upstream-archive/stable-diffusion.cpp
1
0
Fork 0
mirror of https://github.com/leejet/stable-diffusion.cpp.git synced 2026-07-29 14:20:57 -05:00
Code Issues Packages Projects Releases Wiki Activity

small tweak; still not getting ID faces

Browse Source
This commit is contained in:
bssrdf
2024-02-14 21:22:40 -05:00
parent 539a94adc7
commit 26c5cca043
3 changed files with 158 additions and 5 deletions
Download Patch File Download Diff File Expand all files Collapse all files

5
clip.hpp
View File

@@ -960,7 +960,8 @@ struct CLIPVisionModel {
int32_t n_layer = num_hidden_layers;
const int d_head = hidden_size / n_head;
ggml_set_name(x, "id_pixel");
// ggml_set_name(x, "id_pixel");
print_ggml_tensor(x, true, "id_pixel");
ggml_set_name(temp, "temp_input");
int64_t* ne = patch_embeddings->ne;
// struct ggml_tensor *patch_embeddings_f16 = ggml_reshape_3d(ctx0, patch_embeddings, ne[0], ne[1], ne[2]*ne[3]);
@@ -980,7 +981,7 @@ struct CLIPVisionModel {
inp = ggml_cont(ctx0, ggml_permute(ctx0, inp, 2, 0, 1, 3));
print_ggml_tensor(inp, true, "inp_cont");
ggml_set_name(inp, "inp_cont");
ggml_set_name(class_embedding, "class_embedding");
// ggml_set_name(class_embedding, "class_embedding");
// concat class_embeddings and patch_embeddings
// struct ggml_tensor * embeddings = ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, hidden_size, num_positions, batch_size);

132
ggml_extend.hpp
View File

@@ -77,6 +77,25 @@ __STATIC_INLINE__ ggml_fp16_t ggml_tensor_get_f16(const ggml_tensor* tensor, int
return *(ggml_fp16_t*)((char*)(tensor->data) + i * tensor->nb[3] + j * tensor->nb[2] + k * tensor->nb[1] + l * tensor->nb[0]);
}
static struct ggml_tensor * get_tensor_from_graph(struct ggml_cgraph * gf, const char *name){
struct ggml_tensor * res = NULL;
for(int i = 0; i < gf->n_nodes; i++) {
// printf("%d, %s \n", i, gf->nodes[i]->name);
if(strcmp(ggml_get_name(gf->nodes[i]), name) == 0) {
res = gf->nodes[i];
break;
}
}
for(int i = 0; i < gf->n_leafs; i++) {
// printf("%d, %s \n", i, gf->leafs[i]->name);
if(strcmp(ggml_get_name(gf->leafs[i]), name) == 0) {
res = gf->leafs[i];
break;
}
}
return res;
}
@@ -669,6 +688,87 @@ struct GGMLModule {
ggml_backend_metal_set_n_cb(backend, n_threads);
}
#endif
#if 0
struct ggml_tensor * imb = NULL;
for (int i = 0; i < gf->n_leafs; i++) {
struct ggml_tensor * t1 = gf->leafs[i];
if(strcmp(ggml_get_name(t1), "id_pixel_values_input") == 0) {
imb = t1;
int64_t stride = imb->ne[0];
float* out_data = new float[ggml_nelements(imb)];
ggml_backend_tensor_get(imb, out_data, 0, ggml_nbytes(imb));
for(int k = 0; k < 3; ++k){
float mi = 100.f, mx= -100.f;
for(int i = 0; i < stride; i++){
printf("[");
for(int j = 0; j < stride; j++){
float val = out_data[k*stride*stride+i*stride+j];
if(mi > val) mi = val;
if(mx < val) mx = val;
printf("%f, ", val);
}
printf("]\n");
}
printf("B. channel, min, max: %d, %f %f \n", k, mi, mx);
}
// printf("[");
// for(int i = 0; i < stride; i++){
// printf("%f, ", out_data[i]);
// }
// printf("]\n");
delete out_data;
}
}
for (int i = 0; i < gf->n_nodes; i++) {
struct ggml_tensor * t1 = gf->nodes[i];
// if(strcmp(ggml_get_name(t1), "embeddings_after_add") == 0) {
if(strcmp(ggml_get_name(t1), "pooled_output") == 0) {
imb = t1;
int64_t stride = imb->ne[0];
float* out_data = new float[ggml_nelements(imb)];
ggml_backend_tensor_get(imb, out_data, 0, ggml_nbytes(imb));
// for(int k = 0; k < 3; ++k){
// float mi = 100.f, mx= -100.f;
// for(int i = 0; i < stride; i++){
// printf("[");
// for(int j = 0; j < stride; j++){
// float val = out_data[k*stride*stride+i*stride+j];
// // if(mi > val) mi = val;
// // if(mx < val) mx = val;
// printf("%f, ", val);
// }
// printf("]\n");
// }
// // printf("B. channel, min, max: %d, %f %f \n", k, mi, mx);
// }
// for(int i = 0; i < 257; i++){
// printf("[");
// for(int j = 0; j < stride; j++){
// float val = out_data[i*stride+j];
// // if(mi > val) mi = val;
// // if(mx < val) mx = val;
// printf("%f, ", val);
// }
// printf("]\n");
// }
// printf("B. channel, min, max: %d, %f %f \n", k, mi, mx);
printf("[");
for(int i = 0; i < stride; i++){
printf("%f, ", out_data[i]);
}
printf("]\n");
delete out_data;
}
struct ggml_cgraph g1v = ggml_graph_view(gf, i, i + 1);
ggml_backend_graph_compute(backend, &g1v);
}
#endif
ggml_backend_graph_compute(backend, gf);
#ifdef GGML_PERF
@@ -679,6 +779,38 @@ struct GGMLModule {
if (output != NULL) {
ggml_backend_tensor_get_and_sync(backend, gf->nodes[gf->n_nodes - 1], output->data, 0, ggml_nbytes(output));
}
// struct ggml_tensor * imb = get_tensor_from_graph(gf, "id_pixel");
// struct ggml_tensor * imb = get_tensor_from_graph(gf, "id_pixel_values_input");
// if(imb != NULL){
// int64_t stride = imb->ne[0];
// float* out_data = new float[ggml_nelements(imb)];
// ggml_backend_tensor_get(imb, out_data, 0, ggml_nbytes(imb));
// for(int k = 0; k < 3; ++k){
// float mi = 100.f, mx= -100.f;
// for(int i = 0; i < stride; i++){
// printf("[");
// for(int j = 0; j < stride; j++){
// float val = out_data[k*stride*stride+i*stride+j];
// if(mi > val) mi = val;
// if(mx < val) mx = val;
// printf("%f, ", val);
// }
// printf("]\n");
// }
// printf("B. channel, min, max: %d, %f %f \n", k, mi, mx);
// }
// // printf("[");
// // for(int i = 0; i < stride; i++){
// // printf("%f, ", out_data[i]);
// // }
// // printf("]\n");
// delete out_data;
// }
// if (output != NULL)
// print_ggml_tensor(output, true, "output_get");
}

26
stable-diffusion.cpp
View File

@@ -526,6 +526,10 @@ public:
std::vector<int>& tokens = std::get<0>(tokens_and_weights);
std::vector<float>& weights = std::get<1>(tokens_and_weights);
std::vector<bool>& clsm = std::get<2>(tokens_and_weights);
printf("tokens: \n");
for(int i = 0; i < tokens.size(); ++i)
printf("%d ", tokens[i]);
printf("\n");
int64_t t0 = ggml_time_ms();
struct ggml_tensor* pooled = NULL;
size_t total_hidden_size = cond_stage_model.text_model.hidden_size;
@@ -833,8 +837,8 @@ public:
diffusion_model.compute(out_cond, n_threads, noised_input, NULL, c, control_net.controls, control_strength, t_emb, c_vector);
}
else{
diffusion_model.compute(out_cond, n_threads, noised_input, NULL, c_id, control_net.controls, control_strength, t_emb, c_vec_id);
// diffusion_model.compute(out_cond, n_threads, noised_input, NULL, c_id, control_net.controls, control_strength, t_emb, c_vector);
// diffusion_model.compute(out_cond, n_threads, noised_input, NULL, c_id, control_net.controls, control_strength, t_emb, c_vec_id);
diffusion_model.compute(out_cond, n_threads, noised_input, NULL, c_id, control_net.controls, control_strength, t_emb, c_vector);
}
}else{
diffusion_model.compute(out_cond, n_threads, noised_input, NULL, c, control_net.controls, control_strength, t_emb, c_vector);
@@ -1539,6 +1543,22 @@ sd_image_t* txt2img(sd_ctx_t* sd_ctx,
// free(cropped_images[i].data);
// cropped_images[i].data = NULL;
// }
// float* out_data = (float *)(init_img->data);
// int64_t stride = init_img->ne[0];
// for(int k = 0; k < 3; ++k){
// float mi = 100.f, mx= -100.f;
// for(int i = 0; i < stride; i++){
// printf("[");
// for(int j = 0; j < stride; j++){
// float val = out_data[k*stride*stride+i*stride+j];
// if(mi > val) mi = val;
// if(mx < val) mx = val;
// printf("%f, ", val);
// }
// printf("]\n");
// }
// printf(" channel, min, max: %d, %f %f \n", k, mi, mx);
// }
auto cond_tup = sd_ctx->sd->get_learned_condition_with_trigger(work_ctx, prompt,
clip_skip, width, height, num_input_images );
@@ -1610,7 +1630,7 @@ sd_image_t* txt2img(sd_ctx_t* sd_ctx,
int start_merge_step = -1;
if(sd_ctx->sd->stacked_id){
float style_strength_ratio = 20.f;
float style_strength_ratio = 30.f;
start_merge_step = int(style_strength_ratio / 100.f * sample_steps);
if(start_merge_step > 30)
start_merge_step = 30;