import genomicsdlarchsandlosses.bpnet.bpnetdefaults as bpnetdefaults import numpy as np from decimal import Decimal from genomicsdlarchsandlosses.bpnet.archs import BPNet from tensorflow.keras.utils import plot_model from bpnet.utils.exceptionhandler \ import NoTracebackException import os import json # motif_module_params = bpnetdefaults.MOTIF_MODULE_PARAMS # syntax_module_params = bpnetdefaults.SYNTAX_MODULE_PARAMS # profile_head_params = bpnetdefaults.PROFILE_HEAD_PARAMS # counts_head_params = bpnetdefaults.COUNTS_HEAD_PARAMS # profile_bias_module_params = bpnetdefaults.PROFILE_BIAS_MODULE_PARAMS # counts_bias_module_params = bpnetdefaults.COUNTS_BIAS_MODULE_PARAMS # change activations to linear # motif_module_params['activation'] = 'linear' # syntax_module_params['activation'] = 'linear' # profile_head_params['activation'] = 'linear' # tasks = { # 0:{ # 'signal': { # 'source': ['unstranded.bw'] # }, # 'loci': { # 'source': ['peaks.bed'] # }, # 'bias': { # 'source': ['bias.bw'], # 'smoothing': [[7.0, 81]] # } # } # } # tasks = { # 0:{ # 'signal': { # 'source': ['unstranded.bw', 'unstranded.bw'] # }, # 'loci': { # 'source': ['peaks.bed'] # }, # 'bias': { # 'source': ['bias.bw', 'bias.bw'], # 'smoothing': [None, None] # } # } # } tasks = { 0: { "signal": { "source": ["/home/ittai/ENCSR000EGM/data/plus.bw", "/home/ittai/ENCSR000EGM/data/minus.bw"] }, "loci": { "source": ["/home/ittai/ENCSR000EGM/data/peaks.bed"] }, "bias": { "source": ["/home/ittai/ENCSR000EGM/data/control_plus.bw", "/home/ittai/ENCSR000EGM/data/control_minus.bw"], "smoothing": [None, None] } } } # tasks = { # 0: { # 'signal': { # 'source': ['plus.bw', 'minus.bw'] # }, # 'loci': { # 'source': ['peaks.bed'] # }, # 'bias': { # 'source': ['genome_wide_bias_plus.bw', 'genome_wide_bias_minus.bw'], # 'smoothing': [None, None] # } # }, # 1: { # 'signal': { # 'source': ['unstranded.bw'] # }, # 'loci': { # 'source': ['peaks.bed'] # }, # 'bias': { # 'source': [], # 'smoothing': [] # } # }, # 2: { # 'signal': { # 'source': ['unstranded.bw'] # }, # 'loci': { # 'source': ['peaks.bed'] # }, # 'bias': { # 'source': ['unstranded_genome_wide_bias.bw'], # 'smoothing': [[7.0, 81]] # } # }, # 3: { # 'signal': { # 'source': ['plus.bw', 'minus.bw'] # }, # 'loci': { # 'source': ['peaks.bed'] # }, # 'bias': { # 'source': ['genome_wide_bias_plus.bw', 'genome_wide_bias_minus.bw'], # 'smoothing': [None, [7.0, 81]] # } # }, # 4: { # 'signal': { # 'source': ['plus.bw', 'minus.bw'] # }, # 'loci': { # 'source': ['peaks.bed'] # }, # 'bias': { # 'source': ['genome_wide_bias_plus.bw', 'genome_wide_bias_minus.bw'], # 'smoothing': [[7.0, 81], [7.0, 81]] # } # } # } def load_json(params_json_file): # make sure the params json file exists if not os.path.isfile(params_json_file): raise NoTracebackException( "File not found: {} ".format(params_json_file)) # load the params json file with open(params_json_file, 'r') as inp_json: try: model_arch_params = json.loads(inp_json.read()) except json.decoder.JSONDecodeError: raise NoTracebackException( "Unable to load json file {}. Valid json expected. " "Check the file for syntax errors.".format( params_json_file)) return model_arch_params # single unstranded task with linear activations # model_v1 = BPNet(tasks, load_json('/users/zahoor/lab_data3/Chromatin-Atlas/GM12878_baseline/baseline_filtered_peaks_500/bpnet_params.json')) model_v1 = BPNet(tasks, load_json('bpnet_params_1_task_old_counts_head.json')) model_v1.summary() plot_model( model_v1, to_file='model_v1.png', show_shapes=True, show_layer_names=True) model_v2 = BPNet(tasks, load_json('bpnet_params_1_task.json')) model_v2.summary() plot_model( model_v2, to_file='model_v2.png', show_shapes=True, show_layer_names=True) def set_all_weights_to_one(model): """ Set weights in all layers of the model to 1 Args: model (tensorflow.keras.Model): model object """ # set all weights (W) and biases (B) <<- (Wx + B) to 1 for layer in model.layers: if layer.count_params() == 0: continue weights_n_biases = layer.get_weights() weights = weights_n_biases[0] biases = weights_n_biases[1] _weights = np.ones(weights.shape) _biases = np.ones(biases.shape) layer.set_weights((_weights, _biases)) def print_all_weights(model): """ Print weights of all layers of the model Args: model (tensorflow.keras.Model): model object """ for layer in model.layers: if layer.count_params() == 0: continue weights_n_biases = layer.get_weights() weights = weights_n_biases[0] biases = weights_n_biases[1] print("{} - Weights \n {}".format(layer.name, weights)) print("{} - Biases \n {}".format(layer.name, biases)) # #print_all_weights(model_v1) set_all_weights_to_one(model_v1) # #print_all_weights(model_v1) # #print_all_weights(model_v2) set_all_weights_to_one(model_v2) # #print_all_weights(model_v2) one_hot_input = np.ones((1, 2114, 4)) profile_bias_input_0 = np.ones((1, 1000, 2)) # profile_bias_input_2 = np.ones((1, 1000, 2)) # profile_bias_input_3 = np.ones((1, 1000, 3)) # profile_bias_input_4 = np.ones((1, 1000, 4)) counts_bias_input_0 = np.ones((1, 2)) # counts_bias_input_2 = np.ones((1, 2)) # counts_bias_input_3 = np.ones((1, 3)) # counts_bias_input_4 = np.ones((1, 4)) model_input = { 'sequence': one_hot_input, 'profile_bias_input_0': profile_bias_input_0, # 'profile_bias_input_2': profile_bias_input_2, # 'profile_bias_input_3': profile_bias_input_3, # 'profile_bias_input_4': profile_bias_input_4, 'counts_bias_input_0': counts_bias_input_0 # 'counts_bias_input_2': counts_bias_input_2, # 'counts_bias_input_3': counts_bias_input_3, # 'counts_bias_input_4': counts_bias_input_4 } predictions_v1 = model_v1.predict(model_input) predictions_v2 = model_v2.predict(model_input) print("Model 1 (profile sum)", Decimal(np.sum(predictions_v1[0]).astype(float))) print("Model 2 (profile sum)", Decimal(np.sum(predictions_v2[0]).astype(float))) print("Model 1 (counts ouput)",Decimal(np.sum(predictions_v1[1]).astype(float))) print("Model 2 (counts ouput)", Decimal(np.sum(predictions_v2[1]).astype(float)))