Spaces:
Running
Running
Update app.py
Browse files
app.py
CHANGED
@@ -49,7 +49,7 @@ idx_to_word_global = None
|
|
49 |
device_global = torch.device("cuda" if torch.cuda.is_available() else "cpu")
|
50 |
model_load_status_global = "Model not loaded."
|
51 |
|
52 |
-
CHECKPOINT_FILENAME = "swck_model_conceptual_app_fulldebug.pth.tar"
|
53 |
|
54 |
MAIN_LOSS_WEIGHT_APP = 1.0
|
55 |
BLOCK_TARGET_ENTROPY_LOSS_WEIGHT_APP = 0.02
|
@@ -84,7 +84,8 @@ def build_vocab_from_corpus_text_app(corpus_text):
|
|
84 |
print(f"App: Built vocab of size {VOCAB_SIZE_APP}")
|
85 |
return temp_word_to_idx, temp_idx_to_word
|
86 |
|
87 |
-
|
|
|
88 |
global swck_model_global, optimizer_global, word_to_idx_global, idx_to_word_global, \
|
89 |
VOCAB_SIZE_APP, model_load_status_global
|
90 |
|
@@ -103,17 +104,19 @@ def initialize_or_load_model_app():
|
|
103 |
'num_sub_modules_per_block': NUM_SUB_MODULES_PER_BLOCK_APP
|
104 |
}
|
105 |
|
106 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
107 |
swck_model_global = SWCKModel(**model_args).to(device_global)
|
108 |
-
#
|
109 |
-
|
110 |
-
|
111 |
-
|
112 |
-
|
113 |
-
if hasattr(swck_model_global, 'seed_parser'): swck_model_global.seed_parser.debug_prints_enabled = True
|
114 |
-
for block in swck_model_global.adaptive_blocks: block.debug_prints_enabled = True
|
115 |
-
swck_model_global.debug_prints_enabled = True
|
116 |
-
print("App: All model component debugs are intended to be ON by default from their init.")
|
117 |
|
118 |
|
119 |
if os.path.exists(CHECKPOINT_FILENAME):
|
@@ -137,21 +140,29 @@ def initialize_or_load_model_app():
|
|
137 |
print("App: Checkpoint vocab seems invalid, using app's rebuilt vocab.")
|
138 |
else:
|
139 |
print("App: word_to_idx not in checkpoint, using app's rebuilt vocab.")
|
|
|
|
|
|
|
|
|
|
|
|
|
140 |
|
141 |
model_load_status_global = f"Model loaded successfully from {CHECKPOINT_FILENAME}."
|
142 |
print(model_load_status_global)
|
143 |
except Exception as e:
|
144 |
-
print(f"App: Error loading model from checkpoint: {e}. Re-initializing new model with debug
|
145 |
swck_model_global = SWCKModel(**model_args).to(device_global)
|
146 |
-
|
147 |
-
|
148 |
-
|
|
|
149 |
optimizer_global = optim.AdamW(swck_model_global.parameters(), lr=0.001)
|
150 |
-
model_load_status_global = "Error loading checkpoint. Using new (untrained) model with debug
|
151 |
else:
|
152 |
-
print(f"App: Checkpoint {CHECKPOINT_FILENAME} not found. Initializing new model with debug
|
|
|
153 |
optimizer_global = optim.AdamW(swck_model_global.parameters(), lr=0.001)
|
154 |
-
model_load_status_global = "Initialized a new (untrained) model with debug
|
155 |
|
156 |
swck_model_global.eval()
|
157 |
return model_load_status_global
|
@@ -191,13 +202,12 @@ def run_short_training_session(num_epochs_app, batch_size_app, learning_rate_app
|
|
191 |
print("\n--- App: Starting Short Training Session (Full Debug ON for ALL batches/epochs) ---")
|
192 |
progress(0, desc="Preparing training data...")
|
193 |
|
194 |
-
|
195 |
-
set_model_debug_prints(swck_model_global, True, True, True)
|
196 |
|
197 |
training_corpus = SEED_PHRASE_APP + " " + EXTENDED_TEXT_FOR_TRAINING_APP
|
198 |
app_dataset = AppSWCKDataset(training_corpus, word_to_idx_global, SEQ_LEN_APP, SOS_TOKEN, EOS_TOKEN, PAD_TOKEN)
|
199 |
if not app_dataset.samples:
|
200 |
-
set_model_debug_prints(swck_model_global, False, False, False)
|
201 |
return "App Training Error: No samples created from the corpus."
|
202 |
|
203 |
app_dataloader = DataLoader(app_dataset, batch_size=int(batch_size_app), shuffle=True, collate_fn=app_swck_collate_fn)
|
@@ -219,8 +229,7 @@ def run_short_training_session(num_epochs_app, batch_size_app, learning_rate_app
|
|
219 |
print(f"\n>>> EPOCH {epoch+1} - Starting with Full Debug for all batches <<<")
|
220 |
|
221 |
for batch_idx, (src_batch, tgt_batch) in enumerate(app_dataloader):
|
222 |
-
|
223 |
-
print(f"\n--- Training Batch {batch_idx+1}/{len(app_dataloader)} ---")
|
224 |
|
225 |
src_batch, tgt_batch = src_batch.to(device_global), tgt_batch.to(device_global)
|
226 |
decoder_input_tokens = src_batch[:, :-1]
|
@@ -268,7 +277,6 @@ def run_short_training_session(num_epochs_app, batch_size_app, learning_rate_app
|
|
268 |
epoch_loss += combined_loss.item()
|
269 |
|
270 |
log_line = f" Epoch {epoch+1}, Batch {batch_idx+1}/{len(app_dataloader)}, Loss: {combined_loss.item():.4f}"
|
271 |
-
# Print every batch to console due to full debug, but maybe less often to UI
|
272 |
print(log_line)
|
273 |
if batch_idx % max(1, len(app_dataloader)//2) == 0 or batch_idx == len(app_dataloader)-1 :
|
274 |
training_log_output += log_line + "\n"
|
@@ -278,8 +286,6 @@ def run_short_training_session(num_epochs_app, batch_size_app, learning_rate_app
|
|
278 |
print(epoch_summary)
|
279 |
training_log_output += epoch_summary
|
280 |
|
281 |
-
# Set debug prints OFF after the entire training session for subsequent operations (like generation)
|
282 |
-
# unless generation itself re-enables them.
|
283 |
print("--- App: Training Session Finished. Setting debug prints OFF by default. ---")
|
284 |
set_model_debug_prints(swck_model_global, False, False, False)
|
285 |
swck_model_global.eval()
|
@@ -307,7 +313,7 @@ def run_short_training_session(num_epochs_app, batch_size_app, learning_rate_app
|
|
307 |
|
308 |
return training_log_output
|
309 |
|
310 |
-
def generate_text_for_app(prompt_str, max_len_gen, temperature_gen):
|
311 |
global model_load_status_global
|
312 |
if swck_model_global is None or word_to_idx_global is None or idx_to_word_global is None:
|
313 |
return "Model not loaded. Please check server logs or try training.", "Model not available."
|
@@ -315,19 +321,18 @@ def generate_text_for_app(prompt_str, max_len_gen, temperature_gen): # Removed d
|
|
315 |
swck_model_global.eval()
|
316 |
swck_model_global.set_wiring_phase(False)
|
317 |
|
318 |
-
# FULL DEBUG ON for generation
|
319 |
print("\n--- App: Generating Text (Full Debug ON) ---")
|
320 |
-
set_model_debug_prints(swck_model_global, True, True, True)
|
321 |
|
322 |
print(f"App: Generating for prompt: '{prompt_str}', max_len: {max_len_gen}, temp: {temperature_gen}")
|
323 |
|
324 |
tokens = [SOS_TOKEN] + [word_to_idx_global.get(w, UNK_TOKEN) for w in prompt_str.lower().split()]
|
325 |
generated_ids_app = list(tokens)
|
326 |
-
debug_info_lines = [f"Prompt tokens: {generated_ids_app}"]
|
327 |
|
328 |
with torch.no_grad():
|
329 |
for i in range(int(max_len_gen)):
|
330 |
-
print(f"\n--- Generation Step {i+1} ---")
|
331 |
context_start_idx = max(0, len(generated_ids_app) - SEQ_LEN_APP)
|
332 |
current_context_ids = generated_ids_app[context_start_idx:]
|
333 |
|
@@ -353,9 +358,9 @@ def generate_text_for_app(prompt_str, max_len_gen, temperature_gen): # Removed d
|
|
353 |
generated_ids_app.append(next_token_id)
|
354 |
|
355 |
current_word = idx_to_word_global.get(next_token_id, UNK_TOKEN_STR)
|
356 |
-
print(f" ==> Generated token {i+1}: '{current_word}' (ID: {next_token_id})")
|
357 |
|
358 |
-
if i < 10 :
|
359 |
overall_ent = entropy_report_infer['overall_output_entropy'].item()
|
360 |
if entropy_report_infer['block_output_entropies'] and len(entropy_report_infer['block_output_entropies']) > 0:
|
361 |
b0_ent = entropy_report_infer['block_output_entropies'][0].item()
|
@@ -377,12 +382,11 @@ def generate_text_for_app(prompt_str, max_len_gen, temperature_gen): # Removed d
|
|
377 |
debug_output_str = "\n".join(debug_info_lines)
|
378 |
|
379 |
print("--- App: Generation Finished. Setting debug prints OFF by default. ---")
|
380 |
-
set_model_debug_prints(swck_model_global, False, False, False)
|
381 |
return final_text, debug_output_str
|
382 |
|
383 |
-
# Initialize model
|
384 |
-
|
385 |
-
initial_load_status = initialize_or_load_model_app()
|
386 |
|
387 |
with gr.Blocks(title="SWCK Conceptual Demo") as demo:
|
388 |
model_status_md = gr.Markdown(value=f"**Model Status:** {initial_load_status}", elem_id="model_status_md_123")
|
@@ -398,9 +402,9 @@ with gr.Blocks(title="SWCK Conceptual Demo") as demo:
|
|
398 |
with gr.TabItem("Generate Text"):
|
399 |
with gr.Row():
|
400 |
prompt_input = gr.Textbox(label="Enter your prompt:", placeholder="e.g., the meaning of existence is", scale=3)
|
401 |
-
# Removed debug checkbox
|
402 |
with gr.Row():
|
403 |
-
generate_button = gr.Button("Generate", scale=1)
|
404 |
with gr.Row():
|
405 |
max_len_slider = gr.Slider(minimum=10, maximum=150, value=50, step=1, label="Max Generation Length")
|
406 |
temp_slider = gr.Slider(minimum=0.0, maximum=2.0, value=0.8, step=0.1, label="Temperature (0 for greedy)")
|
@@ -422,8 +426,8 @@ with gr.Blocks(title="SWCK Conceptual Demo") as demo:
|
|
422 |
return f"**Model Status:** {model_load_status_global}"
|
423 |
|
424 |
generate_button.click(
|
425 |
-
fn=generate_text_for_app,
|
426 |
-
inputs=[prompt_input, max_len_slider, temp_slider],
|
427 |
outputs=[output_text, debug_text_area]
|
428 |
)
|
429 |
|
@@ -435,4 +439,4 @@ with gr.Blocks(title="SWCK Conceptual Demo") as demo:
|
|
435 |
|
436 |
|
437 |
if __name__ == "__main__":
|
438 |
-
demo.launch(debug=True)
|
|
|
49 |
device_global = torch.device("cuda" if torch.cuda.is_available() else "cpu")
|
50 |
model_load_status_global = "Model not loaded."
|
51 |
|
52 |
+
CHECKPOINT_FILENAME = "swck_model_conceptual_app_fulldebug.pth.tar"
|
53 |
|
54 |
MAIN_LOSS_WEIGHT_APP = 1.0
|
55 |
BLOCK_TARGET_ENTROPY_LOSS_WEIGHT_APP = 0.02
|
|
|
84 |
print(f"App: Built vocab of size {VOCAB_SIZE_APP}")
|
85 |
return temp_word_to_idx, temp_idx_to_word
|
86 |
|
87 |
+
# CORRECTED FUNCTION DEFINITION
|
88 |
+
def initialize_or_load_model_app(enable_initial_debug=True):
|
89 |
global swck_model_global, optimizer_global, word_to_idx_global, idx_to_word_global, \
|
90 |
VOCAB_SIZE_APP, model_load_status_global
|
91 |
|
|
|
104 |
'num_sub_modules_per_block': NUM_SUB_MODULES_PER_BLOCK_APP
|
105 |
}
|
106 |
|
107 |
+
if enable_initial_debug:
|
108 |
+
print("App: Initializing SWCKModel with FULL DEBUG ON by default for init...")
|
109 |
+
|
110 |
+
# Temporarily disable sub-component debug before SWCKModel init if enable_initial_debug is False,
|
111 |
+
# so SWCKModel's own init prints don't get mixed with sub-component init prints prematurely.
|
112 |
+
# SeedParser's internal debug_prints_enabled will control its own prints during its __init__.
|
113 |
+
|
114 |
swck_model_global = SWCKModel(**model_args).to(device_global)
|
115 |
+
# Now set the debug states for all components based on enable_initial_debug
|
116 |
+
set_model_debug_prints(swck_model_global,
|
117 |
+
seed_parser_debug=enable_initial_debug,
|
118 |
+
block_debug=enable_initial_debug,
|
119 |
+
model_debug=enable_initial_debug)
|
|
|
|
|
|
|
|
|
120 |
|
121 |
|
122 |
if os.path.exists(CHECKPOINT_FILENAME):
|
|
|
140 |
print("App: Checkpoint vocab seems invalid, using app's rebuilt vocab.")
|
141 |
else:
|
142 |
print("App: word_to_idx not in checkpoint, using app's rebuilt vocab.")
|
143 |
+
|
144 |
+
# Ensure debug states are correctly set after loading
|
145 |
+
set_model_debug_prints(swck_model_global,
|
146 |
+
seed_parser_debug=enable_initial_debug,
|
147 |
+
block_debug=enable_initial_debug,
|
148 |
+
model_debug=enable_initial_debug)
|
149 |
|
150 |
model_load_status_global = f"Model loaded successfully from {CHECKPOINT_FILENAME}."
|
151 |
print(model_load_status_global)
|
152 |
except Exception as e:
|
153 |
+
print(f"App: Error loading model from checkpoint: {e}. Re-initializing new model with debug state: {enable_initial_debug}.")
|
154 |
swck_model_global = SWCKModel(**model_args).to(device_global)
|
155 |
+
set_model_debug_prints(swck_model_global,
|
156 |
+
seed_parser_debug=enable_initial_debug,
|
157 |
+
block_debug=enable_initial_debug,
|
158 |
+
model_debug=enable_initial_debug)
|
159 |
optimizer_global = optim.AdamW(swck_model_global.parameters(), lr=0.001)
|
160 |
+
model_load_status_global = f"Error loading checkpoint. Using new (untrained) model with debug: {enable_initial_debug}."
|
161 |
else:
|
162 |
+
print(f"App: Checkpoint {CHECKPOINT_FILENAME} not found. Initializing new model with debug state: {enable_initial_debug}.")
|
163 |
+
# set_model_debug_prints was already called for a new model above
|
164 |
optimizer_global = optim.AdamW(swck_model_global.parameters(), lr=0.001)
|
165 |
+
model_load_status_global = f"Initialized a new (untrained) model with debug: {enable_initial_debug}."
|
166 |
|
167 |
swck_model_global.eval()
|
168 |
return model_load_status_global
|
|
|
202 |
print("\n--- App: Starting Short Training Session (Full Debug ON for ALL batches/epochs) ---")
|
203 |
progress(0, desc="Preparing training data...")
|
204 |
|
205 |
+
set_model_debug_prints(swck_model_global, True, True, True) # DEBUG ALWAYS ON FOR TRAINING
|
|
|
206 |
|
207 |
training_corpus = SEED_PHRASE_APP + " " + EXTENDED_TEXT_FOR_TRAINING_APP
|
208 |
app_dataset = AppSWCKDataset(training_corpus, word_to_idx_global, SEQ_LEN_APP, SOS_TOKEN, EOS_TOKEN, PAD_TOKEN)
|
209 |
if not app_dataset.samples:
|
210 |
+
set_model_debug_prints(swck_model_global, False, False, False)
|
211 |
return "App Training Error: No samples created from the corpus."
|
212 |
|
213 |
app_dataloader = DataLoader(app_dataset, batch_size=int(batch_size_app), shuffle=True, collate_fn=app_swck_collate_fn)
|
|
|
229 |
print(f"\n>>> EPOCH {epoch+1} - Starting with Full Debug for all batches <<<")
|
230 |
|
231 |
for batch_idx, (src_batch, tgt_batch) in enumerate(app_dataloader):
|
232 |
+
print(f"\n--- Training Batch {batch_idx+1}/{len(app_dataloader)} ---") # Explicit batch print
|
|
|
233 |
|
234 |
src_batch, tgt_batch = src_batch.to(device_global), tgt_batch.to(device_global)
|
235 |
decoder_input_tokens = src_batch[:, :-1]
|
|
|
277 |
epoch_loss += combined_loss.item()
|
278 |
|
279 |
log_line = f" Epoch {epoch+1}, Batch {batch_idx+1}/{len(app_dataloader)}, Loss: {combined_loss.item():.4f}"
|
|
|
280 |
print(log_line)
|
281 |
if batch_idx % max(1, len(app_dataloader)//2) == 0 or batch_idx == len(app_dataloader)-1 :
|
282 |
training_log_output += log_line + "\n"
|
|
|
286 |
print(epoch_summary)
|
287 |
training_log_output += epoch_summary
|
288 |
|
|
|
|
|
289 |
print("--- App: Training Session Finished. Setting debug prints OFF by default. ---")
|
290 |
set_model_debug_prints(swck_model_global, False, False, False)
|
291 |
swck_model_global.eval()
|
|
|
313 |
|
314 |
return training_log_output
|
315 |
|
316 |
+
def generate_text_for_app(prompt_str, max_len_gen, temperature_gen):
|
317 |
global model_load_status_global
|
318 |
if swck_model_global is None or word_to_idx_global is None or idx_to_word_global is None:
|
319 |
return "Model not loaded. Please check server logs or try training.", "Model not available."
|
|
|
321 |
swck_model_global.eval()
|
322 |
swck_model_global.set_wiring_phase(False)
|
323 |
|
|
|
324 |
print("\n--- App: Generating Text (Full Debug ON) ---")
|
325 |
+
set_model_debug_prints(swck_model_global, True, True, True) # DEBUG ALWAYS ON FOR GENERATION
|
326 |
|
327 |
print(f"App: Generating for prompt: '{prompt_str}', max_len: {max_len_gen}, temp: {temperature_gen}")
|
328 |
|
329 |
tokens = [SOS_TOKEN] + [word_to_idx_global.get(w, UNK_TOKEN) for w in prompt_str.lower().split()]
|
330 |
generated_ids_app = list(tokens)
|
331 |
+
debug_info_lines = [f"Prompt tokens: {generated_ids_app}"]
|
332 |
|
333 |
with torch.no_grad():
|
334 |
for i in range(int(max_len_gen)):
|
335 |
+
print(f"\n--- Generation Step {i+1} ---")
|
336 |
context_start_idx = max(0, len(generated_ids_app) - SEQ_LEN_APP)
|
337 |
current_context_ids = generated_ids_app[context_start_idx:]
|
338 |
|
|
|
358 |
generated_ids_app.append(next_token_id)
|
359 |
|
360 |
current_word = idx_to_word_global.get(next_token_id, UNK_TOKEN_STR)
|
361 |
+
print(f" ==> Generated token {i+1}: '{current_word}' (ID: {next_token_id})")
|
362 |
|
363 |
+
if i < 10 :
|
364 |
overall_ent = entropy_report_infer['overall_output_entropy'].item()
|
365 |
if entropy_report_infer['block_output_entropies'] and len(entropy_report_infer['block_output_entropies']) > 0:
|
366 |
b0_ent = entropy_report_infer['block_output_entropies'][0].item()
|
|
|
382 |
debug_output_str = "\n".join(debug_info_lines)
|
383 |
|
384 |
print("--- App: Generation Finished. Setting debug prints OFF by default. ---")
|
385 |
+
set_model_debug_prints(swck_model_global, False, False, False)
|
386 |
return final_text, debug_output_str
|
387 |
|
388 |
+
# Initialize model. Set enable_initial_debug=True for verbose init logs.
|
389 |
+
initial_load_status = initialize_or_load_model_app(enable_initial_debug=True)
|
|
|
390 |
|
391 |
with gr.Blocks(title="SWCK Conceptual Demo") as demo:
|
392 |
model_status_md = gr.Markdown(value=f"**Model Status:** {initial_load_status}", elem_id="model_status_md_123")
|
|
|
402 |
with gr.TabItem("Generate Text"):
|
403 |
with gr.Row():
|
404 |
prompt_input = gr.Textbox(label="Enter your prompt:", placeholder="e.g., the meaning of existence is", scale=3)
|
405 |
+
# Removed debug checkbox from here
|
406 |
with gr.Row():
|
407 |
+
generate_button = gr.Button("Generate (Full Debug to Console)", scale=1) # Updated button label
|
408 |
with gr.Row():
|
409 |
max_len_slider = gr.Slider(minimum=10, maximum=150, value=50, step=1, label="Max Generation Length")
|
410 |
temp_slider = gr.Slider(minimum=0.0, maximum=2.0, value=0.8, step=0.1, label="Temperature (0 for greedy)")
|
|
|
426 |
return f"**Model Status:** {model_load_status_global}"
|
427 |
|
428 |
generate_button.click(
|
429 |
+
fn=generate_text_for_app,
|
430 |
+
inputs=[prompt_input, max_len_slider, temp_slider], # Removed checkbox from inputs
|
431 |
outputs=[output_text, debug_text_area]
|
432 |
)
|
433 |
|
|
|
439 |
|
440 |
|
441 |
if __name__ == "__main__":
|
442 |
+
demo.launch(debug=True)
|