ -518,7 +518,7 @@ Example of custom object detection: `darknet.exe detector test data/obj.data yol
   * increase network resolution in your `.cfg`-file (`height=608`, `width=608` or any value multiple of 32) - it will increase precision
 
 
-  * check that each object is mandatory labeled in your dataset - no one object in your data set should not be without label. In the most training issues - there are wrong labels in your dataset (got labels by using some conversion script, marked with a third-party tool, ...). Always check your dataset by using: https://github.com/AlexeyAB/Yolo_mark
+  * check that each object that you want to detect is mandatory labeled in your dataset - no one object in your data set should not be without label. In the most training issues - there are wrong labels in your dataset (got labels by using some conversion script, marked with a third-party tool, ...). Always check your dataset by using: https://github.com/AlexeyAB/Yolo_mark
 
   * for each object which you want to detect - there must be at least 1 similar object in the Training dataset with about the same: shape, side of object, relative size, angle of rotation, tilt, illumination. So desirable that your training dataset include images with objects at diffrent: scales, rotations, lightings, from different sides, on different backgrounds - you should preferably have 2000 different images for each class or more, and you should train `2000*classes` iterations or more
 

include/darknet.h

@@ -211,6 +211,7 @@ struct layer {
     int peephole;
     int use_bin_output;
     int steps;
+    int state_constrain;
     int hidden;
     int truth;
     float smooth;
@@ -551,6 +552,7 @@ typedef struct network {
     float learning_rate_min;
     float learning_rate_max;
     int batches_per_cycle;
+    int batches_cycle_mult;
     float momentum;
     float decay;
     float gamma;

src/activation_kernels.cu

@@ -210,6 +210,30 @@ __global__ void activate_array_logistic_kernel(float *x, int n)
     }
 }
 
+__global__ void activate_array_tanh_kernel(float *x, int n)
+{
+    int index = blockIdx.x*blockDim.x + threadIdx.x;
+    if (index < n) {
+        x[index] = tanh_activate_kernel(x[index]);
+    }
+}
+
+__global__ void activate_array_hardtan_kernel(float *x, int n)
+{
+    int index = blockIdx.x*blockDim.x + threadIdx.x;
+    if (index < n) {
+        x[index] = hardtan_activate_kernel(x[index]);
+    }
+}
+
+__global__ void activate_array_relu_kernel(float *x, int n)
+{
+    int index = blockIdx.x*blockDim.x + threadIdx.x;
+    if (index < n) {
+        x[index] = relu_activate_kernel(x[index]);
+    }
+}
+
 __global__ void gradient_array_kernel(float *x, int n, ACTIVATION a, float *delta)
 {
     int i = (blockIdx.x + blockIdx.y*gridDim.x) * blockDim.x + threadIdx.x;
@@ -240,6 +264,14 @@ __global__ void gradient_array_logistic_kernel(float *x, int n, float *delta)
     }
 }
 
+__global__ void gradient_array_tanh_kernel(float *x, int n, float *delta)
+{
+    int index = blockIdx.x*blockDim.x + threadIdx.x;
+    if (index < n) {
+        delta[index] *= tanh_gradient_kernel(x[index]);
+    }
+}
+
 __global__ void gradient_array_hardtan_kernel(float *x, int n, float *delta)
 {
     int index = blockIdx.x*blockDim.x + threadIdx.x;
@@ -248,12 +280,23 @@ __global__ void gradient_array_hardtan_kernel(float *x, int n, float *delta)
     }
 }
 
+__global__ void gradient_array_relu_kernel(float *x, int n, float *delta)
+{
+    int index = blockIdx.x*blockDim.x + threadIdx.x;
+    if (index < n) {
+        delta[index] *= relu_gradient_kernel(x[index]);
+    }
+}
+
 extern "C" void activate_array_ongpu(float *x, int n, ACTIVATION a)
 {
     const int num_blocks = get_number_of_blocks(n, BLOCK);
     if (a == LINEAR) return;
     else if(a == LEAKY) activate_array_leaky_kernel << <num_blocks, BLOCK, 0, get_cuda_stream() >> >(x, n);
     else if (a == LOGISTIC) activate_array_logistic_kernel << <num_blocks, BLOCK, 0, get_cuda_stream() >> >(x, n);
+    else if (a == TANH) activate_array_tanh_kernel << <num_blocks, BLOCK, 0, get_cuda_stream() >> >(x, n);
+    else if (a == HARDTAN) activate_array_hardtan_kernel << <num_blocks, BLOCK, 0, get_cuda_stream() >> >(x, n);
+    else if (a == RELU) activate_array_relu_kernel << <num_blocks, BLOCK, 0, get_cuda_stream() >> >(x, n);
     else if (a == SELU) activate_array_selu_kernel << <num_blocks, BLOCK, 0, get_cuda_stream() >> >(x, n);
     else
         activate_array_kernel<<<cuda_gridsize(n), BLOCK, 0, get_cuda_stream()>>>(x, n, a);
@@ -266,8 +309,10 @@ extern "C" void gradient_array_ongpu(float *x, int n, ACTIVATION a, float *delta
     if (a == LINEAR) return;
     else if (a == LEAKY) gradient_array_leaky_kernel << <num_blocks, BLOCK, 0, get_cuda_stream() >> >(x, n, delta);
     else if (a == LOGISTIC) gradient_array_logistic_kernel << <num_blocks, BLOCK, 0, get_cuda_stream() >> >(x, n, delta);
-    else if (a == SELU) gradient_array_selu_kernel << <num_blocks, BLOCK, 0, get_cuda_stream() >> >(x, n, delta);
+    else if (a == TANH) gradient_array_tanh_kernel << <num_blocks, BLOCK, 0, get_cuda_stream() >> >(x, n, delta);
     else if (a == HARDTAN) gradient_array_hardtan_kernel << <num_blocks, BLOCK, 0, get_cuda_stream() >> >(x, n, delta);
+    else if (a == RELU) gradient_array_relu_kernel << <num_blocks, BLOCK, 0, get_cuda_stream() >> >(x, n, delta);
+    else if (a == SELU) gradient_array_selu_kernel << <num_blocks, BLOCK, 0, get_cuda_stream() >> >(x, n, delta);
     else
         gradient_array_kernel << <cuda_gridsize(n), BLOCK, 0, get_cuda_stream() >> > (x, n, a, delta);
     CHECK_CUDA(cudaPeekAtLastError());

src/blas_kernels.cu

@@ -983,7 +983,7 @@ __global__ void fix_nan_and_inf_kernel(float *input, size_t size)
     if (index < size) {
         float val = input[index];
         if (isnan(val) || isinf(val))
-            input[index] = index;  // pseudo random value
+            input[index] = 1.0f / index;  // pseudo random value
     }
 }
 

src/conv_lstm_layer.c

@@ -902,13 +902,14 @@ void forward_conv_lstm_layer_gpu(layer l, network_state state)
         activate_array_ongpu(l.h_gpu, l.outputs*l.batch, TANH);
         mul_ongpu(l.outputs*l.batch, l.o_gpu, 1, l.h_gpu, 1);
 
+        if(l.state_constrain) constrain_ongpu(l.outputs*l.batch, l.state_constrain, l.c_gpu, 1);
         //constrain_ongpu(l.outputs*l.batch, 1, l.c_gpu, 1);
         //constrain_ongpu(l.outputs*l.batch, 1, l.h_gpu, 1);
         fix_nan_and_inf(l.c_gpu, l.outputs*l.batch);
         fix_nan_and_inf(l.h_gpu, l.outputs*l.batch);
 
         copy_ongpu(l.outputs*l.batch, l.c_gpu, 1, l.cell_gpu, 1);
-        copy_ongpu(l.outputs*l.batch, l.h_gpu, 1, l.output_gpu, 1);
+        copy_ongpu(l.outputs*l.batch, l.h_gpu, 1, l.output_gpu, 1); // required for both Detection and Training
 
         state.input += l.inputs*l.batch;
         l.output_gpu    += l.outputs*l.batch;
@@ -1159,5 +1160,8 @@ void backward_conv_lstm_layer_gpu(layer l, network_state state)
 
     copy_ongpu(l.outputs*l.batch, last_output, 1, l.last_prev_state_gpu, 1);
     copy_ongpu(l.outputs*l.batch, last_cell, 1, l.last_prev_cell_gpu, 1);
+
+    // free state after each 100 iterations
+    //if (get_current_batch(state.net) % 100) free_state_conv_lstm(l);  // dont use
 }
 #endif

src/convolutional_layer.c

@@ -1046,13 +1046,9 @@ void backward_convolutional_layer(convolutional_layer l, network_state state)
     }
 }
 
-void update_convolutional_layer(convolutional_layer l, update_args a)
+void update_convolutional_layer(convolutional_layer l, int batch, float learning_rate, float momentum, float decay)
 {
-    float learning_rate = a.learning_rate*l.learning_rate_scale;
-    float momentum = a.momentum;
-    float decay = a.decay;
-    int batch = a.batch;
-
+    //int size = l.size*l.size*l.c*l.n;
     axpy_cpu(l.n, learning_rate / batch, l.bias_updates, 1, l.biases, 1);
     scal_cpu(l.n, momentum, l.bias_updates, 1);
 
@@ -1067,6 +1063,7 @@ void update_convolutional_layer(convolutional_layer l, update_args a)
 }
 
 
+
 image get_convolutional_weight(convolutional_layer l, int i)
 {
     int h = l.size;
@@ -1101,7 +1098,7 @@ void rescale_weights(convolutional_layer l, float scale, float trans)
 
 image *get_weights(convolutional_layer l)
 {
-    image *weights = calloc(l.n, sizeof(image));
+    image *weights = (image *)calloc(l.n, sizeof(image));
     int i;
     for (i = 0; i < l.n; ++i) {
         weights[i] = copy_image(get_convolutional_weight(l, i));

src/detector.c

@@ -666,7 +666,7 @@ float validate_detector_map(char *datacfg, char *cfgfile, char *weightfile, floa
         char *train_images = option_find_str(options, "train", "data/train.txt");
         valid_images = option_find_str(options, "valid", train_images);
         net = *existing_net;
-        //remember_network_recurrent_state(*existing_net);
+        remember_network_recurrent_state(*existing_net);
         free_network_recurrent_state(*existing_net);
     }
     else {
@@ -1077,8 +1077,8 @@ float validate_detector_map(char *datacfg, char *cfgfile, char *weightfile, floa
     if (existing_net) {
         //set_batch_network(&net, initial_batch);
         //free_network_recurrent_state(*existing_net);
-        //restore_network_recurrent_state(*existing_net);
-        randomize_network_recurrent_state(*existing_net);
+        restore_network_recurrent_state(*existing_net);
+        //randomize_network_recurrent_state(*existing_net);
     }
     else {
         free_network(net);

src/image_opencv.cpp

@@ -1081,7 +1081,7 @@ void draw_train_loss(mat_cv* img_src, int img_size, float avg_loss, float max_im
             cv::putText(img, char_buff, cv::Point(10, 28), cv::FONT_HERSHEY_COMPLEX_SMALL, 0.7, CV_RGB(255, 255, 255), 5, CV_AA);
             cv::putText(img, char_buff, cv::Point(10, 28), cv::FONT_HERSHEY_COMPLEX_SMALL, 0.7, CV_RGB(200, 0, 0), 1, CV_AA);
 
-            if (((int)(old_precision * 10) != (int)(precision * 10)) || (max_precision < precision) || (current_batch - text_iteration_old) >= max_batches / 10) {
+            if ((std::fabs(old_precision - precision) > 0.1)  || (max_precision < precision) || (current_batch - text_iteration_old) >= max_batches / 10) {
                 text_iteration_old = current_batch;
                 max_precision = std::max(max_precision, precision);
                 sprintf(char_buff, "%2.0f%% ", precision * 100);

src/network.c

@@ -95,7 +95,7 @@ float get_current_seq_subdivisions(network net)
 {
     int sequence_subdivisions = net.init_sequential_subdivisions;
 
-    if (net.policy)
+    if (net.num_steps > 0)
     {
         int batch_num = get_current_batch(net);
         int i;
@@ -145,11 +145,20 @@ float get_current_rate(network net)
         case SIG:
             return net.learning_rate * (1./(1.+exp(net.gamma*(batch_num - net.step))));
         case SGDR:
+        {
+            int last_iteration_start = 0;
+            int cycle_size = net.batches_per_cycle;
+            while ((last_iteration_start + cycle_size) < batch_num)
+            {
+                last_iteration_start += cycle_size;
+                cycle_size *= net.batches_cycle_mult;
+            }
             rate = net.learning_rate_min +
-                        0.5*(net.learning_rate-net.learning_rate_min)
-                        * (1. + cos( (float) (batch_num % net.batches_per_cycle)*3.14159265 / net.batches_per_cycle));
+                0.5*(net.learning_rate - net.learning_rate_min)
+                * (1. + cos((float)(batch_num - last_iteration_start)*3.14159265 / cycle_size));
 
             return rate;
+        }
         default:
             fprintf(stderr, "Policy is weird!\n");
             return net.learning_rate;

src/parser.c

@@ -200,7 +200,7 @@ layer parse_crnn(list *options, size_params params)
     int batch_normalize = option_find_int_quiet(options, "batch_normalize", 0);
     int xnor = option_find_int_quiet(options, "xnor", 0);
 
-    layer l = make_crnn_layer(params.batch, params.w, params.h, params.c, hidden_filters, output_filters, groups, params.time_steps, size, stride, padding, activation, batch_normalize, xnor);
+    layer l = make_crnn_layer(params.batch, params.h, params.w, params.c, hidden_filters, output_filters, groups, params.time_steps, size, stride, padding, activation, batch_normalize, xnor);
 
     l.shortcut = option_find_int_quiet(options, "shortcut", 0);
 
@@ -260,8 +260,9 @@ layer parse_conv_lstm(list *options, size_params params)
     int xnor = option_find_int_quiet(options, "xnor", 0);
     int peephole = option_find_int_quiet(options, "peephole", 1);
 
-    layer l = make_conv_lstm_layer(params.batch, params.w, params.h, params.c, output_filters, groups, params.time_steps, size, stride, padding, activation, batch_normalize, peephole, xnor);
+    layer l = make_conv_lstm_layer(params.batch, params.h, params.w, params.c, output_filters, groups, params.time_steps, size, stride, padding, activation, batch_normalize, peephole, xnor);
 
+    l.state_constrain = option_find_int_quiet(options, "state_constrain", params.time_steps * 32);
     l.shortcut = option_find_int_quiet(options, "shortcut", 0);
 
     return l;
@@ -668,7 +669,8 @@ void parse_net_options(list *options, network *net)
     net->batch = option_find_int(options, "batch",1);
     net->learning_rate = option_find_float(options, "learning_rate", .001);
     net->learning_rate_min = option_find_float_quiet(options, "learning_rate_min", .00001);
-    net->batches_per_cycle = option_find_int_quiet(options, "sgdr_cycle", 500);
+    net->batches_per_cycle = option_find_int_quiet(options, "sgdr_cycle", 1000);
+    net->batches_cycle_mult = option_find_int_quiet(options, "sgdr_mult", 2);
     net->momentum = option_find_float(options, "momentum", .9);
     net->decay = option_find_float(options, "decay", .0001);
     int subdivs = option_find_int(options, "subdivisions",1);
@@ -721,40 +723,44 @@ void parse_net_options(list *options, network *net)
     if(net->policy == STEP){
         net->step = option_find_int(options, "step", 1);
         net->scale = option_find_float(options, "scale", 1);
-    } else if (net->policy == STEPS){
+    } else if (net->policy == STEPS || net->policy == SGDR){
         char *l = option_find(options, "steps");
         char *p = option_find(options, "scales");
         char *s = option_find(options, "seq_scales");
-        if(!l || !p) error("STEPS policy must have steps and scales in cfg file");
+        if(net->policy == STEPS && (!l || !p)) error("STEPS policy must have steps and scales in cfg file");
 
+        if (l) {
             int len = strlen(l);
             int n = 1;
             int i;
-        for(i = 0; i < len; ++i){
+            for (i = 0; i < len; ++i) {
                 if (l[i] == ',') ++n;
             }
             int* steps = (int*)calloc(n, sizeof(int));
             float* scales = (float*)calloc(n, sizeof(float));
             float* seq_scales = (float*)calloc(n, sizeof(float));
             for (i = 0; i < n; ++i) {
+                float scale = 1.0;
+                if (p) {
+                    scale = atof(p);
+                    p = strchr(p, ',') + 1;
+                }
+                float sequence_scale = 1.0;
                 if (s) {
-                seq_scales[i] = atof(s);
+                    sequence_scale = atof(s);
                     s = strchr(s, ',') + 1;
-            } else
-                seq_scales[i] = 1;
                 }
-        for(i = 0; i < n; ++i){
                 int step = atoi(l);
-            float scale = atof(p);
-            l = strchr(l, ',')+1;
-            p = strchr(p, ',')+1;
+                l = strchr(l, ',') + 1;
                 steps[i] = step;
                 scales[i] = scale;
+                seq_scales[i] = sequence_scale;
             }
             net->scales = scales;
             net->steps = steps;
             net->seq_scales = seq_scales;
             net->num_steps = n;
+        }
     } else if (net->policy == EXP){
         net->gamma = option_find_float(options, "gamma", 1);
     } else if (net->policy == SIG){